Temporary bond and debond runs once per layer, before any final-joining technology is chosen. Each layer added on the way to 20-hi adds a pass, so the step compounds. Schematic: chokepoints.ai.

In 60 seconds

A finished HBM stack is not allowed to be thicker than the wafer its chips were cut from. That is a JEDEC rule, and it forces every memory die in the stack to be thinned to a fraction of its original height before it can be stacked at all. Past a certain thinness a wafer can no longer be handled unless it is glued to a carrier and released again afterwards, a cycle called temporary bond and debond. SUSS MicroTec builds the machines for both halves of it, and is one of only two toolmakers qualified to sell them into HBM today.

The cycle runs once for every layer, and layer count is climbing from twelve toward twenty. Tokyo Electron sells the same step, and sells more of it. That leaves one question: is the second seat at a step nobody can skip worth owning at €79.80, after the shares have tripled off last September’s low?

Playing devil’s advocate

The strongest case against us says SUSS is worth its margin trough, and that the market has already paid for the physics. Tokyo Electron leads HBM temporary bonding on two 2024-vintage broker reads, it owns the SK Hynix line growing fastest, and it has just put a five-year, ¥500bn number on the category; EV Group’s laser tools become a live threat once stacks reach 16-hi. Neither risk resolves in public sourcing, and the shares have doubled this year, so nobody is being handed a discount. We grant all of it. Our answer is narrow: the step is forced by arithmetic that no joining technology relieves, and a starter is how we hold that without pretending the entry is cheap.

Subscribe now

01 What a temporary bonder actually is

A temporary bonder glues a silicon wafer face down onto a rigid carrier so the back of it can be ground away without the wafer shattering. A debonder takes the carrier off again afterwards. SUSS sells both halves of that pair, and together they are most of the company.

The step exists because of a thickness budget. A 300mm silicon wafer leaves the fab about 775 microns thick, roughly eight sheets of paper. JEDEC, the body that writes memory standards, caps the total height of a finished HBM4 stack at the same 775 microns. Sixteen memory chips, a logic chip beneath them and all the interconnect between have to fit inside the thickness of one raw wafer. Each chip therefore comes down to somewhere between 30 and 40 microns. At that point silicon stops behaving like a slab. It flexes, and it cracks under its own handling, so it cannot cross the rest of the line unsupported.

The fix is to give it something to lean on. The wafer gets a release layer and an adhesive, is bonded face down onto a carrier of glass or silicon at low force, then cured. Now it can be ground, etched and metallised as a rigid object, even though the device layer underneath is thinner than a human hair. SUSS sells that machine as the XBS300.

Getting the carrier back off is the harder problem, and it is a different machine, the XBC300 Gen2. It has to release a wafer of 50 microns or less, at room temperature, across 300 millimetres, without cracking something that by then carries the full value of a finished HBM layer. Where a mechanical peel would be too violent, SUSS sells the ELD300, which fires a 308 nanometre excimer laser through the glass carrier to break the adhesive optically and lift the wafer away untouched.

Two consequences follow. A JEDEC standard fixes the thickness budget, so no fab can choose to skip the step. And it repeats: every layer that needs thinning is bonded and debonded once, so a 16-high stack runs the cycle more times than a 12-high one.

02 SUSS MicroTec, briefly

Karl Süss started the company in 1949 in Garching, outside Munich, selling other people’s optical instruments: Leitz microscopes, cameras, laboratory equipment. It became a manufacturer on commission. In 1963 Siemens asked it to build a simple photolithography tool, and that contract pulled it out of distribution and into engineering. Karl Süss and his son Ekkehard both died in 1994. The business was reorganised, renamed and listed in Frankfurt in 1999.

It reports in two segments today. Advanced Backend Solutions holds imaging, coating and bonding systems, and carries most of the revenue. Photomask Solutions is the smaller half, cleaning and processing the masks that carry circuit patterns into a lithography scanner. Revenue was €446.1 million in 2024 and €503.2 million in 2025. Burkhardt Frick has been chief executive since 11 September 2023.

SUSS employed 1,498 people at the end of 2024, and 1,497 at the end of 2025, while revenue rose almost 13 percent. Fewer than fifteen hundred people in a Munich suburb, and every HBM stack shipped this year went through the step their machines perform. Section 05 settles how much of it went through their machines and how much through a competitor’s.

03 Where the step sits in the line

All of that happens before the dies are joined to one another. The joining step is a different machine, sold by different companies, and it has the industry’s attention.

The consensus read on advanced packaging is that hybrid bonding is coming: it welds dies together with direct copper contact and no solder, and it makes the tools built for the old way of joining chips obsolete. That read is right about the joining step. Hanmi, ASMPT, Besi and EV Group are fighting over it, hybrid bonding against thermocompression against mass reflow, and whoever loses that fight has a real problem.

SUSS is not in that fight. Thinning happens one station earlier, and the height budget forces it. The eventual method of attaching the dies does not enter into it. Hybrid bonding joins dies that have already been thinned. It does not thin them. Whichever way the joining fight ends, the winner’s tool sits downstream of a wafer that was bonded to a carrier, ground, and debonded first.

Arithmetic protects the step. Stack height forces the thinning, no joining technology relieves it, and a second source still has to win its qualification at each memory maker, one customer at a time.

The tape has been arguing the wrong question. The step survives either way. SUSS’s share of it is the live question.

04 Why layer count is the demand driver

The height budget also sets the growth rate. JEDEC has relaxed the ceiling twice, and the room per layer shrank both times.

Through HBM3E the ceiling was 720 microns. JEDEC raised it to 775 for HBM4 at the memory makers’ request, and that relaxation lets a 16-high stack ship on existing bonding instead of forcing hybrid bonding. The industry is now weighing 825 to 900 microns for 20-high parts. Divide each ceiling by the dies it has to hold, plus the base logic die, and the budget per layer falls from about 55 microns at 12-high, to 46 at 16-high, to 43 at a 900-micron 20-high. Interconnect comes out of the same allowance. Layer count rises faster than the ceiling does, so relaxing the standard does not relieve the thinning. It tightens it.

Each relaxation of the ceiling has been more than absorbed by the extra layers it enables, so the thickness budget per layer has tightened at every step. Cap divided by the memory dies plus the base logic die; interconnect comes out of the same allowance. Source: JEDEC JESD270-4 via TrendForce (21 Apr 2025); 20-high thickness discussion per TrendForce (6 Mar 2026). Arithmetic and chart by chokepoints.ai.

Every layer added on the way from 12-high to 20-high is another die that has to be thinned, and another bond-and-debond pass. Stacks per accelerator are fixed by interposer area, so a taller stack raises core-die count per accelerator close to one for one, and demand for the step tracks that die count.

That linearity is the easy half of the argument. The hard half is SUSS’s share of it, and two things bite into that share: Tokyo Electron’s SK Hynix lock and EV Group’s laser-debond lead.

Subscribe now

05 Who buys the step, and where SUSS ranks

The memory makers stacking on this node split SK Hynix 60 percent, Samsung 30 percent, Micron 10 percent (Mirae Asset Securities, 15 Jun 2026): that is who buys the output, not who sells the tools. On the tool side, Jefferies scopes SUSS’s position to its two core customers at roughly 55 percent of Samsung and Micron’s combined temporary-bond spend, real share counted against a narrower base than the whole HBM pie.

SUSS is strong at two of the three makers and holds only an R&D toehold at the largest, SK Hynix, where Tokyo Electron is entrenched. The single biggest pool of demand is the one where SUSS’s seat is thinnest, so the SK Hynix qualification, and the process SK Hynix ships at 16-hi, carry more weight than the headline share suggests.

Broker estimates of SUSS’s share of the toll span a wide range. On HBM temporary bonding, Tokyo Electron is the consensus’s clear leader: Bernstein estimates 60 to 70 percent (6 Dec 2024), Jefferies a narrower c.50 percent (20 Sep 2024). SUSS is the clear number two, with broker estimates spanning 10 to 20 percent (Deutsche Bank, 11 Jul 2024) to 40 to 50 percent (Metzler Capital Markets, 28 Oct 2024); Berenberg (29 May 2024) puts it near 35 percent while naming SUSS the leader in the specialised temporary-bonding segment, a narrower cut than the whole pie.

EV Group sits in a different market. Its widely-cited 65 to 70 percent share figure describes wafer-to-wafer permanent bonding, where Bernstein (19 Nov 2025) puts EVG’s footprint in high-volume HBM lines below both TEL’s and SUSS’s. In the market that figure does measure, the split runs EVG 60 to 70 percent, TEL a little over 20 percent, SUSS 10 percent (Bernstein, 13 Mar 2025), the same challenger share the house map carries for SUSS in hybrid bonding.

The axes a process engineer screens a vendor on. TEL leads HBM temporary bonding outright; SUSS is a corroborated number two; EVG is a real incumbent in the adjacent permanent-bonding market, not this one, yet. Ranges above are broker estimates, not point figures.

06 Two cuts and a record year

Burkhardt Frick has run SUSS since 11 September 2023, and his 2025 delivered two margin cuts and a record year inside the same twelve months. Order intake fell 16.4 percent year over year to €354.3 million even as revenue grew on backlog conversion; in equipment, the order book leads and revenue lags. Frick cut margin guidance twice: first on 28 July 2025, gross margin to 37 to 39 percent and EBIT margin to 13 to 15 percent, then on 27 October 2025 on Q3 preliminary figures that missed hard, 33.1 percent gross margin against a 38.1 percent consensus. Sales guidance was never touched either time, and FY2025 sales of €503.2 million beat the original range.

On the 30 March 2026 call Frick named the cause himself: many customers, he said, are “placing orders later and ordering different solutions compared to the peak of the first AI waves, when demand for temporary bonders and debonders was particularly high.” He was blunter about the credibility cost: “as we saw last year, where we had to go in and correct twice our guidance. This is something we don’t want to repeat.”

07 The order book and the margin disagree

Q1 2026, reported 7 May, delivered the reversal in the order book: intake of €149.3 million, a record, up 69.5 percent year over year; order book €330.1 million, up 23.7 percent in the quarter. The EBIT margin, though, was 4.3 percent, against 18.0 percent in the same quarter a year earlier, still below the low end of FY2026’s 8 to 10 percent guided range. The margin trails both intake and revenue because the cost base is fixed and the converting backlog has not yet covered it.

FY2025’s order-intake decline led the year’s guidance cuts by two to three quarters; Q1 2026’s record intake has not reached the margin yet. The H1 2025 bar is a half-year and sits above the single quarters around it. Source: SUSS MicroTec reported quarters and half-years, compiled 17 Jul 2026. Chart by chokepoints.ai.

Asked on the same call about SK Hynix qualification, Frick said the orders meant SUSS is now “in all 3 major memory makers,” and, pressed on whether that meant repeat business, drew the line himself: “we are not talking volume orders here, but at least we have our hardware placed now in the most recent HBM R&D line.” That puts SUSS inside the door at the one memory maker the market had already priced as closed, with hardware in an R&D line and no volume behind it.

COO Thomas Rohe put a number on part of the cause: the new Zhubei, Taiwan site was running at roughly 70 percent utilisation as of the March call, still absorbing double rent against the old Hsinchu-area sites it replaces, though Q1 2026 is flagged as the last quarter carrying that specific drag. The path back to the guided floor is backlog conversion and the Zhubei ramp, not a new pricing cycle.

None of this is a China manufacturing story. SUSS’s entire production footprint sits in Germany and Zhubei, Taiwan; its China presence is a sales and service subsidiary only. The China exposure that does show up is demand-side: photomask order intake fell €43.5 million in FY2025, €31 million of it attributed to lower orders from Chinese customers, a cyclicality risk rather than a supply-chain one, and one Frick flagged as partially reversing late in Q4 2025. The balance sheet carries no draw against the €115 million syndicated facility signed in February 2026, which management frames as a cyclicality buffer.

08 The delay the market sold and we bought

We disagree with the market about one specific day, and that disagreement is the position. On 7 July 2026 TrendForce reported that Samsung and SK Hynix were reconsidering the hybrid-bonding timeline, with 16-hi HBM4E the earliest realistic adoption. SUSS closed down 12.3 percent that day, €87.45 to €76.65, its worst session of the year and the largest single move inside the fall from June’s high. The tape read that headline as bad news for SUSS. We read it as good news, and both cannot be right.

Our reading is that pushing hybrid bonding out to HBM4E extends the life of the thermocompression and mass-reflow flows, and each of those runs temporary bond and debond on every layer. The bear reading is that a hybrid-bonding delay signals the whole 16-hi node slipping, which defers every tool in the line, temporary bonders with it. Both fit the headline. They differ on whether what slipped is the joining technology or the layer count itself.

That distinction is the cleanest test of the thesis, and it resolves on a date. If 16-hi HBM4 volume arrives roughly on schedule and only the joining method moves out, the delay lengthens SUSS’s runway and the 12 percent sell-off was an overreaction. If the 16-hi ramp itself slipped, the layer-count compounding behind our base case slips with it and the market priced it correctly on the day. The H2 2026 process disclosures settle which, and so we hold a starter instead of a full position while a 12 percent tape print disagrees with us.

09 What would break this

The kill chain has two links. Either one trims SUSS’s share on its own; both together cap it, and that conjunction carries the 30 percent bear weight. The first link is process: SK Hynix’s mass-reflow flow needs fewer temporary bond and debond cycles per finished stack than the thermocompression flow Samsung and Micron run. No public source quantifies that difference, and we flag the gap instead of filling it with an assumption. The second link is competitive: EV Group’s laser-debond platform wins at 16-hi HBM4 volumes, a scenario Kepler Cheuvreux (21 May 2025) calls a future wildcard that could take share from TEL and SUSS if its laser proves superior. Both links holding sends the layer-count growth in our base case to someone else’s income statement, and makes the 2025 to 2026 margin trough correct pricing.

Three variables will tell us which way it breaks, in order of how soon we will know:

• SUSS’s own book-to-bill through H2 2026. The next hard read is the H1 2026 report, expected early-to-mid August. The tell is order intake converting into revenue; backlog alone proves nothing.

• Which final-joining process each memory maker ships at 16-hi. JEDEC’s thickness relief has already pushed hybrid bonding’s realistic debut to at least HBM4E, per TrendForce (7 Jul 2026), a delay that buys the temp-bond thesis time either way.

• Whether EV Group’s laser tools win a qualified design slot at Samsung, Micron or SK Hynix, beyond demo showcases. At that point the wildcard is a real threat.

The SK Hynix qualification headline therefore matters less than the process choice sitting behind it.

10 What €79.80 already demands

SUSS’s enterprise value sits close to its market cap: €1.526 billion on 19.12 million shares, no material net debt (yfinance, 23 Jul 2026). Grow that EV forward at a 9.5 percent WACC, a reasonable small-cap German-listed semicap discount rate, over four years to 2030, and it compounds to roughly €2.19 billion. Then ask what 2030 revenue, at management’s guided 20 to 22 percent EBIT margin, would justify that compounded EV at a given exit multiple. The base year matters as much as the multiple: SUSS guides FY2026 sales of €425 to €485 million, down from the €503.2 million FY2025 record, and every growth rate below runs off that midpoint of about €455 million.

To justify today’s price at a mid-case fourteen-times exit, SUSS needs about €746 million of 2030 revenue, a 13 percent CAGR off the FY2026 guided midpoint of €455 million. Management guides €750 to €900 million by 2030, so the mid-case bar sits just below the bottom of the guided band. Source: SUSS FY2026 guidance and 17 Nov 2025 Capital Markets Day targets, chokepoints.ai reverse-DCF, 23 Jul 2026. Chart by chokepoints.ai.

Management’s Capital Markets Day on 17 November 2025 guided €750 to €900 million of 2030 sales at a 9 to 13 percent CAGR off its 2025 base. The mid-case exit lands just under the bottom of it, so €79.80 pays for the floor of that range and nothing above it, and because 2026 is a guided down year, reaching even the floor takes a 13 percent CAGR off the FY2026 midpoint, the top of management’s own band. On this arithmetic the valuation is support and not an edge, which puts the whole upside on the share SUSS captures. We read the comps the same way:

The trailing-P/E column shows the trough trap: a 41 to 151 times spread across the set means none of these businesses trades near normalised earnings, and it says nothing about whether SUSS is cheap or its peers dear. Forward P/E is the less-distorted read, and even there SUSS sits below every direct peer. Berenberg’s own analyst, Amelie Dueckelmann-Dublany, raised her target to €125 on 19 June 2026, the day SUSS closed at its 52-week high of €111.70, arguing that market participants have not yet priced the ongoing momentum. The most bullish sell-side read on the name was published at the top, and it still sits inside our chokepoint-holds ceiling.

Multiples from yfinance at the 23 Jul 2026 close. Forward P/E moves with consensus estimates, which SUSS’s FY2026 guidance has already cut once.

11 Three scenarios, three different drivers

Probability-weight the three and the fair value is €95, about 19 percent above spot , with the base case near €100. The distribution is two-sided: the bull pays 88 percent and the bear costs 50, so we watch and size a starter. The single driver is the share SUSS captures as layer count climbs.

Sell-side targets, comps and the upper legs of our own DCF all sit above spot; only the bear leg sits below it. The 52-week range spans €24 to €118, so spot sits nearer the top of the market’s own band than the bottom. Source: public market data and chokepoints.ai model, 23 Jul 2026. Chart by chokepoints.ai.

12 Up 365, then down 29

The 52-week range runs €23.99 to €118.40, and the order of those prints matters more than the spread. SUSS bottomed at a €24.03 close on 2 September 2025, then rallied to €111.70 on 19 June 2026, a 365 percent move, and has given back 29 percent since. At €79.80 the stock is up 104 percent for 2026 and sits nearer the top of its own 52-week band than the bottom.

Both 2025 margin cuts land before the September low, so they cannot explain a derating that began in June 2026. The worst single day of the fall is 7 July, on a headline we read the other way. Source: XETRA closing prices via yfinance, company releases and TrendForce, compiled 24 Jul 2026. Chart by chokepoints.ai.

That ordering rules out the easy explanation. Both 2025 margin cuts landed before the September low, so they sit eight months and a 365 percent rally behind the current derating; they cannot be its cause. The fall from June’s high is a 2026 event, and its largest single day carries a date and a headline. Spot still sits below the comps-implied band and the sell-side range, with only our own bear leg beneath it, but a €40 bear case is a round trip to February 2026 rather than a catastrophe. A starter is sized for that, and the entry price is the weakest leg of this case.

13 The call

SUSS sells a step that every added HBM layer runs again, at a price that buys the bottom of its own management’s plan and nothing above it, after tripling off last September’s low. The call stays at watch with an accumulate tilt: a starter position, not a full one, until the H1 2026 report shows the margin closing and not only the order book improving. SK Hynix’s process choice, more than its supplier relationship with Tokyo Electron, decides whether the share caps, and until that prints, sizing past a starter prices a physics argument as a fact.

14 Scorecard

15 What moves the call

16 How we would express it

1\. Starter long SUSS, add on confirmation. Buy a starter near €80, sized well below a full position, and add toward the mid-€50s to €65 or on a second quarter of intake above revenue. The entry pays for the physics without needing the bull, and without pretending it is cheap after a triple off the low.

• Risk:* the margin trough proves structural, and the H1 report shows the order book improving with no margin following.

2\. Pair trade: long SUSS, short an expensive comp. SUSS trades at about 21 times forward earnings against 30 to 36 times for Besi, ASMPT, Aixtron and Camtek. A small, beta-aware long-SUSS, short-Besi pair expresses the cheapest seat in the set without a view on memory’s direction.

• Risk:* the comps are not a clean hedge, since their end-market mix differs from SUSS’s HBM concentration, and a sector de-rate moves both legs together.

Treat a Tokyo Electron or EV Group design-slot win at Samsung or Micron as the signal to cut, not to average down: the cleanest evidence that the share is capping.

Subscribe now

17 Dates to watch

The events that resolve the call over the coming months:

• Early-to-mid August 2026: H1/Q2 2026 report. The first hard read on whether Q1’s 4.3 percent EBIT margin was a fixed-cost trough or the FY2026 run rate.

• H2 2026: HBM4 16-hi process disclosures at SK Hynix, Samsung and Micron. Which final-joining technology each maker ships, the direct test of the kill chain.

• Ongoing: SUSS book-to-bill. Whether Q1 2026’s record order intake converts into revenue or piles up as backlog.

Two predictions we will mark, dated. First, SUSS’s reported H1 2026 EBIT margin prints above 6 percent, a narrowing from Q1’s 4.3 percent toward the guided floor, resolved against the H1 report by mid-August 2026. Second, SK Hynix converts its R&D-line toehold into a repeat or volume order by 31 December 2026, resolved against company disclosures and trade press. We expect the first and not the second; the second landing early would take conviction above Medium.

Not investment advice. For information and discussion only, and not a personal recommendation, offer or solicitation. Capital is at risk, and investments can fall as well as rise. The scenario fair values and price levels here are the publisher’s own estimates, not a forecast or a guarantee. Do your own work.

All figures come from SUSS MicroTec’s public disclosures, public market and industry data, and the publisher’s own calculations. Sell-side and industry analyst material is attributed by name and date wherever it informs the discussion above; it does not back any published figure or price level beyond what is explicitly sourced to it. © 2026 chokepoints.ai, Issue 007, SUSS sells the step that makes each HBM layer thin enough to fit inside the stack’s own height limit, and it sells the second-most of it. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

In 60 seconds

Powertech (PTI) does the last step in making a chip: it takes the finished silicon and packages and tests it so a computer can use it. It is Taiwan’s second-largest at that and the world’s fifth, and the leading specialist in packaging memory chips.

The AI twist is newer. The tightest bottleneck in an AI accelerator is no longer the transistors; it is the advanced packaging that wires the chip out to the rest of the system. TSMC’s version of that step, chip-on-wafer-on-substrate or CoWoS, is sold out and mostly goes to NVIDIA. Powertech is building a credible alternative on large panels, and AMD and Broadcom are reported (in trade press, not confirmed by either company) to be evaluating it, so it could work as an overflow valve for the AI demand TSMC cannot meet. Trade reporting puts that capacity booked through 2027, behind a record 2026 capital budget of about NT$50bn (US$1.6bn).

The market had paid for most of it, then gave a large piece back. Between 15 and 17 July the whole chip-equipment cohort de-rated on cycle-peak fears, and PTI fell with it: from NT$316.5 on 16 July to NT$266.5 by the 20 July close, about 31% below the NT$387 high and roughly 3.4 times its net asset value on an 11% return on equity. The question is no longer whether the chokepoint is real. It is whether enough has now come out of the price. Our answer: at NT$266.5, yes, in starter size. The selloff has put the price below every central value we compute, and inside the accumulate zone this note has carried since June.

We run this newsletter on a structurally bullish view of the AI build-out. At the June price our hard test failed: even with our own worldview in the valuation, the price already paid for the ramp, and we said so. The cohort selloff has changed the arithmetic, and the same model that said avoid at NT$333.5 now says the market prices less growth than our own base case. The discipline that kept us out is the discipline that now puts us in.

Subscribe now

01 Stack: the #2 memory OSAT, and a CoWoS overflow valve

A short primer. OSAT stands for outsourced semiconductor assembly and test. Chip designers design the silicon; foundries print the transistors; OSATs do everything after the wafer leaves the cleanroom. They cut the wafer into individual die, mount each die on a substrate or interposer, wire it up, encapsulate it, and run the burn-in and final test. The packaging step is increasingly where the value lives, because as transistors get smaller the bottleneck migrates upstream into how you wire a chip to the outside world.

PTI is Taiwan’s #2 OSAT and the world’s #5, with about 5.5% of the global OSAT market in 2024. It is the leading specialist in memory-IC packaging and test for DRAM and NAND, and its customer base is anchored by all three global DRAM leaders, Samsung, SK Hynix and Micron, who outsource portions of their memory packaging to it.

Two things to note about that customer mix. First, the DRAM leaders do not outsource everything; the highest-value memory packaging, particularly HBM stacking, stays largely in-house. PTI does not own the HBM TSV stack. Second, Micron acquired PTI’s Xi’an back-end facility in 2023 to internalise its China memory packaging. That cuts against a simple single-customer Micron-dependency story.

What is the chokepoint, then? Two of them. The first is conventional memory packaging, cyclical and competitive, riding the DRAM and NAND volumes of the AI server build-out. The second is the interesting one: advanced packaging for AI logic chips. When an AI accelerator needs to be packaged, die-on-interposer, with thousands of connections per mm², the dominant solution is TSMC’s CoWoS. CoWoS is sold out. It is the binding constraint on AI accelerator supply today. NVIDIA is the dominant customer; AMD, Broadcom and the custom-ASIC programmes form the queue behind it, and that non-NVIDIA queue is the natural overflow pool. Anyone building a credible alternative is, in effect, building an overflow valve. That is what PTI has done.

02 Technology: PiFO is a credible second source with no monopoly attached

The technology is the moat, so here is the technical bit. Fan-out wafer-level packaging embeds the die in a moulded wafer with redistribution layers, fine copper wiring that fans out from the die’s tiny contacts to larger package pads. Shorter interconnects, better electrical performance, lower cost at scale. Panel-level packaging (FOPLP) does the same on large square panels rather than circular 300mm wafers. The area-efficiency gain is enormous, but panel processing is hard; yields collapse if the panel warps, and only a handful of players are credible at it.

PiFO is PTI’s FOPLP architecture, with redistribution layers on both sides of the die. Trade reporting frames it as rivalling TSMC’s CoWoS-L; be precise about that. PiFO is a mould-compound panel fan-out, not a silicon interposer, so it is better read as a CoWoS-S/R-class alternative for AI-logic packages that do not need a large interposer, while the very largest accelerators stay on CoWoS-L. The distinction sets the runway: as flagship accelerators get bigger, more of the package volume migrates to CoWoS-L, the lane PiFO does not serve. Trade reporting put trial yields at 90% in September 2025 and at 95% by June 2026, with customer validation due in 2H 2026. Two caveats belong next to that figure: no broker source we have seen corroborates it independently, and a qualification yield on early product is not a volume yield; peer-reviewed work on large panels puts die-shift near 19 microns at the panel edge, which is exactly the failure mode that decides whether fine-pitch panel packaging survives mass production. The number that matters is sustained production yield at volume, and nobody outside the company has printed it.

On the demand side, the public reporting is that several major US AI chipmakers have turned to Powertech for the overflow, and that AMD and Broadcom back its FOPLP, with monthly FOPLP revenue targeted toward about NT$3bn. At announcement, in late 2025, that capacity was reported fully booked with demand near twice planned supply. The freshest broker channel checks, from spring 2026, are softer: clients adjusting quarterly orders, and FOPLP revenue estimates for 2026 and 2027 being trimmed. Booked is not the same as shipped, and the order book has already breathed once.

Three honest caveats. The AMD and Broadcom interest is reported, and brokers still class the datacenter engagements as market speculation rather than wins; the one confirmed AMD project is a console gaming APU, and Broadcom’s TPU packaging sits with TSMC and MediaTek. The order book is only partly visible; the full customer mix is undisclosed, so take it seriously but do not size it precisely. FOPLP is contested, not a monopoly: ASE is also expanding FOPLP capacity and TSMC has been preparing a CoPoS pilot line for 2026. PTI is one credible second source among several, which caps the long-run margin pool even as it leaves the next 18 months intact. And the HBM nuance matters: the HBM TSV and 3D-stacking process stays largely captive at the memory makers. The memory TSV and 3D-stacking packaging market is expected to grow from roughly US$2 to 3bn in 2024 toward US$10 to 15bn by 2028, but that growth accrues primarily inside the memory makers, not at PTI.

03 Numbers: the memory upcycle is already in the prints

The conventional half of the business is firing. Q4 2025 revenue was NT$21.41bn, up about 25.2% year on year, with net income of NT$1.86bn, up about 22.3%, its strongest quarter in three years. Q1 2026 sales were NT$21,314m, up about 37% year on year, with net income of NT$1,844m versus NT$1,175m a year earlier, up about 57%. April 2026 revenue reached NT$7.58bn, up about 32.5% year on year, and May followed at NT$7.92bn, up 30.2%. June came in at NT$7.62bn, up 21.7%, the sixth straight month in which the growth rate has come down: still a clean upcycle print, with the deceleration now impossible to ignore.

Be honest about the shape. Q1’26 net income (NT$1,844m) was essentially flat on Q4’25 (NT$1.86bn). The big year-on-year numbers are real, but they lap a weak 2025 base; this is a strong upcycle level; a sequential inflection has yet to show. The inflection the bulls need is the FOPLP ramp in the second half.

The context is a memory market in full upcycle. The HBM market is projected to grow from roughly US$35bn in 2025 toward US$45 to 55bn in 2026, with estimates varying by research house. PTI is a geared way to play that cycle without backing a single memory maker, because it packages and tests for all three. A reported dividend yield of about 2.6% is a modest cash return while the capex bet is in flight, and FOPLP spending could pressure even that if free cash flow stays negative through the ramp. The risk on this leg is duration, not direction: the prints are clean now, the question is how long the cycle holds in 2026.

04 Capex: a foundry-grade bet funded on OSAT earnings

This is where the hold gets uncomfortable. Management is putting real money behind PiFO: PTI raised its 2026 capex guidance to about NT$50bn (~US$1.6bn) in April 2026, up from an earlier NT$40bn plan and roughly double the prior year, primarily to expand FOPLP. Here is the number that should give a value-conscious investor pause. NT$50bn of capex against a revenue run-rate around NT$85 to 90bn is roughly 45 to 55% of revenue. That is at or above the top of foundry-grade capital intensity: across the recent build-out TSMC has run about 33 to 48% of revenue, peaking near 48% in 2022, and OSAT peers run 10 to 18%. PTI is spending like a leading-edge foundry on the earnings base of an OSAT.

The new panel lines in Hsinchu, P11 plus P12 (a former AUO display fab acquired for NT$6.9bn), carry the build-out, with a target of about 6,000 panels per month. The timing needs stating precisely, because the market has at times traded this as a mid-2026 story: first-phase panel equipment arrives over the summer of 2026, the chairman puts customer qualification in the second half, and volume shipments start from 1H 2027, building through that year. FOPLP is a 2027 revenue story with a 2026 proof-of-life. A capex programme of this size at a company earning roughly NT$8 to 9bn a year in net income is a heavy lift: NT$50bn is close to six years of current net income in a single year. Free cash flow will be negative through the heavy ramp in 2026, and the gap has to be funded, by the balance sheet, by debt, or eventually by equity, none of which the bull case prices. The whole ROE bridge depends on that NT$50bn converting into revenue at FOPLP margins. If it converts, the spend buys a seat in the AI advanced-packaging pool. If it does not, it widens an asset base that still earns an OSAT return, and the multiple compresses.

05 Valuation: every lens lands below today’s price

PTI closed at NT$266.5 on 20 July 2026, a market capitalisation of about NT$197bn on 738.8m shares, still more than double its 52-week low of NT$116 but now about 31% below the all-time high near NT$387. The run up was a brutal re-rating, from pricing PTI as a memory OSAT to pricing it as an AI advanced-packaging capacity owner; the give-back has come in two legs, the late-June round trip (NT$387 to NT$311 and back above NT$330) and then the 15 to 17 July cohort selloff that took the whole chip-equipment complex down together. Book value per share is about NT$79, putting the stock at roughly 3.4× book on an ~11% ROE. That multiple still needs the capex bet to lift ROE materially. The difference is that the market is no longer paying for the answer in advance.

Our cross-check runs FY27E EPS against a justified exit multiple, discounted one year at 10%. These EPS figures are judgemental, a scenario frame rather than a bottom-up segment model, and the biggest swing factor is the exit multiple:

Prob-weighted at 25/50/25, the EPS-model target is about NT$272, now roughly 2% above spot; the cohort selloff has closed the gap this lens spent all of June flagging. If you believe PiFO earns a durable second-source re-rating and holds 28 to 30× on bull EPS, the upside tail is roughly NT$560 to 600; we exclude that from the headline because we do not think a panel fan-out second source holds a foundry multiple, but an honest reader should know the tail exists.

Implied share price by method. The central estimate of every lens, the worldview DCF (NT$296), the Monte Carlo median (NT$285), the EPS cross-check (NT$272) and the Street average (NT$277), now sits above the NT$266.5 close. Source: public market data and chokepoints.ai model, 20 Jul 2026 close. Chart by chokepoints.ai.

The Street picture has inverted with the price. Consensus targets range from about NT$125 to NT$362.5, average near NT$277, on 6 buy ratings and 0 sells, and at NT$266.5 the price now sits below the average for the first time since the re-rating began. Zero sells still means no downgrade cushion on a print miss. The difference is that a buyer here is no longer paying above what the Street itself thinks the shares are worth.

5b | Worldview: even our bullish DCF lands below spot

Here is the honest objection to everything above. We run on a structurally bullish view of the AI build-out, a front-loaded, decade-long trajectory rather than a cycle. Is the verdict just an artefact of conservative, mean-reverting assumptions? If we let our own worldview drive the numbers, what does PTI become?

We tested exactly that. Instead of an exit multiple, we built a multi-stage free-cash-flow model and put the worldview into the demand path: structural revenue growth, margins expanding as the FOPLP mix rises, and, crucially, capex modelled explicitly, gliding from roughly 55% of revenue in 2026 down toward 15% as the build-out matures. That is the right way to value a capex-heavy name. The bet is spend now, harvest the structural demand later, and a DCF is the only honest way to see whether it pays at NT$266.5.

Weighting these 30/45/25, deliberately tilted toward our structural view, gives an intrinsic value of about NT$296, roughly 11% above spot. A Monte Carlo over the drivers (growth, margin, capex normalisation, discount rate) puts the median at about NT$285 and gives a 56% probability that intrinsic value clears today’s price, up from 36% at the June anchor.

Monte Carlo over growth, terminal margin, capex normalisation and WACC on the base-case path. About 56% of outcomes clear NT$266.5. The market is pricing roughly a 13.0% five-year revenue CAGR, below our 15.4% structural base for the first time. Source: chokepoints.ai worldview DCF on public figures, 20 Jul 2026 close. Chart by chokepoints.ai.

Then the clincher, run in reverse. At NT$266.5 the market is pricing roughly a 13.0% five-year revenue CAGR, below our structural base-case path of about 15.4% for the first time in this note’s life. In mid-June, at NT$363.50, the market priced more growth than even our bullish base case; at the July peak-fear close it prices less growth than we ourselves underwrite. That inversion is the whole reason the verdict moves. The margin of safety is real but not lavish: 56% of Monte Carlo outcomes clear the price, the median sits 7% above it, and the name still burns cash for two years on foundry-grade capex while sharing the durable rent with a bigger overflow winner. That mix is what starter size is for.

06 Market tape: louder demand, a cheaper stock, a more crowded packaging race

A note is anchored to a moment, so we read the live tape for the memory cycle, the advanced-packaging race and PTI itself, and map each event to the call.

The demand leg, the part we called the free option, got distinctly louder. TrendForce has conventional DRAM contract prices rising 58% to 63% in Q2 2026 after a 93% to 98% jump in Q1, with NAND up as much as 75%, and Micron’s 24 June quarter printed record revenue of about $41.5bn, more than four times the year-earlier figure, at a record gross margin near 85%, a once-in-cycle print, with supply called tight into 2028. PTI is the number-two memory packaging and test house, and that volume is already in its prints: April 2026 revenue rose 32.5% year-on-year and Q1 sales about 37%. The conventional memory-OSAT business that does the work while FOPLP is the upside is geared straight to this, and a cycle called tight into 2028 lengthens its runway.

The competitive leg, which is where our bear case lives, got more crowded in the same window. TSMC is expanding CoWoS and SoIC hard and has a CoPoS panel-level pilot line completing around June 2026, with mass production targeted for 2028 to 2029; ASE is accelerating a 310mm fan-out panel line, PTI’s most direct competitor. Read together they set the clock precisely: PTI has roughly a two-year FOPLP window before TSMC’s panel-level packaging and a larger rival compete head-on, which is exactly the booked-through-2027, unproven-beyond-it risk in the bear column. The reported AMD and Broadcom interest is the bull’s answer, a reason PTI might hold a seat outside TSMC’s walled garden, but it rests on trade-press reports of evaluations, with no confirmed win, and it does not extend the window.

Then the tape broke, and not on anything PTI did. A rout in US semiconductor shares on 15 July spread through Asia; Tokyo Electron and Advantest each fell about 9% on 17 July as the Philadelphia index touched bear-market territory intraday, with global semiconductor market value down an estimated US$3.3tn from its 22 June peak. The proximate trigger was TSMC’s own 16 July print: a strong quarter with a raised 2026 capex guide, read by a nervous market as evidence the cycle had peaked rather than as good news. PTI fell from NT$316.5 on 16 July to NT$266.5 by the 20 July close, in line with its equipment-cohort neighbours and on no company-specific news; its own June revenue (NT$7.62bn, up 21.7%) had landed a week earlier, and on 17 July its board approved a US$400m joint venture with Broadcom in Singapore for fine-geometry redistribution-layer substrates, the first time the Broadcom relationship has appeared as a signed structure rather than a trade-press evaluation. A selloff that lands a company inside our accumulate zone in the same week its most-caveated customer story hardens is the setup this note was built to catch. Three cautions keep the enthusiasm honest.

• The HBM stack is not PTI’s to package. The biggest memory dollars sit in HBM, and the HBM stack stays captive at the memory makers. PTI benefits through conventional memory packaging and test and the broader logic-packaging pull; the HBM stack itself passes it by, so the loudest part of the cycle flows to it the least.

• The second-source race tightened inside a fortnight. Trade reporting on 25 June has ASE’s automated FOPLP line targeting mass production by the end of 2026, against PTI’s qualification-then-1H27 schedule, roughly two quarters ahead; on 18 June TSMC and Amkor signed a definitive ten-year Arizona packaging agreement; and on 1 July ASE raised CoWoS and FoCoS quotes by more than 20%. Overflow pricing power is real and being printed, but the bigger rivals are monetising it first, and broker research now scores ASE and Amkor as the winners of the high-end overflow race while PTI’s lane runs through ASICs, CPUs and memory-adjacent work. One offset runs the other way: mid-June checks have TSMC’s CoPoS slipping, with volume production now unlikely before 2029 to 2030, which lengthens the panel window for everyone already standing in it.

• A cyclical tailwind is not the structural re-rating. The free option getting louder does not prove the FOPLP capex bet works; that is still PTI’s own execution, the qualification-and-ramp schedule we track at the end. A peak keeps the prints strong for a few quarters. It is no reason to pay past our margin of safety below about NT$285.

The tape has now done what we asked of it: the price crossed our accumulate line, and the demand half of the call strengthened while it happened. The durability half is still open, and the print that moves our number next is PTI’s own second-quarter result, due in late July, where what matters is whether the FOPLP mix lifts margins toward the high teens. Another revenue beat on its own tells us little.

07 Both sides: the price is paying for the bull

Both halves are real. The whole call is which half the price is paying for, and at NT$266.5 it no longer fully pays for either: the price sits below the base case on every lens we run. That is what an entry looks like.

08 Scorecard

Best-in-class AI advanced-packaging beneficiary, and the chokepoint survives scrutiny. For the first time since we opened coverage, the value leads the price: the cohort selloff has delivered the entry the June note said to wait for. We accumulate in starter size at NT$266.5 and add on P11 margin proof, sized for the NT$114 bear tail that has not gone anywhere.

09 Triggers: what would move us to buy or sell

10 The one number: the FOPLP ramp, qualification in 2H26, volume from 1H27

The FOPLP ramp: customer qualification in 2H 2026, then monthly panel output against the ~6,000-panel build-out through 1H 2027. We are precise here because the tape has not been: first-phase panel equipment arrives this summer, qualification is a second-half event, and volume is a 2027 event, so a mid-2026 “ramp” headline in either direction is noise. This schedule is the cleanest proof point that the capex bet is converting into capacity and that the CoWoS overflow thesis has physical inventory behind it. If qualification lands in 2H 2026 and output scales on the 1H 2027 schedule, the bull case is intact and the starter position earns its adds. If it slips, the multiple starts to look foolish, and ASE will be shipping panels first; the position is sized so that outcome hurts rather than wounds. The accumulate line we spent June waiting for is the price on the screen.

Step back once before the desk section, because the shape of this trade is the shape of the whole build-out. On our worldview the binding constraint keeps migrating down the stack: transistors were the chokepoint until 2024, CoWoS packaging is the chokepoint now, and the next squeeze is already visible in panel formats, HBM test intensity and, from late 2027, hybrid bonding. Each migration re-prices a new layer, and each opens a window for whoever built capacity just before the constraint arrived. PTI is a bet on exactly one of those windows. Windows are real money, but they are rented, not owned: the durable rent lands with whoever holds the qualified interface when the constraint moves again, which is why the equipment and materials one layer further down (the moulding presses, the substrates, the glass cores) keep showing up as the better chokepoints in this series. Pay for a window only at a price that does not need the window to last. That is the whole discipline behind the NT$272 to NT$285 line, and the 20 July close is the first print inside it.

11 If you want to trade it: expressions, not chases

The verdict above is the position: the bid we advertised at NT$272 to NT$285 has been filled from below, and the book expression is a starter long at NT$266.5. For readers who run books rather than watchlists, our desk pass (computed 2 July; the tape has moved, the structure of each idea has not) turned up expressions worth testing, with the numbers computed deterministically from the tape (realised volatility, pair z-scores, betas), not eyeballed. Treat them as ideas to test, never recommendations, and size for a stock whose 30-day realised volatility was running near 91% before the cohort selloff added to it. There are no listed options on this line, so every expression is cash equity.

• The core expression is the starter long, with the adds pre-committed. Own it at NT$266.5, add toward NT$230 to 250 on further cohort weakness or on FOPLP margin proof, and trim back into the NT$333.5 to NT$362.5 zone where every lens again sits below price. The reference levels: DCF-weighted NT$296, Monte Carlo median NT$285, Street average NT$277, all above spot. Risk: the bear case is NT$114; a memory-cycle turn plus a FOPLP slip gets there, which is why the starter is a starter.

• The relative trade, long TSMC short PTI, has had its reset. TSMC’s 16 July print (strong, higher capex guide) was the event the 2 July desk pass said to wait through, and the cohort selloff has since moved the spread violently in the short leg’s favour. With PTI now below every central lens we compute, the risk-reward of being short it against TSMC is gone; we retire the idea rather than update it. Risk: none retired ideas carry; the note is flat the pair.

• The statistical anomaly we are NOT taking. PTI versus ASE screens at a weekly z-score of −2.41, PTI cheap against its own history with its closest rival, which mechanically argues long-PTI short-ASE. We decline it, and the reason is the point: the fundamentals just moved against the statistics. ASE’s panel line targets volume two quarters ahead of PTI’s and it is already printing 20%-plus price rises, so some of that spread is a re-rating that deserves to stick, not noise that mean-reverts. When the quant signal and the second-source race disagree, the race wins.

• The patience play is a real position. PTI’s own second-quarter date is not yet confirmed, and its expected earnings-day move has historically been small (about 1.6% mean absolute over the last eight prints) against a 91-vol tape, so the event is not the edge. Cash, plus the alerts below, is the fourth expression.

For good order: none of this is investment advice, and the desk numbers above are the publisher’s own calculations from public price data.

12 Dates to watch

The late-July calendar is dense, and almost all of it is other people’s prints that read through to PTI. Dates are source-traced; PTI’s own results date was unconfirmed at publication.

• 10 July, landed : June monthly sales printed NT$7.62bn, up 21.7%, the sixth consecutive deceleration. The 10 August July print shows whether the rate stabilises.

• 16 July, landed : TSMC’s second-quarter beat with a raised capex guide was read as cycle-peak evidence and set off the cohort selloff that delivered our entry. CoWoS commentary stays the overflow complex’s metronome.

• 27 July : Amkor second-quarter results. The US OSAT read-across, now carrying the ten-year TSMC Arizona agreement.

• 30 July : ASE second-quarter results. The single most important cross-read: the rival panel line’s ramp commentary and whether the 20%-plus quote rises are holding.

• Late July, unconfirmed : PTI second-quarter results, the margin print our number waits on.

• 23 September : Micron FQ1 results. The memory-cycle read for the conventional half of the business.

One layer further out, the schedule that decides the thesis: FOPLP customer qualification in 2H 2026, volume shipments from 1H 2027, ASE’s rival line at end-2026, and TSMC’s CoPoS now checking toward 2029 to 2030. The window is long enough to pay the patient and short enough to punish the late.

13 Go deeper

The public sources behind this note, if you want to take the thesis apart yourself. Each opens in a new tab.

The thesis: CoWoS overflow to Powertech.

• TrendForce: TSMC CoWoS crunch pushes US AI chipmakers to Powertech through 2027, the core overflow-valve report.

• Taipei Times: Powertech board approves US$400m Broadcom joint venture in Singapore, the customer story turning into a signed structure, 17 July 2026.

• TrendForce: AMD Zen 7 with Powertech FOPLP reportedly under evaluation, the AMD customer signal.

• TrendForce: ASE targets FOPLP mass production by end-2026, the rival line running two quarters ahead.

• TrendForce: ASE raises advanced-packaging quotes by more than 20%, overflow pricing power, printed.

• Taipei Times: chairman Tsai on the FOPLP schedule, qualification 2H26, shipments from 2027, from the company itself.

The technology and the competition.

• TrendForce: FOPLP heats up, ASE and Powertech expand, TSMC preps a 2026 CoPoS pilot, the case that this is a second source with no monopoly.

• TrendForce: FOPLP trial yields reportedly reach 90% at Powertech, the yield question that decides the ramp.

The capex bet and the roadmap.

• DigiTimes: Powertech lifts capex to US$1.6bn for AI packaging, the April 2026 guidance raise.

• DigiTimes: AMD and Broadcom back Powertech’s FOPLP; monthly revenue toward NT$3bn, the named backers.

• DigiTimes: Powertech secures full FOPLP bookings ahead of its expansion, the P11 6,000 panels/month target.

• DigiTimes: Powertech to accelerate FOPLP, mass production targeted 1H27, the P12 roadmap.

The company, the cycle and the worldview.

• Yahoo Finance: Powertech (6239.TW), live price, valuation and financials.

• Taipei Times: Micron’s acquisition of Powertech’s Xi’an plant, the 2023 Xi’an transfer.

• Situational Awareness (Leopold Aschenbrenner), the structural AI-build-out worldview this note reasons from.

Not investment advice. For information and discussion only, and not a personal recommendation, offer or solicitation. Capital is at risk, and investments can fall as well as rise. The scenario fair values and price targets here are the publisher’s estimates, not a forecast or a guarantee. Do your own work.

All figures come from Powertech’s public disclosures, peer public market data, and the publisher’s own calculations. No third-party research is quoted or reproduced. Sources: Situational Awareness; Powertech public disclosures; chokepoints.ai scenario and DCF work. © 2026 chokepoints.ai, Issue 004, Powertech sits at the advanced-packaging layer of the AI chokepoints map. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

Compression moulding seals the finished stack in epoxy; a hybrid-bonded stack still gets encapsulated, but the job shrinks and moves away from the moulding press TOWA sells. Source: facts-public.json and TOWA IR; schematic by chokepoints.ai, 21 Jul 2026.

In 60 seconds

Eight or nine of every ten HBM stacks are sealed in their epoxy shell on a TOWA press. Compression moulding is the step, TOWA is the default supplier, and the AI build-out pulls the tool directly. That much is not in dispute. Between 15 and 17 July the shares fell 17.9 percent in two sessions to ¥2,791, as the whole Japanese chip-equipment cohort de-rated together; Tokyo Electron fell 9 percent and Advantest 9.4 percent the same week on fears that AI capital spending has peaked. Nothing TOWA-specific caused it. The stock has since recovered part of the move, to ¥2,848.

At that price the gap between what the market pays and what we can defend has narrowed but not closed. A reverse-DCF says the tape still prices about 17.2 percent revenue growth a year for five years; our defensible base is 8.8 percent, and even management’s own plan delivered in full runs at about 12.8 percent on revenue. Close that gap and the weighted fair value sits 34 percent below spot, though for the first time the bull case, management’s own Vision 2032 plan executed in full, now clears spot rather than falling just short of it. We hold the chokepoint and wait for a better price.

Playing devil’s advocate

The strongest case against us says the correction did the valuation work already: a stock that fell 21 percent from its high in a week is no longer priced for perfection, and buying chokepoints into a sector panic is usually the trade. We grant the premise. The cohort-wide fall is exactly the kind of indiscriminate selling that creates entries in names whose fundamentals never changed, and TOWA’s own order book, per the May 2026 results call, was running at a record for compression-mould dies on AI data-centre demand. The arithmetic still answers the case: even after a 21 percent drawdown, the weighted fair value is ¥1,872, a third below the new spot, and the market-implied growth rate of 17.2 percent a year is still double our 8.8 percent base and above the 7 to 8 percent the moulding-equipment market itself compounds at. A cheaper multiple is not the same as a cheap stock when the cash flows have not moved. The correction closed part of the gap. It did not close it.

01 TOWA owns the seal every HBM stack needs before it ships

A finished HBM stack is eight to sixteen DRAM dies threaded with through-silicon vias on a logic base, and it cannot survive board assembly unmoulded. Compression moulding wraps the stack in a thermoset epoxy shell that protects it mechanically, blocks moisture, helps move heat and leaves a flat surface for the next step. JEDEC has already lifted the HBM4 module-height limit from 720 to 775 micrometres and is weighing a move toward roughly 900 for HBM4E; the compound has to flow into a taller stack, encapsulate the connections, and cure without a void, a bow or a cracked die. Transfer moulding, the older process, cannot do this reliably at that height. Compression moulding can, and TOWA holds 80 to 90 percent of the units doing it.

The moat is switching cost, not a patent wall. TOWA makes its own moulds by electroforming, a process that duplicates a mould in about a day, not weeks of machining; its BANCERA coating cuts release force and cleaning frequency; its CBN end mills finish pockets to a few nanometres of roughness. A rival would need years to qualify an equivalent chain at a memory maker. That defence protects a step for as long as the step exists, which is the subject of section 05. The schematic at the top of this note lays out the full chain: dies, TSV, the moulding step TOWA owns, and the dashed hybrid-bonding bypass that section 05 covers in full.

02 A focused tool-maker with a record order book going into the print

TOWA is a Kyoto tool-maker founded in 1979, built almost entirely on one process it won early: transfer moulding first, then compression moulding as packages shrank and stacks grew taller. Semiconductor Manufacturing Equipment is the business; Medical Device and Laser Processing are minor arms. Leadership passed cleanly in April 2025, Hirokazu Okada to Chairman and CEO, Muneo Miura to President. It is a razor-and-blade model: systems (compression and transfer moulding presses, the LGS1040 laser singulation platform) are lumpy and cyclical, an advanced compression system running upwards of US$1.5m a unit against management’s guide of 30 to 45 HBM-related systems a year; the aftermarket (precision moulds, BANCERA coatings, release films, field service) carries higher margins and renews on a cadence that outlives any single order cycle.

The fiscal year ended March 2026 was the trough for margins and the record for orders at the same time: sales grew 1.7 percent while operating profit fell, but AI data-centre demand pushed compression-mould die orders to a record and management guides FY3/27 sales up 17 percent. The company also introduced INNOMOS, a next-generation compression system claiming roughly half the mass-production cost and double the productivity of older lines, positioned as the mechanism for the Vision 2032 margin target. None of this is the cause of July’s fall; TOWA’s own next print, Q1 FY3/27, is not due until 6 August 2026. The order book going into that print is the strongest evidence the bull case has for durability; the trough margin is the strongest evidence against it, and section 04 takes both in turn.

03 The 15 to 17 July fall was a cohort de-rating that TOWA rode down

TOWA fell 6.9 percent on 16 July to ¥3,165, then a further 11.8 percent on 17 July to ¥2,791, a two-session fall of 17.9 percent, before recovering to ¥2,848 on 21 July. No TOWA-specific news explains it. What explains it is a sector-wide repricing of AI-capital-spending risk that hit the entire Japanese chip-equipment cohort at once. A rout in US semiconductor shares on 15 July, Micron falling on intensifying competition fears from Chinese memory makers, dragged the Philadelphia Semiconductor Index sharply lower and pulled South Korea’s KOSPI down more than 7 percent intraday. Tokyo followed on 16 July: the Nikkei 225 reversed 1,916 points, with Advantest the single biggest drag on the index. The rout deepened on 17 July, when the Nikkei fell further and the Philadelphia index touched bear-market territory intraday, more than 20 percent below its June peak, before paring the fall by the close; Tokyo Electron fell about 9 percent and Advantest about 9.4 percent that day alone. Two forces compounded the move: TSMC had just raised its 2026 capital-spending guidance, and the market read a strong print with a higher guide as evidence the cycle had already peaked; separately, a new model from the Chinese AI lab Moonshot narrowed the perceived gap with the leading US labs, reviving doubts about how much compute the buildout still needs. Global semiconductor market value fell by an estimated US$3.3tn from its 22 June peak. TOWA fell harder than the megacaps, 17.9 percent over two sessions against single-day falls of about 9 percent at Tokyo Electron and Advantest, which is what a ¥214bn small-cap’s beta does inside a cohort selloff; Hanmi, its closest small-cap comparator in the HBM tool chain, fell a similar 17 percent over the same fortnight.

Hanmi’s fall came on the same cohort-wide repricing, with no bad news of its own either. It is the clean test of the sector-beta read. A company-specific shock would have hit TOWA and spared its neighbours; instead the whole cohort, moulding and bonding vendors alike, fell together. This matters for the call: a fundamental shock would force us to re-underwrite the fair values in the table above. A multiple compression that leaves the cash flows untouched does the opposite; it makes the entry better without changing what TOWA is worth. If the peak-capex read behind the selloff proves right, it reaches TOWA through the FY3/27 guide, and the Cyclical case (¥1,098, weighted 18 percent) is where the model already prices it; the 6 August print is the first test.

The 52-week range and the four confirmed July closes are the only exact points; the path between them is illustrative. Source: facts-public.json (52-week range) and desk-verified yfinance closes, 21 Jul 2026.

04 The margin bet is the same bet it was, just at a lower entry

TOWA’s profitability is the swing factor in its own valuation. FY3/26 was a trough: the trailing net margin compressed to 8.4 percent from 15.2 percent, against an FY3/25 that earned a 16.6 percent operating margin on ¥53.47bn of sales. Vision 2032 targets ¥100bn of sales at a 25 percent operating margin, with a guided recovery toward 16 percent as soon as FY27. The distance between an 8.4 percent trough and a 25 percent target is the distance the bull case has to travel, and INNOMOS is management’s named mechanism for travelling it, not evidence it already has.

The workforce shape is the quiet tension underneath the plan. TOWA runs about 2,099 consolidated staff at March 2025, rising toward 2,211 a year later, 1,906 of them in semiconductors, average age 39.5 and average tenure 11.3 years. That depth is precisely what a chokepoint needs to defend a step against qualification by a rival. It is also a thin base from which to deliver a near-doubling of revenue at a record margin while standing up INNOMOS and adjacent product lines. The bull case in the signal box, a 24 percent terminal margin, sits just under the 25 percent Vision 2032 target; it assumes the plan lands close to as written.

05 Hybrid bonding is the expiry date the price still does not carry

Every defence in section 01 protects a step that could stop existing, and the market has priced that risk as remote instead of merely deferred. Hybrid bonding joins dies copper-to-copper with no bump and no gap to underfill, and a stack bonded that way needs a smaller, different encapsulation job. TOWA has no public hybrid-bonding hardware. Besi leads the field with 200-nanometre placement accuracy at three sigma; ASMPT has taken its first HBM hybrid-bonder order and targets 35 to 40 percent of chip-to-substrate bonding; Hanmi has taken a US$30m-scale SK Hynix order for its Griffin thermocompression line, a bonder win rather than a hybrid one, and has its own hybrid bonder on a roadmap with a dedicated factory due in H1 2027. A step that evolves rewards TOWA’s installed base. A step that is removed hands the work to a different machine from a different vendor, and TOWA has no seat at that table.

The saving grace is timing. SK Hynix has said it will keep MR-MUF while it refines hybrid bonding, and all three memory makers are expected to stay on microbump-plus-underfill through HBM4. The threat most likely arrives at HBM5 or HBM6, perhaps from around 2031, with no published date. Nothing in the news sweep behind this re-anchor changes that picture: no maker has announced an election either way since the prior anchor. The Kill case prices the full chain: a memory maker elects hybrid bonding at HBM5 or HBM6, new lines carry no moulding step to sell into, systems demand decays toward the aftermarket, the terminal operating margin settles near 11 percent, and fair value lands at ¥929, the bottom line of the scenario table. It carries a 22 percent weight because it is deferred, and it is the one line this note will not soften on a price move alone.

06 A short customer list is pricing power and hostage risk at once

Downstream, the customer pyramid is short: an OSAT layer led by ASE (44.6 percent of a top-ten that turned over US$41.56bn in 2024), and a memory layer of SK Hynix, Samsung and Micron, with SK Hynix holding the majority of HBM. TOWA sits on consortia that write the standards, one of 27 members of Resonac’s JOINT3 panel-level group and a member of the US-JOINT advanced-packaging consortium launched in April 2026. SK Hynix’s ¥19 trillion packaging-and-test build at Cheongju, due around end-2027, is a direct downstream pull on the tool line. That concentration is why TOWA can charge what it charges, and why the next order pause, with 88.4 percent of sales overseas, will originate abroad and arrive without much warning; the 15 to 17 July fall shows what that kind of surprise looks like when it arrives market-wide.

Upstream the risk is mild: the bought-in content of a press is standard industrial automation from Yaskawa, Mitsubishi, FANUC and Nidec, and the epoxy moulding compound comes from a plural field led by Sumitomo Bakelite. No single supplier can hold TOWA to ransom; compound pricing is simply one channel into the margin story in section 04.

07 TOWA owns the narrow step almost outright and competes for the rest

The near-monopoly label is correct only on the narrowest of three markets. TOWA holds 80 to 90 percent of HBM compression moulding, a genuine near-lock; 64.8 percent of the broader moulding-equipment market in 2024, up from 59.3 percent, still gaining but plainly contestable; and roughly 22 percent of compression-moulding equipment of every kind, not dominance at all. The Japanese trio of TOWA, Yamaha and Takara holds about 58 percent of compression-moulding capacity between them, with Besi near 14 percent. Flattening these three numbers into one figure is the single most common error in the bullish telling.

08 Four risks sit ahead of everything else in this note

• The hybrid-bonding tail. Besi, ASMPT and Hanmi own the technology that could end the franchise; TOWA has no product answer if a memory maker elects it at HBM5 or HBM6.

• A cohort-wide de-rating with no cushion. TOWA’s relative cheapness inside its cohort is no defence against the cohort re-rating together, which is what just happened and could happen again.

• A margin plan that is still a guide. Vision 2032 and INNOMOS are the mechanism management names; the FY3/27 Q1 print on 6 August is the first hard evidence either is landing.

• Customer-calendar concentration. SK Hynix, Samsung and Micron set the order book between them, and 88.4 percent of sales sit overseas with a currency exposure attached.

09 The new price still asks for growth well above the moulding market

Two lenses value TOWA and they still point in opposite directions. On the relative lens, the selloff compressed the whole cohort, so the comparison below uses one consistent basis: trailing multiples at the 17 July post-selloff closes, all from the same pull. The AI-equipment names now sit in a tight band, Tokyo Electron and Disco at 52.1 times, Advantest at 53.5, Kokusai at 68.5, with SCREEN, the broad-equipment bracket, at 33.9. TOWA at ¥2,848 pays 46.6 times its FY3/26 earnings of ¥61.1 a share, or 37.6 times consensus forward earnings of ¥75.70: still the cheapest of the HBM-exposed set, though by a narrower margin than the pre-selloff numbers suggested. Applied to TOWA’s trailing earnings, the cohort multiples imply ¥3,184 to ¥4,187; SCREEN’s implies ¥2,072. The relative lens still brackets spot from above.

The intrinsic lens is ours. A reverse-DCF, run at the new ¥2,848 close on the same public spine as the prior anchor (launch revenue ¥54.37bn, WACC 9.0 percent, terminal growth 2.0 percent, ¥20bn of net cash added to enterprise value), backs out market-implied growth of about 17.2 percent revenue a year for five years, down from 19.1 percent at the old anchor but still double our defensible base of 8.8 percent and above the moulding-equipment market’s own 7 to 8 percent CAGR. The bull case’s own revenue path, the one that assumes Vision 2032 lands, runs at about 12.8 percent, itself below what the market prices in; the bull case only reaches ¥3,137 through a 24 percent terminal margin, not through faster growth than the market already assumes. This is the sharpest way to see the gap: even crediting management’s full plan on revenue, the market still wants more.

The bull case reaches its ¥3,137 fair value through a 24 percent terminal margin, not through revenue growth above what the market already assumes. Source: dcf_worldview.py public spine, recomputed at spot ¥2,848, 21 Jul 2026.

Peer trailing multiples computed close over trailing EPS at the 17 July post-selloff closes (yfinance, pulled 21 Jul 2026), applied to TOWA’s FY3/26 EPS of ¥61.1. One basis across all rows; the prior draft mixed forward and trailing multiples of different vintages.

10 Four scenarios, and for the first time spot clears one of them

The scenarios each flex a different driver instead of the same thesis dialled up and down. The spread is informative rather than a menu for exactly that reason. None of the fair values below has moved since the prior anchor; only the gap to spot has. The Kill weight is the call’s key sensitivity; section 05 carries the evidence for holding it at 22 percent, and a confirmed election at any memory maker moves it, and the verdict, immediately.

Probability-weight all four (bear Kill at 22 percent, SA-base at 35 percent, SA-bull at 25 percent, and an 18 percent-weighted Cyclical air-pocket case at ¥1,098, not pictured above) and the fair value is ¥1,872, 34 percent below the new spot. A 20,000-draw Monte Carlo over growth, margin, WACC and the yen puts the median at ¥1,689 and the tenth-to-ninetieth band at ¥869 to ¥3,148; at the prior anchor only about 10 percent of draws beat spot, and at the new ¥2,848 close, where spot now sits between the published P75 (¥2,435) and P90 (¥3,148), perhaps 15 to 20 percent of draws clear it, still a minority. A separately built FCFF engine (the Sturzo MIT model, fed our public spine) lands at ¥1,671, a 41 percent discount to the new spot, with its own median at ¥1,776. Two models built differently still agree within about 10 percent of each other.

The intrinsic lenses (worldview scenarios, Monte Carlo, Sturzo cross-check) sit below spot; the relative lenses (peer multiples, Street consensus) sit above it. That split is the entire call. Source: dcf_worldview.py, dcf_crosscheck.py and facts-public.json, recomputed at spot ¥2,848, 21 Jul 2026.

Street consensus range of ¥3,100 to ¥4,000 (four-analyst average ¥3,425) per Investing.com, 21 Jul 2026. We could not independently confirm a current ¥4,000 target as Jefferies’ own; Jefferies has repeatedly revised its TOWA target through 2025 and 2026 and we did not treat any single bank’s number as load-bearing. We use the same analyst pool’s forward EPS and reject its average target: the disagreement is the multiple, roughly 45 times forward earnings in the ¥3,425 average against the 25 to 30 times our weighted case implies, and the scenario table prices that disagreement through margins and growth rather than assuming it away. A separate four-analyst average of ¥12,500 to ¥13,800 reported by one data aggregator has the signature of a data error, not a Street view, and we still do not use it.

11 The market has closed part of the gap it never should have priced

Spot fell 21 percent from the 52-week high of ¥3,595 to the new close of ¥2,848, entirely inside a two-session cohort de-rating with no TOWA-specific cause. The effect on the call is arithmetic, not qualitative: every fair value in this note is unchanged, and every gap to spot has shrunk. The weighted fair value moved from 41 percent below the old anchor to 34 percent below the new one; the bull case moved from 1 percent below the old anchor to 10 percent above the new one. A cohort re-rating that leaves the cash flows untouched makes the entry better without making the story better; the entry price is the whole difference between this note and the one it replaces.

12 We hold the chokepoint and wait for a better price

Both halves are real, and the reverse-DCF still answers which one the new price is paying for: mostly the bull, with less spare room reserved for the bear than a month ago. We hold, add only into the ¥1,872 to ¥1,957 intrinsic zone, and trim on either trigger named in section 14: a confirmed hybrid-bonding election, or the 6 August print missing the margin recovery while orders roll over. Between those levels the position waits.

13 The scorecard shows a real chokepoint still priced above its own cash flows

14 What moves the signal from here

15 How we would express it

1\. Hold the position, add only on a further pullback. The bull case now clears spot; the weighted case does not. Size for a further leg down toward the ¥1,957 base case before adding, and treat ¥2,848 as a level to watch, not to buy.

Risk: a confirmed hybrid-bonding election or a second cohort-wide leg down would take the shares through the base case before the entry improves further.

2\. Relative-value pair: long TOWA, short the more expensive cohort names. TOWA’s own multiple (46.6 times trailing) sits below Disco and Tokyo Electron (52.1), Advantest (53.5) and Kokusai (68.5) at matched post-selloff closes; a pair expresses the cohort-discount thesis without taking a view on whether the whole group re-rates further.

Risk: if the cohort recovers together, TOWA’s laggard beta could mean the short leg outperforms the long leg even though the relative call is correct.

16 Dates to watch

The events that resolve the call over the coming months:

• 6 August 2026: FY3/27 Q1 results. The first hard evidence on whether the margin recovery and the INNOMOS-led plan are converting the record order book into an operating margin above the 8.4 percent trough.

• Ongoing: any HBM-generation election of hybrid bonding over compression moulding. Not a roadmap slide, a real order line or a credible channel check. This is the one event that would force a re-underwrite of every fair value in this note.

• Ongoing: the cohort. A further AI-capex de-rating across Tokyo Electron, Advantest and Disco would confirm the 15 to 17 July fall as the first leg of a longer repricing, not an air-pocket; a recovery across the cohort would confirm the opposite.

TOWA still owns the step; the market has simply stopped paying for a flawless decade of it in full. Not investment advice. For information and discussion only, and not a personal recommendation, offer or solicitation to buy or sell any security. Capital is at risk, and investments can fall as well as rise; past performance does not indicate future results. The scenario fair values here are the publisher’s own estimates, not a forecast or a guarantee.

All figures are from TOWA’s public disclosures, public market and industry data, and the publisher’s own calculations, current to 21 Jul 2026 unless otherwise dated. No third-party research is quoted or relied upon beyond what is explicitly attributed by name and date above. © 2026 chokepoints.ai, Issue 006, TOWA sits at the moulding sub-step of the AI advanced-packaging stack. The chokepoint is real. The price is closer to fair, not fair yet. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

The chokepoint comes from indispensability: an AI cluster cannot run without memory bandwidth and no substitute exists at the bandwidth it needs. Three firms supplying almost all of it is what concentrates that dependence in Micron’s hands. The catch is that its earnings swing harder than almost anything else in the chain. Source: public disclosures and industry data; schematic by chokepoints.ai.

In 60 seconds

Micron makes memory chips: dynamic random-access memory (DRAM), the working memory in every computer, and high-bandwidth memory (HBM), the stacked DRAM that sits next to every AI accelerator. An AI-driven shortage has sent memory prices to records, and Micron just printed a blowout: quarterly revenue of $41.46bn, record gross margins, and earnings many times the prior year’s. In the fortnight since the print, profit-taking, an antitrust suit against the three DRAM makers and a rival’s imminent US listing have cut the shares about 25 percent from their high, to $938.38, still up more than 700 percent in a year at a market value above $1tn.

At about ten times this year’s earnings the stock now trades at a normal through-cycle multiple on peak earnings. Memory is cyclical because high prices fund new factories, the capacity lands in a lump roughly two years later, and supply overshoots demand. The market knows this, which is why memory earns its lowest multiples at the very top. It comes down to one number: whether a gross margin in the mid-80s can hold, or whether the wave of new factories now under construction tips the market back into oversupply around 2028. At $938 the price matches our probability-weighted fair value; we hold, and we would add near $750.

Playing devil’s advocate

The strongest case against us now comes from the dip-buyers. Nothing in the business earned a 25 percent slide: the quarter was a record, the guide went higher, HBM is sold out through 2026, DRAM contract prices rose again in June, and the revenue sits on $22bn of customer commitments under take-or-pay contracts (the customer pays whether or not it takes delivery). The selling came from rebalancing after a huge first half, a lawsuit that will take years, and nerves ahead of a rival’s US listing; several banks have already called the pullback temporary. On this reading, ten times peak earnings for the strongest memory cycle on record is the entry the June price never offered. We grant the mechanics: the selling changed the price, and it left the margin outlook exactly where it was. Our answer is what the price now matches. At $938 the market is paying for our base case, the one where demand stays strong and margins still ease toward the mid-60s. Fair value is a fair deal. The price we want pays us for the bear, and it sits near $750.

01 Micron holds a real chokepoint on the most cyclical layer of the stack

Micron sits on the memory rung of the AI compute stack, the layer that supplies the DRAM and the high-bandwidth memory stacked beside every accelerator. It is a genuine chokepoint, and the definition matters: an AI training or inference cluster cannot run without memory bandwidth, no substitute exists at the bandwidth AI needs, and the supply is controlled by just three companies. The dependence creates the chokepoint; the three-firm structure decides who collects the rent. Micron, SK Hynix and Samsung together hold about 95 percent of the DRAM market, which is as concentrated as any layer in the chain outside the foundry itself.

A genuine chokepoint inside a tight oligopoly, and the most cyclical link in the stack. After the slide, the price pays for a soft landing in the margin.

The difference between memory and the logic foundry is the cycle. Leading-edge logic is concentrated and its pricing grinds higher; memory is an oligopoly whose product is closer to a commodity, and prices swing violently because capacity is added in multi-billion-dollar fab increments that take about two years to build, so supply arrives in lumps and regularly overshoots demand. Right now that balance is the tightest it has been in a decade, AI has pulled an entire generation of memory into the data centre, and Micron is earning record margins. The chokepoint itself is not in question. What matters is how long the margin it throws off can last.

02 The record quarter has been reported, and the market has traded through it

Unlike the other names in this set, Micron has already delivered its make-or-break print. The fiscal third quarter, reported in late June, was a record on every line, and the fourth-quarter guide went higher still:

The catch is what the market did with it. The quarter was extraordinary, and within two weeks the shares traded below their pre-earnings level anyway. The next print, on 23 September, carries one job: show the margin is holding. The bull case rests on a peak margin proving durable rather than cyclical, so the gross-margin guide is the number to watch, well ahead of the headline.

03 The contract book is the genuinely new feature of this cycle

Three things make this cycle different from the ones before it. The first is structural: a three-maker DRAM oligopoly that has, for now, chosen supply discipline over share wars, which is what lets prices run. Conventional DRAM contract prices roughly doubled inside two quarters at the start of 2026, and industry price trackers expect them to keep climbing through the year.

The second is the HBM seat. High-bandwidth memory is the part of the market AI needs most, and Micron now has a place at the table: it began volume shipments of its newest HBM for Nvidia’s next platform early in 2026. The standing is still modest. SK Hynix holds well over half the HBM market and Samsung is second; Micron’s share, low single digits a year ago, is now around a fifth and climbing fast. The seat is real and improving, but a long way from a lead.

The third is the contract book, the genuinely new thing. Micron has signed sixteen long-term, take-or-pay agreements running three to five years, and added two more since the June print, with General Motors on 1 July and Ford on 7 July, extending the model into automotive memory. The sixteen carry about $22bn of near-term financial commitment and roughly $100bn of revenue through 2030 (the gap is timing, the first a down-payment on the second), backed by around $18bn of customer cash deposits, and they lock in roughly a fifth of Micron’s DRAM and a third of its NAND flash, the storage memory it sells alongside DRAM. That genuinely changes the shape of the downside. It has also never been tested in a real downturn, and that test is the whole question.

04 Record cash is funding the capacity that will end the boom

Micron makes money by selling memory, and at this point in the cycle it is making an enormous amount of it. The third quarter threw off $25.39bn of operating cash flow and about $18.3bn of adjusted free cash flow against $7.1bn of capital spending. Data-centre revenue rose more than seven-fold from a year earlier, and cloud memory revenue more than quadrupled. These are the numbers of a company at the very top of its cycle.

The problem is where the cash is going. The boom that produces the record cash is paying for the capacity that will end it: Micron is lifting capital spending sharply, including a new megafab, and its two Korean rivals are expanding in parallel. That spending is committed at the top of the cycle, which is why a memory balance sheet looks its best just before the turn.

05 The call turns on whether the gross margin holds

Everything comes back to one line: the gross margin. Micron’s reached the mid-80s last quarter and is guided near 86 percent next. The bull case needs that level, or close to it, to hold for years. The bear case is just the industry’s own history: memory has never sustained a margin near today’s for long. Prior peaks topped out far lower, in the 50s and 60s, and even those faded within six to eight quarters as new supply caught up. This run is already that far in. The swing factor is how far the margin falls when the cycle turns, and when that is.

Prior memory peaks topped out well below today’s mid-80s and faded within a couple of years. Our base case has the margin easing toward the mid-60s by fiscal 2028. Source: company disclosures and chokepoints.ai estimates, 29 Jun 2026. Chart by chokepoints.ai.

The timing evidence cuts toward caution. Micron’s own chief executive expects industry supply to improve gradually in 2028, which is management itself putting a clock on the tightness. Independent forecasters see DRAM and NAND prices peaking around the middle of 2026 and beginning to roll over as early as 2027, about a year sooner than they had thought. The supply agreements cushion that roll, because a fifth of DRAM and a third of NAND now carry minimum pricing; they cannot prevent it. The question on the September print, and every print after it, is whether the guide defends the mid-80s or begins the walk down toward the 60s our base case assumes.

06 HBM demand runs through a few platforms, and China remains a live risk

The concentration that matters most is in HBM, where the demand runs through a very small number of accelerator platforms. Micron’s HBM fortunes are tied closely to Nvidia’s roadmap, and the competitive set is shifting underneath it: one large accelerator programme has named Samsung rather than SK Hynix as its lead HBM partner, with Micron a smaller participant. A seat that depends on one or two platform winners is a narrower base than the headline memory share suggests.

The contract book concentrates the other way, and more favourably: the take-or-pay agreements spread roughly $100bn of future revenue across sixteen named customers through 2030, backed by cash. That is real diversification of the downside. Geography is the live risk on the other side. Micron’s China business, once around a tenth of revenue and worth roughly $3.4bn, was cut sharply after a server-chip ban, and the memory names remain exposed to export politics on both the demand and the equipment side. The contract book is the strongest concentration story Micron has ever had; the customer and the geopolitics are where it is thin.

07 The market is paying a $1tn-plus valuation for the number-three HBM position

The competitive set is the oligopoly itself, and the positioning differs by product:

Spot pays about ten times peak earnings, the through-cycle norm; the June high paid nearly thirteen. The question has moved from the multiple to the earnings under it. Source: public market data and chokepoints.ai model, 7 Jul 2026. Chart by chokepoints.ai.

That ranking matters for the multiple. The market is paying Micron a market value above $1tn, roughly three times what it was worth at the last cycle peak, for the number-three position in HBM and a strong but not leading position in DRAM. The discipline among the three is the single most important variable in the cycle: it is what is holding prices up, and it is exactly the thing that has broken in every prior memory downturn when one player chose share over price. It now faces a test from outside the industry as well: a class action filed in a California federal court in late June alleges the three coordinated to restrict DRAM supply and inflate prices. A complaint is not a finding, and cases of this kind run for years, but it puts a legal price on behaviour the market values as if it were free.

08 The structural risk is a capacity wave landing in 2027 and 2028

The demand backdrop is genuinely strong: AI memory demand is real, HBM is sold out through 2026, and the contract book pulls a large slice of future revenue forward under minimum pricing. The risks sit elsewhere, and there are four of them.

• The capacity wave, with a date on it. Micron is spending above the mid-forty-billion range in fiscal 2027, its Korean rivals are spending in parallel, and the combined industry budget runs well above $54bn for the year, almost all of it new fabrication capacity; Micron broke ground on a western-Japan expansion in the first week of July. It lands in a cluster around 2027 and 2028: management guides supply improving gradually in 2028, and independent forecasters now see prices rolling over as early as 2027. The take-or-pay agreements soften the blow, but a fifth of DRAM under contract leaves the majority exposed to the spot and short-contract market, where a glut shows up first.

• The courtroom. The class action filed on 25 June attacks the supply discipline directly. Damages are one risk; the larger one is behavioural, if a defendant adds capacity or cuts price to blunt the coordination claim.

• Geopolitics. Export rules can shut a market overnight, as China’s server-chip ban already did.

• Valuation. At a $1.06tn market value the stock still has a long way to fall if the cycle turns out to be a cycle; the slide from the June high, 25 percent with no change to guidance, showed how fast the repricing arrives.

None of these resolves in a single quarter, which is why one print can re-weight the call without settling it.

09 A through-cycle valuation now lands just above the price

The reverse lens is the heart of the case. At $938.38 on about 1.13bn shares, Micron carries a market value near $1.06tn. On this year’s peak earnings of roughly $97 a share that is about 9.7 times, down from nearly thirteen at the June high and in line with the multiple memory earns through the cycle. At ten times, the work the multiple used to do falls on the earnings: the price assumes the earnings hold roughly where they are. The test that matters is what a through-cycle valuation produces.

Implied price = exit multiple times this year’s peak earnings of about $97 a share. Spot pays about 9.7 times; the base case adds the mild margin slide the scenarios carry.

Hold earnings near this year’s peak, the contract book and rising volumes roughly offsetting the margin slide, and apply a normal ten times, and the base case is $975, just above today’s price. A sterner nine times, the kind memory earns through the cycle rather than at the peak, is $873, 11 percent below. Cross-checks on sales and on cash earnings land near $1,000 to $1,060, just above spot. The market’s own target sits far higher, anchored on the supply agreements re-rating the whole cycle, the bull scenario rather than the base. Valued through the cycle, the stock trades at roughly fair value; the selloff closed that gap in the first week of July.

10 The scenarios turn on how far the margin falls by 2028

The three scenarios each turn on a single driver: how far the gross margin normalises by fiscal 2028, and what multiple the market pays on the other side.

Probability-weight the three and the fair value is $991, a few percent above spot , with a simulation median of $925 a few percent below it. The distribution is genuinely two-sided: the bull pays 60 percent, the bear costs 23, which is why this is a hold rather than a short, and the weight of probability now sits at the price rather than below it. The single driver that moves the call is the gross-margin trajectory; the nearer test is the 23 September guide.

The weighted value ($991) and the base ($975) sit a few percent above the $938 spot, the simulation median ($925) just below it, and the sell-side crowd far above. Source: public market data and chokepoints.ai model, 7 Jul 2026. Chart by chokepoints.ai.

11 A 25 percent slide took the price below its pre-earnings level

The two weeks around the print were violent. Micron fell more than thirteen percent on 23 June after a Korean regulator warned on memory-linked exchange-traded funds, rose fifteen percent on the 25 June print to an intraday record of $1,255, then gave it all back and kept going: the shares slid to $975 by 2 July, their lowest since mid-June, and a modest Monday recovery left them at $938.38 on 7 July, about 25 percent below the high and beneath their pre-earnings level.

Three forces drove the second leg. The first was mechanical: profit-taking and half-year rebalancing after a first half in which the main semiconductor ETF rose 72 percent, with hedge funds trimming chip exposure for a month. The second was legal: the class action filed on 25 June. The third was competitive: SK Hynix prices a roughly $29bn Nasdaq listing this week, giving US investors a direct route into the HBM leader, and Michael Burry disclosed a short position in Micron as the slide gathered pace. The consensus target still sits about 50 percent above spot, pricing the bull case as the base. The call turns on the margin, and the next few prints answer it.

• The market traded through the good news. A record print and a raised guide preceded a slide below the pre-earnings level; the burden of proof now sits on margin durability rather than the next beat.

• The crowd is still above us. A consensus average near $1,486 across 45 analysts pays for the supply agreements to re-rate the entire cycle. We treat that as the bull case rather than the base, and the gap closes downward if the margin clock runs out.

12 Verdict: hold at fair value, and wait for $750

We hold pieces of both, and the gross-margin trajectory decides which confirms. At $938 against a $975 base and a $991 weighted value, the arithmetic carries no lean: the price sits just below fair value, the bull pays 60 percent, the bear costs 23. We stop at hold rather than accumulate because fair value is a fair price, and the entry that compensates for a 2028 glut sits near $750, close to the bear. A hold here means hold what you own rather than buy more. A guide that defends the mid-80s margin would move us toward accumulate; a roll toward the 60s, the first sign of the 2028 glut, or discovery in the antitrust case that reaches supply decisions, would move us toward an outright sell.

13 Scorecard

14 The gross-margin guide decides the next move

The one number we track: the gross-margin trajectory through fiscal 2027, and the first sign of the 2028 supply wave. The chokepoint and the contract book are settled questions. Whether the peak margin lasts is the open one, and it is what moves Micron from here. We hold, with the add level parked near $750, and revisit on the September guide.

15 If you want to trade it

A hold offers few natural trades, but the analysis points to some ways to express it. These are ideas to test, not recommendations.

1\. Hold within a band: trim toward $1,500, add toward $750. Even after the slide the stock is up more than 700 percent in a year, and the price now matches a base case in which margins ease while volume and the contract floor hold earnings roughly flat. The real test plays out over several prints, and a hold means waiting for the band: trim into strength toward the $1,500 bull case, and add only on weakness toward the $750 re-entry level, near the bear.

• Risk:* the contracts hold pricing and the cycle re-rates before any pullback, leaving a patient buyer behind.

2\. Pair trade: long SK Hynix, short Micron (relative value). The clean memory pair arrives this week. SK Hynix lists American depositary receipts on the Nasdaq around 10 July in a roughly $29bn offering, which removes the access friction that has kept the pair off most US books. Hynix holds about 60 percent of the HBM market against Micron’s fifth and the lead seat on the dominant accelerator platform, so a small, beta-aware long-Hynix, short-Micron pair expresses the HBM ranking directly, without a view on memory’s direction. The old alternative, fading Micron against TSMC and Nvidia, has weakened: the slide has already pulled much of that premium in.

• Risk:* Micron is the share gainer in HBM, so the gap the pair rests on narrows over time; listing-week flows are noisy in both directions; and both legs are defendants in the same antitrust suit, so the case moves them together.

3\. Buy a put spread into the 23 September print (high conviction only). The case here has weakened: at $938 a bearish structure bets on the bear scenario itself rather than on a reversion the market has already delivered. Around the print Micron’s realised volatility ran near 120 percent annualised, and it routinely moves ten percent or more on a result, so an outright short is dangerous and the borrow is unforgiving. An investor who shares the downside tilt can buy a put spread into the 23 September print, capping the cost in advance. One caveat on the data: the reported implied volatility looks stale and the historical earnings-move sample is too small to trust, so price the structure off how violently the stock actually moves, not off a screen.

• Risk:* a blowout guide that defends the margin can gap the stock straight up through the structure; defined risk caps the loss, but the bull tail toward $1,500 is real.

Above all, size it small. Micron carries a beta near one and a half to the chip index and moves violently around prints, and the debate it turns on, whether a peak margin can hold into 2028, resolves over several prints rather than one. The first idea fits a hold best; the pair is tactical; the options route is for high conviction only.

16 Dates to watch

The events that move Micron or its read-through over the coming weeks:

• 10 Jul 2026: SK Hynix lists on the Nasdaq. A roughly $29bn ADR offering puts the HBM leader on a US exchange for the first time; the debut tests whether money rotates out of Micron or the listing lifts the whole memory complex.

• 29 Jul 2026: SK Hynix results. The HBM leader and the clearest read on memory pricing, HBM demand, and whether the three makers are still holding supply discipline.

• 26 Aug 2026: Nvidia results. The customer whose accelerator roadmap sets the HBM order book; the demand read that underwrites the whole memory cycle.

• 23 Sep 2026: Micron fiscal Q4 results. The main event. The beat itself is expected; the read is the gross-margin guide and the first fiscal-2027 signal, the test of whether the peak margin holds.

• through 2026 and 2027: spot DRAM and NAND contract prices, and the first hard signs of the 2027 to 2028 capacity wave, the deciding variable for the call.

Dates from company IR calendars and reported schedules as of 7 Jul 2026; confirm against each company’s IR page, as dates can shift.

Not investment advice. For information and discussion only, and not a personal recommendation, offer or solicitation. Capital is at risk, and investments can fall as well as rise. The scenario fair values and price targets here are the publisher’s estimates, not a forecast or a guarantee. Do your own work.

All figures come from Micron’s public disclosures, public market and industry data, and the publisher’s own calculations. No third-party broker or expert-network research is quoted or reproduced. Sources: Situational Awareness; Micron public disclosures; chokepoints.ai scenario and valuation work. © 2026 chokepoints.ai, Issue 005, Micron is the US memory maker at the top of an AI super-cycle; a July slide took the price to a through-cycle multiple, and the call now rests on whether peak margins hold into 2028. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

In 60 seconds

Every AI rack is held together by copper. The GPUs talk to each other over high-speed cable and backplane connectors, and that in-rack wiring is a two-horse market: Amphenol and TE Connectivity. Every accelerator that ships needs it, and there is no third supplier of consequence.

TE holds a strong position in the custom-silicon programs at the major cloud operators, and its data-centre business is growing fast. The other half of the company, mostly tied to cars, is not. That split is the whole story, and it raises a simple question: is the market paying for the car business and getting the AI business close to free?

That last question is the whole case. Amphenol and TE sell into the same racks, to the same buyers, yet Amphenol trades at 32 times forward earnings and TE at 18. The spread has stayed open long enough that the “second-source discount” is starting to look less like a discount than a misprint.

Price targets

On the base case we are close to the Street. Where we differ is the shape of the distribution around it.

The weighted $262 is almost exactly the Street’s consensus: a $262 average, a $190 to $302 range, and a Buy rating. We agree on the base. Where we part company is the spread. We think the re-rating optionality in the bull case ($357) is underpriced, and that the floor sits higher than the Street’s $190 low. The drivers behind these numbers sit in the bridge table in section seven, and they are worth reading before the targets.

“Copper will continue to be the workhorse in the rack. It’s not copper or optical, it’s copper and optical. The orders that we have received are building backlog into 2027”. Terrence Curtin, CEO, TE Connectivity, Q2 FY2026 earnings call, 22 April 2026

That one sentence carries both the bull case and the main risk. (The rack is the cabinet that holds a cluster of GPUs and wires them together into a single machine; copper is the short, fast link inside it.) On the bull side, copper is still the dominant in-rack link, demand is strong enough that the company is already booking 2027, and the orders are real: a record $5.3B in Q2 at a book-to-bill of 1.12. On the risk side, Curtin reaches for “copper and optical” in the same breath, and TE has just bought into leading-edge optical technology to widen its roadmap. The hedge tells you the optics question is live. So the investment case turns on how much of the company is the copper workhorse, and how much has to become the optical hedge.

By the numbers · Q2 FY2026

01 Where TE sits in the stack

AI runs on a physical supply chain. The layers below are the chokepoints atlas: L0 at the base (raw materials) up to L8 (the apps people pay for). Red layers are real chokepoints, meaning one or two suppliers, one region, or multi-year lead times. TE sells at L4, in high-speed connectors and cable. That is a genuine chokepoint, but a shared one: TE and Amphenol both supply it, so TE owns half of the bottleneck rather than all of it.

Top: closer to user and revenue Bottom: scarcer and harder to replace

In-rack copper for scale-up systems is close to a duopoly. Amphenol and TE supply most of it, and you cannot build a dense GPU rack without it. Optics will take the longer links between racks; copper holds the short, fast links inside the rack. For an investor the distinction matters, because TE owns half of a chokepoint rather than a monopoly. Its edge is price, not a moat nobody else has. No third credible supplier has broken in at the leading edge, so the duopoly itself is the protection, and the second seat inside it is what you are buying at a discount.

02 What a connector actually is

The investment case stands or falls on the product, so it is worth understanding how a connector actually works. A connector joins two segments of wire or two circuit boards so an electrical signal can pass between them, reliably and cheaply at scale. The hard part in a modern AI server is that the signals are very-high-frequency analogue waveforms running at tens of gigabits per second, and the connector has to preserve the shape of those waveforms as they pass through. If the signal degrades, the GPU at the other end receives a corrupted bit.

Stopping that is signal-integrity engineering, and it is genuinely difficult: pin geometry, the dielectric around the pins, how the cable is terminated, how the assembly is routed inside the rack. The big AI platforms need three things in particular:

• Backplane connectors that join the main board to plug-in cards in the server chassis.

• Direct-attach copper (DAC) cables that link one GPU directly to the next, or to a switch, in a fixed-length assembly built to tight tolerance.

• High-speed cable assemblies that fan out from a switch to many GPUs, each cable tuned to the same electrical length.

The canonical example is Nvidia’s GB200 NVL72 rack, which wires together its 72 GPUs with about 5,000 copper NVLink cables. The design is copper rather than optical because an all-optical version would draw far more power. Over the distances inside a single rack, copper wins on the things that decide the design: cheaper per port, lower latency, lower power.

Where TE’s connectors and cables sit inside an AI server rack

The three things TE makes for an AI rack, and where each sits: backplane connectors inside the chassis, direct-attach copper between GPUs, and high-speed fan-out from the switch. Nvidia’s GB200 NVL72 is the canonical case, about 5,000 copper cables in one rack. Schematic by chokepoints.ai.

Once a connector is designed into a platform, it tends to stay. The hyperscaler qualification cycle is long, running through electrical simulation, mechanical and thermal testing, pilot builds, and then full production, and it can take well over a year. After that the connector ships for the life of the platform. A customer will not swap a qualified connector for a cheaper one lightly, because the connector is a tiny fraction of the cost of the rack and the cost of getting it wrong is enormous. That is the moat, and it is structural: a multi-year qualification cycle and a track record on the platforms that matter, rather than a brand or a patent.

The argument between bulls and bears is copper versus optics, and really it is an argument about where the line between them sits. Consensus today puts it at the edge of the rack. The risk is that the line moves inward over time, and that risk now has a date on it. Nvidia plans optical scale-up for systems of more than 1,000 GPUs by about 2028, and has invested billions in optical suppliers. Copper is expected to stay the in-rack workhorse through the near-term Vera Rubin generation in 2026 and 2027. TE’s answer is that it is not a pure copper bet; it has bought leading-edge optical technology to strengthen its passive optical roadmap. The hedge is real, and it costs the bull case some of its simplicity.

03 The company, briefly

TE makes connectors and sensors. Those two words cover an enormous range of products: the small terminals inside a car dashboard, the heavy blocks that route power through a factory, and the precision backplanes and DAC cables that link GPUs inside a server rack. The lineage runs back to 1941, when AMP Incorporated was founded in Pennsylvania. It was spun out of Tyco in 2007, renamed TE Connectivity in 2011, and later redomiciled to Ireland, which is where the “plc” comes from. It employs about 90,000 people across roughly 130 countries.

• 1941 · AMP founded in Pennsylvania, building the early connector industry around automotive and industrial customers.

• 1999 · AMP acquired by Tyco International, folded into a diversified conglomerate.

• JUL 2007 · Tyco break-up. The connectivity business is spun off as Tyco Electronics.

• MAR 2011 · Renamed TE Connectivity. NYSE-listed, with industrial and transportation as its twin pillars. Later redomiciled to Ireland.

• 2010s · The auto franchise. Sensor and connector content rises as cars electrify, and TE becomes a backbone supplier to global auto platforms.

• 2026 · AI acceleration. DDN orders up about 60% YoY, AI applications around 70% of the segment, and the full-year AI revenue target raised mid-year.

Today TE sells into three places: cars (the Transportation segment, mostly automotive sensors and harnesses), factories (industrial automation, energy, aerospace), and data centres (the Digital Data Networks unit inside Industrial). The split between the first two and the third is what defines the investment case.

One half is accelerating and the other is flat. The blended company grew 7% organic, and the market prices the blend rather than the half that is winning. The opportunity is in that gap.

04 How it makes money

The model is simple to describe and hard to compete with.

1 · Design in. TE’s engineers specify a connector, backplane or cable for the customer’s AI server, car or factory system, to electrical, mechanical, thermal and cost targets.

2 · Qualify. The part runs the customer’s qualification cycle: signal-integrity and reliability testing, then pilot builds. Often more than a year for a new AI platform.

3 · Ship for life. Once designed in, it ships in production volume for the life of the platform. Volume is visible, because orders track the customer’s build plan.

4 · Content rises. Each GPU generation uses more high-speed links per rack. Even when the unit price falls, the content per rack rises.

The flywheel is the whole model, and the hard part is the front of it. The segment numbers show the cycle in action: Industrial, which contains DDN, was up 27% reported, with DDN itself up 46% organic, while Transportation was down 0.5% organic.

Dividend up, buyback expanded

In March 2026 TE raised its quarterly dividend to $0.78 from $0.71 and expanded the buyback authorisation by $3.0B, to $22.25B in total. A company asking the market to re-rate it on AI orders is, in the same quarter, handing real cash back while the multiple is still wide. Read that as confidence in the cash flows behind the order book.

05 The AI engine: Digital Data Networks

DDN is the reason to look at the stock now. The unit sits inside Industrial and houses TE’s data-centre and AI business: the high-speed connectors, cable assemblies and backplanes used inside AI server racks.

The macro picture is simple. Hyperscalers are still building AI capacity faster than the supply chain can keep up, and TE is one of only two suppliers they can call. TE raised its $2.4B AI target mid-year rather than setting it at the start, and Curtin said the orders already in hand stretch into 2027. That is more visibility than most AI-exposed suppliers can claim. Orders hit a record $5.3B in the quarter, up 25% on the year and running ahead of sales, so the demand is booked rather than hoped for.

The lever to watch is the mix. AI is only about 13% of TE’s roughly $18.7B in annual sales today. The bull case is that this share climbs into the high teens or low twenties over the next two to three years, the point at which the slow car business stops dragging down the valuation. That share is the single number to track.

06 Cheap against what?

Valuation only means something in comparison, so here are three: peers now, TE against its own history, and the cash flows the market is discounting.

Forward P/E: TE sits with the auto names, not the AI-interconnect premium

TE is priced like the auto-component names rather than the AI-interconnect ones. Yet it just posted 24% adjusted EPS growth, a 46% organic DDN print, and a 1.12 book-to-bill. The peer it resembles on growth is not the peer it is priced like. Source: public market data, 16 Jun 2026; chokepoints.ai. Chart by chokepoints.ai.

TE also looks cheap against itself. Its own trailing P/E sits at 21.9×, against a ten-year mean near 28× and a five-year range of 12.9× to 32.5×. So TE today is roughly 22% below its own long-run average. Amphenol has gone the other way: it re-rated up on AI and still trades at a premium to its own median near 30×. The gap is wide three ways at once, against Amphenol, against the AI-interconnect group, and against TE’s own past.

Putting the methods on one page, the implied price band by approach comes out as:

• Scenario fair value (bear / base / bull): $157, $266, $357.

• Connector-peer forward P/E (17× to 32×) on NTM EPS: $201 to $378.

• TE’s own ten-year P/E (12.9× to 32.5×) on TTM EPS: $126 to $318.

• EV/EBITDA re-rate (14.4× to 25.6×): $218 to $399.

• Street consensus range: $190 to $302, average $262.

Implied value by method: scenario, peer P/E, own history, EV/EBITDA, Street

The $214.50 price sits below the midpoint of every method except the most pessimistic bear case. That is the trade in one chart. Source: chokepoints.ai scenario work on public figures. Chart by chokepoints.ai.

The cash flows tell the same story. At $214.50, a reverse-DCF says the market is discounting only about 4.4% free-cash-flow growth a year for ten years, on a 9% cost of capital and a 2.5% terminal rate. A company with that order book and 46% organic growth in its AI line is not a 4.4% grower, unless the car business drags everything down forever. Re-rate EV/EBITDA to 20×, still below the twin’s 25.6×, and the implied value is about $308, or +44%; at Amphenol parity it is about $399, or +86%. The bet is not on a boom. It is that the discount is too wide.

07 How the three cases add up

The honest way to bridge “should be worth more” and “$266” is to show the maths. The lever is not capacity. It is the exit multiple. AI revenue and the auto trajectory set the EPS, and the re-rating does the rest.

Each fair value is the exit multiple applied to FY27 adjusted EPS, then discounted back about a year at a 9% cost of capital. That is why the base reads $266 rather than the naive $290 you get from 22× times $13.2. The discount keeps the base case honest. Read the table one way and the point jumps out: the swing factor is the exit multiple, not the EPS. Across all three cases the EPS barely moves, from $12.2 to $13.9, while the multiple swings from 14× to 28×. The value is made in the re-rating, which is also why the trade is easy to underprice.

08 Pumps and corrections

This is a re-rating stock, so it trades on the multiple and on expectations as much as on the fundamentals. That makes the tape do counter-intuitive things. The cleanest example is that TE has, more than once, fallen on a quarter that was a clear beat with record orders. The reason was not the print. It was the bar. The market had already priced a strong quarter, so meeting it rather than clearing it was enough to trigger a sell-off.

Illustrative single-day reactions: what pumps the stock and what corrects it

What actually moves the stock, in rough order of magnitude. Illustrative; chokepoints.ai. Chart by chokepoints.ai.

The practical lesson on entry is to size and time around the expectations bar, not the print. A good quarter after a strong run can still correct, while a good-enough quarter on soft consensus is often where the multiple re-rates. The fundamentals here are solid. The work is in the timing.

09 Re-rating needed to make the upside real

Here is the tension we will not paper over. At a conservative 22× exit, the base-case upside (+24%) roughly matches the bear-case downside (−27%), so on the multiple alone the trade is close to a coin flip. It turns clearly asymmetric only once the exit multiple clears about 24×, which is still below both TE’s own history and Amphenol’s multiple. The call therefore rests on the re-rating, and the reverse-DCF is what tips the odds in its favour: at today’s price the market already assumes almost no growth.

The other thing we will not pretend away is that the gap to Amphenol is not random. Amphenol leads on speed, holds the larger Nvidia position, and earns its higher multiple for those reasons. The 43% discount is a real second-source discount, not a misprint. Our case is that the gap is too wide for what is, at the platform level, a very similar product. The bear case is that the gap is simply fair.

10 The two-sided case

11 Scorecard

12 What moves the signal

The one number we track from here is AI revenue as a share of the group. It is in the low teens today. When it clears roughly 15%, the auto drag stops setting the price, and the re-rating case stops being a forecast and starts being a fact.

Not investment advice. For information and discussion only, and not a personal recommendation, offer or solicitation. Capital is at risk, and investments can fall as well as rise. The price targets here are the publisher’s scenario-based fair-value estimates, not a forecast or a guarantee. Do your own work.

All figures come from TE Connectivity’s public disclosures, peer public market data, and the publisher’s own calculations. Broker research informed our internal conviction but is not quoted here. Sources: TE Q2 FY2026 results (PR Newswire); earnings-call transcript (Motley Fool); Stock Analysis and Companies Market Cap; The Register on Nvidia optical scale-up. © 2026 chokepoints.ai · Issue 003 · TE sits at L4 (compute hardware, high-speed interconnect) in the chokepoints AI-compute map. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

We like SK Hynix. We think memory is one of the genuine chokepoints of this AI build-out, and we hold that view more firmly than most of the desks now upgrading the stock. We are still not buying it at ₩2,521,000. It is up 248% this year and 587% over twelve months, and the easy money has been made.

Both of those things are true at once. The demand for high-bandwidth memory is real, it is locked in under contract, and it is sold out for 2026. None of that is in dispute. The question we actually care about is narrower: at this price, what is the market already paying you to expect? Our answer is that it is paying for the peak: peak earnings, peak valuation, and peak market share, all in the same quarter.

So here is what we would do instead of buying it today. We are not chasing it, and we are not shorting a business this good. We wait. Section 06 sets out the price at which we become buyers; Section 07 sets out the two things that would make us sell rather than buy. Until one of those prints, the position is simple: add on weakness, not a fresh buy at today’s quote, and not a lazy HOLD.

New to memory? Start here

Every computer chip does two jobs: it calculates, and it remembers. The remembering is handled by memory chips. The everyday workhorse is DRAM, the same broad type that has been in PCs and phones for decades. Three companies make almost all of it: SK Hynix, Samsung, and Micron.

AI created demand for a premium version called HBM: high-bandwidth memory. It is DRAM stacked into tall towers and bonded right next to the AI processor (the GPU), so data can move between the two at enormous speed. Every high-end AI chip needs a block of it, and it sells far above ordinary DRAM. This is the product the whole note is about.

It matters because modern AI is held back as much by memory as by raw calculating power: a GPU starved of fast memory simply sits idle. SK Hynix makes more HBM than anyone else today, which is how a memory company became one of the most important names in the AI supply chain. (NAND, which appears later, is the other big memory type, the flash storage in drives, and it is a sideshow to this story.)

In 60 seconds

The short version: SK Hynix makes the specialised memory that sits next to every AI chip, it dominates that niche today, and it is having the best year in its history, which is exactly what makes the stock risky from here.

SK Hynix holds an estimated ~62% of the global HBM market, has effectively sold out HBM3E for 2026 at a ~50% price premium over conventional DRAM, and supplies roughly two-thirds of NVIDIA’s HBM4. Q1 2026 was the most extreme quarter in memory history: revenue of ₩52.6tn, operating profit of ₩37.6tn at a 72% margin, net profit of ₩40.3tn at a 77% net margin. The single quarter nearly matched all of FY2025 (operating profit ₩47.2tn at a 49% margin).

The bear case is just as real. NVIDIA certified all three HBM vendors for the Vera Rubin platform on 5 June 2026. Samsung retook the overall DRAM share lead in Q1 2026 (38% versus SK Hynix at 29%). The HBM lead is on a documented downslope: 62% in Q2 2025, ~57% in Q4 2025, with three-vendor HBM4 capacity landing through 2026 and 2027. The forward P/E of ~6.6× sits on a cycle peak. The whole trade is a referendum on whether HBM has structurally de-cyclicalised memory, or whether 2026 is the top of a wave that reverts as three-vendor capacity lands at once.

We believe the first. We are not buying the second.

01 We believe memory is the chokepoint of the build-out

Every note we write reasons from the same prior: the AI build-out is a default trajectory, not a forecast. Effective compute (very roughly, the useful computing power thrown at AI) is rising about half an order of magnitude a year. Annual AI investment is climbing fast toward a $7.6T baseline over 2026 to 2031, with hyperscaler capex alone guided to ~$700B in 2026. The surge is front-loaded into this decade. We pass through more OOMs (orders of magnitude: 10× jumps in scale) between 2024 and 2030 than in all the decades that follow. The exponential is in motion, the clusters are already signed, and we put the burden of proof on the sceptic.

The spending behind the compute: hyperscaler capex has run ~4.5× in three years

The spending behind the compute. Combined annual capex at the four largest US hyperscalers (Microsoft, Amazon, Alphabet and Meta) has run roughly 4.5× in three years, with 2026 guided to ~$700–725B. This is the demand curve SK Hynix sells into. Calendar-year capex; 2026 is company guidance. Sources: company filings, Financial Times, Goldman Sachs.

Inside a build-out this large, the money is made at the bottlenecks. Whenever one component is the thing everyone is short of, whoever controls it earns outsized profits, for exactly as long as they remain the only real source. First it was the GPUs themselves. Now, increasingly, it is the memory that feeds them.

Memory is now a chokepoint, by name. Situational Awareness names HBM and CoWoS advanced packaging as “already key bottlenecks… the primary constraint,” and projects memory rising to potentially more than 20% of cluster capex after the GPU itself. We hold that view. SK Hynix sits at the memory layer, in the HBM sub-layer, and it owns the largest single share of the most strategic product in the AI supply chain.

That is the bull case. The honest question, and the one this note actually answers, is what comes after. A chokepoint is only as good as its monopoly. The rent gets competed away wherever the constraint is contestable. So for any memory name the decisive question is not “is demand strong,” because we assume it is, but “how single-sourced is the specific thing this company sells, and for how long?”

For SK Hynix, the answer is: very single-sourced today, but visibly less so tomorrow. That is the entire note.

02 SK Hynix’s HBM lead is already melting

This is the chart that matters. SK Hynix’s HBM bit-share has already gone from 62% in Q2 2025 to ~57% in Q4 2025, and every force pulling it lower from here is now public. The same sell-side models that have SK Hynix leading today have its share eroding further as Samsung’s HBM4 ramps. The lead is melting on the industry’s own numbers, not on bear-case speculation.

The melt: SK Hynix dominates HBM today, yet the lead is slipping

The melt, in two cuts. SK Hynix dominates HBM today (~62% bit-share versus Micron and Samsung), yet Samsung already leads overall DRAM (38% versus 29%). The HBM lead has slipped from 62% to 57% in two quarters as three vendors ramp HBM4. Source: chokepoints.ai. Chart by chokepoints.ai.

Three things put the melt in motion. First, Samsung began mass-producing HBM4 in February 2026 and has shipped early samples to NVIDIA. Second, Micron sold out its 2026 HBM capacity and is ramping HBM4 a quarter early in Q2 2026. Third, NVIDIA certified all three vendors for HBM4 supply to the Vera Rubin platform on 5 June 2026. The stock fell roughly 10% on the multi-sourcing news.

None of this means SK Hynix is a bad business. It means the lead is a lead, not a lock. The difference matters for valuation. A moat is a permanent feature of a business; a lead is a snapshot, and snapshots decay.

03 The durable rent is migrating upstream, to TSMC and ASML

The lazy version of the chokepoint thesis says: SK Hynix makes the memory, SK Hynix owns the chokepoint, the rent accrues to SK Hynix. The honest version says the chokepoint is a triad, and the durable rent is migrating upstream.

The triad: the most durable rent sits upstream of the part SK Hynix sells

The triad. SK Hynix owns the HBM layer (~62% today, three vendors, and slipping). But the packaging and the base die beneath it are TSMC’s, the lithography beneath all of it is ASML’s, and NVIDIA sets the specification the whole stack must meet. The most durable rent sits upstream of the part SK Hynix actually sells. Source: chokepoints.ai. Diagram by chokepoints.ai.

Here is that picture in detail. HBM is built from three layers. The HBM stack itself, made by SK Hynix, Samsung, and Micron. The CoWoS advanced packaging that bonds the HBM to the logic die, made almost entirely by TSMC, fully booked with 52 to 78 week lead times, scaling from ~35,000 wafers per month in late 2024 to a projected 130,000 WPM by end-2026. And the logic base die that the HBM sits on, also made by TSMC, since HBM4 uses a 12nm TSMC-fabricated die designed by NVIDIA. NVIDIA occupies more than 70% of TSMC’s CoWoS-L capacity.

The rent in the triad splits roughly like this. SK Hynix captures the HBM bit-share and the premium pricing that bit-share commands. But the CoWoS gate, the base-die fab, and the EUV scanners behind all of it sit with TSMC and the equipment oligopoly (ASML, principally). NVIDIA, as monopsonist buyer, designs the HBM4 specification and dictates the base die. The bargaining power is asymmetric and tilts against SK Hynix over time. Downstream, hyperscaler long-term agreements pre-empt roughly half of SK Hynix’s revenue at contracted prices. That is a floor on volume, but it is also a ceiling on rent: contracted pricing renegotiates in the customer’s favour more often than the seller’s.

For readers who want the chokepoint with the least contestable rent, the upstream layers are arguably the cleaner trade. CoWoS and the base die sit with TSMC. EUV sits with ASML. Both have longer lock-in, fewer competitors, and less customer-concentration risk than any single HBM vendor. We hold SK Hynix for the idiosyncratic kicker, the purest HBM read-through, Korea-beta exposure, and a possible US-listing catalyst, with eyes open to share decay.

04 The real risk: memory-per-chip may not keep climbing

The Situational Awareness prior is a compute prior, not a memory-intensity prior. It says effective compute is rising 0.5 OOMs per year. It does not say HBM per unit of compute is rising. That part is a bet.

Two forces pull in opposite directions on how much HBM each unit of computing work needs. The bearish force is algorithmic efficiency: the steady progress that lets each new generation of models do more with less hardware. KV-cache compression, paged attention, multi-latent attention, mixture-of-experts, 4-bit and 8-bit quantisation: each of these is a technique that squeezes the memory needed per unit of work. A model that runs on half the HBM is, from SK Hynix’s point of view, a model that buys half the HBM.

The bullish force is Jevons. Long-context models, agentic workloads, reasoning chains, multi-modal inference. Each raises HBM per query, because the system has to hold more state in working memory while it thinks. The bet is that Jevons wins, that the demand for intelligence per dollar rises faster than algorithmic efficiency per dollar. We hold that bet, with caveats.

The honest horizon is not “structural for the decade.” It is structural 2025 to 2028, with a high-probability air-pocket, then cyclical re-assertion. The sell-side’s own DRAM price models already price the cycle turning by 2029, with DRAM ASPs modelled to fall as three-vendor capacity lands at once. If algorithmic efficiency compounds faster than Jevons, the air-pocket arrives earlier. If a hyperscaler capex pause of more than six months lands on the trip-wire of the ~3x capex-to-AI-services-revenue ratio, the air-pocket arrives in 2027. We are not pricing 2030. We are pricing the next three years.

This is the part of the chokepoint thesis that the consensus does not write about. The demand for memory is real. The demand for more memory per unit of compute is a separate question, and it is not guaranteed.

05 The price is paying for three peaks at once

The price is paying for the peak on three independent axes.

Peak earnings. Q1 2026 operating profit of ₩37.6tn nearly matched all of FY2025 (₩47.2tn). DRAM ASPs rose mid-60% quarter on quarter in Q1; NAND ASPs rose mid-70%. The Q1 print is the most extreme in memory history, and the entire 6.6× forward P/E sits on a denominator the company itself will almost certainly not repeat. Market cap of ~₩1,786tn implies forward net income of roughly ₩271tn (mcap / 6.6×) versus ~₩79tn trailing (mcap / 22.5×). The price requires net income to roughly triple off the trailing base and hold there. The forward P/E is a peak-earnings denominator, not a cheap multiple.

Peak P/B. ~10.9× trailing book, outside SK Hynix’s historic ~1 to 2.5× band by a factor of four or more. The normalised ROE of 53.18% must persist to justify the multiple, and ROE is a mean-reverting series in this industry. Forward P/B re-rates to roughly 2.1× as the FY2026E book builds (current ~₩232k BVPS plus retained earnings), which is more defensible, but still assumes that ROE persists. Re-rating from trailing to forward is a real effect, but it does not save the entry: it makes the over-statement less severe.

Peak HBM share. ~62% in Q2 2025, ~57% in Q4 2025, with three-vendor HBM4 capacity ramping through 2026 and 2027. The lead is melting. SK Hynix is about to start earning peak margins on a falling share of the most strategic product it sells.

Quarterly revenue and operating profit, Q1 2024 to Q1 2026: the denominator the 6.6x forward P/E sits on

Quarterly revenue and operating profit, Q1 2024 to Q1 2026. The Q1 2026 bar dwarfs everything before it. This is the denominator the 6.6× forward P/E is sitting on. Source: SK Hynix public disclosures; chokepoints.ai. Chart by chokepoints.ai.

Blended scenario fair value: today’s price only clears the bull case, by 9%

Blended scenario fair value. Current price ₩2,521,000. Bear ₩1,003,000 (−60%). Base ₩1,735,000 (−31%). Bull ₩2,740,000 (+9%). Today’s price only clears the bull case, by 9%. Source: chokepoints.ai scenario work on public figures. Chart by chokepoints.ai.

The sensitivity grid is the cleanest tell. Our base case sits ~31% below the current price. The bull case barely clears it (+9%). The bear is severe (−60%). And the bulls’ own published bear scenarios, including the larger Korean and global houses, sit at ₩1.2M to ₩2.0M, all below the current price. You can be right on the thesis and lose 30% on the entry. The asymmetry is the wrong way.

06 We will buy the pullback, not the peak

ACCUMULATE means we have a price at which we will be a buyer, and we have published it.

Trigger to add: a pullback toward the ~₩1.7M base case (a drawdown of roughly a third from the current price), provided two conditions hold. (1) SK Hynix’s HBM4 share is still at or above 40%, indicating the moat has not collapsed. (2) No HBM contract-price cut has printed, indicating the supercycle has not ended.

The next two earnings prints (Q3 and Q4 2026) are the binary read on HBM4 share, not on 2027 numbers. We will be buyers at the price the oligopoly produces, not at the price the consensus is paying. Pay the base case for the structural bull’s upside.

What it is worth now, and what a re-rating would change. On our base case the shares are worth about ₩1.7M today. At ₩2,521,000 the market is paying roughly 45% above that (the same gap the sensitivity grid shows as our base sitting ~31% below the current price), and only the bull case (₩2,740,000) justifies it. A genuine re-rating would move our number, not just the price: if HBM holds its premium, the US listing lands, and the market accepts that memory has partly de-cyclicalised, fair value migrates up toward the bull. We put ~60% on the base-case path (an air-pocket and a pullback toward ₩1.7M inside 12 to 24 months) and ~40% on the re-rate path. At those odds, today’s price is already paying you for the re-rate before it has happened.

07 Two events would make us sell, and one has half-fired

We will change the position the day any of the following fires. We are stating it now, in writing, so that the discipline is public.

• (a) First HBM contract-price cut, quarter on quarter. Not a volume miss. Not a mix shift. Not a guidance update. A real reported HBM ASP reduction, in earnings disclosures or in sell-side channel checks. Price is the cycle.

• (b) Samsung wins majority HBM4 share at NVIDIA. Not a hair over half. A clear majority on Vera Rubin volumes, confirmed in earnings or in disclosed allocation.

The kill-switch has already partially fired. NVIDIA’s 5 June 2026 three-vendor certification and Samsung’s Q1 2026 DRAM share lead (38% versus SK Hynix at 29%) are facts, not forecasts. We are not exiting. We are flagging that the conditions for a re-rating are now on the page, not in the future.

Why hold, and not move to underweight today? Three reasons. First, the business is genuinely excellent and the order book is sold out, so shorting it now is a bet against contracted demand. Second, the timing of the air-pocket is unknowable, and being early on a cyclical short is indistinguishable from being wrong. Third, the two events that would actually break the thesis have not printed: the conditions for a re-rating are visible, but they are not yet facts. Against the other ideas on our sheet, SK Hynix earns its place as the cleanest, highest-share read on the memory chokepoint, with an idiosyncratic US-listing catalyst on top. What it does not earn, at this entry, is fresh capital.

If (a) or (b) prints, we move to UNDERWEIGHT on the position, sell the cyclical, and revisit only when the air-pocket has cleared.

08 The bull and the bear are both right

Both halves are real. The whole call is which half the price is paying for.

09 The scorecard: real chokepoint, contestable rent, peak entry

10 Watch one number: the HBM contract price

HBM ASPs. Not revenue. Not share. Not earnings. The contract price of HBM, quarter on quarter. Volume misses can be explained by mix, customer timing, or inventory. A price cut means the cycle has turned.

The first house or channel check that reports an HBM ASP cut QoQ is the print we publish a position change on. That is the discipline. The chokepoint is real, the rent is contestable, and the entry is the peak. We will be buyers at the price the oligopoly produces, not at the price the consensus is paying.

Not investment advice. For information and discussion only. Not a personal recommendation, nor an offer or solicitation to buy or sell any security. Capital is at risk; investments can fall as well as rise; past performance does not indicate future results. The scenario fair values and price targets in this note are the publisher’s estimates, not a forecast or a guarantee. All figures are from public disclosures and the publisher’s own calculations. No third-party research is quoted or relied upon.

All figures come from SK Hynix’s public disclosures, peer public market data, and the publisher’s own calculations. No third-party research is quoted or reproduced. Sources: Situational Awareness; SK Hynix public disclosures; chokepoints.ai scenario work. © 2026 chokepoints.ai · Issue 003 · SK Hynix sits at the memory layer, in the HBM sub-layer, in the AI chokepoints map. The chokepoint is real. The price is the peak. We are patient. All issues

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

Jean-Claude-Richard de Saint-Non, Ancient Ruins: Ancient Aqueduct (1756). Cleveland Museum of Art, CC0.

In 60 seconds

Every company building AI, from a two-person startup to Microsoft, needs large clusters of Nvidia chips to train and run its models. Most cannot build that capacity fast enough on their own. Nebius builds the data centres, fills them with GPUs, wires up the power, and rents the capacity out. Think AWS, but built specifically for AI.

It was carved out of Yandex, Russia’s dominant search and technology group, in 2024. It kept the engineers and has scaled fast: run-rate revenue went from $121M to $1.92B in six quarters, and it holds close to $50B of contracts with Microsoft and Meta. The question is not whether AI demand is real. It is whether Nebius can build and fund fast enough to justify a ~$56B valuation.

Each issue of this newsletter examines one point in the AI build-out where demand has run well ahead of supply, and asks which companies benefit. Nebius Group sits on one of those points. It rents out GPU capacity at gigawatt scale, and on its first-quarter call management described the bottleneck plainly.

For every GPU coming online, roughly four customers are competing for capacity. Time-to-power is critical for both revenue conversion and contract momentum. Paraphrasing Nebius management, JMP Securities, 20 May 2026

That quote holds both the bull case and the central risk. Demand is secured: Nebius has signed close to $50bn of contracts with Microsoft and Meta. What remains open is execution: can a company whose last data-centre build finished over a decade ago in Finland turn contracted power into connected , revenue-earning power fast enough, while funding a $20–25bn capital programme without heavy dilution? This issue is our first pass at that question.

01 Where Nebius sits in the stack

AI runs on a physical supply chain. The layers below are the chokepoints atlas: L0 at the base (raw materials) up to L8 (the apps people pay for). Red layers are real chokepoints, meaning one or two suppliers, one region, or multi-year lead times. Nebius sells at L6 (neocloud, segment L6.1.2) and builds down into L3, where it owns its sites as a facility operator. That position shapes both its opportunity and its risk: it depends on every bottleneck beneath it and holds no monopoly of its own.

Stack layers adapted from the chokepoints.ai value-chain map.

02 What Nebius is

Nebius is what was left of Yandex once the Russian search business was carved off; it re-listed on Nasdaq in late 2024 and reinvented itself as a *neocloud*: a specialist that does nothing but rent AI compute, vertically integrated from the concrete and the power contract up through Nvidia GPUs to its own software and inference stack. Growth since the relisting has been rapid: AI-cloud revenue grew 841% year on year to $390m in Q1 2026, 98% of the group total, and annualised run-rate revenue hit $1.92bn, up 674%.

Management guides to full-year 2026 revenue of $3.0–3.4bn and, more ambitiously, an exit run-rate of $7–9bn. That exit number is where bulls and bears diverge: hitting it means compounding the run-rate roughly 54–67% every quarter for three more quarters, the bulk in Q4. The growth so far supports the trajectory, but the implied step-up is far larger than anything the company has yet delivered in a single quarter.

Background

• 2000 — Yandex founded by Arkady Volozh. Russia’s dominant search engine, then a full cloud, ads and ride-hail group.

• 2011 — Lists on Nasdaq as Yandex N.V., a Dutch holding company.

• FEB 2022 — Trading suspended. Sanctions after Russia’s invasion freeze the shares.

• JUL 2024 — The split. Russian businesses sold to local investors. The group keeps its non-Russian assets and over 1,000 engineers, renames itself Nebius Group. Volozh returns as CEO.

• OCT 2024 — Relists on Nasdaq (NBIS), moves fully into AI cloud.

• 2025–26 — Scale arrives. A $17.4B Microsoft deal, up to $27B with Meta, a $2B Nvidia investment, and run-rate revenue through $1.9B.

What the group contains

Nebius AI cloud is ~95% of the story. Around it:

• ClickHouse : ~25% stake marked at ~$15bn round (a ~$780m non-cash gain ran through Q1 2026).

• Toloka : AI data-labelling; Bezos Expeditions invested.

• Avride : self-driving vehicles, partnered with Uber.

• TripleTen : edtech reskilling, ~88% YoY growth.

Headcount ~1,371, HQ Amsterdam, listed New York. Net loss ~$447M in FY2025; Q1’26 profit was a one-off accounting gain, not operating cash.

The bars show the track record; the dashed box is the target, with most of 2026 back-half loaded. Source: Nebius Q1’26 disclosures; broker estimates (Singular Research, JMP). Chart by chokepoints.ai.

03 The backlog: Microsoft and Meta

In eight months Nebius signed two of the largest GPU-capacity contracts in the market. Microsoft came first, September 2025: a five-year deal worth $17.4bn, with options to $19.4bn, anchored on the Vineland, New Jersey site. Meta followed with a small $3bn deal late in 2025, then a far larger arrangement in March 2026 worth up to $27bn. The Meta contract is structured unusually.

The $27bn Meta contract includes $12bn of dedicated capacity from early 2027, plus $15bn that Nebius can allocate to Meta at pre-agreed terms or sell to other AI-cloud customers at market prices. Broker synthesis from broker and expert-network research

Add the smaller deals and the backlog lands near $48–50bn: multi-year revenue visibility from investment-grade counterparties, which is rare at this scale. It is already converting to cash: deferred revenue (customer prepayments) jumped to $4.8bn at end-March from $1.6bn three months earlier. Customers are paying up front to hold their place in the queue.

Nebius is the clear number-two neocloud by backlog, with two customers making up essentially the whole book. Source: company disclosures; Evercore ISI, Wolfe Research, United Securities. Chart by chokepoints.ai.

04 How it makes money

The model is straightforward to describe and very capital-intensive to run: buy chips, house them, rent them. What matters is utilisation, price per chip-hour, and how long the chips stay economically useful.

1. Raise capital ($20–25B). FY26 capex. Funded by equity, convertibles, customer prepayments and, soon, asset-backed debt.

2. Build the kit. About 80% of spend is Nvidia GPUs; ~20% is data centre, power and cooling.

3. Rent it out. Reserved multi-year capacity (Microsoft, Meta) plus on-demand for enterprises and AI-native customers, which carries higher margin.

4. Earn per MW. Management cites up to ~$20M of annual revenue per megawatt on optimised inference workloads.

One number decides whether the model works: GPU useful life. At 5 years , the economics are solid. At 3 years , they shrink fast. Nebius extended its assumed life from 4 to 5 years in January 2026, lifting Q1 profit by about $42M. Hyperscalers use 5–6 years; bulls call it reasonable, bears call it the single biggest lever on reported earnings quality.

05 Contracted power versus connected power

Contracted power is enormous: over 3.5 GW secured by May 2026, with management ratcheting the year-end contracted target between 2.5 and 4 GW. Connected power (built, energised, earning) is a sliver: guidance is 800 MW–1 GW by year-end, and the company exited 2025 with only ~0.2 GW live. Roughly a tenth of what’s signed is generating revenue today.

That gap is the centre of the near-term debate, and broker coverage treats it as the main risk. Nebius owns more than three-quarters of its contracted power rather than leasing, which means higher returns if it works and higher execution risk if it slips, across a pipeline of 16–17 sites. The cautions are specific:

We see moderate risk to the timing of planned builds for five of its largest sites… its last data-centre build completed over a decade ago in Finland. We expect the majority of FY26 capacity from colos, with greenfield primarily impacting the FY27 ramp. Sell-side capacity review, May 2026

The gap: only ~5–10% of contracted power earns revenue today. Watch the gigawatt sites in Pennsylvania (1.2 GW) and Independence, Missouri (1.2 GW, first 200 MW late 2026). Also note: Vineland, NJ (the Microsoft anchor site) is facing a community noise lawsuit. One mitigant is a **328 MW** Bloom Energy on-site fuel-cell deal that lets Nebius sidestep slow grid interconnection. Source: Nebius guidance; capacity reviews; public filings. Chart by chokepoints.ai.

06 Funding the build

Cash plus customer prepayments do not cover the 2026 build. Nebius carried $9.3bn of cash at end-March (up from $3.7bn) and $4.8bn of prepayments, call it $14bn of resources against a capex guide of $20–25bn, roughly four-fifths of it Nvidia silicon. The gap is filled with debt and equity-linked paper, and that’s the dilution overhang.

Already raised: a $2.0bn Nvidia pre-funded warrant (strike $94.94) and $4.34bn of convertibles, both March 2026, plus an at-the-market equity programme (~$2.9bn unused). Risk: if the gap is plugged with parent-level converts and equity rather than asset-backed project debt, shareholders wear the dilution. One bear scenario in broker research puts it at ~17% a year through 2030. Source: Nebius filings; financing notes. Chart by chokepoints.ai.

07 Operating leverage, and a flattered bottom line

The operating story is strong. Gross margin went from 51% to 74% in a year; group adjusted EBITDA turned positive and reached $130m at a 32% margin, with the core AI-cloud segment at a 45% EBITDA margin. When the fleet fills, the unit economics work.

The headline $621m of Q1 net income is mostly an accounting effect: a $781m non-cash gain from marking up the ClickHouse stake (which raised at a ~$15bn valuation in January). Strip it out and the underlying operation ran a $128m operating loss. Two more flags worth noting: the GPU useful-life assumption was stretched from four to five years this quarter (which flatters depreciation), and operating cash flow is driven by customer prepayments, not yet by profit.

Every cost line falls as a share of revenue as the fleet fills. This is the core bull point: unit economics work once capacity is used. Source: Nebius Q1’26 6-K; chokepoints.ai calculations. Chart by chokepoints.ai.

08 Competition, and a closing window

Nebius is the clearest number-two to CoreWeave. CoreWeave is bigger and faster on the top line ($2.1bn Q1 revenue, +112%, on a ~$99bn backlog) but carries over $20bn of debt and a US-centric footprint. Nebius is smaller but better-funded for its size, net-cash-ish, spread across the US, Europe and the Middle East, and pushing into software to make the offering stickier than raw GPU rental.

Nebius is viewed as a key player alongside CoreWeave, Lambda and Crusoe. It differentiates by investing in software and inference platforms to increase stickiness and avoid commodity GPU-reseller status. Expert Insight, 16 Apr 2026

The category’s pricing power has a time limit. It exists because hyperscalers are short of compute, and Bernstein and others expect that shortage to ease toward 2030 as AWS, Microsoft and Google bring their own capacity online. At that point raw compute commoditises, and the neoclouds that survive will be the ones that moved up the stack in time. CoreWeave targets 8 GW of active power by 2030; Nebius is pursuing the same scale with materially less debt.

09 Valuation

At roughly $222 a share the equity is worth about $57bn, already down ~20% from its $279 high. On enterprise value over run-rate revenue, you pay about 29× today’s ARR, or about 7× the exit-2026 target if the build lands on guide. The valuation is a bet that the power gap in section five closes on schedule.

83% of gross receivables sat with Microsoft at 31 March, up from 59% a year earlier. That gives multi-year visibility, but also binary concentration and renewal risk. Source: Nebius 20-F / 6-K; EV from public market data; multiples are chokepoints.ai’s own. Concentration corroborated via broker research.

10 Three scenarios

11 The two-sided case

12 Scorecard

13 What moves the signal

The number we will track from here is connected megawatts.

Not investment advice. For information and discussion only; not a personal recommendation, offer or solicitation. Capital is at risk. The “signal” is an analytical view from public fundamentals and licensed research read in summary, not a price target. Do your own work.

© 2026 chokepoints.ai · Issue 001 · Nebius sits at L6 (cloud & infrastructure software) in the chokepoints AI-compute map. All issues ·

Read the full interactive atlas of the AI compute stack at chokepoints.ai.

The chokepoints.ai atlas: the AI-compute value chain, layer by layer.

When SpaceX listed, up to a third of the float was earmarked for retail. By then the prize had already been settled: early investors and more than 4,000 employees were millionaires before the first public share traded, around 400 were centimillionaires, and a handful were billionaires. That is how most of the obvious AI winners have gone too. The gains were made in private, years before anyone outside could buy in.

Which makes the useful question a different one: where did the prize go for investors who could only buy public stocks? The market answers it cleanly.

01 Where the prize went

If you’d invested in the S&P in 2013, you’d have 5x’d your money in 13 years. Not bad, but a more concentrated NASDAQ bet would’ve yielded 8.1x.

All sounds great, until you see that investing in an index of the top 10 stocks would have yielded 113x, a 40+% IRR, easily beating most top-quartile hedge funds. (Obviously you’d have had to keep tracking the top 10 as it changed over the years, but it’s useful as a thought exercise.)

Almost a dream: 113x on $10k makes you a millionaire in 13 years. That is, unless you’d gone all in on Nvidia and made a staggering 684x in that time frame.

Growth of $100 from Jan 2013 · log scale

The whole market returned 5×; the single chipmaker at the chokepoint returned 684×. Note the log scale: each gridline is 10× the last. Source: public market total-return data, Jan 2013 to Jun 2026; top-10 basket equal-weighted, rebalanced as constituents changed. Chart by chokepoints.ai.

The effect only amplifies from 2025 onwards: investing in the AI picks and shovels has made Leopold Aschenbrenner 12x on his money (1000% net of fees), and a SanDisk pick would’ve yielded 34x in the same period.

Growth of $100 from Apr 2025 · log scale

The same pattern, faster: SanDisk is a holding in Aschenbrenner’s book. Log scale. Source: public market total-return data, Apr 2025 to Jun 2026; fund figure as reported. Chart by chokepoints.ai.

Notice the direction of travel. The whole market made you 5x, big tech 8x, the ten giants 113x, the single chipmaker 684x, and the infrastructure names since 2025 more still. The alpha was never in the apps. It compounded one layer down: in the compute, the power, the memory. When compute was the chokepoint, Nvidia and its virtual monopoly won. When memory became the chokepoint, SanDisk won. When electricity was the chokepoint, Bloom Energy won.

We’re sceptics of our own convictions and quite reserved investors, but even then we’re up 80% YTD on actual portfolio net worth, not just a tiny pot.

The authors’ portfolio, year-to-date rate of return, account value redacted. Brokerage screenshot, 2026. Past performance is not indicative of future results.

Reading Issue 000? chokepoints.ai is a weekly research note on the AI-compute supply chain: where the bottlenecks are, and which companies benefit from them.

Get the Briefing · No spam · unsubscribe any time

02 What this is

Though the majority of foundation model opportunities have stayed private for a while, there are thousands of public companies further down the AI supply chain that you could’ve bought into, and we’re here to help you find the next batch.

With the chokepoints.ai tool and newsletter, we’ve mapped the end-to-end AI value chain, its 2,376 dependencies, and the 17,500 companies within it, so that you can understand what the core drivers, barriers and opportunities are. More importantly, we’ve highlighted the 95 chokepoints: the areas within AI where the bottlenecks are greatest. Over the next few months, we’ll be dissecting these even further to build a more comprehensive view of both the strength of each chokepoint and its geopolitical status.

Our objective is NOT to

• Help degens yolo into stocks.

• Promise 1,000% returns year on year.

• Give Twitter influencers ammo to copy.

We serve only to

• Educate you on what the full AI stack looks like.

• Open the discussion on interesting investment ideas in the space.

• Give the world a better dataset for the space: which skills, sectors and geographies are worth knowing.

03 Who we are

As co-authors, we’ve been fortunate to live in this world for a while now, one of us as a machine learning researcher at one of the UK’s leading AI labs, the other active in the Effective Altruism and AI safety movements since first year of undergrad at Oxford.

The chokepoints thesis has shaped bigger decisions than where we invest:

• Building and running the AI team at a top-decile private equity fund.

• Getting deeper into the x-risk AI safety movement, working with a range of organisations.

04 Why we bother

A word on why we bother, because this is a pillar of the whole project and not an aside.

There is an evolutionary mismatch between genes that change over generations and the exponential rate of progress in AI, and it is the core driver of humanity losing its status as the most powerful entity in the world. Just as we evolved from primates and eventually overpowered them, AI has evolved from humans and may overpower us.

The chokepoints are where the leverage is: the points at which this technology can still be steered, slowed, or governed. They are also where the returns are, and that is no coincidence. Scarce, hard-to-replace positions are what markets price and what policy pulls on. Understanding them is the best way to make sure that you benefit from the technology rather than lose out to it.

We take that seriously. It is precisely why we would rather you understood this infrastructure than looked away from it.

Whether you’re just getting started in AI or investing, looking to start or buy a business, or managing billions, we’re here to give you a real understanding of the picks and shovels of AI, one level deeper than the chips: the places the market has yet to price properly.

Subscribe for the map.

Welcome to Chokepoints.

Subscribe for the map One company from the AI-compute supply chain examined each week: what it controls, what it depends on, and where the risk sits.

Get the Briefing · No spam · unsubscribe any time

Disclaimer. Chokepoints is for informational and educational purposes only. Nothing in this newsletter or the chokepoints.ai tool constitutes financial, investment, legal, or tax advice, an offer or solicitation to buy or sell any security, or a recommendation of any investment strategy. Always do your own research and consult a qualified financial adviser before making investment decisions. The authors may hold positions in companies or assets mentioned, and may buy or sell them at any time without notice. Parts of our research and writing are produced with the assistance of AI tools, which can make mistakes. Despite our review, figures, facts, and sources may contain errors, so verify anything material before acting on it. Past performance is not indicative of future results.

About the charts. Growth-of-$100 figures use public total-return market data; the top-10 basket is equal-weighted and rebalanced as constituents changed, as a thought exercise rather than an investable strategy; multiples are rounded. The fund figure is as publicly reported. The portfolio figure is the authors’ own brokerage screenshot, June 2026. Every chart reconstructed by chokepoints.ai.

© 2026 chokepoints.ai · Issue 000 · Welcome. All issues · Subscribe