AI chip financing at 11% while chip stocks slide

In one week this October, two readings landed side by side. In the public market, memory and chip stocks fell 6 to 9 percent over two days. In the private market, a $5bn loan raised to buy AI chips left $3.7bn ununderwritten, and paid close to 11 percent. Two markets priced the same assets in opposite directions for different reasons: one was pricing the future price of memory, the other the terms on which the money is raised. That second reading belongs on a buyer’s diligence sheet.

The selling side spoke first

On 12 October in Singapore, Broadcom chief executive Hock Tan said many investors do not really know what they are investing in, that they set extremely high expectations and end up in something close to a casino. His sharper line was that financing may be a bigger problem than the chip shortage, and that model developers need customers who actually make money from deployment rather than users burning cash on tokens for experiments. He compared enterprise token experiments to a child in an arcade burning ten dollars of tokens in ten minutes and learning nothing.

Intel chief executive Lip-Bu Tan, at the same event, said the buildout means hardware stays in short supply for two to three years, and that memory may be the first segment to lose momentum.

The market moved quickly. In Tuesday trading in Asia, the Kospi fell as much as 6.2 percent, Samsung Electronics and SK Hynix each dropped more than 7 percent, and Micron had closed 3.9 percent lower the previous session. The index is about a third below its June high, and foreign investors sold more than $6bn of Korean equities in October. One trigger that day was reporting that China’s CXMT had made progress on a memory architecture and planned higher-density server memory by the end of 2026, on top of Broadcom’s remarks.

Keep the speaker attached: Broadcom and Intel sit on the selling side of this chain. Their comments are judgments about demand. They are worth hearing and they are not a measurement of it. And when a price decline gets expected, what gets repriced first is something else — the terms on which the money was raised.

On the other side, a loan nobody would underwrite

The second reading came from a syndicated loan. Of Volta Infrastructure’s $5bn financing, JPMorgan and Goldman Sachs underwrote only $1.3bn of term loan; $3.7bn of chip loan went ununderwritten. The blended yield was around 11 percent, roughly 625 to 650 basis points over the benchmark, at 97 to 98 cents on the dollar, with no rating. No rating typically excludes the largest buyers, the collateralised loan obligations, from the pool.

Set the two readings together and the interesting part is not the yield. It is who is holding the risk. Underwriting means the banks take the exposure onto their books and then look for buyers. Leaving it ununderwritten means the paper either finds a buyer at that price or returns to the borrower. $3.7bn sits on the second side of that line.

One reading in the same week pointed the other way. DayOne, a Singapore data centre developer, raised about $3.6bn for a Malaysian campus. That money is a different animal: a multi-currency dual tranche including MYR 7.5bn of Islamic financing and an offshore term loan, and the deal had been in the market for over a year, closing only now. The arrangers say it was more than twice oversubscribed, and the campus is intended to carry a gold green-building certification.

Two discrepancies need flagging. On size, Reuters reports about $3.67bn while the arranger’s announcement says $3.5bn; the offshore term loan figures also differ between the two. We do not pick either as the definitive number.

Read together, the conclusion is that money has not become too expensive to raise, but the way it is raised has changed. The deal that closed got there by splitting its structure — different rule systems, currencies and certifications — not by accepting a lower price.

Why this is procurement, not market news

For a C-suite or transformation lead, the compute checklist had three lines: specification, price, delivery date. This week’s readings say a fourth belongs there — the terms on which the supply behind the capacity was funded.

The reason is mechanical. A data centre or a batch of chips takes one to two years from contract to useful work, and the money has to go out first. Whoever fronts it is relying on expectations of future demand. When those expectations get discounted, the first thing to move is not the price of chips but the terms of the funding. The buyer at the end of the chain feels a delivery delay or a changed clause, and the cause sits upstream.

The same mechanism yields a testable prediction. If expectations keep being discounted, the first visible change is not cheaper chips but restructured financings: more currencies, more certifications, more tranches, each splitting risk across a different rule system. If such transactions return to simple structures that are heavily oversubscribed, expectations have come back.

A funding checklist you can fill in

In practice, six questions belong in vendor diligence. Was the loan behind this capacity underwritten, and who carries the risk if it does not sell? Is it rated, and if not, how much does that shrink the buyer pool? How are maturities stacked, and when does the money need refinancing? How much capacity is promised to customers without long-term contracts? If demand expectations are discounted, which clause moves first — price, delivery date or service level? And what exactly are our renewal price and delivery obligations tied to?

Companies building their own capacity need one more layer: whether the funding structure counts as delivery risk. Structures that bundle equipment with leases are efficient when demand is certain and leave very little room to adjust when it is not. A project funded without third-party ratings, or by carving out the asset’s revenue rights into a separate vehicle, carries a thicker buffer against a demand swing.

Nothing on this list needs a new tool. Its value is turning something normally discussed only after a problem appears into a line in the contract.

Three structures, three different readings

To fill the checklist in, a buyer has to read the structure, and these deals come in roughly three forms. The first is a plain syndicated term loan, where the risk sits on bank balance sheets. That is the layer that tightened here, which is why the money came in through other doors.

The second splits the raise across different rule systems, such as an Islamic financing tranche or one attached to a green certification. The point is not a lower rate but a wider buyer pool: capital that cannot hold the asset under one mandate can participate under another. The fact that such structures get done is itself a usable piece of information about what is replicable.

The third carves out the asset’s revenue rights into a separate vehicle, common in lease-style financing of data centres and compute clusters. Its efficiency comes from separating the physical asset from the credit of the project owner, and the price is complexity and little room to adjust. For a buyer, the implication is blunt: the vendor’s price and delivery commitments are attached to a shell that is thinner than the vendor itself.

What not to write into the analysis

Three brakes. First, executive comments are usable only with the speaker attached, as one company’s judgment about demand. Second, index and share moves are market reactions, not fundamental readings; two factors drove that day, and attribution stops there rather than reaching a structural conclusion about a segment.

Third, the financing that closed had been in the market for over a year. It shows the structure works; it does not show that pricing has eased. Merging those two claims is the most common error in this kind of analysis.

Then separate what was said from what was verified. The oversubscription multiple, the certification plan and the conflicting size figures all come from parties to the transaction or from wire copy, and carry that status into any document rather than being flattened into fact.

Our judgment

Compute is turning from a commodity into a leveraged asset, and the buyer’s risk has moved from whether the capacity can be bought to who carries the interest rate and the maturity on this chain. That is a structural change, not a week of price movement.

Three things can be done now. Add the funding structure to vendor diligence as a line next to price. Write delivery obligations and funding conditions together, so it is clear how delivery clauses move if the terms behind that loan change. If you build your own capacity, put the funding structure and delivery risk on one page instead of leaving them to finance alone.

This judgment does not forecast prices, and it does not claim any particular supplier is in trouble. It changes one thing: what the buyer asks first. The question to leave with — if you signed a three-year compute contract tomorrow, could you say who is putting up the money behind it, and on what terms?

Sources

Broadcom chief executive Hock Tan and Intel chief executive Lip-Bu Tan, remarks in Singapore, 2026-10-12

Korean chip stocks and the Kospi intraday reading, 2026-10-13, which also reports the day’s triggers

Underwriting and pricing of Volta Infrastructure’s $5bn loan: $3.7bn of chip loan ununderwritten, blended yield about 11 percent, no rating

Second report on the same loan, giving the $1.3bn underwritten tranche

DayOne’s Malaysian financing, Reuters, total of about $3.67bn

Arranger announcement, self-reported total of $3.5bn and a claim of more than twice oversubscribed; conflicts with the wire figure and is not presented here as settled