By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Invest Like the BestWhy the Markets Are Pricing AI Wrong | Gavin Baker
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
Jump to publisher timestampsChapter summaries
Pressure-testing the AI thesis after a sharp selloff
Gavin Baker and Patrick O’Shaughnessy set the frame: AI-related stocks have sold off hard despite what Gavin describes as accelerating fundamental demand signals. Gavin’s stated goal is to find genuinely negative, quantitative evidence that would make him more cautious.
- •Gavin says AI names are down sharply in a straight line, yet he hasn’t found a clear negative quantitative metric
- •He argues the underlying AI fundamentals look like they’re improving rather than deteriorating
- •Gavin notes NVIDIA is at its lowest forward P/E in the last 10 years as of recording, implying the market expects over-earning
- •Patrick observes the AI model-release cadence is compressing, changing how quickly narratives and data shift
- •They set up the episode as a search for what the market is missing versus what it’s pricing in
Reconstructing “2022 in a month”: narratives that drove the drawdown
Gavin walks through the sequence of stories that, in his view, piled up to create the selloff. He argues several widely-circulated interpretations were directionally wrong, except for credit-related stress, which he treats as more legitimate.
- •He explains why the market read Meta potentially renting out compute as bearish, and why he thought that interpretation was incorrect
- •He describes a market freak-out around open source and a dip/flattening in a token index, tying it to mix shift rather than true demand weakness
- •He argues open source taking share mainly shifts margin away from frontier model providers and toward infrastructure, while still consuming similar compute per token
- •He describes China DUV-related news as another catalyst that hit semiconductor equipment stocks via basket trading
- •He flags rising real yields, wider spreads, and CDS moving as the most concrete macro/credit concern in the stack of narratives
Financing the AI buildout: operating cash flow vs. debt dependence
They focus on whether the AI buildout will require substantial credit, and what happens if credit markets tighten. Gavin argues consensus models underestimate monetization of new compute, and that repricing compute could reduce the need for debt financing.
- •Patrick frames the classic capital-cycle risk: debt-fueled buildouts can unwind quickly when supply and demand shift
- •Gavin says he “went deep on credit” because widening spreads and higher real yields are undeniable signals
- •He claims consensus estimates effectively model new generations of compute monetizing at older-generation rates, which he views as unlikely
- •He argues stronger monetization and repricing of installed compute could materially reduce implied credit demand
- •He ties the key question to whether on-the-ground demand signals persist so operating cash flows can fund most or all of the buildout
Why GPU rental prices rising changes the ROI debate
Gavin shares specific anecdotes suggesting GPU rental pricing is rising rather than falling, contrary to many expectations. He argues this implies hyperscalers may be under-earning relative to spot dynamics and that contract roll-offs could lead to higher repricing.
- •He says many expected GPU rental prices to decline, but claims even older GPU pricing is rising sharply in 2026
- •He gives an example of a startup re-renting a similar Blackwell cluster at a much higher $/GPU-hour months later
- •He references an inference cloud comment that it planned to pay substantially more when contracts expire
- •Gavin argues the gap between contracted compute and spot market pricing is central to the next phase of monetization
- •Patrick offers one potential negative datapoint he’s heard: third-party data suggesting Anthropic’s curve may have softened
Claude as “Walter Cronkite”: AI-driven interpretation and fast cycles in markets
Gavin argues markets may be reacting faster and more uniformly because many investors funnel news through the same AI tool. He connects this to rapid boom-bust “micro-cycles” in related equities that occur before fundamentals meaningfully change.
- •Gavin says many investors feed news into Claude (including agentic setups), reducing diversity in interpretation
- •He argues probabilistic AI interpretation can be smart but not always right, yet can still drive correlated trading
- •Patrick and Gavin discuss how this could compress the time scale of market narratives
- •Gavin cites an example of a “capacitor cycle” playing out in weeks rather than years
- •They link the month’s volatility to narrative reflexivity rather than clear fundamental deterioration
Technical progress that could shift compute needs: continual learning and sample efficiency
Patrick asks about innovation that might materially change inference/training economics. Gavin highlights chatter about continual learning and sample-efficient learning, noting it could alter training demand dynamics, while still being “amazing for the world.”
- •Gavin says people seem to feel close to solving continual learning and sample-efficient learning
- •He contrasts smaller token training regimes with much larger ones for frontier-scale models, suggesting a potential discontinuity if learning becomes more efficient
- •He suggests training demand may become a smaller share of overall compute demand over time
- •He notes uncertainty about timelines, referencing new labs focused on these problems
- •He adds that NVIDIA is heavily involved with many of these startups, influencing his view of ecosystem momentum
What would break the bullish thesis: cash flow deceleration, GPU oversupply, lab slowdowns
Patrick presses Gavin to define the conditions that would genuinely scare him. Gavin identifies a short list: failure of operating cash flow acceleration, sustained GPU price contraction, and meaningful demand plateaus at major labs—while noting he hasn’t heard anyone claim they have too many GPUs.
- •Gavin says a failure of hyperscaler operating cash flow to keep accelerating would be negative
- •He says a sustained, dramatic contraction in GPU prices and easing of GPU scarcity would be worrisome
- •He emphasizes he hasn’t heard anyone say they have excess GPUs; he’s only heard the opposite
- •He says lab demand (Anthropic, OpenAI) and broader ecosystem demand (Cursor/Grok and open source) matter for cash flow outcomes
- •He argues multi-model routing (open source plus frontier) can lower customer spend while still increasing underlying compute consumption
Who pays for AI: token spend vs labor spend, productivity vs substitution
They explore where the money ultimately comes from if AI infrastructure monetization rises. Gavin frames the answer as either productivity-driven growth or labor substitution, and they discuss early signals from “AI-native” companies and spending metrics.
- •Gavin says operating cash flow must ultimately come from either faster growth via productivity or labor substitution
- •He argues AI-native companies often aren’t firing people but are hiring fewer humans while leaning into tokens
- •Patrick shares ranges he’s seen for token spend as a percent of compensation spend at “pilled” companies
- •Gavin mentions examples of higher reported percentages in specific cases discussed on the show
- •They discuss evidence (as characterized by Gavin) that heavier AI spend correlates with faster growth, while noting skeptics’ concerns about confounds
The memory supply war and LTAs: trading upside for durability
Gavin argues the market is shifting from near-term “crushing numbers” to long-term agreements (LTAs) that reshape incentives—especially in memory. He describes game theory where breaking an LTA could risk future allocation and competitive standing in a supply-constrained world.
- •Gavin says everything is in shortage and that near-term weakness is often about energizing power fast enough
- •He describes long-term agreements with features like prepayment and floors/ceilings, emphasizing their growing importance in memory
- •He argues memory quantity per compute is pivotal for token output, making supply allocation strategically central
- •He outlines game-theoretic risks of breaking LTAs: future allocations could be punished if leverage shifts back to suppliers
- •He contrasts this dynamic with prior eras where dominant buyers could pressure suppliers with fewer consequences
NVIDIA’s new playbook: credit wrapper, revenue share, and stronger positioning
Gavin describes what he sees as a misunderstood evolution in NVIDIA’s approach: structures that help buyers finance compute and give NVIDIA a revenue share when prices are above a floor. He argues it can increase NVIDIA’s revenue per gigawatt while also reinforcing its competitive moat.
- •Gavin argues NVIDIA GPUs are the most financeable compute asset, which matters as credit markets tighten
- •He describes a “credit wrapper” concept paired with a revenue share/royalty-like structure tied to GPU economics
- •He distinguishes this from vendor financing, emphasizing that third parties provide capital while NVIDIA participates via terms and alignment
- •He suggests memory companies might logically pursue similar structures to stabilize outcomes amid credit stress
- •He argues these moves both boost NVIDIA’s economics and make it harder for smaller competitors facing higher costs and weaker financing terms
China DUV reactions and the open-source-driven software renaissance
They discuss the market’s reaction to reported China DUV progress and how hard it is to calibrate long-run implications. The conversation then shifts to open source and inference-cloud tooling as a “godsend” for software and AI-native defensibility through customization and routing.
- •Gavin says China DUV progress can be meaningful even if it’s far behind the frontier, and he thinks the market likely overreacted short term
- •He emphasizes “learning by doing” and argues you can’t simply skip the manufacturing learning cycles
- •He describes decoupling as a self-reinforcing process on both sides
- •He argues open source nearing the frontier plus tooling (like model customization and routing) helps software companies reduce costs and become more defensible
- •They discuss an approach where frontier models plan and smaller models execute tasks, improving efficiency while expanding usage
Data centers and regulation: political narratives, PR failures, and local backlash risk
Gavin calls regulation the biggest risk and cites a data center moratorium as an example that can’t be ignored. He argues the AI industry has done a poor job communicating realities about electricity, water, jobs, and community benefits, allowing misinformation to spread.
- •Gavin says regulation is the most obvious major risk and highlights the threat of restrictive policy moves
- •He argues political narratives focus on data centers raising power prices, consuming water, and taking jobs
- •He claims behind-the-meter deals and community contributions can make local economics better, including persistent blue-collar work
- •He discusses how a widely-cited water-usage error propagated despite being admitted and debunked, illustrating misinformation dynamics
- •He suggests the industry needs coordinated storytelling (potentially via a foundation/PAC-style effort) to correct perceptions and reduce regulatory risk
SpaceX as a compute player and the possibility of orbital compute
They examine how SpaceX’s compute efforts have been digested by markets and whether its role is underappreciated. Gavin also describes why orbital compute feels increasingly plausible, referencing partnerships and investment signals he views as a sanity check, before closing the conversation from Benchmark’s office.
- •Gavin argues SpaceX’s ability to bring on compute quickly and at low cost has been demonstrated, and demand absorbed its added supply without much disruption
- •He references a report claiming SpaceX could bring on a very large amount of compute, calling it difficult to believe while still refusing to “bet against” Elon
- •He says he sees a short case centered on compute spot prices collapsing, but he thinks little upside is priced in for SpaceX’s potential compute scale
- •He says “orbital compute feels more real every day,” mentioning an orbital compute company SpaceX is partnering with and why that matters to him
- •They note they recorded in Benchmark’s office and end by emphasizing that time will reveal which expectations were right
