Can AI answer the $3 trillion question?

Three years ago, Sequoia partner David Cahn was among the earliest voices to put numbers around Silicon Valley’s extraordinary AI infrastructure push. Back in 2023, he started with Nvidia’s reported annual GPU revenue of $50 billion and worked outward from there. Once the cost of running data centers and the profit expectations of their operators were added, his conclusion was striking: the industry would need roughly $200 billion in revenue just to support that initial level of spending.

At the time, his message was essentially a challenge to founders and operators: create AI products capable of absorbing all that compute capacity and turning it into real business income. Now, after several more years of rapid expansion, the scale of that challenge looks much larger. Cahn’s latest estimate puts AI infrastructure spending in 2026 at $1.5 trillion.

The size of the payoff AI now needs

Using that spending trajectory, Cahn estimates the AI sector would ultimately need to generate about $3 trillion in revenue to justify what is being poured into chips, data centers, and related infrastructure. He also suggests that even this figure could prove too low. Costs tied to memory, along with growing use of more specialized hardware such as exotic or inference-focused chips, could drive the required return even higher.

“Recently, the required revenue per GW of CapEx has sharply increased due to these bottleneck dynamics and rising costs of construction.”

That frames the central question facing the industry: can demand for AI products rise quickly enough to support an infrastructure buildout of this magnitude?

Strong revenue growth, but still far from the target

There are signs of major commercial progress. Anthropic is believed to have reached $60 billion in annual recurring revenue, while OpenAI reportedly generated $13 billion in 2025 and is assumed to be earning more this year. Those are large numbers by any ordinary standard, and they show that AI demand is no longer theoretical.

Even so, they remain small compared with the revenue levels required to match the sector’s enormous capital bill. The gap between current earnings and the scale of infrastructure investment is still substantial, which is why the economics of AI remain under such close scrutiny.

Why hyperscalers remain confident

Despite the size of the challenge, the largest cloud and platform companies continue to signal confidence. In a recent note, Apollo chief economist Torsten Slok pointed out that the hyperscalers — Google, Meta, Microsoft, and Amazon — are all projecting a sharp improvement in free cash flow by 2028.

That matters because these companies are the ones making many of the biggest infrastructure commitments. Their outlook suggests they believe the returns from their heavy chip buying and data center expansion are still ahead, not behind.

The threat from cheaper AI

But the optimistic scenario depends on those returns arriving on time. If they do not, the economics could look far less favorable.

Slok highlights a risk that is already emerging: many organizations are increasingly willing to use lower-cost open-weight models, including Chinese offerings, instead of relying solely on systems from leading frontier labs. At the same time, token prices continue to fall. OpenAI CEO Sam Altman has said the company’s latest model is 54% more token-efficient on coding tasks.

For users, this is clearly attractive. Lower token costs can make AI agents and other applications more practical and affordable. For companies trying to build large token-based businesses, however, falling prices are more complicated. If cheaper usage is not matched by a dramatic surge in total demand, revenue growth may not keep pace with the infrastructure already being financed.

Why this may become a broader market issue

Slok argues that the consequences of a slower AI payoff would likely extend well beyond the technology sector. Because so much investor attention and market value are concentrated in a small group of hyperscalers, disappointment in their cash-flow expectations could have wider economic effects.

“With so much riding on so few names, a slower payoff wouldn’t just be a sector problem, it would risk tipping the economy into recession and the S&P 500 into a correction.”

That is what makes the AI infrastructure debate more than just a question about computing capacity or startup opportunity. It has become a test of whether the industry can convert historic capital spending into equally historic revenue.

As more businesses shift AI workloads toward cheaper tokens and lower-cost models, that debate only becomes more urgent. The core issue remains unchanged: the technology is advancing rapidly, the infrastructure buildout is massive, and the financial returns now need to catch up.