The US AI Labs, Part 2: The Suppliers Are Financing Their Own Customers

On September 22, 2025, NVIDIA and OpenAI signed a letter of intent. NVIDIA would invest up to $100 billion in stages as OpenAI brought new data-centre capacity online, and OpenAI would fill that capacity with NVIDIA systems. The first $10 billion was to arrive once the two companies signed a definitive agreement for the purchase.

That agreement was never signed. By February 20, 2026, Reuters reported, it had taken much longer than either side expected, and NVIDIA was instead close to putting $30 billion into OpenAI's next funding round as an ordinary investor. The same source said OpenAI would spend much of the new money on NVIDIA chips. The round closed at the end of March at $122 billion, valuing OpenAI at $852 billion once the money was in.

The paperwork changed. The direction of the money did not.

Part 1 ended on the point that the suppliers are paid first. This part measures what they paid to get there and what they now hold. The filings show stakes that are small next to the orders that come back, so what the suppliers mostly own is a claim on the labs' ability to keep paying. Lucent owned the same kind of claim in 2000, and one difference decides most of the comparison.

Two thirds of OpenAI's round came from its suppliers

NVIDIA was not the largest supplier in that round. Amazon committed $50 billion, $35 billion of it conditional, and in the same announcement OpenAI expanded its AWS contract by $100 billion over eight years. Counting both commitments in full, the two companies that sell OpenAI its computing supplied $80 billion of the $122 billion.

SoftBank and the venture funds made up the rest. They have nothing to sell OpenAI, and they set the price alongside the suppliers. Still, most of the cash behind the $852 billion mark came from companies that expect much of it back as revenue.

Each dollar of supplier equity buys up to $22 of orders

OpenAI's terms are mostly reported. Anthropic's are filed, which makes it the better place to measure what a supplier's dollar buys. Its S-1 lists the stock each supplier agreed to buy this year and the spending Anthropic committed in return, and the table below puts them side by side. Amazon's terms are the steepest. In April it agreed to buy about $5 billion of Anthropic preferred stock, and in the same month Anthropic committed to spend at least $110 billion with Amazon over ten years, "regardless of actual usage."

Anthropic promised $22 of orders for each dollar Amazon put in. Google's figure is about 11. AMD's is the only one in the filing written as a direct swap: AMD will buy up to $5 billion of Anthropic stock, drawable only as its GPUs are delivered and deployed, against hardware Anthropic expects to cost more than $20 billion.

At those ratios the equity works as a discount on the sale, paid in shares instead of price.

One arrangement in the filing is a different instrument. In March Anthropic signed an equipment lease with Broadcom for specialised AI accelerators, and alongside it an agreement to issue up to $42 billion of convertible notes, to Broadcom or a financial partner, to pay for that lease. Anthropic describes the notes as available if needed. Here the supplier is not buying stock. It is lending the customer the money for the customer's own purchase, which is vendor financing in the textbook sense and the instrument Lucent used.

The risk has moved from the buyer's cash to the price of a token

At the end of 2000 Lucent had committed about $7.3 billion of credit to its own customers, including a $1.6 billion loan it was still finalising, and had guaranteed another $1.8 billion of their debt. Its quarterly filing explained how those sales were booked. Lucent counted the revenue when the customer's notes were "reasonably assured of collection," judged partly on whether Lucent could sell the notes on.

In December, three months after its fiscal year closed, Lucent cut $679 million from that year's revenue, which Time described as on-credit purchases that fell through. In January it reported quarterly sales down 28 percent and took a $1.6 billion restructuring charge. By February the SEC was looking at how it booked sales. The buyers Lucent had lent to were, in Time's phrase, start-ups and emerging companies, and when their funding dried up the loans and the sales went together.

Anthropic is not that customer. It brought in $16.29 billion of revenue in the first half of 2026, ended June with $131.65 billion in cash and short-term investments, and lost $693 million at the operating line. Nothing in the filing points to a collection problem today.

Further out, the comparison holds. The S-1 says Anthropic's compute commitments generally have to be paid whatever its revenue or utilisation, and the longest run to 2045. Its three cloud contracts alone, with Amazon, Google and Microsoft, total $269.4 billion. Spread evenly over their terms they come to about $33 billion a year, roughly 71 percent of the run-rate Anthropic cited in May. The contracts are priced in capacity and dollars, not tokens, so if the price of a token falls faster than usage grows, the same capacity earns less against the same bill.

A take-or-pay contract leaves the lab carrying the price of its own product, and the supplier carries it only if the lab can no longer pay.

The first disclosure to watch is a drawdown on the Broadcom notes, which would mean Anthropic is paying for hardware with its supplier's credit rather than its own cash. A supplier commitment rewritten the way NVIDIA's was in February would mean another link proved too hard to keep. The clearest signal would be a true-up payment, a lab paying for compute it did not use, because it would show usage falling behind the contracts.

If, on the other hand, committed spend shrinks as a share of revenue quarter after quarter, the labs are outgrowing their contracts, and the shares the suppliers took in place of a discount will be worth more than the discount was.

Until one or the other shows up, the suppliers hold what Lucent held: a book of orders whose value depends on the buyer. This buyer is far stronger, and the book is far larger. Lucent's customer credit came to $7.3 billion; Anthropic's three cloud contracts come to $269.4 billion.