Over the weekend, the memory chip industry received the strongest confirmation of demand it has ever been given. Samsung Electronics and SK Hynix signed supply agreements worth roughly $950 billion in total — SK Hynix around $750 billion to American companies including Nvidia, Samsung a further $200 billion or so to Broadcom. The subject was high bandwidth memory, the components that sit inside every AI accelerator. And these were no longer the annual contracts that have been standard in this industry for decades, but five-year agreements running to 2030.
On Tuesday morning, Samsung Electronics fell 13.4 percent, its worst single trading day in almost two decades. SK Hynix lost 14.7 percent. South Korea’s Kospi closed 10.8 percent lower at 6,023.66, its weakest finish since April, with the exchange halting trading for twenty minutes — the eighth circuit breaker of the year.
This is not a contradiction. It is the answer to a different question. The market has stopped asking whether demand for AI infrastructure is real. It is now asking whose balance sheet that demand is standing on.
What was actually reported on Monday evening
While the Korean order numbers still filled the headlines, the Wall Street Journal reported that Nvidia is negotiating a guarantee of roughly $250 billion. The subject is not a chip contract but a lease: OpenAI wants to rent a data center campus with ten gigawatts of power capacity in southern Ohio, built by SB Energy, the energy arm of Japan’s SoftBank, on the site of a decommissioned uranium enrichment facility about fifty miles south of Columbus. Total project costs including the chips are expected to exceed $500 billion, which would make it the largest data center project ever announced anywhere.
Nvidia would therefore be standing behind the lease and construction debt of its own customer. A parallel negotiation covers financing for OpenAI’s chip purchases, on a scale of up to $350 billion. The reason for the structure is stated in the reporting itself, and it is the heart of the story: OpenAI is not profitable and therefore cannot obtain an investment-grade credit rating. Lenders want someone who will step in.
Investor Michael Burry compressed the reaction into a single post on X: “Around and around we go. Nvidia to guarantee $200 billion of ChatGPT’s spending on $NVDA chips.” Nvidia itself fell 4.99 percent to $196.51 on Monday, dropping back below the $200 mark. AMD lost more than five percent, Micron five, Sandisk around twelve.
The $950 billion that is not money
This is worth looking at carefully. The $950 billion out of Seoul is not a transfer, not an equity stake and not an investment. It is the projected value of committed forward purchases over five years. An order book, in other words — and an order book is worth precisely as much as the creditworthiness of whoever signed it.
That is exactly where the two stories connect, despite appearing unrelated on the surface. On the same weekend that the memory makers collected five-year commitments, it became public that the strongest name in that supply chain is prepared to stand behind the obligations of the weakest. Both describe the same chain, viewed from opposite ends.
The line worth taking away from this Tuesday is therefore this: an order book and a guarantee for that same order book are not two assets. They are one risk, counted twice. Anyone reading the Korean contracts as confirmation while treating the Nvidia guarantee as a separate annoyance is counting the same promise once as revenue and once as collateral.
What a guarantee tells you about credit
The question of how solid those commitments are can also be answered with publicly available numbers. OpenAI reached an annualized revenue run rate of about $25 billion in February, up from $20 billion at the end of 2025 — roughly $2 billion a month, of which about $17 billion comes from ChatGPT subscriptions, $6.5 billion from the programming interface and $1.5 billion from video products and licensing.
Against that stand an expected loss of around $14 billion for 2026 and cash burn of roughly $27 billion, with about $63 billion already projected for 2027. In the first quarter of 2026 the company burned $3.7 billion against $5.7 billion of revenue. On current projections it does not turn cash-flow positive until around 2030. But the decisive figure is a different one: against $25 billion of annual revenue sit compute commitments of $80 billion to $100 billion per year over the next five years.
Seen against that backdrop, Nvidia’s willingness to guarantee is not a footnote to the financing. It is the most precise public statement yet of how an exceptionally well-informed participant assesses the creditworthiness of its own largest growth driver. You do not guarantee for someone who can comfortably pay the bill.
The second trigger: China is financing the supply side
It would be too simple, however, to trace the entire decline back to Ohio. On Monday, China delivered two pieces of news that hit the same nerve — but from the supply side.
First, memory maker CXMT debuted on Shanghai’s STAR market with a gain of 466 percent: a close at 49 yuan against an offer price of 8.66 yuan, with an intraday high of 55.03. The 57.92 billion yuan raised, about $8.6 billion, made it the largest semiconductor listing mainland China has ever seen. A market capitalization of roughly 3.3 trillion yuan — some $480 billion — makes CXMT the most valuable company on China’s A-share market, ahead of the major banks and internet groups.
Second, The Information reported that China has begun mass production of its own DUV immersion lithography machines, with first deliveries to SMIC, Hua Hong and CXMT expected this year. That matters strategically, because lithography has been the most reliable chokepoint in Western export controls.
So within a single trading session, two opposing financing models for the same bet stand side by side: the West is financing demand with guarantees, China is financing supply with equity. A company that manufactures memory is handed 466 percent in one day in Shanghai. A company that buys memory needs a guarantor in Ohio.
The weakest of the three arguments
A third motive circulated on Tuesday, and it is the one that holds up least well. The claim is that cheap Chinese models such as Kimi K3 from Moonshot AI — released on July 17, the largest open model ever built at 2.8 trillion parameters and the first to top the Code Arena rankings — will require fewer high-end chips and therefore undermine demand.
The actual record argues the opposite. Within 48 hours of release, demand so far exceeded expectations that Moonshot AI found its compute at cluster limits, suspended activation of new memberships and made capacity expansion the priority. Cheaper inference has so far produced more demand for compute, not less. Trading this thread means trading a reflex rather than a finding.
What the price action itself reveals
More informative than the explanations is the structure of the move. In US pre-market trading, Dow futures stand at 52,733, up 351 points or 0.67 percent, while Nasdaq 100 futures are down 1.10 percent at 27,880.50; S&P 500 futures are 0.12 percent lower. The VIX volatility index sits at 19.10 — elevated, but not panicked. Gold is down 1.21 percent at $4,027.70, and in a genuine flight to safety it would be rising.
On Monday the Dow had already added 262.83 points, or 0.51 percent, to 52,210.08, while the Nasdaq Composite lost 0.18 percent to 24,932.08 and the S&P 500 closed essentially unchanged at 7,413.18. Yields are calm too: ten-year Treasuries at 4.624 percent, thirty-year at 5.116 percent, both marginally lower.
This is not an exit from equities. It is the repricing of one credit chain. Money is not leaving the market; it is leaving a particular financing structure and landing with companies whose customers pay their invoices out of current earnings. Fittingly, Coca-Cola and Boeing report before the bell today, with Visa and Seagate after the close.
Where American investors are actually exposed
For US investors the distinction matters more than the headline percentages. Nvidia is the name at the center, and the risk is not that the Ohio guarantee is called tomorrow but that the market begins to apply a discount to revenue that arrives with a contingent liability attached. Broadcom is the quiet counterexample: it signed a $200 billion memory commitment with Samsung, but its custom accelerator business rests on hyperscaler customers who generate cash today.
Micron, Sandisk and Western Digital are the direct American analogues to the Korean pain, exposed both to the credit question and to the Chinese capacity question — Sandisk fell around twelve percent on Monday alone. Seagate reports tonight and will be read as an early tell on whether storage pricing power is holding.
The rotation trade deserves equal attention. Homebuilders remain the cleanest derivative of the ten-year yield, and consumer staples such as Coca-Cola are where the Dow’s 351-point futures gain is coming from. For taxable accounts, the familiar distinction applies: positions held longer than a year qualify for long-term capital gains treatment, while a reflexive exit from a chip position bought during the spring rally would realize short-term gains at ordinary income rates. In a week containing a Fed decision and four megacap reports, that is not a trivial consideration.
The counterarguments worth taking seriously
Four points argue against the pessimistic reading. First, a guarantee is not a payment. It is a contingent liability, triggered only if the debtor defaults, and Nvidia generates the cash to carry such a risk. Morningstar explicitly considers the circularity concerns legitimate while still viewing the stock as undervalued.
Second, the Ohio talks are, by all accounts, unfinished and subject to change. On a Tuesday morning the market priced an intermediate negotiating stage as though it were a signed document.
Third, China’s lithography progress is modest in volume: the reporting describes roughly five machines this year and about twenty in 2027, still behind the state of the art on performance and reliability and requiring further testing. The selloff prices in an import substitution that will take a decade.
Fourth, the skepticism cuts both ways. If the $950 billion is softer than the headline suggests because of the buyers’ credit, then the order book was never as robust as it looked — including before the crash. Both sides of this debate lose something at the same point, just with opposite signs.
What decides this week
The resolution comes quickly and all at once. The Federal Reserve begins its two-day meeting today and decides on Wednesday; the policy rate stands at 3.50 to 3.75 percent, and market pricing sits at roughly two-thirds for a hold and one-third for a 25 basis point increase. That is the denominator of every valuation.
The numerator arrives immediately afterwards: Microsoft and Meta on Wednesday, Apple and Amazon on Thursday. Around a third of the S&P 500 reports this week, including nine Dow members and four of the largest technology companies. After this Tuesday, the relevant question for each of them is no longer how large the capital spending budget will be, but what share of it is paid for out of their own earnings and what share rests on third-party commitments. SK Hynix reports quarterly figures this evening.
Analyst Matt Simpson described the mood in Asia by saying the market appeared to be in the despair phase of the selloff, with technology investors rushing for the exit. That may well be right. What remains remarkable is that the strongest demand confirmation in the industry’s history and its worst trading day in twenty years occurred within 72 hours of each other. When a five-year, $950 billion commitment cannot hold a share price up for a single hour, the market is no longer pricing the commitment. It is pricing whoever made it.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
