Citadel founder Ken Griffin told Bloomberg this week that investors are “almost certainly” in the middle of an AI bubble. Speaking at Citadel’s Future of Global Markets conference in New York, he said radical technological shifts have usually come with bubbles attached, and that the current round of spending has not yet produced clear proof of durable productivity gains or pricing power.
He compared GPUs to airline seats. Both have a very low marginal cost once the capacity exists. An empty airline seat sold for $20 is still profitable; once too many seats exist, fares collapse. Griffin said the same dynamic could apply to compute. “If we reach a point where the provisioning of GPUs exceeds demand for GPUs, marginal pricing will actually go into freefall,” he said. He pointed to the telecom fibre build-out around 2000 as a precedent, when overcapacity drove prices down sharply.
Watch Ken Griffin on X:
Griffin was not arguing that AI is a mirage. He said the companies that survive bubbles of this kind tend to become the market leaders for decades. His point was narrower: the market is currently pricing in outcomes that are not yet proven. A large part of the near-term demand supporting those prices is still coming from venture capital rather than from customers paying with operating cash flow.
Venture capital is still the main source of demand
PitchBook and the National Venture Capital Association reported that AI companies captured 82.7% of all US venture deal value through the first three quarters of 2026. Total US venture deal value reached $515.8 billion in that period, already above the previous full-year record, with roughly $426 billion going to AI companies.
Concentration is extreme. OpenAI and Anthropic alone accounted for $217 billion, or 43% of all startup funding in the first half of 2026, according to Crunchbase. Deal count has fallen while total dollars have risen sharply, meaning the money is flowing into a smaller number of large rounds. Non-AI startups raising Series A this year are landing median valuations roughly half those of AI companies at the same stage.
Investor Chamath Palihapitiya has noted that a large share of current AI spending is still coming from unprofitable, venture-backed companies. The PitchBook and Crunchbase concentration figures point in the same direction: capital is flowing into a small number of large rounds, many of them still dependent on external funding rather than operating cash flow.
This is the pattern that shaped early cloud demand as well. A large share of AWS’s first customers were venture-backed startups. Those revenues only became durable after a few clearing events sorted out which business models could pay for the service without continuous external funding. The difference now is the absolute size of the commitments and the fact that much of the capacity being ordered will only arrive in 2027 and 2028.
Those venture-backed orders are one reason the infrastructure numbers have grown so large.
How much is being spent
The scale of the infrastructure build-out is not in dispute. TrendForce estimates that the nine largest cloud service providers will spend $957 billion on capital expenditure in 2026 and more than $1.5 trillion in 2027. AI systems account for roughly 81% of that spending.
Morgan Stanley has projected AI-related compute investment of about $1.58 trillion in 2027 across the hyperscalers and neocloud providers. Hyperscalers had already issued roughly $255 billion in unsecured bonds by mid-October 2026. The five largest US platforms have guided to between $660 billion and $690 billion in capex for 2026 alone. These figures cover capacity that will largely come online in 2027 and beyond.
A substantial portion of the near-term demand justifying that capacity is still coming from companies that raised venture capital and are now spending it on compute. Those companies will eventually need either durable customer revenue or continued external funding to keep ordering at the current pace.
Productivity gains have not yet closed the gap
Surveys of enterprise adoption show widespread use but limited financial impact so far. McKinsey’s latest global survey found that 37% of respondents attribute at least some positive EBIT impact to AI, essentially unchanged from the prior year. Only about 6% of organisations qualify as high performers, meaning AI accounts for at least 5% of their EBIT. Eighty percent of respondents said AI improved their individual productivity, but those gains have not yet translated into broad profit-and-loss results.
A Wharton and GBK Collective study of 1,416 US leaders and employees found that 80% of leaders report positive returns from AI, but only 7% of organisations clear all three of the report’s markers of success. Departments that track growth metrics rather than hours saved report stronger results, yet the majority of organisations are still measuring efficiency rather than revenue.
These findings do not mean AI is failing. They show that the productivity gains being reported inside companies have not yet scaled to the level of revenue or profit growth required to underwrite the current volume of infrastructure spending without continued reliance on venture capital.
The transition that has not happened yet
Griffin’s airline-seat comparison and the venture capital concentration data point to the same open question. If GPU supply begins to exceed demand before enterprise spending has shifted from experimental budgets and venture capital to recurring, profitable usage, the pricing assumptions behind much of the current build-out would change. The companies that survive that adjustment would still have the advantage of having built the infrastructure, but the path between the current spending commitments and that outcome is not yet reflected in reported productivity or profit figures.
For now the market is hearing both messages. Griffin says there is almost certainly a bubble inside this moment. The funding data show that a large share of the demand supporting that spending is still coming from venture capital rather than from a broad base of customers paying with operating cash flow. The VCs backing those companies are effectively underwriting a bet that the business models will evolve into durable demand before the capital runs out.
Author: Andy Samu
See Also:
Who Owns Europe’s AI Champions? – MRKT3.0
What Is the Scaleup Europe Fund? Who Backs It, and How Startups Get In – MRKT3.0
