\n<\/p>\n
TOPSHOT – He Xiaopeng, cofounder and chairman of Chinese electric vehicle maker Xpeng, launches Xpeng’s next-gen Iron humanoid robot during AI Day press conference at its headquarter in Guangzhou, in southern China’s Guangdong province on November 5, 2025. (Photo by Jade GAO \/ AFP) (Photo by JADE GAO\/AFP via Getty Images)<\/span><\/p>\n AFP via Getty Images<\/small><\/div>\n<\/div>\n<\/figure>\n The traditional software playbook promised that more users meant better margins. Generative AI breaks that rule entirely; every new user requires burning expensive GPU compute, which means the application layer is effectively subsidizing a semiconductor monopoly<\/u><\/a> that collects the lion’s share of AI’s explosive revenue growth.<\/p>\n That is the central thesis of MS&E 435, “Economics of the AI Supercycle<\/u><\/a>,” a spring 2026 seminar at Stanford University taught by Apoorv Agrawal<\/u><\/a>, a Partner at Altimeter Capital<\/u><\/a> who led the firm’s investments in OpenAI and Glean. The course, now publicly available on YouTube, is arguably the clearest institutional map of where AI economics actually stand, and why the standard software venture thesis does not yet apply.<\/p>\n Agrawal frames the current market structure as an “inverted triangle.” In prior technology cycles, including the internet, mobile, and cloud, value migrated upward over time to software and applications. Distribution was nearly free, and gross margins of 80% to 90% became the rule. AI works in reverse. At the base of the inverted triangle sits the semiconductor layer, where Nvidia’s fiscal year 2025 gross margins reached 75%<\/u><\/a> on the back of near-monopolistic control over data center GPU supply. Above that, the application layer is fighting to achieve margins anywhere between 0% and 30%. The infrastructure middle layer, meanwhile, sustains the highest competitive intensity as hyperscalers and well-funded startups battle for inference dominance.<\/p>\n Over the past two years, the AI ecosystem added roughly $350 billion in new revenue. Approximately 75% of that went straight to semiconductors, according to Agrawal’s analysis of the value chain. Application companies grew heroically in user terms, but the overall shape of value distribution barely moved.<\/p>\n The user monetization gap is where the stakes become concrete for investors. ChatGPT now reaches roughly 900 million weekly active users<\/u><\/a>, a scale that rivals the largest social platforms in history. But OpenAI monetizes those users at approximately $10 per user annually, while Alphabet monetizes its users at closer to $100 annually<\/u><\/a> and Meta at roughly $49 per daily active user<\/u><\/a>. That gap is not primarily a product failure. It reflects a structural difference in business model: search and social run on advertising, and AI does not, yet.<\/p>\n Agrawal attributes part of this monetization ceiling to what he calls the “active work” bottleneck. Messaging apps and social feeds are passive; users absorb content without effort. AI requires users to formulate queries, think through prompts, and actively engage. That friction limits AI’s path to truly universal scale. WhatsApp and Chrome became mandatory apps because users do not have to try to use them. Generative AI, as it stands, remains a tool for people who want to actively query technology, which is a much smaller global population.<\/p>\n That ceiling, Agrawal argues, will force AI applications toward advertising. His case is counterintuitive: critics assume ads will not work inside personal AI chats. But the opposite may be true. Because AI platforms engage users in deep, logged-in conversations, they understand intent with a precision that search and social cannot match. The early mobile skeptics who doubted that ads would work on small screens were proven wrong within a decade. Agrawal sees similar dynamics here, and suggests there is meaningful alpha for investors who understand the ad model transition before it becomes consensus<\/u><\/a>.<\/p>\n The infrastructure layer presents a different set of risks for venture investors. Agrawal describes it as the “feature, not a platform” trap. Inference startups that win early adoption remain vulnerable to absorption by hyperscalers who treat inference as a native cloud service. Every infrastructure startup building on top of AWS, Google Cloud, or Azure is building on a platform whose operators have both the incentive and technical capability to replicate what the startup does. Until a major tech company achieves breakout success with custom silicon, specifically ASICs developed at Google or Meta scale, the power will remain concentrated at the Nvidia layer and the inference middle will continue churning.<\/p>\n None of this makes the current cycle a bad investment. Agrawal’s analogy is railroad construction: the infrastructure must be laid before the freight business becomes viable. Amazon Web Services broke ground in 2004<\/u><\/a> but did not reach commercial dominance until 2012. The investors who understood that eight-year gap made generational returns. The investors who dismissed the infrastructure spend as irrational never recovered the opportunity.<\/p>\n What the MS&E 435 lecture clarifies is where that railroad analogy breaks down. The cloud cycle eventually inverted, with software capturing the surplus. AI may stay inverted longer because inference compute costs are structural, not temporary. For founders building at the application layer, the playbook is not “grow users and the margin follows.” It is “grow users and then figure out an advertising model before the VC math catches up.” For investors, the near-term bet remains silicon. The longer-term bet is on whoever cracks the monetization model that closes the gap with Alphabet and Meta. That course is publicly available on YouTube<\/u><\/a> and the alpha is in watching it.<\/p>\n<\/div>\n