Gavin Baker says Google's temporary edge as lowest-cost token producer let it drain the AI ecosystem with negative 30% margins while Blackwell deployments lagged.
Gavin Baker says Gemini 3 confirms pre-training scaling laws are intact, though nobody can explain why.
Baker cites C.H. Robinson quoting 100% of requests in seconds and rising 20% on earnings as proof AI ROI is unambiguously positive.
Why Google's cheap AI tokens squeezed the whole ecosystem
Google bought a temporary edge in AI by selling tokens below cost. Its TPUs made it the lowest-cost producer of intelligence, and it used negative margins to take share while Nvidia's Blackwell transition dragged on. Baker argues the constraint has moved to physics, where racks, power and memory set the pace, and that the returns on AI are real even though they arrive after the spending.
Takeaways
Pre-training scaling holds while two new laws emerged. Gemini 3 confirmed that pre-training scaling laws are intact, which the lab stated unequivocally, though nobody understands why they hold. Progress in 2024 and 2025 came from reinforcement learning with verifiable rewards and test-time compute, which took ARC-AGI from 8% to 95% in three months.
Hopper to Blackwell is the hardest transition. Racks went from about 1,000 pounds and 30 kilowatts to 3,000 pounds and 130 kilowatts, which forced liquid cooling. Blackwells take six to nine months to perform at Hopper's level, so the chips have only just started deploying at scale.
Google's cheap tokens squeezed the ecosystem. Google's TPUs made it the lowest-cost token producer, letting it run negative 30% margins in principle and suck oxygen from the rest of the ecosystem. Baker says the window closes once Blackwell supply loosens.
ASIC economics favour Broadcom and Google. Broadcom earns 50 to 55% gross margins on TPUs, so $50bn of TPU revenue in 2028 would send $25bn its way. Baker expects few ASICs beyond Trainium and TPU, and calls MediaTek a warning shot.
AI returns are real but ROIC lags. C.H. Robinson, a freight forwarder, now quotes 100% of requests in seconds and rose 20% on earnings. Baker warns that Blackwell spending lands before inference revenue arrives, creating an air gap in return on invested capital.
Edge AI is the scariest bear case. Within three years a bigger, bulkier phone could run a pruned Gemini 5 or Grok 4 at 30 to 60 tokens per second for free, which would undercut the data centre build-out.
What Baker expects next
The first models trained on Blackwell arrive in early 2026 from xAI, and Baker expects Nvidia's lead to widen with Rubin, which he calls a far easier transition than Blackwell. AMD's MI450 also enters the race.
The first true DRAM cycle since the late 1990s could act as a natural governor that slows the build-out.
Rare earths get solved faster than most expect, because DARPA and Defense Department programmes fund alternatives and friendly countries with deposits do the refining.
Data centres in space become plausible. A satellite in the right orbit gets 24-hour sun at six times the irradiance of a terrestrial solar farm, and cooling is free with a radiator pointed at the dark side.
Tap a name to go there · tap a line to see why they are linked