AI Revenue Has Quintupled, but Semiconductors Still Capture the Profit

Apoorv Agrawal frames the central economic question as a mismatch between where AI revenue sits today and where value accrued in the cloud cycle. His comparison places cloud’s estimated annual revenue at roughly $600 billion for applications, $300 billion for infrastructure, and $80 billion for semiconductors. AI, by contrast, is shaped like an inverted triangle: about $60 billion in applications, $75 billion in infrastructure, and $300 billion in semiconductors.

Agrawal’s explanation is not simply that AI applications are early. The economics of serving an AI user differ from those of conventional software: software could reach millions of users at marginal operating costs close to zero, supporting gross margins of 80% or 90% in some cases. AI applications consume compute as usage rises.

That helps explain, in Agrawal’s account, why some large AI businesses can remain unprofitable even at billion-dollar revenue scale. The current allocation reflects an early applications market, concentrated compute supply, and the different “physics” of inference. The central question is whether applications eventually take a much larger share of value, as they did in cloud software, or whether the economics remain concentrated in the lower layers.

Agrawal estimates that AI revenue grew from about $90 billion in the first quarter of 2024 to $435 billion in the first quarter of 2026: roughly fivefold growth in two years. Yet the distribution among applications, infrastructure, and semiconductors changed little.

By his calculation, roughly 75% of the approximately $350 billion added over those two years accrued to semiconductors. Applications grew more than tenfold, but not enough to materially alter the shape of the ecosystem. The slide attributes semiconductor totals to recent Nvidia, Broadcom, and Arm results, while treating AI application and infrastructure figures as internal estimates.

Most of the $300 billion semiconductor layer, he says, is Nvidia. The applications layer is also concentrated: two companies account for roughly 90% of it, though he does not name them in this discussion. Infrastructure has the greatest competitive intensity, with startups competing against one another and hyperscalers seeking control of the same layer. Agrawal describes it as the most unstable equilibrium in the stack.

Profit is even more concentrated than revenue. His slide estimates that semiconductors accounted for 87% of AI gross profit in 2024 and 79% in 2026. The cloud-software comparison assigns 70% of gross profit to applications.

Agrawal puts Nvidia’s data-center gross margin at about 75%, with the caveat that the figure could vary by a few percentage points. He estimates some application businesses at gross margins between 0% and 30%, depending on the company. The gap reflects the market structure he describes: a dominant semiconductor supplier, recurring compute costs for applications, and a heavily contested infrastructure layer.

A student proposed that spending at the bottom of the AI triangle may be buying capacity for future application revenue. Agrawal agreed that there is a timing mismatch. Semiconductor and data-center investments are made for multiyear operating lives, while application revenue is measured in the present. That makes the lower layers cyclical, more like laying railroad track than operating a steady-state software business.

He points to the mobile supercycle as an earlier case in which capex-heavy businesses received inflated valuations in the first phase. He also uses AWS as a longer reference point: AWS started in 2004, had Netflix as its first customer in 2010, and Amazon shifted fully to AWS in 2012. The infrastructure buildout took eight years from the first investment cycle, amid arguments that Amazon’s spending could lead to bankruptcy.

Agrawal suspects AI may take at least as long to find a stable economic arrangement, and perhaps longer, because the underlying substrate is difficult to build. He calls AI unlikely to be a fad or an unsuccessful endeavor, but does not claim to know whether the present triangle flips in five years, 10 years, 15 years, or ever.

He identifies two possible catalysts for a repricing. A breakout custom-silicon program—Google’s TPU, Meta’s MTIA, or efforts at Amazon, OpenAI, Microsoft, and other labs—could alter the semiconductor layer substantially. A second signal would be a retreat in hyperscaler capital-expenditure guidance. If the largest cloud companies stop forecasting major capex commitments, Agrawal says, that could indicate that the current equilibrium is not working.

That is why he recommends listening to quarterly earnings calls. Four times a year, he says, public-company executives discuss their biggest questions and concerns; in this cycle, capex guidance is among the clearest public indicators of whether the buildout continues.

The mix between training and inference matters to the future allocation of AI value. Nvidia’s earnings calls are closely watched, Agrawal says, for indications of inference’s share of its fleet. His last reading was roughly 40% inference and 60% training, assuming full utilization. He expects inference to become a larger share over time, but does not forecast when.

Training and inference have different operating profiles. Training is predictable: it runs at high utilization for a bounded period. Inference is burstier and harder to forecast because it follows user activity. It rises when people are awake and falls around holidays such as Thanksgiving and Christmas. If agents become more autonomous, Agrawal suggests, inference could become closer to a 24/7 workload; for now, it remains less even than training.

That variability matters for the infrastructure layer. Startups may solve genuine inference problems, but hyperscalers have incentives to build or absorb those capabilities. Agrawal offers a practical test for whether an infrastructure business can stand independently:

A chip startup faces a different form of concentration. Agrawal says that, according to Jensen Huang’s disclosures, roughly half of the $300 billion semiconductor figure is associated with the large hyperscalers. A new chip company would therefore confront a small number of very large prospective buyers rather than the broad customer base available to a consumer or enterprise-software business. Its first strategic question, he says, is which of those few buyers it can sell to.

Agrawal separates Google into business units when placing it on the stack: TPUs belong in semiconductors, Google Cloud Platform in infrastructure, and Gemini in applications. But the fact that those units sit inside one company raises a larger question: whether AI value will accrue to the best specialist at each layer or to companies able to control several layers at once.

He sees precedent for the latter possibility. He calls Google the biggest winner of the internet supercycle, citing a roughly $3 trillion market capitalization and near-90% search share. In his description, Google integrated from servers through search and advertising to the user experience. Apple was the principal mobile-cycle winner; Meta was the leading social winner, though less vertically integrated. Cloud, by contrast, produced a more heterogeneous oligopoly of AWS, Google Cloud, and Azure rather than one dominant integrated company.

Nvidia has also tried to extend its reach through DGX Cloud and vertical applications. The unresolved AI question is not merely which layer earns the best margins, but whether a single player can sustain control over chips, infrastructure, models, distribution, and the application experience at once.

Agrawal identifies consumer AI, outside coding, as the largest AI market today. Its usage is substantial, but its monetization remains small compared with the largest consumer internet franchises. He says roughly 95% of ChatGPT users are free.

The source’s weekly-active-user comparison places YouTube, Chrome, and WhatsApp in a “core utility” category; Facebook, Instagram, and TikTok in a social category; and Spotify, Amazon, and X in a niche category. On that framing, ChatGPT has just overtaken the niche tier, while Gemini remains below it. Both are growing, but neither yet resembles a core utility at the scale of Chrome or WhatsApp.

Agrawal puts Alphabet at roughly four billion users monetized at about $100 per user annually, and Meta at about 3.5 billion users monetized at roughly $70 per user. He estimates ChatGPT at around one billion users and about $10 per user annually.

The constraint, he argues, may be the nature of knowledge work itself. An AI assistant is not yet where people message friends, manage an inbox, or seek passive entertainment. It requires users to ask questions and do active work. Getting from one billion users to four billion may require going beyond knowledge work.

The monetization problem is separate. Agrawal doubts subscriptions alone can move AI revenue per user from roughly $10 toward $100. He suspects advertising will become an important unlock, even though DeepMind has said it was not planning to use ads as a revenue model for Gemini.

His case for AI advertising is that an assistant could understand intent, operate in a logged-in environment, offer stronger attribution, and hold a higher-trust relationship with the user. Those properties could support better-priced advertising. The counterargument is that users may resist commercial interruptions in a personal conversation. Agrawal does not claim to know the eventual format. He compares the uncertainty to skepticism around Facebook’s mobile advertising at the time of its IPO, when critics saw no room for ads on a phone and the market ultimately found room.