OpenAI has its own chip, Nvidia buys models | Future Intelligence #24
AWS has ordered another 2 million GPUs from Nvidia for 2027-2028 across three chip generations that largely don't exist yet - and today this capacity is financed by debt, not operational cash flow. It also became clear where the real bottleneck is: Nvidia's long-term purchase commitments jumped to $279 billion from $119 billion, mainly due to memory, whose shortage according to Micron will last until 2028. And OpenAI, with the first benchmarks of the Jalapeño chip, showed that compute performance per kilowatt can be obtained outside Nvidia's rack systems.
Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically every Sunday morning by email - in full length, including specific scenarios and market implications. Permanent access to the report is obtained through Bulios Black membership.
Key points
Why AWS is ordering two million GPUs for hardware that largely doesn't exist yet
What it means that Nvidia locked $279 billion in long-term commitments and what's behind it
How memory became the new AI bottleneck and why Micron is doubling investments
What Nvidia gains by buying Hugging Face and why it's reaching for a layer above its own silicon
Why Salesforce and Okta moved by a fifth and what OpenAI's inference chip brought
AWS ordered 2 million GPUs. Capacity is now paid by debt, not cash flow
Amazon and Nvidia announced an expansion of their collaboration, under which AWS will deploy an additional 2 million Nvidia GPUs across its global infrastructure in 2027-2028. These will be Blackwell Ultra, Rubin, and Rubin Ultra chips, three successive generations - the order covers hardware that largely doesn't exist yet.
The agreement also includes the arrival of Nvidia Vera CPUs to AWS as a compute option for agentic AI and the construction of AI factories for the U.S. government, including 100,000 GPUs on AWS's secured infrastructure. At the same time, Amazon is developing its own Trainium chips; customers will be able to choose between Nvidia GPUs, Trainium, or both.
Why two chip generations are bought in advance
An order three years ahead is not a sign of demand certainty, but queue management. Production of advanced accelerators is planned years in advance - from advanced packaging capacity through HBM memory to transformers and cooling. Whoever doesn't reserve a place in the queue today gets capacity in 2029, by which time the cycle will already be decided.
The second layer of the story is a shift from training to inference, i.e., to running finished models. Nvidia said that the Groq racks it acquired in December for about $20 billion will launch this year together with Vera CPUs and Rubin GPUs at the neocloud company Nebius.
Groq is built on an architecture optimized for fast response generation, not for training models. It was Nvidia's largest acquisition to date, surpassing the purchase of Mellanox in 2019 for under $7 billion. The economics of AI are shifting to where you charge per token, not per training.
The cash flow that disappeared
The bill is visible on the balance sheet. Amazon reported second-quarter revenue of $200.6 billion (+20%), AWS grew by 37% - the fastest since late 2021 - and operating profit of $27.5 billion (+43%). However, the $62.6 billion profit included $53.4 billion in pre-tax gains, mostly from its investment in Anthropic.
Free cash flow over the past 12 months fell to -$7.6 billion from an inflow of $18.2 billion a year earlier, even though operating cash flow rose 33% to $161.4 billion. The difference is capital expenditures, which increased by $66.1 billion year over year. Operations are earning record amounts, but everything is immediately poured into buildings and chips.
It's not a one-off fluctuation. At its July results, Amazon raised this year's capital expenditure guidance to about $220 billion; Morgan Stanley expects free cash flow around -$17 billion, Bank of America a deficit of $28 billion. According to J.P. Morgan, the five largest U.S. hyperscalers are set to spend nearly $700 billion this year.
Who's financing it
Debt is replacing the missing cash flow. Hyperscalers issued about $121 billion in bonds in 2025, more than four times the five-year average, and total debt issuance for data centers almost doubled to $182 billion. Morgan Stanley expects hyperscaler issuance of $250 to $300 billion this year.
The second tier is private credit. The $7.5 billion CoreWeave facility led by Blackstone showcased a model of loans secured by GPUs and customer contracts with an average rate of about 11% and repayments starting January 2026. The collateral is hardware whose lifespan and residual value are the least certain variable of the entire cycle.
That's where the $5 billion debt package that $JPM is arranging for the construction of Volta Infra Holdings data centers is headed. The bank also estimates capital expenditures of the five largest hyperscalers this year at $697 billion and emphasizes that the cash flow profile of these builds differs from classic investment grade.
Where it shows up in valuations
For $AMZN, this changes the nature of the stock: from a cash generator to a capital-intensive infrastructure operator. The growth rate justifies it for now - TD Cowen expects AWS revenue of $165 billion in 2026 (3% above consensus) and $222 billion in 2027, implying acceleration to 28% and 34%, respectively.
For $NVDA, the two-million order provides revenue visibility through 2028, but also a reminder of concentration: buyers increasingly pay on credit. According to media reports, the company also suspended part of its credit support program for AI clouds in exchange for a revenue share due to antitrust concerns - and may revamp it.
The most sensitive to this structure are the intermediaries between capital and silicon: neoclouds like $NBIS and $CRWV, whose margins depend on the difference between the cost of credit and the price of leased capacity. As long as contracts grow faster than interest expenses, the model holds. This ratio, not GPU orders, is the true indicator of the cycle's health.
Memory is the new AI bottleneck: Nvidia locked $279 billion, Micron doubles capex
Nvidia said in its quarterly report that its long-term supply and purchase commitments rose to $279 billion from $119 billion in the previous quarter. CFO Colette Kress clarified that the increase is mainly related to memory purchases. It's not cautious inventory building, but capacity reservation for years ahead.