Nvidia v. Google, Scaling Laws, and the Economics of AI
What this covers
This Invest Like the Best conversation features Gavin Baker working through the hardware and business mechanics behind the current AI cycle. The spine is chip history. A single pre-training cluster tops out near 200,000 Hopper GPUs that share coherence, so further scaling had to wait for Blackwell. He calls that the most complex product transition in technology history: liquid cooling, racks going from about 1,000 to 3,000 pounds, and power density rising from 30 to 130 kilowatts. He reads Gemini 3 as evidence that pre-training scaling laws still hold, while admitting nobody knows why they work. Reasoning models, he says, filled the gap in between.
The range is wide. On semiconductors, he argues TSMC has smoothed supply so well that no true capacity cycle has occurred since the late 1990s. He also argues a competitive custom chip takes at least three generations, with TPUs only becoming vaguely competitive around v3 or v4. On economics, he says AI's need to recompute every answer pushes gross margins toward 40 percent, and compares SaaS incumbents to retailers that resisted e-commerce. His evidence for returns includes ROIC at large GPU spenders, AI handling over half of customer support at some companies, and C.H. Robinson quoting every inbound freight request within seconds. He also names a risk: an "ROI air gap" while Blackwell capacity goes to training.
The most contested stretch concerns orbital data centers. Baker argues cooling is nearly free in space, sunlight is constant and far more intense, and lasers in vacuum could outperform fiber. He thinks the idea suits inference better than training. Meta's missed prediction of AI leadership gets a brief mention.
Baker argues AI progress now rests on post-training reinforcement learning and test-time compute layered on still-intact pre-training scaling, and that the winners will be the lowest-cost producers of tokens, with returns on AI already visible in public company financials.
- Pre-training scaling held with Gemini 3, and reasoning (RL with verifiable rewards plus test-time compute) supplied the progress from late 2024 onward while Blackwell was delayed
- Low cost per token has become a decisive competitive advantage in AI, unlike in prior tech eras, so Google and vertically integrated players gain share
- Measured ROIC and revenue acceleration at large GPU spenders show AI returns are already positive, though the ROI air gap during Blackwell training is a real risk
Frontier AI progress is limited by chip generations and models lacking unique data depreciate
- Hopper is like a World War II-era P-51 Mustang, Google's 2024-2025 TPUs were F-4 Phantoms, and Blackwell is an F-35, so Google trained Gemini 3 on chips that were two generations behind.
“Because Blackwell was such a complicated product and so hard to ramp, Google was training Gemini 3 on '24 and '25 era TPUs”
- A single pre-training cluster cannot exceed roughly 200,000 Hoppers that are coherent, meaning all GPUs share memory and act as one system, so progress requires waiting for the next chip generation to scale further.
“you really can't get more than 200,000 Hoppers coherent.”
- Foundation models without unique data and internet-scale distribution are the fastest-appreciating assets in history in reverse, meaning they depreciate quickly, a view Baker modified from Eric Vishria's claim that they appreciate fastest.
“I said, 'Foundation models without unique data and internet scale distribution are the fastest-appreciating assets in history.'”
AI and GPU efficiency gains show up as company growth and founder productivity
- Moving recommender systems from CPUs to GPUs drove large efficiency gains at public internet companies, which helps explain why their revenue growth accelerated.
“moving the big recommender systems that power the advertising and the recommendation systems from CPUs to GPUs, and you've had massive efficiency gains.”
- Young founders are becoming very polished because they use AI to prepare for investor pitches, handle HR situations, and improve sales, which is why their companies show such large AI productivity.
“These young CEOs, they're just so impressive in all ways and they get more polished faster.”
Baker found his edge in investing after trying and abandoning other pursuits
- Baker dedicated himself to rock climbing in college, doing homework at the gym, and considered becoming a ski bum, river worker, or novelist before an internship changed his path.
“rock climbing is by far the most important thing in my life. I dedicated myself to it completely.”
- Baker says he never became good at anything other than investing, despite trying chess, poker, and skiing, and that competitive investing is the one domain where he has excelled.
“And to this day, this is the only thing I'm good at.”