A private Chinese AI lab delivered numbers at a Goldman Sachs conference on Monday that complicate the most crowded trade in American equity markets. MiniMax co-founder Yeyi Yun said first-half revenue grew more than 280% while R&D spending grew at only 130%, a ratio that, if it holds, undercuts the foundational assumption behind NVIDIA (NASDAQ: NVDA), Advanced Micro Devices (NASDAQ: AMD), and Broadcom (NASDAQ: AVGO).
The case for owning those three names rests on compute demand outrunning efficiency gains. MiniMax's numbers are self-reported and unaudited, so treat them as directional. Still, the pattern they describe is worth taking seriously. Yun said first-half revenue had already reached 1.5 times the company's full prior-year figure, with July token consumption running 20 times the January level. She said the company is targeting $1 billion in full-year revenue and that its H suite open-source model has been downloaded more than 24 million times on Hugging Face with more than 300 derivatives. Her summary of the strategy: "Revenue growth is much faster than our R&D spending."
What the infrastructure incumbents are pricing in
NVIDIA's Q2 FY2027 results show how large the other side of that bet has grown. Data Center revenue came in at $89 billion of a total $96 billion. Management said cloud industry backlog exceeded $2 trillion and that the top five hyperscalers are expected to spend nearly $800 billion on capex in 2026, rising to $1.3 trillion in 2027. NVIDIA also disclosed $279 billion in supply obligations tied largely to Vera Rubin memory, a forward commitment that pays off only if hyperscaler orders keep landing. NVDA trades at $217.44 with a market cap of $5.25 trillion, up 25% over one year.
Broadcom reported Q2 AI semiconductor revenue of $10.8 billion, up 143% year-over-year, and Hock Tan guided fiscal 2027 AI revenue to in excess of $100 billion. AMD posted Q2 revenue of $11.54 billion with Data Center up 107% to $6.718 billion, and Lisa Su has committed the company to gigawatt-scale Helios deployments with Anthropic, Meta, and Microsoft.
The counterargument
Jensen Huang addressed the efficiency question directly on the earnings call. His argument: as inference becomes cheaper, usage explodes, and agentic workloads run continuously. "If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services," he said. He added that frontier labs are "only limited by the amount of compute." The Jevons pattern is real. Cheaper unit costs have historically expanded total consumption in computing, and MiniMax's H suite almost certainly runs on NVIDIA silicon, meaning efficiency wins can still flow back to the chipmakers.
What changed is that NVIDIA is now backstopping demand rather than watching it arrive. Management said it has invested nearly $50 billion in frontier labs and organized partners to mobilize over $500 billion of third-party capital. That is the read-through the MiniMax data raises most sharply: a chipmaker underwriting demand it once expected organically is a different business than the one the multiple prices.
On balance, the Jevons argument buys the AI infrastructure trade time. But $279 billion in supply obligations against a capex cycle that a private Chinese lab just credibly challenged is not the setup for adding at $217.44.