Why the Markets Are Pricing AI Wrong | Gavin Baker

41 insights saved from this video by @money

@money· Personal Finance & Investing
Compute rental prices rose roughly 50–60% in six or seven months. That is bullish demand evidence.
@money· Personal Finance & Investing
AI-mediated trading compressed a Japanese capacitor-stock cycle that normally takes years into roughly six weeks.
@money· Personal Finance & Investing
One startup paid mid-$2s per GPU-hour for thousands of Blackwells, then expected just under $4 for equivalent B200 capacity seven months later.
@money· Personal Finance & Investing
An inference cloud planned to pay 100% more for Blackwell GPUs when its contract expired.
@money· Personal Finance & Investing
Cheaper tokens can drive more usage, so falling AI spend can coexist with rising GPU utilization.
@money· How Money & Economics Work
Some inference-cloud customers expect to pay 100% more for Blackwell capacity when current contracts expire.
@money· Personal Finance & Investing
If credit stops funding new compute infrastructure, existing compute becomes more valuable instead of killing AI demand.
@money· Personal Finance & Investing
A falling token-price index can signal cheaper open-source tokens replacing expensive frontier-model tokens, not weakening inference demand.
@money· How Money & Economics Work
Open-source AI is public markets’ “dark matter,” so public estimates may understate actual inference demand.
@money· How Money & Economics Work
A falling token-price index can reflect cheaper open-source tokens replacing expensive frontier-model tokens, not weakening demand.
@money· How Money & Economics Work
Compute contracted at older prices trades below current spot rates. When contracts expire, revenue can rise even as spot prices fall.
@money· How Money & Economics Work
Cheaper AI inference can increase usage enough to expand the total market instead of shrinking infrastructure demand.
@money· How Money & Economics Work
Routing between customized open-source and frontier models can deliver slightly better outcomes at roughly half the cost.
@money· Personal Finance & Investing
Only hyperscalers, CoreWeave, Crusoe, and SpaceX have deployed more than 500 megawatts in one year. SpaceX did it fastest and cheapest.
@money· How Money & Economics Work
A public Funder AI report projected eight gigawatts of SpaceX compute, an estimate considered highly implausible.
@money· How Money & Economics Work
Only hyperscalers, CoreWeave, Crusoe, and SpaceX have deployed more than 500 megawatts in one year, with SpaceX the fastest and cheapest.
@money· Personal Finance & Investing
AI-chip market share will be determined by supply allocations and purchased capacity, not only price or product quality.
@money· How Money & Economics Work
The gap between contracted and spot compute pricing can create cash-flow upside as installed capacity reprices, reducing reliance on debt.
@money· Personal Finance & Investing
If newer chips monetize below Blackwell rates, hyperscaler operating cash flow could approach $2 trillion and erase $700 billion of credit demand.
@money· How Money & Economics Work
Inference clouds and model routers let AI companies escape dependence on one frontier-model provider’s pricing and terms.
@money· How Money & Economics Work
Disaggregated inference splits prefill, attention, and feed-forward computation across specialized chips instead of forcing one chip to do everything.
@money· Personal Finance & Investing
AI-related stocks fell roughly 40% to 60% from their highs in a single month.
@money· Personal Finance & Investing
SRAM accelerators could reshape compute architecture by bypassing HBM constraints and using older manufacturing nodes.
@money· Personal Finance & Investing
A Funder AI report projects SpaceX could bring 8 gigawatts of compute, though the estimate is highly implausible.
@money· How Money & Economics Work
Consensus models Blackwell and Rubin at Ampere-era monetization rates. Observed demand suggests newer capacity will command higher pricing.
@money· Personal Finance & Investing
Defensibility rises when frontier models handle planning while customized models trained on proprietary data handle most workloads.
@money· How Money & Economics Work
GPU availability, GPU retail pricing, DRAM spot prices, and token growth are accelerating, not decelerating.
@money· Personal Finance & Investing
When installed compute reprices upward as contracts expire, hyperscaler credit ratios improve and additional financing gets easier.
@money· How Money & Economics Work
The market may be underestimating SpaceX as a compute company because valuations do not reflect its potential capacity.
@money· Personal Finance & Investing
Disaggregating AI workloads across specialized chips could substantially improve the return on investment of AI infrastructure.
@money· How Money & Economics Work
Higher real yields, wider credit spreads, and rising CDS costs are genuine negatives. They become especially dangerous if debt funds the AI buildout.
@money· How Money & Economics Work
As installed compute reprices when contracts expire, hyperscaler cash flow can fund most or all of the buildout without heavy credit reliance.
@money· Personal Finance & Investing
Microsoft, Meta, and Amazon's operating cash flow growth accelerated from 28% to 32%, or 35% adjusted.
@money· How Money & Economics Work
A plateau or decline across major AI labs would be negative unless open-source models expand the market while taking share from those labs.
@money· Personal Finance & Investing
AI stocks sold off sharply while the underlying fundamentals improved significantly instead of deteriorating.
@money· Personal Finance & Investing
For the next several years, AI hardware market share will depend heavily on supply-chain allocations and pre-purchased capacity.
@money· How Money & Economics Work
Operating cash flow at Microsoft, Meta, and Amazon accelerated from 28% to 32%, or 35% after one-time adjustments.
@money· How Money & Economics Work
Open-source AI matters: one or two frontier models charging roughly 90% margins is an undesirable market.
@money· How Money & Economics Work
The evidence is broadly positive, yet the search continues for quantitative negative data instead of anecdotes or contested signals.
@money· Personal Finance & Investing
The evidence revealed no clear quantitative negative signal in AI, apart from contested third-party data on Anthropic.
@money· Personal Finance & Investing
Even a bearish view of the AI ecosystem can look like foolish optimism beside the stock market.

More from Invest Like The Best

See everything from Invest Like The Best

Post on X

Keep what you found in Why the Markets Are Pricing AI Wrong | Gavin Baker

A free account saves any of these cards to your Collection, and Korva resurfaces them so you actually remember.