Stanford Lecture by Prominent PE Partner: Where Is Capital Flowing in the AI Era?

Altimeter Capital Partner Apoorv Agrawal pointed out in a Stanford University public lecture that the current AI industry exhibits an "inverted pyramid" profit structure, starkly different from the traditional internet. Of the roughly $300 billion in ecosystem revenue, as much as 75% flows to the semiconductor layer, which enjoys gross margins of about 75%; the application layer, constrained by high marginal compute costs, sees gross margins of only 0% to 30%. To break the bottleneck of a mere $10 per user per year (ARPU), he asserted that ChatGPT and other AI applications moving into advertising will be a major breakthrough for the industry this year.

The $300 Billion "Inverted Pyramid": 75% of Revenue Flows to Semiconductors

Looking back at the internet, mobile, and cloud computing revolutions of the past 25 years, their monetization models typically formed a pyramid—infrastructure at the bottom earned relatively thin profits, while applications at the top captured the richest margins. The current AI ecosystem, however, presents the exact opposite shape.

Agrawal stated bluntly in his lecture that the generative AI ecosystem is fundamentally different from past cloud computing ecosystems, forming an "inverted pyramid." To explain the underlying logic, he contrasted it with the view of renowned Silicon Valley investor Marc Andreessen:

"As Marc Andreessen said, software eats the world because I can develop software and distribute it to millions of people, and the marginal cost of running that software is close to zero, so software companies enjoy gross margins of 80% or even 90%. But the economics of artificial intelligence are not like that."

The root of this difference lies in compute costs. "Incremental users of AI applications are not free, because you need to consume massive GPU resources; the cost of having AI users is much higher." Agrawal noted that of the roughly $300 billion in new revenue generated by the AI industry over the past two years, "about 75% went directly to the semiconductor industry; the application business grew more than tenfold but still hasn't made a significant impact."

Across the entire technology stack, the semiconductor layer is currently the most profitable. Agrawal offered a striking margin comparison:

Catalysts to Break the "Equilibrium": Big Tech's In-House ASICs and Capex Inflection

Market opinions diverge sharply on how long this Nvidia-dominated, bottom-heavy "inverted pyramid" will persist. Agrawal offered his own forecast: "I guess it will stay in its current state in the cloud longer than I might have expected, about a decade. I feel it could last even longer, because reaching a stable equilibrium is just too difficult."

However, he also pointed investors to two core catalysts that could break this structure:

Breakthroughs in cloud giants' in-house chips: If a hyperscaler's ASIC project achieves a breakthrough—whether Google's TPU, Meta's MTIA, or projects at Amazon or Microsoft—it could be a game changer. A breakthrough would trigger the biggest repricing at that layer.

Shifts in big tech capex guidance: Another catalyst is hyperscaler capital expenditure (capex) guidance. If the giants stop making aggressive capex forecasts, it may signal that the current equilibrium is not working.

Additionally, the shifting structure of compute demand deserves close attention. Currently, about 60% of GPUs sold by Nvidia are used for training and 40% for inference. Agrawal believes that "over time, the share of inference workloads will gradually increase," but inference workloads are highly volatile, which will not only test hardware utilization but also reshape the competitive landscape of the middleware layer (inference intermediary layer).

The Ultimate AI Monetization Hypothesis: "We Have to Move into Advertising"

On the application side, consumer AI led by ChatGPT faces the contradiction of massive users and inefficient monetization. Agrawal shared a key data point: ChatGPT has about 1 billion users (comparable to Facebook's early scale), but "about 95% of users are free users."

Compared with current internet giants' average revenue per user (ARPU), the gap for AI applications is obvious:

[Alphabet] $100 / year (~4 billion users) 

[Meta]  $70 / year (~3.5 billion users)

[ChatGPT]    $10 / year (~1 billion users)

After ruling out the possibility that a pure knowledge-work subscription model could support a trillion-dollar valuation, Agrawal delivered his ultimate prediction for AI commercialization: "The second question is: How do we raise revenue from $10 per user per year to $100? I'm not sure advertising and subscriptions alone can do it. The answer is, we have to move into advertising."

He believes large AI models have inherent advantages in ad placement:

Better pricing: Large models can understand users' true intent more precisely.

Excellent attribution analysis: They can accurately track conversion paths.

Higher trust: Conversational interaction builds strong user trust.

Although the market is pessimistic or skeptical about "inserting ads into private AI conversations," Agrawal likened it to the bearish arguments against Facebook's mobile ads a decade ago. "I think this will be another major story this year, it will definitely be a big deal... I'm optimistic that we'll find the answer, and it will be a major breakthrough for the existing economic model."