Making Sense of This Week’s AI Market Crashes | #x307

This week’s selloffs reflected a market trying to price both AI’s potential and the enormous investment required to realize it. U.S. markets were down about 3%–3.5% on the day, a move Dave said had become frequent, while Alphabet lost 13% and $500 billion in market value and Intel fell 28%. Dave argued that technology companies are still committing tens of billions to data centers, often borrowing because cash flow is not enough, and warned that volatility could intensify before returns arrive. He still sees the investment as potentially highly rewarding, but questioned whether enough capital exists to bridge the gap.

Alphabet’s decline came despite revenue rising 25% to $127 billion, Google Cloud growth of 100%, and the Gemini app reaching 1 billion users. Investors focused on slower search-ad growth: revenue was up 12%, but growth had slowed by 17% from the prior quarter and 19% from earlier in the year. Alphabet reported a $9 billion free-cash-flow burn for the quarter and maintained annual capital-expenditure guidance of $195–$205 billion. The hosts saw Google’s search cash flow, TPUs, models, and training data as a powerful integrated position, but questioned whether AI agents buying products could undermine advertising aimed at people. Search has higher margins than Cloud, and Google’s ability to secure scarce TSMC capacity could also constrain Cloud growth.

Google’s Gemini 4 Carbon announcement was presented as evidence of serious progress, but not yet a directly comparable commercial product: it was restricted to trusted cyber defenders through Fairwind, with no broader release date announced. Google reported a score of 59 on the Artificial Analysis Intelligence Index, detection of 88% of vulnerabilities on SEC Bench Pro, exploitation of 55% of flaws on Exploit Bench, and working patches for 71%. Alex cautioned that Carbon was not generally available like Opus 5.5 and might no longer be at the frontier when released. He said it appeared on a similar intelligence-versus-cost frontier but at higher reasoning cost; its reported 23% lower price did not resolve questions about restricted access or margins.

The scale of AI spending is also colliding with supply constraints and uncertain economics. Dave linked shortages to rising RAM and high-bandwidth-memory costs, saying fab capacity was not built for current AI demand; expensive memory was also weighing on PC sales, according to Intel’s CFO. One speaker attributed $300 million of Intel’s earnings miss to its fab in Israel being bombed. Dave said costs are shifting toward capital expenditure and away from employee spending at hyperscalers, though he described that reading of Microsoft’s voluntary employee exits as preliminary. Amad expects CPUs to become a more important constraint in four or five years as agent use grows, while Alex questioned whether stronger models might reduce the number of agents needed; he cited a contrast between OpenAI’s reported use of 10,000 agents and a later result achieved by one agent working for three hours.

The hosts saw expanding AI capabilities as a reason to invest, but not proof that returns will come quickly. Google’s Genie Four can generate interactive 3D worlds with audio and simulated physics, which speakers said could help train robots and model factories, construction, and operations. They described a possible convergence of Gemini and Genie as a route to real-world orchestration, while noting that the architecture was unpublished and its application remained uncertain. Sam Altman said an OpenAI model solved many difficult research problems; he estimated about $10 million, possibly more, for one proof and confirmed about $100 million for broader mathematical work. These claims point to costly research with potentially valuable results, but do not settle how quickly the products will generate revenue.

Robotics and autonomous vehicles offer another possible return on AI investment, with demand and deployment constraints still unsettled. Tesla says its Optimus line is designed for one million robots a year, but Elon Musk said reaching that rate will take at least a year, and customer deliveries are expected in 2027. Speakers disagreed on whether industrial buyers or households would drive demand first, and emphasized safety and the risk that Tesla factories alone could not absorb output. Waymo, meanwhile, was said to have 4,000 robotaxis, 500,000 paid rides a week, and service in 15 cities; speakers cautioned that per-mile estimates omit maintenance, remote assistance, cleaning, and downtime. Regulation, chip supply, and city limits could slow deployment even if the technology improves.

Crypto added a separate source of market anxiety. Chamath argued that Bitcoin’s capped supply, limited buyer pool, and inability to be meaningfully held by central banks left it vulnerable to AI and quantum computing and could eventually drive it to zero. Others disagreed: Emaid said cryptography could be upgraded and Bitcoin could retain historical value, while Alex identified coordination on an upgrade as the central difficulty. Speakers also described a reported $6.3 billion Binance theft, with roughly $3 billion frozen; the attackers allegedly compromised an engineer’s machine, while claims of North Korean involvement were unconfirmed. Emaid noted that traceable transaction histories and issuers’ ability to freeze stablecoins such as Tether and Circle can limit what attackers can do.