The AI industry's $600B revenue problem

Nvidia is making historic amounts of money selling chips, but the companies buying them are running out of excuses for where the returns are.

According to a widely circulated analysis by Sequoia Capital partner David Cahn, the artificial intelligence industry now needs to pull in roughly $600 billion in annual revenue just to pay off its staggering infrastructure spending.

The only problem is that the actual revenue coming out of the ecosystem is nowhere near that mark.

The double-and-double math

Cahn reaches the $600 billion threshold using a straightforward formula based on hardware spending.

First, he takes Nvidia's run-rate revenue and doubles it. The assumption is that GPUs account for only half of total data center costs, with the remainder consumed by energy, real estate, cooling, and backup generators. Last year, Nvidia pulled in $47.5 billion from data centers alone, and hauled in another $19 billion in a single quarter this year.

Next, he doubles that combined total again. This step assumes that end users, such as startups and corporate customers leasing cloud compute from Microsoft Azure or AWS, need roughly a 50% gross margin to run viable businesses.

Against this requirement, current revenues look paper-thin.

OpenAI has indeed seen explosive expansion, scaling revenue from $1.6 billion in late 2023 to $3.4 billion this year. But it stands virtually alone, while most AI startups are still fighting to break the $100 million mark.

Even under an aggressive scenario where Microsoft, Google, Apple, and Meta each generate $10 billion in AI revenue, and players like ByteDance, Tencent, Tesla, and Oracle bring in $5 billion each, the total still leaves a gaping $500 billion shortfall.

The depreciation trap

Unlike traditional bridges or fiber optic cables, AI silicon devalues almost immediately.

Nvidia’s upcoming B100 and B200 chips promise 2.5 times the performance for just 25% higher cost. That jump will almost certainly trigger another procurement frenzy while instantly degrading the commercial value of existing H100 clusters.

At the same time, custom accelerators from Google and Meta, alongside chips from AMD and Intel, are commoditizing inference work and threatening to spark brutal price wars.

Not everyone buys Cahn's model. Commentators on forums like Tom's Hardware pointed out that blanket assumptions like 50% data center overhead and 50% software margins oversimplify enterprise economics, noting that tech giants are willing to accept depressed margins while building out generational platforms.

Why it matters

We think this report marks an important shift in the AI narrative from raw compute capacity to ruthless unit economics.

For enterprise buyers and developers, the honeymoon phase of subsidized compute is drawing to a close. CFOs will no longer approve six-figure software contracts just to experiment with generative features unless they clearly eliminate headcount or drive direct top-line growth.

For the broader industry, it exposes the massive divergence between cloud hyperscalers and venture-backed startups. Tech giants with fortress balance sheets can afford to treat multi-billion-dollar GPU clusters as strategic long-term investments, but application startups that cannot bridge this monetization gap face a painful reckoning.