TPU
Tensor Processing Unit—Google's custom AI accelerator chips, including specialized inference-only variants.
As defined in lex_demis-hassabis-2
Where it comes up · 19
- Nvidia v. Google, Scaling Laws, and the Economics of AI5 claims
- Inside the Trillion-Dollar AI Buildout4 claims
- Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI4 claims
- AMD: How Chips Are Changing3 claims
- Dwarkesh Patel Podcast - Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat3 claims
- 5 Companies Making Real Products with 3D Printing2 claims
- Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480]2 claims
- George Hotz #32 claims
and 11 more
Most important claims · 10 of 37Search all claims for “TPU” →
- Google has a structural infrastructure advantage because NVIDIA chip margins are insane; Google can develop its full stack top-to-bottom with TPUs to avoid that margin, and combined with a head start in data centers and the high lead times and hard margins involved, this gives Google a historical advantage in scaling AI.Ai Sota #2026
- ML accelerators are fundamentally a software problem, not a hardware problem; companies like Graphcore, Cerebras, TensTorrent and Groq built good chips but terrible software, and if you can't write a Torch-level performance stack on Nvidia GPUs down to the driver, you certainly can't write one for your more specialized chip.George Hotz #3
- It takes at least three generations to produce a competitive custom chip; TPU v1 was an achievement just to ship, and TPUs only became vaguely competitive around TPU v3 or v4.Nvidia v. Google, Scaling Laws, and the Economics of AI
- Google said its TPU AI chip allowed it to cut the number of data centers it had to build by 50%, when it was building one roughly every nine months at about $1 billion each.AMD: How Chips Are Changing
- Google has the biggest cluster overall using TPUs, but spreads it across multiple data center super-regions ~30 miles apart connected by high-bandwidth fiber, whereas Elon's 200,000-GPU Memphis cluster is the largest fully connected single-site cluster; Google doesn't sell TPUs because they're optimized for internal workloads and selling would cannibalize its search profit DNA.Deepseek Dylan Patel Nathan Lambert
- Because chips deployed this year, such as GB300, TPUv7 and Trainium3, deliver 3-5x more performance per watt than prior generations, each new gigawatt adds far more effective compute than older ones, so labs' share of usable FLOPs compounds beyond their share of watts.Dwarkesh Patel Podcast - Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
- Over the last decade, general-purpose CPU-based machine scaling has slowed dramatically: fabrication process improvements now take three years instead of two, and multi-core architectural improvements no longer give the boosts seen 20-to-10 years ago, while specialized ML accelerators like TPUs and ML-focused GPUs now deliver high performance and efficiency for modern computations.Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
- At Google scale, where every search runs through multiple AI iterations for billions of users, a special-purpose ASIC doing AI math is more efficient than a GPU, so data centers will increasingly use custom chips for AI.AMD: How Chips Are Changing
- Google moved from asynchronous CPU training to fully synchronous TPU training because synchronous runs are replicable—results no longer depend on incidental factors like a web crawler running on the same machine—making it an easier mental model where only your algorithm fights you rather than asynchrony plus the algorithm both battling you, though at greater scale some asynchrony may return.Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
- The only remotely successful ASIC is Google's TPU, and the only reason it succeeded is that Google wrote a machine learning framework (TensorFlow); you have to write a competitive ML framework in order to be able to build a viable ASIC.George Hotz #3