Hock Tan, the CEO of Broadcom, told a Singapore conference that most AI investors "have no clue what they are investing in" and are operating like casino gamblers with "extreme expectations." He pointed to OpenAI's $50 billion in annualized revenue as evidence of potential but noted the company is still losing money because of training costs. Enterprise AI revenue, he said, is still substantially composed of pilot projects — companies paying to experiment, not committing to production deployments. The remarks landed on a market already jittery about AI capital expenditure. But the accelerant was a separate report: China's CXMT, which debuted on the Shanghai exchange in late July, has reportedly made significant progress on a new DRAM architecture that would boost memory density. DRAM is the workhorse memory that AI chips draw on for calculations, and the global market is dominated by three companies — Samsung, SK Hynix, and Micron. CXMT's advance threatens that oligopoly directly. The market response was severe and geographically concentrated. South Korea's Kospi plunged 6.25%, with Samsung down 7.4% and SK Hynix cratering 10%. Japan's Nikkei fell 1.8%, with Tokyo Electron dropping 7.15% and SoftBank losing 8.5%. Taiwan's exchange shed nearly 2%, dragging TSMC down with it. US futures pointed to a mixed open after the Nasdaq had already slipped 0.4% on Monday, with Micron, Nvidia, AMD, and Broadcom all falling. The dual trigger — a credibility critique from inside the industry plus a competitive threat from China — exposes the structural fragility of the AI trade. The entire rally rests on a small number of semiconductor companies delivering on promises that their own customers are still only piloting. As IG's Chris Beauchamp put it, CXMT's progress "reminds investors of the huge competition in the sector" — great for customers who get falling prices, dangerous for stocks priced for monopoly-grade margins. Tan's comments are notable because Broadcom is not a skeptic or short-seller. It is one of the primary beneficiaries of AI infrastructure spending. When the CEO of a company that sells the picks and shovels says the gold rush has too many blind prospectors, the signal is qualitatively different from an outside critic making the same observation. His analogy — enterprise AI pilots as a kid burning through $10 in video arcade tokens in 10 minutes and learning nothing — is pointed. The deeper question is capital allocation. AI companies are spending enormous sums on training infrastructure, funded substantially by debt, while the revenue base remains experimental. OpenAI's $50 billion run rate sounds impressive until you note it doesn't cover costs. If enterprise customers convert pilots to production at high rates, the spending is justified. If they don't, the semiconductor companies supplying the infrastructure are overbuilt for demand that never materializes at the priced-in scale. China's CXMT adds a second vector of risk. Even if AI demand materializes as promised, the margin assumptions embedded in chip stock valuations depend on limited competition. A viable Chinese DRAM competitor reprices that assumption directly. The market is now processing two uncomfortable possibilities simultaneously: demand may be softer than priced, and even if it isn't, supply competition may compress margins. Both paths lead to lower semiconductor valuations than current prices imply.
