Gavin Baker says July’s AI selloff conflicts with rising GPU prices, cash flow, and token demand

*Invest Like The Best* is a podcast, and in this episode host Patrick O’Shaughnessy talks with investor Gavin Baker about the July 2026 selloff in AI and semiconductor stocks. Baker argues that the market’s panic does not match what he sees in hyperscaler cash flow, GPU rental prices, memory contracts, and token demand. They ask what evidence would prove Baker wrong, including tighter credit, better training efficiency, China’s chip progress, and public resistance to data centres. The conversation ends with Nvidia’s financing playbook, SpaceX’s data-centre ambitions, and the increasingly weird prospect of orbital compute.

Baker says the selloff conflicts with rising GPU prices, token demand, and hyperscaler operating cash flow, while public comparisons miss private AI labs and inference clouds.

Watch the related Gavin Baker clip on X

• AI stocks fell roughly 40% to 60% in a month as Baker’s main demand measures accelerated.
• Microsoft, Meta, and Amazon’s combined operating-cash-flow growth rose from 28% to 32%, or about 35% after unusual items.
• He says open models reduce frontier-model margins but increase token demand on the same underlying infrastructure.

Higher real yields and wider credit spreads could threaten a debt-heavy buildout, but Baker expects new compute revenue to cover much of the spending.

Watch the related Nvidia clip on X

• Weak Meta bond pricing, higher default-swap costs, and wider spreads show that financing conditions have deteriorated.
• Baker estimates Blackwell and Rubin capacity could lift operating cash flow from $1.3–$1.4 trillion to about $2 trillion.
• Private demand must hold and contracts must reprice; falling stocks may still signal an unseen problem because prices kept dropping despite stronger data.

Identical Blackwell clusters are repricing sharply higher even as public markets value infrastructure providers as though their earnings will fall.

• One B200 cluster was expected to rise from the mid-$2 range to nearly $4 per GPU-hour within seven months.
• Another operator reportedly expected its Blackwell price to double, suggesting hyperscalers currently undercharge for scarce capacity.
• Anthropic’s third-party use may be slowing, but Baker says OpenAI and open-model activity could offset it.

Easy GPU access or falling lab demand would contradict him; cheaper tokens would not if open models expand total compute use.

• Customized open models can cut customer bills through lower margins without reducing the GPU hours they consume.
• AI-native companies spend more on tokens and hire fewer people, while adoption at established companies remains limited.
• He estimates only 250,000 to 500,000 people use agentic AI and says broader spending needs productivity or labor savings.

Long-term memory agreements now shape future capacity, so a buyer that leaves during oversupply may lose allocations and AI market share when supply tightens.

• The agreements can require prepayment and set both a price floor and ceiling.
• More memory for the same compute can raise token output and reduce the cost per token.
• Memory suppliers can redirect scarce volume toward competing chip platforms or a buyer’s AI rival.

Under Baker’s description, another lender finances the GPU buyer while Nvidia receives revenue above a price floor.

• Because a third party supplies the loan, Baker distinguishes the structure from conventional vendor financing.
• Nvidia can combine equity upside and recurring revenue while customers wait for operating cash flow to catch up.
• Its scale also improves access to foundry capacity, memory pricing, and financing relative to chip startups.

China’s reported DUV progress could matter despite lagging EUV systems, while customizable open models give AI-native companies more control over cost and proprietary data.

• He calls the alleged capability a phase change but says effects on ASML orders may take about five years.
• Limited visibility and learning-by-doing cycles make the timing unusually uncertain.
• Open models and routers can handle cheaper work while frontier models still capture most economic value.

Regulation is Baker’s biggest risk, and he says the industry has failed to explain local power, water, employment, and health effects clearly.

Watch the related Gavin Baker clip on X

• Current agreements can lower nearby power prices through behind-the-meter deals and fund hospitals, schools, police, and fire stations.
• Data centres create continuing work for plumbers, electricians, and HVAC contractors after construction ends.
• The speakers cite a corrected 10,000-fold water-use overestimate that continued shaping public belief.

Splitting prefill, attention, and feed-forward work across specialized chips could improve returns because workloads need different mixes of compute and memory.

• One design puts prefill, HBM-heavy attention, and SRAM-heavy feed-forward work on separate chips.
• Changing workloads make a fixed ratio of compute, HBM, DRAM, and SRAM hard to choose.
• Micron is his dark-horse candidate to become central to AI economics.

In Baker’s view, SpaceX probably will not deliver eight gigawatts on the reported timeline, but its construction record and outside investment make terrestrial and orbital compute plausible.

• SpaceX has added more than 500 megawatts in a year, faster and cheaper than most operators.
• Compute prices could still fall, and Baker calls eight gigawatts implausible and GPU energization difficult.
• Benchmark’s StarCloud investment and possible access to Starlink lasers provide an independent check on orbital compute.