The Three Waves of AI and the $250 Billion Question

The Week's Three Data Points

On July 22, Alphabet reported second-quarter revenue of $119.8 billion, up 24% year over year, with Google Cloud revenue growing 82% to $24.8 billion, the segment's fastest growth on record (the acceleration path over the past year and a half: 28%, then 48%, then 63%, now 82%), with cloud operating income roughly tripling. Management raised 2026 capital-expenditure guidance to $195-205 billion and signaled significant further growth in 2027. The market's response was to sell the stock on the spending, a reaction worth sitting with: the most profitable buyer of AI infrastructure on earth reported that demand is accelerating, and was punished for investing to meet it.

On July 23, Intel reported second-quarter revenue of $16.1 billion, up 25%, its fastest growth since the third quarter of 2011 and roughly $1.8 billion above its own April outlook. Data Center and AI segment revenue rose 59% to $6.3 billion, and management said its data center operations cannot keep up with orders. AI-driven businesses grew more than 70% year over year and, per management, now contribute roughly 70% of total revenue. Even the foundry business, still loss-making, narrowed its operating loss to $2.1 billion from $3.2 billion a year earlier as the 18A process ramped. The shares rose double digits after hours. The significance for the broader thesis is breadth: the AI buildout is now reviving demand well beyond the accelerator leader, because every accelerated data center also needs host processors, networking, and the industrial base to make them. When the number-three beneficiary of a spending cycle prints its best growth in fifteen years and reports demand it cannot fully serve, the cycle is not narrow and it is not finished.

Then, on July 27, the Wall Street Journal reported that NVIDIA is in talks to provide a roughly $250 billion backstop for financing tied to OpenAI's planned 10-gigawatt data center campus in Ohio, a project developed with SoftBank whose all-in cost, including chips, could exceed $500 billion. Because OpenAI lacks an investment-grade credit rating, NVIDIA's guarantee would let the project borrow at favorable terms. Reports describe a separate discussion about financing chip purchases that could total roughly $350 billion more. NVIDIA shares fell about 3.5% on the report, and prominent skeptics moved quickly: one well-known short seller added to his position, summarizing the structure as "around and around we go."

These reactions cannot all be right at once. If AI demand is strong enough that Alphabet's cloud can grow 82% at scale, and strong enough to hand Intel its best quarter since 2011, the marginal gigawatt of compute is scarce and valuable. If the marginal buyer of that compute requires its chip supplier to co-sign its debts, perhaps the demand is more fragile than it appears. Resolving that tension requires a framework, not a headline. Here is ours.

The Three Waves

Wave 1 (now): the race toward maximum capability

The current era is organized around a single question: how powerful can artificial intelligence be made, and can it reach what researchers call artificial superintelligence? Every frontier lab is running that experiment at maximum intensity. This is a wave of capability proofs and benchmark races, funded not by the economics of average tasks but by the size of the tail prize: leadership in machine intelligence that could compound across science, engineering, medicine, and industry.

Two things about this wave are underappreciated. First, not every task needs superintelligence, and that does not matter to the wave's economics; a handful of major breakthroughs would justify the buildout on their own. Second, the spending at the buyer layer is being validated by cash flow, not hope. Alphabet's cloud growing 82% with tripling segment profits is not speculative demand; it is paid demand. We also observe enterprises working hard to minimize their token spend. In our view that efficiency effort does not change the long-term picture: every unit-cost decline in AI to date has expanded total consumption faster than it reduced unit revenue. Google disclosed processing 3.2 quadrillion tokens per month this spring, roughly a 330-fold increase in two years, across a period when token prices collapsed by orders of magnitude. Cheaper intelligence has meant more spend on intelligence, not less.

Wave 2 (beginning): smart, cheap, and fast

As frontier capability converges among several providers, the competition shifts to delivery: who can serve intelligence that is simultaneously smart, cheap, and fast, to everyone, everywhere. Speed is the most interesting of the three axes, because modern reasoning models convert computing time into answer quality. A provider with ten times the serving speed can afford ten times the reasoning on every query at the same response time, which means the fastest infrastructure does not merely answer sooner: it answers better. Speed is becoming intelligence.

The decisive fact about this wave is that it requires more compute than the capability race, not less. Serving near-frontier intelligence to billions of people at low latency and low cost is a larger industrial problem than training frontier models. This is the wave that turns today's 10-gigawatt campus announcements from apparent excess into baseline capacity.

Wave 3 (the big one): physical AI

The third wave is intelligence entering the physical world: robotics in the broadest sense, from humanoids to autonomous logistics, manufacturing, agriculture, and construction. Prior waves manipulated information; this one manipulates matter, and its addressable market is measured against global labor and physical-economy spending, categorically larger than anything software touches. Its compute signature is also different: training world models on video and simulation dwarfs text-scale data, and every deployed machine becomes a recurring compute node. Demand from this wave adds to the data center economy; it does not substitute for it.

The $250 Billion Question

With the framework in place, the NVIDIA question becomes answerable. We present both cases as strongly as we can, because we hold positions in this space and readers deserve the argument, not the conclusion.

The Moat Question: Who Else Could Do This?

There is a reading of the backstop that the sell-off ignored, and it requires stating precisely, because the honest answer is not "only NVIDIA."

The hyperscalers can do this, and one already has. In the spring, Google agreed to backstop lease payments underpinning a roughly $35 billion chip-financing arrangement for Anthropic: a special-purpose vehicle syndicated by Apollo and Blackstone buys Google TPUs and leases them to Anthropic, with Google guaranteeing the payments at five data centers and Broadcom providing a residual-value guarantee on much of the senior debt. The strategic consequence is the point: the deal deepened Anthropic's commitment to Google's custom silicon for years. Balance-sheet sponsorship of AI labs is not a NVIDIA invention; it is the new competitive currency of the platform war, and Google fired first.

What is true is narrower and, for NVIDIA, more important: no merchant chipmaker can answer. Intel is in the middle of an impressive operational turnaround, but it remains a net recipient of outside capital: the United States government took an equity stake in 2025, SoftBank invested, and NVIDIA itself invested $5 billion, while foundry capex above $20 billion this year runs against a segment still losing money. Intel cannot backstop anyone; it is the one being backstopped. AMD is executing well in accelerators, but its annual free cash flow is a single-digit fraction of NVIDIA's roughly $100 billion; a commitment at this scale would be existential for AMD's balance sheet, while for NVIDIA it is a contingent liability against a fortress.

Which reframes the backstop as necessity, not indulgence. The Google-Anthropic deal demonstrates that the custom-silicon threat now arrives bundled with financing: an AI lab choosing its long-term silicon platform is partly choosing which balance sheet will sponsor its buildout. Had NVIDIA declined to match, it would have conceded that dimension of competition to the hyperscalers entirely, and with it, over time, its most important customers. The backstop converts NVIDIA's greatest static asset into dynamic customer lock-in at exactly the moment the threat demands it.

It is fair to dislike the credit exposure. It is not accurate to say the transaction has no strategic logic. The companies that could not do this include every merchant competitor NVIDIA has; the companies that could are precisely the ones NVIDIA is defending against.

What Would Prove Us Wrong

A framework that cannot be falsified is a slogan. Ours has explicit break conditions, and we publish them:

• The efficiency ratio turns. If total token consumption growth falls below token price deflation for four consecutive quarters, the volume flywheel that underpins waves one and two is failing, and the compute buildout is overbuilt.
• The profitable buyers blink. If Alphabet, Microsoft, Amazon, and Meta begin cutting AI capital spending while only speculative, credit-dependent buyers keep building, demand quality is deteriorating from the top down.
• The financing spreads. One guarantee for a flagship customer is strategy. Guarantees for the third, fourth, and fifth customers would mean the demand curve as a whole cannot fund itself: that is the telecom-era pattern, and it would change our view regardless of reported earnings.
• The lock-in fails anyway. If OpenAI proceeds with large-scale custom silicon despite the financing relationship, the credit risk was assumed without securing the strategic benefit.

None of the four conditions is present in the current data. We monitor all of them quarterly.

Our Opinion on NVIDIA

Sources