On Monday 14 September 2026 the Philadelphia Semiconductor Index fell about 5 per cent after a weekend in which the head of Anthropic called for a slowdown in frontier model development. Korea fell with it. The next day, with American chip stocks still going down, Samsung Electronics closed up 0.7 per cent and SK hynix up 1.82 per cent. The reason offered for that split, in Korean coverage and in English, is that the two Korean makers have locked in long-term supply contracts. That is the weakest of the three explanations available, and the one we can already show is doing less work than it appears to.
Last verified: 17 September 2026, correcting the version published on 16 September. Built from TrendForce’s public press releases of 2 June, 30 July and 4 August 2026, from S&P Global capital expenditure and cash flow estimates as reported on 6 September 2026, from Moody’s May 2026 hyperscaler forecast, from the index and share price moves of 14 and 15 September 2026 as reported, from published fab construction schedules as reported in February 2025 and March 2026, and from two Korean research reports on long-term supply agreements supplied to us, dated 30 July and 3 August 2026. The share price moves are settled. The 2027 supply and spending figures are estimates and will be revised, in some cases substantially.
Editor’s opinion
Right outcome, wrong reason
Our view is that Korean memory has a real defence against an artificial intelligence spending slowdown, and that it is not the one being cited. It is not the contracts. It is that 2027’s supply is already fixed by decisions taken two and three years ago, and no capital expenditure decision made this month can add a meaningful bit to it.
The contract argument is stronger than we first wrote, and we have corrected this section. Two Korean research reports we have since read describe the new memory agreements as carrying minimum committed volumes, a price floor and a ceiling, deposits and supply priority, with repricing taking place inside that band rather than against an open market. Samsung has said publicly that it sets a floor for commodity products. A contract that reprices annually inside a floor is not the same instrument as one that reprices without one, and our first reading conflated the two.
What survives is narrower, and it is still the point. On the same reporting SK hynix has said it designed its contracts to cope with price volatility but has not disclosed the structure. Samsung’s floor is a statement on a call rather than a filed term. Micron’s floor sits in its accounts with more than 22 billion dollars of customer deposits behind it. The gap between Korea and Micron, as we put it in our 10 August piece, is no longer a gap in what the contracts do. It is a gap in what a reader can check. That is a smaller claim than the one we published on 16 September, and it is the one the documents support.
Andrew Lee. This section argues a view rather than establishing a fact. It is not a forecast and not a recommendation, and the material it draws on is listed in the sources below.
What happened on the fourteenth and the fifteenth
Over the weekend of 12 and 13 September, Anthropic’s chief executive Dario Amodei published an essay calling for stronger safety measures and a slowdown in model development, after demonstrations of what current models can do in offensive cyber operations. Sam Altman and others backed the position. On Monday the American market turned that into a trade, and the trade was not a general risk-off move. It was a rotation out of the companies that sell the equipment and into the companies that buy it.
The Philadelphia Semiconductor Index fell close to 5 per cent. Intel, Advanced Micro Devices and Marvell each lost 5 to 6 per cent, and Nvidia more than 3 per cent. On the same day Alphabet rose nearly 2 per cent, Microsoft 1.6 per cent and Meta roughly 1.4 per cent. Gil Luria of D.A. Davidson gave the reasoning in plain terms: if artificial intelligence progress slows, the hyperscalers can stop adding capacity and harvest the returns on what they have already built, digesting the estate rather than extending it. Spending falls, free cash flow rises, and the people selling the picks lose.
Seoul took the same blow on the same day. The KOSPI closed down 3.26 per cent, Samsung Electronics down 4.05 per cent and SK hynix down 6.35 per cent, a deeper fall than the American chip index itself. On Tuesday the two separated. American chip stocks continued lower, in a session in which the ten-year Treasury yield went above 5 per cent for the first time since 2007 and crude rose above 107 dollars a barrel after a Saudi pipeline closure, while Samsung Electronics closed up 0.7 per cent and SK hynix up 1.82 per cent.
One reading of that is simply that Seoul absorbed the shock first and would not sell the same piece of news twice without evidence that orders, server deployments or memory consumption had actually weakened. That reading is probably right about the mechanics of the two sessions. It says nothing about whether the position survives the next three months, which is the question worth asking.
The bear case is arithmetic, not sentiment
It would be easy to file the Amodei essay under sentiment and move on. The reason not to is that the spending it calls into question is already projected to exceed what the companies doing it can fund from operations. S&P Global’s estimates, reported on 6 September, put six large infrastructure buyers at 470 billion dollars of capital expenditure in 2025, 870 billion in 2026 and 1.3 trillion in 2027. On the same estimates, five of the six are projected to run negative free cash flow in 2027.
2027 capital expenditure against 2027 free cash flow
Six large buyers of artificial intelligence infrastructure, 2027 estimates in billions of dollars, on S&P Global figures as reported 6 September 2026
| Company | Capital expenditure | Free cash flow |
|---|---|---|
| Alphabet | 357.0 | −82.7 |
| Amazon | 319.1 | −60.1 |
| SpaceX | 197.2 | −114.4 |
| Microsoft | 189.0 | +33.6 |
| Meta Platforms | 164.0 | −3.5 |
| Oracle | 95.0 | −41.6 |
These are estimates, not guidance. A different house counting a different set of companies gets a different total: Moody’s, in May 2026, put its six — the same list with CoreWeave in place of SpaceX — at 785 billion dollars for 2026 and close to 1 trillion for 2027, after marking its own March forecast up by 85 billion. The two sets are not comparable and should not be added together or averaged.
A company whose 2027 plan implies burning 82.7 billion dollars has a standing, arithmetic reason to look for something to cut, and Moody’s said as much in May when it warned that higher capital intensity and debt could force a reassessment of hyperscaler creditworthiness if profit growth does not arrive. That is what makes the slowdown trade something other than a mood. It is not that investors believe artificial intelligence has stopped working. It is that the buyers have been handed a respectable reason to do the thing their balance sheets were already pushing them towards.
The reason Korea was given, and what is wrong with it
The defence offered for Korean memory on Tuesday was contracts. Samsung told analysts on 30 July 2026 that it had signed long-term deals with the five largest data centre operators, was close to agreements with five more, and intended to place 60 to 70 per cent of its memory capacity under multi-year supply contracts on a rolling five-year basis. SK hynix has said it completed multi-year agreements with around ten key customers. Coverage published on 15 September repeated those July figures. They are the same figures, not new ones, and we went through them at length in our 10 August piece on what Micron discloses and Korea does not.
What is new is a piece of mechanism, and it cuts less cleanly than we first reported. In its 2 June 2026 release TrendForce explained that high bandwidth memory contract prices have not kept up with the rise in conventional memory prices because the three large suppliers negotiate those contracts on an annual cycle. Taken alone that sounds like a contract which resets in whichever direction the market has moved. Taken with what the makers have since described on their earnings calls, the reset happens inside a band.
Two Korean research houses have since set the structures side by side. A traditional memory contract ran on near-month volumes that could be renegotiated before the purchase order and on a price linked to the market and reset quarterly. The new agreements run about five years with rolling extension, convert a customer’s demand forecast into a take-or-pay obligation, set an upper and lower price band with repricing inside it, and are secured by deposits. Micron sets a cap and a floor and reprices quarterly within the band, with a premium for new products. Samsung has said it sets a floor for commodity products. SK hynix has said it designed its contracts for price volatility and has not described how. On one house’s reading the penalty for walking away runs to a share of total contract value, and the customer that walks loses allocation priority in the next shortage, so the binding force does not rest on the legal clause alone.
That is a materially different instrument from the one Micron describes in its FY2024 Form 10-K, where “pricing, quantity, and other terms will be periodically negotiated to reflect market conditions” with nothing underneath. A floor is a brake. We said on 16 September that the annual reset put Korea’s agreements in the older family, and on the evidence we have now that was wrong.
We should be careful about how far that goes. TrendForce is describing how the industry prices high bandwidth memory, not reading the text of any particular Samsung or SK hynix agreement, and a long-term deal with a hyperscaler could carry volume commitments or deposits that a pricing cycle says nothing about. But neither Korean company has published terms, and that absence is now in its second quarter. The people arguing that contracts will cushion the cycle and the people arguing they will not are both working from the same spoken summaries.
What is actually holding it up is the 2027 supply book
The stronger defence is physical, and it does not depend on anybody’s good faith. High bandwidth memory consumes wafer area out of all proportion to the bits it yields, because the stacking, the logic die and the yield loss all come out of the same wafer budget as ordinary memory.
High bandwidth memory as a share of the memory industry
Bits produced against wafers consumed, top three suppliers, on TrendForce estimates of 2 June 2026
| Share of industry total | 2025 | 2026 | 2027 |
|---|---|---|---|
| Bits produced | 8% | 9% | 13% |
| Wafers consumed | ≈18% | ≈22% | ≈30% |
By 2027 the three large suppliers are estimated to be spending close to a third of their wafer starts to produce an eighth of their bits. This is why conventional memory stays tight in a year when high bandwidth memory is the only thing anyone discusses, and why the Korean makers’ earnings in this cycle are not explained by high bandwidth memory share alone.
The rest of the supply picture follows from that. TrendForce put the 2026 memory sufficiency ratio at roughly minus 1 to minus 2 per cent, meaning demand exceeds supply, and expects the gap to widen through 2027 as demand growth continues to outrun capacity. Server shipments grew 17 per cent in 2026 and are expected to grow faster in 2027. High bandwidth memory bit shipments are forecast to rise 50 to 60 per cent in 2027 and still fall short of demand. New capacity, on the same 30 July assessment, will not ramp meaningfully until the second half of 2027 because of construction, equipment installation and materials lead times, with substantial output only in 2028.
That last sentence is the whole defence, and it is worth stating bluntly. If every hyperscaler decided this morning to cut its 2027 budget, the memory available to buy in 2027 would be almost exactly what it is today, because the fabs that would have served the extra demand do not exist yet and the fabs that will serve 2027 were committed years ago. A capital expenditure slowdown decided in September 2026 lands on 2028 and 2029. It does not land on the year the market spent Monday repricing.
The published construction schedules say so directly. SK hynix broke ground on the first fab of its Yongin cluster in February 2025, a 9.4 trillion won project it then described as due for completion in May 2027; by March 2026 that had been pulled forward, with the building finishing early in 2027, equipment going in during the second quarter and the first phases running later the same year. Samsung’s P5, the largest single addition either company is making, resumed construction in 2026 on an accelerated schedule and is still not expected to reach mass production until the latter part of 2028.
What arrives, and when
Korean memory capacity additions on published schedules, as reported to March 2026
| Site | Construction | Reaches output |
|---|---|---|
| Samsung P4 Pyeongtaek | Converting existing lines to the 1c process | Already contributing |
| Samsung P5 Pyeongtaek | Resumed 2026; building complete first half 2027, then equipment | Mass production late 2028 |
| Samsung P6 Pyeongtaek | Begins third quarter 2028 | Beyond this horizon |
| SK hynix M15X Cheongju | Expanding inside an existing building | Already contributing |
| SK hynix Y1 Yongin | Broke ground February 2025; equipment second quarter 2027 | First phases late 2027 |
| SK hynix Y2 Yongin | Begins third quarter 2028 | Beyond this horizon |
Only one line on this table adds a new building to 2027, and it was in the ground before the argument about artificial intelligence spending began. Everything decided in the second half of 2026 appears in the last two rows. Reported schedules also move: Yongin Y1 was announced for May 2027 completion and had gained three months by March 2026.
Not all memory is in that position. On the same analysis, flash memory goes the other way, with its sufficiency ratio turning positive in 2027 as higher-layer migration and new fabs arrive. The two halves of the memory business are separating, which is a reason to be careful with any argument that treats Samsung and SK hynix as a single exposure.
Nvidia is already rationing, and it cuts both ways
The most useful single fact in this argument has been public since early August and has not been connected to it. TrendForce reported on 4 August 2026 that Nvidia had begun, early in the third quarter, to evaluate alternatives to the twelve-high HBM4e stack that had been the baseline for Rubin Ultra, opening the specification to eight-high HBM4e, twelve-high HBM4 and eight-high HBM4. Two reasons were given: the memory shortage expected in 2027 limits the wafer capacity suppliers can turn over to high bandwidth memory, and the validation schedule and yield ramp for twelve-high HBM4e are uncertain.
Read as a bull point, that is about as strong as evidence of scarcity gets. The largest customer in the market is redesigning its flagship part around what it expects to be unable to buy. No forecast of tightness carries the weight of a purchaser acting on it.
Read the other way, it disposes of the simple version of the bear case. Memory content per accelerator is not a constant. It is a variable that the customer is currently holding down because supply is short. If artificial intelligence spending genuinely slows and the pressure on wafer capacity eases, the first thing that happens is not that memory orders fall by the amount that unit demand falls. It is that configurations go back up, because a twelve-high stack was what Nvidia wanted before it was told what it could have. Part of the demand loss gets absorbed inside the bill of materials before it reaches a memory maker’s order book.
This is the reason we think “artificial intelligence slows, therefore memory falls” is not an argument in its own right. It is a claim about unit volumes that has to survive a second step about content per unit, and right now that second step runs against it. It does not run against it forever. Once configurations have normalised, the cushion is spent.
What would show this to be wrong
Three things are checkable, and two of them have dates.
Micron reports on 30 September 2026. It is the only one of the three large makers that puts contract structure into auditable form, and its remaining performance obligation, customer deposits and any commentary on whether Strategic Customer Agreement coverage is still growing will say more about the contract question in one filing than another quarter of Korean earnings calls will.
Samsung Electronics files its half-year report and SK hynix its quarterly report with the Korean regulator, and any supply agreement large enough to require a single-contract disclosure would appear there as well. If either company puts pricing mechanism, volume commitment or deposit terms on the record, the contract argument becomes something a reader can check instead of something a reader has to accept. We will go through those filings and say what is in them. Anyone who wants to do the same can start with our guide to reading Korean filings in English.
The third is the supply forecast itself. If high bandwidth memory bit shipment growth for 2027 is revised down from the 50 to 60 per cent range, or if the memory sufficiency ratio turns positive earlier than expected, the physical argument in this piece weakens directly and the contract question stops being academic. That is the point at which it would matter a great deal whether Korea’s agreements reprice annually or not, and the point at which two companies that have published no terms would be asked for them.
What Tuesday showed is narrower than the headlines suggested. Korean memory did not decouple from artificial intelligence. It declined to sell a Monday story twice, and it has a genuine reason to hold through 2027 that almost nobody named, because that reason is a groundbreaking in February 2025 rather than a signature in 2026.
Sources
We take facts from these and write our own sentences. Where a figure originates in a research house estimate we name the house, and where it originates in a company statement we say which statement. We link to a document only where we hold the exact address for it.
Research houses
- TrendForce, “Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM”, 2 June 2026. Source for the annual contract negotiation cycle, and for high bandwidth memory at 8, 9 and 13 per cent of bit supply and roughly 18, 22 and 30 per cent of wafer input across 2025 to 2027.
- TrendForce, “Diverging Memory Market Outlook in 2027”, 30 July 2026. Source for the minus 1 to minus 2 per cent sufficiency ratio, the widening 2027 gap, the second-half 2027 ramp and 2028 output, the 17 per cent 2026 server shipment growth, and the flash memory divergence.
- TrendForce, “DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra”, 4 August 2026. Source for the Rubin Ultra specification evaluation and for 50 to 60 per cent high bandwidth memory bit shipment growth in 2027.
- S&P Global capital expenditure and free cash flow estimates for six infrastructure buyers, as reported by The Motley Fool, 6 September 2026. Source for the table of 2027 capital expenditure and free cash flow and for the 470 billion, 870 billion and 1.3 trillion dollar totals. We have not read the underlying S&P report.
- Moody’s hyperscaler capital expenditure forecast of 14 May 2026, as reported by Data Center Dynamics. Source for the 785 billion dollar 2026 figure, the approach to 1 trillion in 2027, the 85 billion dollar markup and the creditworthiness warning.
Reporting
- Fortune, “Wall Street’s AI doomsday trade is here”, 14 September 2026. Source for the 14 September moves in the Philadelphia Semiconductor Index, Nvidia, Intel, Advanced Micro Devices, Marvell, Alphabet, Microsoft and Meta, for the Amodei essay as trigger, and for the Gil Luria reasoning quoted.
- Invezz, “Samsung and SK Hynix defy a 6% US chip rout”, 15 September 2026. Source for the KOSPI, Samsung and SK hynix moves on 14 and 15 September, and for the argument that Seoul had priced the concern a day earlier.
- Aju Press, “AI slowdown may put a damper but not dent in Korean HBM demand”, 15 September 2026. Source for the restatement of Samsung’s July contract remarks and SK hynix’s roughly ten customers. The underlying Samsung figures are from the company’s second-quarter earnings call of 30 July 2026, not from new disclosure.
- Korea JoongAng Daily, “Samsung and SK are expanding fast, but why is memory still in short supply?”, 12 March 2026. Source for the fab schedule table: Samsung P4, P5 and P6, SK hynix M15X and Yongin Y1 and Y2, and for the point that near-term spending goes to process conversion and cleanroom infrastructure before it goes to wafer output.
- The Korea Times, “SK hynix breaks ground on its 1st fab in Yongin semiconductor cluster”, 25 February 2025. Source for the February 2025 groundbreaking, the 9.4 trillion won figure and the original May 2027 completion target.
- Quartz, on the 15 September session. Source for the ten-year Treasury yield above 5 per cent for the first time since 2007 and crude above 107 dollars after a Saudi pipeline closure.
Broker reports supplied to us
- Growth Research, “LTA 톺아보기” long-term agreement industry report, analysts Han Yong-hee and Kim Ju-hyung, 3 August 2026. Source for the side-by-side comparison of traditional and long-term contract structures, for Micron’s price band with quarterly repricing inside it, for Samsung’s stated floor on commodity products, for SK hynix having described no structure, and for the deposit figure of more than 22 billion dollars. We do not reproduce its tables or estimates.
- Hanwha Investment & Securities, SK hynix company report, analyst Park Jun-young, 30 July 2026. Source for the description of the new agreements as carrying minimum committed volumes, a price floor and ceiling, deposits and supply priority, and for the argument that the penalty for breaking one runs to a share of total contract value and costs the customer allocation priority in the next shortage. Its earnings and margin estimates are that house’s and we do not carry them.
Material we could not open
- TrendForce, HBM Industry Analysis 3Q26, report code RP260805KC3. This is the subscription report that carries the supplier-by-supplier share table and the 2027 bit supply and demand growth rates. We could not open it, we do not reproduce its contents, and where this article gives a direction rather than a percentage that is why.
- The Mirae Asset Securities note underlying the Google accelerator memory configurations cited in Korean coverage. We could not establish its date or author and have left those figures out.
Corrections
17 September 2026. We rewrote the contract section and the part of the opinion that rested on it. The version published on 16 September argued that because high bandwidth memory contracts reprice annually, Korea’s long-term agreements belonged to the price-renegotiable generation Micron describes in its FY2024 Form 10-K. Two Korean research reports we have since read describe the new agreements as repricing inside a floor and a ceiling, with minimum committed volumes and deposits behind them, and record Samsung stating publicly that it sets a floor on commodity products. Repricing inside a band is not the same as repricing without one, and the original claim overstated the case. What survives is the narrower point that SK hynix has disclosed no structure and that Samsung’s floor is a spoken statement rather than a filed term. Nothing in the supply section changed, and the conclusion of the article is unaffected.
If you can point to a primary document that contradicts anything above, we will correct it and say what changed.
Disclosure. The author holds no position in any company named in this article as of the date of publication. This article is general information and analysis about memory semiconductor supply, demand and customer contracts. It is not investment advice, not a recommendation to buy or sell any security, and not a forecast of any price or index level. The section headed Editor’s opinion is comment: it argues a view, and we fence it off from the sourced material for that reason. We have not tailored anything here to your circumstances, and only a qualified adviser in your jurisdiction can do that. We are not registered as an investment adviser in any jurisdiction, and this site has no commercial relationship with any company, exchange or research provider named above. See the Editorial & Sourcing Policy and Disclaimer.