The 70% Revenue Dependency: Why AI's Real Risk Is a Two-Customer Price War
AI demand is real, but two private model labs may drive too much cloud growth. The first risk is a price war, not a capex collapse.
The market has spent the summer asking whether the artificial-intelligence build-out has become too expensive. A more immediate question is hiding in the revenue line: how much of the cloud economy is ultimately a bet on two private model companies? Steve Eisman, the former Neuberger Berman portfolio manager whose skepticism helped define the 2008 trade, estimates that OpenAI and Anthropic account for roughly 70% of AI-related revenue at Microsoft, Amazon, Alphabet and Oracle, and between 25% and 35% of their cloud revenue. The number is an estimate, not an audited disclosure, but the risk it describes is concrete. AI demand can be genuine, earnings can be rising and the trade can still be vulnerable to a two-customer price war.
That distinction matters because the current debate is pointed at the wrong transmission mechanism. Investors are trying to decide whether the data-center build-out will generate sufficient returns over time. The first break, if one comes, may arrive earlier through pricing: a model customer delays a training run, shifts workloads between vendors, or cuts inference prices to defend market share. The servers do not disappear. The contracted revenue, utilization assumptions and return on capital attached to them can nevertheless change quickly.
Eisman laid out the concentration argument on CNBC Fast Money, in comments reported Aug. 13, 2026. “The futures of these massive companies, in a sense, are a bet that OpenAI, Anthropic are going to succeed,” he said. His point was not that Microsoft, Amazon, Alphabet or Oracle lack diversified businesses. It was that the incremental growth investors are paying for is increasingly linked to the spending decisions and competitive fortunes of a very small group of model developers.
The near-term numbers make the bet look rational. Anthropic’s preliminary second-quarter 2026 revenue exceeded $11.5 billion, according to Bloomberg and CNBC, compared with $787 million a year earlier and $4.73 billion in the first quarter. CNBC reported positive adjusted operating income, while stressing that the figures were preliminary. Bloomberg separately reported an OpenAI revenue run rate above $40 billion. Those figures are extraordinary, but extraordinary growth is not the same as diversified demand. A large model lab can be a powerful customer and still be exposed to a concentrated set of end users, a fast-moving technology curve and an aggressive need to buy compute before the revenue base is fully durable.
For the hyperscalers, the visible result is an impressive run of cloud numbers. AWS revenue was reported at $42.2 billion, up 37%, while Azure growth was 43% and annual sales exceeded $100 billion. Google Cloud revenue reached $24.8 billion, up 82%. The figures show that customers are paying for cloud capacity. They do not, by themselves, show how much of the spend is recurring enterprise consumption, how much is model-lab expansion, or how much is a temporary race to secure scarce capacity.
This is where the accounting language can become more reassuring than the economics. A cloud provider can report strong revenue while its largest AI customers are simultaneously raising capital, signing capacity agreements and spending ahead of their own monetization. That circularity is not necessarily fraudulent or even unhealthy. It is a financing structure that moves the point of maximum sensitivity from current demand to future demand. If the future arrives, the infrastructure earns its keep. If pricing falls first, the same structure can turn operating leverage into a valuation problem.
The capex debate is still real, but it is not the first alarm
There is no shortage of scale behind the build-out. Goldman Sachs puts its baseline estimate for AI capital expenditure at $765 billion in 2026, rising to $1.6 trillion annually by 2031. The bank has also estimated that global AI investment could exceed $1 trillion in 2026, with the United States accounting for just under $600 billion. CNBC reported Aug. 14 that Nvidia’s financing push could mobilize more than $500 billion in third-party capital, while Goldman analysts estimate combined hyperscaler lease commitments at $1.5 trillion, up from $200 billion five years ago.
Lotfi Karoui, multi-asset credit strategist at PIMCO, captured the scale in PIMCO commentary published Aug. 11 and quoted by CNBC on Aug. 14: “The AI capital expenditure cycle remains solidly on track to eclipse the telecom boom of the late 1990s and become the largest investment cycle since the railway buildout of the 19th century in inflation-adjusted terms.” Karoui also described the scale as deeply uncertain. The phrase is useful because it separates two questions that are often blended together: whether the investment cycle is historically large, and whether the cash flows that justify it are already visible.
Credit investors are watching the second question through leases, project finance and the growing use of outside capital. Jensen Huang, Nvidia’s chief executive, has described chips as an “investable infrastructure asset,” a formulation that invites more balance-sheet participation. That may broaden the funding base, but it also broadens the number of investors who must believe that model demand will persist through several hardware generations. The risk is not limited to a single company’s debt ratio. It is the repricing of a chain of contracts whose economics depend on high utilization and continued willingness to pay.
Sahil Mahtani, director of the Investment Institute at Ninety One, told CNBC on Aug. 14 that the danger is “primarily an expectations problem rather than a leverage problem.” That is a helpful correction to the most dramatic version of the bear case. The question is not whether every AI investment becomes stranded at once. It is whether an expectation of perpetual high growth is embedded in the price of cloud, chip and infrastructure assets, leaving little room for a normalisation in model pricing or customer concentration.
Eisman’s sharper warning is about competition. “The Achilles’ heel of this whole story ... is if something bad happens to Anthropic and OpenAI ... the Chinese open-end models, open-weight models are much cheaper. And if they start really taking a lot of market share ... you could have a big price war. And then we have a problem,” he said in the same CNBC appearance. This is not a claim that Chinese models have already displaced the leading US labs. It is a warning that lower-cost or open-weight alternatives could reset the price of inference faster than the infrastructure stack can reset its fixed costs.
What investors should measure next
The most useful indicators are therefore less theatrical than the daily capex headlines. Watch the share of cloud growth coming from the largest model customers, the duration and termination terms of capacity contracts, and the percentage of AI revenue that is tied to external enterprise users rather than model-company spending. Watch inference prices and token economics, not just training announcements. A falling price can be bullish for adoption while bearish for the revenue pool captured by the infrastructure providers.
There is also a difference between a model lab raising revenue and a model lab generating durable economic demand. Anthropic’s reported jump from $787 million to more than $11.5 billion in a year is evidence that the market is willing to pay for capability. It is not proof that the current price of that capability is sustainable after subsidies, promotional credits, internal transfers and the next wave of competition. The same logic applies to OpenAI’s reported run rate. Scale reduces some risks, but it can also increase the amount of capacity that must be filled every quarter.
Our view is not to short the AI complex simply because two customers matter so much. The more disciplined trade is to underwrite the dependency directly. Companies with diversified enterprise demand, visible cash generation and flexible capacity should command a premium over businesses that require a handful of model labs to keep spending at peak intensity. The next market signal to watch is not a single data-center cancellation. It is a change in the language around price, utilization or customer concentration in the next cloud and model-company updates.
If AI adoption broadens, the current concentration will fade and the infrastructure cycle may earn its scale. If adoption stays concentrated while open-weight competition compresses prices, the market can discover that the principal risk was never a shortage of enthusiasm. It was too much revenue riding on too few buyers.
This note is for informational purposes only and does not constitute investment advice. Sources: CNBC, Aug. 13, 2026, https://www.cnbc.com/2026/08/13/big-short-investor-steve-eisman-sees-an-achilles-heel-in-the-ai-boom.html; CNBC, Aug. 14, 2026, https://www.cnbc.com/2026/08/14/ai-infrastructure-debt-leverage-risks.html; CNBC, Aug. 15, 2026, https://www.cnbc.com/2026/08/15/anthropic-revenue-jumps-to-over-11point5-billion-in-q2-report.html; Bloomberg, Aug. 14, 2026, https://www.bloomberg.com/news/articles/2026-08-14/anthropic-revenue-ahead-of-ipo-surges-over-14-fold-in-second-quarter; Goldman Sachs, https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out; PIMCO commentary, Aug. 11, 2026, https://www.financialinvestigator.nl/en/nieuws-detailpagina/2026/08/11/pimco-the-ai-split-between-us-dollar-and-euro-investment-grade.
