All podcasts / Invest Like the Best / Summary

Gavin Baker - Nvidia v. Google, Scaling Laws, and the Economics of AI (EP.451)

2025-12-09 - 89 min - source - Read full transcript

Patrick O'Shaughnessy (host)Gavin Baker

Key insights

Gemini 3 was the first confirmation since Hopper that pre-training scaling laws remain intact, closing an 18-month period in which reasoning alone (not chip-driven scaling) carried all visible AI progress.

Baker argues that because Blackwell's product transition was the most complex in Nvidia's history (air-cooled to liquid-cooled, 30kW to 130kW racks), Google trained Gemini 3 on 'F-4 Phantom'-era TPUs rather than 'F-35'-era Blackwell chips, yet still proved pre-training scaling holds. Had reasoning (RLVR and test-time compute) not emerged when it did to bridge the gap, he says AI progress would have flatlined from mid-2024 through today.

scaling-laws-and-frontier-progress

Reasoning created a genuine data flywheel for frontier labs for the first time, which is why Meta, Microsoft, and Amazon all failed to build a top-tier model despite trying hard and spending heavily.

Before reasoning, a pre-trained model was static once released; verified user feedback on reasoning answers can now be fed back into training, replicating the product-to-data-to-better-product loop that made prior internet giants durable. Baker cites Zuckerberg's January 2025 prediction that Meta would have the best AI by year-end (it didn't crack his top 100), Microsoft's failed bet on Inflection AI's team, and Amazon's Nova models (outside the top 20 despite acquiring Adept AI) as evidence this is much harder to replicate than assumed a year ago.

scaling-laws-and-frontier-progress

Edge AI - a 'good enough' model running locally on a phone at roughly 30-60 tokens per second - is Baker's single most plausible bear case for the AI infrastructure trade.

He expects phones within about three years to run a pruned-down frontier-class model locally and for free (Apple's implied strategy), calling the cloud only for harder queries. If a modest-IQ local model proves sufficient for most use cases, he says that undercuts the case for ever-larger cloud compute demand; he considers this a more serious risk than a slowdown in the pre-training scaling laws themselves.

scaling-laws-and-frontier-progress

Nvidia's GB300, being drop-in compatible with existing GB200 racks, should let Nvidia customers reclaim the low-cost-producer position from Google in 2026, which should force Google to reconsider running AI at a roughly negative 30% margin.

Baker frames Google's current strategy - undercutting on price to 'suck the economic oxygen' out of the ecosystem and starve capital-constrained competitors - as rational only as long as Google is the lowest-cost token producer. Once GB300 racks (which reuse GB200 power and cooling infrastructure companies already know how to operate) scale into training and then inference, he expects that cost advantage, and the strategic calculus behind Google's negative-margin pricing, to flip.

nvidia-vs-google-chip-economics

Google's conservative TPU design choices are, in Baker's telling, a direct consequence of the roughly $15-25 billion a year it pays Broadcom for ASIC back-end design and Taiwan Semi management, and bringing that fully in-house (as Apple did with its chips) is economically inevitable past a certain scale.

Broadcom earns an estimated 50-55% gross margin on TPU back-end work while its entire semiconductor division's opex is only about $5 billion; Baker's math implies Google could roughly double or triple Broadcom-equivalent engineers' pay for less than what it currently pays as margin. He reads Google's decision to bring in MediaTek as a second ASIC partner as the 'first warning shot' in a negotiation, not yet a full in-house move.

nvidia-vs-google-chip-economics

It takes about three chip generations for an ASIC program to become genuinely competitive with merchant GPUs, which is why Baker expects only TPU and Trainium to matter among AI accelerators, and even expects both to eventually offer customer-owned tooling.

He cites Amazon's Trainium/Inferentia team, which he calls the best ASIC design team among the hyperscalers, as still only reaching rough parity around its third generation; TPU followed the same arc, becoming 'even vaguely competitive' only at v3/v4. He calls this inevitable economics, not a company-specific choice, meaning most other ASIC efforts are unlikely to matter.

nvidia-vs-google-chip-economics

China's push to force domestic AI development onto Huawei chips by refusing US Blackwell imports is, in Baker's view, a strategic mistake that will widen the US-China AI gap and hand America real geopolitical leverage.

He points to DeepSeek's own V3.2 technical paper citing insufficient compute as a reason it struggles to match American frontier labs as an implicit admission that Chinese labs need Blackwell access. He expects the compute gap to become undeniable to Chinese policymakers by late 2026, at which point rare-earth supply (which he argues is 'not actually rare,' just costly to refine) becomes a bigger point of American leverage as DARPA-backed refining alternatives come online.

nvidia-vs-google-chip-economics

Data centers in orbit are, per Baker, 'the most important thing that's going to happen in this world in the next three to four years,' because space offers roughly six times Earth's solar irradiance, free radiative cooling, no battery requirement, and faster inter-node networking than terrestrial fiber.

A satellite can stay continuously sunlit for 24-hour solar capture at 30% higher intensity than on Earth, eliminating the battery cost that dominates terrestrial data-center economics; radiative cooling on a satellite's dark side is free versus the HVAC and liquid-cooling infrastructure that makes up most of a rack's mass on Earth; and laser links between satellites (already proven by Starlink) travel faster through vacuum than light through fiber optic cable. He expects inference workloads to migrate to orbit well before training, given training's larger scale requirements, and ties the buildout directly to SpaceX/Starship launch cadence and the reported convergence of SpaceX, Tesla, and xAI.

orbital-compute

The ROI on AI capital spending is already empirically positive by return-on-invested-capital measures, and Q3 2025 was the first quarter non-tech Fortune 500 companies reported concrete, quantified AI-driven earnings uplift.

Baker points to C.H. Robinson, which moved from quoting only 60% of inbound truck-availability requests within 15-45 minutes to quoting 100% of requests within seconds using AI, posting a roughly 20% earnings beat and a matching stock pop. He argues this validates VCs' greater bullishness on AI relative to public-market investors, since venture portfolios already show falling headcount-per-revenue-dollar trends that public reporting is only now starting to confirm.

ai-roi-and-enterprise-adoption

A potential near-term 'ROI air gap' - heavy capex on Blackwell for training with no matching inference revenue yet - is a real risk Baker was worried about, evidenced by Meta's declining ROIC in a recent quarter tied to its lack of a frontier model.

Training-only GPU deployment produces capex without matching token revenue until models are actually shipped to users; Baker says this dynamic already dented one hyperscaler's reported returns and that the C.H. Robinson-style Fortune 500 examples are important because they suggest the broader economy may be able to navigate this gap rather than stall out.

ai-roi-and-enterprise-adoption

SaaS incumbents are repeating brick-and-mortar retailers' e-commerce mistake by refusing to accept AI-agent gross margins of roughly 35-40% in order to protect legacy 70-90% software margins, even as AI-native competitors already access their own customer data through agents.

Baker's analogy: retailers dismissed e-commerce because of its lower apparent margin structure, and Amazon ultimately achieved higher margins in North American retail than many of the incumbents that refused to invest. He argues any SaaS company (naming Salesforce, ServiceNow, HubSpot, GitLab, and Atlassian) could build a competitive agent product by automating its own core customer function and selling it at 10-20% of the value delivered, but that protecting legacy gross margins is a 'life or death' decision most of them are failing, with Microsoft as the main exception via its GitHub Copilot distribution strategy.

saas-application-layer-disruption

Books referenced

What Technology Wants - Kevin Kelly - Patrick invokes Kelly's concept of the 'Technium' (technology as a self-propagating force) when Baker observes that AI has gotten whatever it needed to keep growing - reasoning arriving just as Blackwell was delayed, public opinion on nuclear power flipping almost overnight, orbital compute emerging as power becomes a bottleneck.
Peter Lynch's books - Peter Lynch - The first investing books Baker bought and read in two days during his DLJ internship, kicking off the self-directed reading binge that changed his college major from English/history to history/economics.
Market Wizards - Jack D. Schwager - Read during the same internship binge as Peter Lynch's and Warren Buffett's books, part of Baker's crash course in how skilled investors think.
Warren Buffett's Letters to Shareholders - Warren Buffett - Baker read the full set of letters twice during his DLJ internship, forming an early backbone of his investing framework.
Why Stocks Go Up (and Down) - William H. Pike - The book Baker used to teach himself accounting after his internship reading binge; he calls it out by name as foundational to his investing education.

Media referenced

DeepSeek V3.2 technical paper - paper - Baker cites DeepSeek's own admission in the paper that insufficient compute is a reason it struggles to compete with American frontier labs, which he reads as evidence China's rare-earth leverage play against Nvidia access was a strategic mistake.
Rounders - movie - The movie that got Baker and his college rock-climbing friends into playing poker, part of his framing of investing as 'a game of skill and chance.'

Companies

Nvidia - Central to the episode's framing: Blackwell's product transition (air-cooled to liquid-cooled, 30kW to 130kW racks) was the most complex in the company's history and delayed deployment by roughly 18 months, but Baker expects the GB300 (drop-in compatible with GB200 infrastructure) to restore Nvidia customers' low-cost-producer status over Google in 2026.
Google - Currently the low-cost producer of AI tokens because Blackwell's complexity forced it to train Gemini 3 on 2024-25-era TPUs while still proving pre-training scaling laws intact; Baker argues its rational strategy of running AI at a roughly negative 30% margin to starve competitors of capital will have to change once Nvidia customers regain the cost advantage.
OpenAI - Described as the highest-cost producer of tokens among the four leading labs because it pays a margin to third parties (the Stargate structure) for compute rather than running it directly; Baker connects this to the reported 'Code Red' internal memo and the jump to $1.4 trillion of spending commitments.
Anthropic - Credited with burning dramatically less cash than OpenAI while growing faster, benefiting from its Google TPU and Amazon Trainium relationships; its $5 billion Nvidia deal is read by Baker as a savvy hedge that gives Nvidia a third 'fighter' (alongside xAI and OpenAI) in the fight against Google.
xAI - Expected to release the first model trained on Blackwell in early 2026 because, per Jensen Huang, 'no one builds data centers faster than Elon'; xAI's speed in deploying Blackwell at scale also helps Nvidia work out early-generation bugs for everyone else. Grok is also cited as having outsized OpenRouter API token share relative to its overall scale.
Meta - Cited as proof frontier-model development is harder than it looks: Zuckerberg predicted in January 2025 that Meta would have the best AI by year end, but Baker doesn't think Meta's models cracked the top 100; he says Nvidia is likely helping Meta's infrastructure quietly and that Chinese open-source models serve as a bootstrap checkpoint for Meta's own training.
Microsoft - Also failed to build a competitive internal model despite acquiring Inflection AI's team and predicting it would run more workloads on its own models; Baker cites Microsoft's roughly six-week 'blink' earlier in 2025 (participating less aggressively in the AI capex race) as a decision he believes the company now regrets.
Amazon - Its Nova models rank outside the top 20 despite acquiring Adept AI's team; separately praised as having the best ASIC design team in semiconductors via Trainium and the Graviton CPU program, illustrating that strong chip execution and strong frontier-model execution are different skill sets.
Broadcom - Earns an estimated 50-55% gross margin on the back-end design and Taiwan Semi management work it does for Google's TPU program; Baker estimates this costs Google roughly $15-25 billion a year at 2027-28 TPU revenue scale, which is the economic driver behind Google's move to bring more of the process in-house and to add MediaTek as a second ASIC partner.
MediaTek - Brought in by Google as a second TPU back-end partner, which Baker calls Google's 'first warning shot' to Broadcom over pricing.
TSMC - Framed as one of two 'natural governors' (with a potential DRAM cycle) preventing an AI compute glut; Baker says TSMC is deliberately cautious about expanding capacity because it fears an overbuild, and recalls TSMC executives dismissing Sam Altman's early compute requests.
Intel - Discussed as a beneficiary of the compute shortage: new CEO Lip-Bu Tan is 'reaping the benefits' of predecessor Pat Gelsinger's foundry strategy, and Intel's empty fabs are likely to get filled as TSMC capacity stays constrained; Baker calls Gelsinger's firing 'shameful' in hindsight.
AMD - Named alongside Trainium as one of the only other accelerator programs Baker expects to matter, with the MI450 generation cited as a coming 'big quantum' of AI spend alongside Rubin and TPU v9.
C.H. Robinson - Baker's headline example of Fortune 500 AI ROI: the freight brokerage moved from quoting only 60% of inbound truck-availability requests in 15-45 minutes to quoting 100% in seconds using AI, reporting a roughly 20% earnings beat and a 20% stock pop that he calls the first clear non-tech-sector proof of AI-driven revenue uplift.
Fortel - A hearing-technology company Baker and Patrick have both personally invested in, cited as an example of AI being used to design a genuinely novel product rather than just cut costs (transcribed name may be approximate).
Tesla - Named by Baker (via Elon Musk's own comments) as converging with SpaceX and xAI: Optimus will run on xAI intelligence with Tesla Vision as its perception system, giving xAI a built-in customer relationship and giving Tesla a data-center-in-space compute source.
SpaceX - Baker's data-centers-in-space thesis rests on SpaceX/Starship being the only economically viable launch platform at the volume required, and on Starlink's proven direct-to-cell laser and satellite-networking technology, which Baker argues makes orbital inference both cheaper and lower-latency than terrestrial data centers.
Salesforce - Named, with ServiceNow, HubSpot, GitLab and Atlassian, as a SaaS company that could and should build a native agent business, but is instead protecting 70-90% gross margins instead of matching AI-native competitors' roughly 35-40% agent economics.
Adobe - Cited as the historical precedent for the market tolerating gross-margin (and even revenue) compression during a business-model transition, referring to its on-premise-to-SaaS conversion.
Caterpillar - Cited as evidence the power supply chain is responding to AI demand: the company announced it will increase turbine manufacturing capacity by 75% over the next few years.
DeepSeek - Its V3.2 paper's admission of a compute shortfall is Baker's evidence that China's push to force domestic AI development onto Huawei chips (by refusing US Blackwell imports) was a geopolitical misstep that will widen the US-China frontier gap.
Huawei - China's domestic chip alternative to Nvidia, described by Baker as 'okay' but well behind Blackwell, underlying his argument that China's chip-sovereignty strategy backfired.
The Motley Fool - An early source Baker read obsessively as a young investor; he credits the site with being early to popularize return on invested capital (ROIC) as a key metric.

Techniques and frameworks

Scaling laws for pre-training, RLVR, and test-time compute - Baker's three-layer framework for AI progress: pre-training scaling (confirmed intact by Gemini 3), reinforcement learning with verified rewards (RLVR) in post-training, and test-time/inference compute. He argues nearly all visible AI progress from late 2024 through Gemini 3 came from the latter two because Blackwell's delay stalled pre-training gains.
Anything verifiable can be automated - Baker's extension of Andrej Karpathy's line that anything specifiable can be automated with software; with AI, anything with a verifiable right/wrong outcome (accounting reconciliation, sales conversion, game wins/losses) can be automated via reinforcement learning.
The reasoning data flywheel - Baker's explanation for why frontier labs suddenly have durable moats: verified user feedback on reasoning-model answers can now be fed back into training in a way pre-reasoning RLHF never captured well, replicating the product-data-product flywheel that made Amazon, Netflix, and Google hard to compete with.
Intelligence-to-usefulness handoff - Baker's framework for where AI value creation goes next: raw model intelligence gains are becoming hard to perceive for non-experts, so the next value unlock is 'usefulness' (consistent, reliable task execution held together by long context and memory of user preferences), followed eventually by scientific breakthroughs.
Investing as the search for truth - Baker's closing definition of his career: investing is a game of skill and chance (like poker) where edge comes from finding a truth about the future that others haven't yet seen, built by intersecting deep historical knowledge with an accurate read of current events.

Summary

In this sixth-ish return visit with Patrick O'Shaughnessy, Atreides Management's Gavin Baker works through why Gemini 3 mattered, why Nvidia's cost advantage over Google should reassert itself in 2026, and why he thinks data centers belong in space. He opens on process: how he personally tracks AI progress (paying for the top-tier subscription rather than judging models on their free tiers, following the roughly 500-1,000 people worldwide who actually understand frontier research on X, and listening to lab researchers on podcasts). From there he explains why Gemini 3 was significant less for beating benchmarks than for confirming pre-training scaling laws are still intact after an 18-month stretch in which Blackwell's extraordinarily complex product transition (from air-cooled 30kW racks to liquid-cooled 130kW racks) stalled Nvidia-side scaling entirely; reasoning (RLVR and test-time compute) was the only thing carrying AI progress during that window, and without it, he says, the field would have flatlined.

The heart of the conversation is Baker's Nvidia-versus-Google framework. Google currently holds a real, temporary cost advantage as the low-cost producer of tokens, which he says explains its rational strategy of running AI at a roughly negative 30% margin to starve capital-constrained competitors of oxygen. He expects that advantage to flip once GB300 (drop-in compatible with existing GB200 infrastructure) scales, restoring Nvidia customers' cost leadership and forcing a change in Google's pricing calculus. He layers in a granular account of why Google's TPU program is structurally conservative: paying Broadcom an estimated $15-25 billion a year for ASIC back-end design creates strong economic pressure to bring that work in-house, which he reads Google's addition of MediaTek as an early signal of. He extends the same logic to explain why he expects only TPU and Trainium to matter among ASIC alternatives to Nvidia GPUs (it takes roughly three chip generations to become competitive) and why he thinks China's refusal to import Blackwell chips, in favor of forcing domestic development onto Huawei silicon, was a geopolitical misstep that DeepSeek's own technical papers now implicitly admit.

Baker's most expansive riff is on data centers in space, which he calls the most important development of the next three to four years: orbital solar delivers roughly six times Earth's irradiance with no need for batteries, radiative cooling is free, and laser links between satellites are faster than fiber optic cable on Earth. He ties this to the reported convergence of SpaceX, Tesla, and xAI (Optimus running on xAI intelligence with Tesla Vision perception, powered eventually by SpaceX's orbital compute) and frames Starship's launch cadence as the binding constraint. On demand, he argues AI's ROI is already empirically positive and points to Q3 2025 as the first quarter non-tech Fortune 500 companies reported concrete AI-driven earnings uplift, led by freight broker C.H. Robinson's jump from quoting 60% of truck-availability requests in minutes to 100% in seconds. He is candid that a near-term "ROI air gap," where Blackwell capex goes toward training with no matching inference revenue yet, is a real risk he has watched play out in weaker ROIC at labs without a frontier model.

The conversation closes with two structural bets and a personal story. Baker argues SaaS incumbents are repeating brick-and-mortar retailers' e-commerce mistake: refusing to accept roughly 35-40% AI-agent gross margins to protect legacy 70-90% software margins, even as venture-funded AI-native competitors already access their customer data through agents and will eventually cut them out entirely. He also makes the case that the semiconductor venture ecosystem's resurgence, after decades of dormancy, is structurally necessary because no single company - not Nvidia, not Google, not AMD - can execute an annual chip-refresh cadence alone; a rack has thousands of parts, and all of them have to accelerate together. He closes, in response to a question about how to describe his career to his host's young son, by tracing his path from a history-obsessed, athletically middling Dartmouth rock-climber and aspiring ski bum to investor: a DLJ mailroom internship that turned into a reading binge (Peter Lynch, Market Wizards, Buffett's letters, Pike's "Why Stocks Go Up and Down"), reframing investing as a lifelong search for hidden truth in a game of skill and chance.

Notable Quotes

"Foundation models without unique data and internet-scale distribution are the fastest depreciating assets in history. And reasoning fundamentally changed that." - Gavin Baker

"It's fucking gold. In space, cooling is free. You just put a radiator on the dark side of the satellite." - Gavin Baker

"If you're trying to preserve an 80% gross margin structure, you are guaranteed that you will not succeed in AI. Absolute guarantee." - Gavin Baker

"I have just been fascinated that for the last two years, whatever AI needs to keep growing and advancing, it gets." - Gavin Baker

"Investing, I kind of conceptualized it as a game of skill and chance, kind of like poker... the way you got an edge was you had the most thorough knowledge possible of history, intersected with the most accurate understanding of current events." - Gavin Baker