Gavin Baker - AI Market Jitters - [Invest Like the Best, EP.485]

Summary

Gavin Baker says the recent AI stock drawdown looked like “2022 in a month,” but he sees fundamentals improving beneath the surface: GPU scarcity remains real, token demand is accelerating, and private AI companies are largely invisible to public markets. He argues open-source models do not reduce compute demand so much as shift margin from frontier labs to infrastructure providers, because a token still requires the same flops, memory, watts, and cooling. Baker also thinks GPU contracts are repricing higher as older commitments roll off, and that AI capex can increasingly be funded by operating cash flow rather than debt. The main macro watchpoint is whether hyperscaler cash flow growth continues fast enough to avoid credit stress. He closes by saying regulation is the biggest risk, while SpaceX and orbital compute remain underappreciated as potential infrastructure powerhouses.

Notes

• July felt like a compressed 2022-style reset for AI names, with many public stocks down 40% to 60% from highs.
• Baker says he has not seen a single quantitative sign of AI deceleration.
• He says GPU availability, GPU rental pricing, DRAM spot prices, and token growth are all accelerating.
• Public markets still lack visibility into Anthropic, OpenAI, and inference clouds like Fireworks and Modal.
• His core mechanism: a token is a token, so open-source share does not change the underlying compute required per output token.
• Open-source adoption mainly shifts margin dollars away from frontier model labs and toward infrastructure providers.
• He thinks market panic was amplified by Meta compute interpretation, China semiconductor equipment headlines, and higher real yields.
• Higher yields matter because a larger share of the buildout is being financed with credit, even if operating cash flow still covers most of it.
• He flags that Meta’s bond priced wider than investors would typically expect for Meta.
• Installed compute appears to be repricing upward as old contracts roll into the spot market.
• That example implies roughly 50% to 60% higher GPU-hour pricing in six or seven months.
• Baker sees accelerating hyperscaler operating cash flow as the key variable for funding the AI buildout internally.
• He says the adjusted acceleration is even stronger once unusual items are stripped out.
• Consensus appears to underwrite AI monetization at something like Ampere-like economics, which he thinks is too conservative.
• In his bullish case, hyperscaler operating cash flow could reach around $2 trillion and remove about $700 billion of credit demand.
• He says there are no signs anyone has too many GPUs, which he treats as a key tell for whether demand has stalled.
• A sustained drop in GPU prices would worry him, because it would imply demand is weakening.
• Continual learning and sample-efficient learning could create a temporary discontinuity in training demand.
• He notes the possibility that a future system could learn from far fewer tokens than current frontier models.
• Enterprise and AI-native companies are already spending a material share of compensation on tokens.
• He says 30% token spend is already visible at some companies and he has even heard of 50%.
• He frames token spend as a proxy for either labor substitution or faster economic growth.
• LTAs are central to supply-chain planning because customers prepay and contracts can have floors and ceilings.
• He argues the competitive set at scale is Amazon, Google, AMD, and Nvidia.
• Memory is the biggest lever for token output per unit of compute.
• That memory intensity also lowers unit costs, which he says helps explain why demand has not fallen.
• Nvidia is rolling out a credit-wrapper style financing model with revenue share.
• He thinks the model could quickly become a large royalty-like cloud business.
• Baker says the credit-wrapper approach helps alleviate the cash-flow mismatch in GPU finance.
• He would pursue the same financing approach if he were Nvidia.
• He views the structure as an extension of LTAs that trades some upside for durability.
• He argues that open-source and custom models make AI-native software more defensible.
• Using frontier models for only part of token consumption and routing the rest to custom RL models can improve defensibility.
• He thinks open-source tokens could become the majority of tokens processed while frontier models still capture most economic value.
• That dynamic is still bullish for infrastructure demand because compute intensity per token stays similar.
• He says regulation is the biggest AI risk.
• He specifically points to New York’s data-center moratorium as a warning sign.
• He rejects the common anti-data-center narrative around electricity and water usage.
• He says data centers create durable blue-collar employment in trades like plumbing, electrical, and HVAC.
• He thinks the industry is under-telling the story of AI’s scientific and medical benefits.
• He sees disaggregated inference as a major architectural opportunity.
• His ideal stack splits prefill, attention, and feed-forward onto different chips.
• He says SRAM-based accelerators are attractive because they are not constrained by HBM DRAM.
• SpaceX is underappreciated in public markets as a compute infrastructure player.
• He cites a public report suggesting SpaceX could bring on 8 gigawatts of compute over the next 18 months.
• He says only a few companies have ever brought more than 500 megawatts of power online in a year.
• He says SpaceX has been the fastest and lowest-cost at bringing power online.