Google Sells TPU Full Racks Externally, Challenging NVIDIA on Inference Cost-Performance
nashnova research
Google is selling complete TPU systems to external customers for the first time, shifting from cloud-only rental to full-rack hardware sales; the move helped push Google Cloud revenue to $30.7 billion in Q3, up 103% year-on-year, and benchmark tests show TPU beating Nvidia on cost per token in some inference workloads.
Why is Google suddenly selling TPU hardware outright?
Until this year, Google's TPU chips were rent-only — customers accessed them through cloud services but could never buy the hardware. That changed: Google now ships complete TPU systems including compute, storage, networking hardware, and software.
This means → Google has moved from "landlord selling compute time" to "equipment vendor competing head-to-head with Nvidia."
According to Quantumbit (量子位), a senior executive who joined Google less than three years ago championed the decision internally, arguing that Google should "sell TPUs the way Nvidia sells GPUs."
What is Google's logic — how does "sell to fund R&D" work?
Internally, Google frames the shift as a "sell-to-fund-research" structural swap: external orders share the capital cost of data centers, and scale demand drives down unit costs for in-house chip development.
In plain terms = sell the capacity you can't use yourself, let outside buyers subsidize your costs, then plow the revenue back into the next generation of chips.
Chip chief Amin Vahdat publicly stated that "the future belongs to specialization," positioning purpose-built TPUs against Nvidia's narrative that one general-purpose GPU can do everything.
How fast is the revenue growing?
Google Cloud posted $30.7 billion in Q3 revenue, up 103% year-on-year. Operating profit jumped from $3.594 billion to $11.121 billion — more than tripling.
Alphabet's latest 10-Q filing explicitly lists "product sales, primarily TPU systems" as a separate line item under Google Cloud revenue, with management forecasting further expansion in Q4.
This reflects a business that has moved well past the pilot stage — it is large enough to warrant its own financial disclosure. But most contract revenue won't hit the books until 2027, so today's numbers understate the real scale.
Can TPU actually beat Nvidia on inference cost?
Semiconductor research firm SemiAnalysis published agent-inference benchmarks showing Google's seventh-generation TPU Ironwood delivering higher tokens per dollar than Nvidia GPUs in certain tests using models like Muse Spark.
This means → for the first time, Nvidia faces a cost-performance challenge from a completely different architecture — not a rival GPU from AMD, but a purpose-built TPU.
Nvidia responded fast: after Google submitted its AgentX benchmark results, Nvidia engineers worked through the night until 2 a.m. Over four rounds of optimization, the GB300 NVL72 FP4 clawed back a 1.5× to 2.6× lead in tokens per dollar on one model test.
Is the software moat being dismantled?
Google and Meta's PyTorch team have jointly developed TorchTPU, now open-sourced, which lets developers run PyTorch models on TPUs with minimal code changes.
In plain terms = Nvidia's deepest moat is not hardware — it is the CUDA software ecosystem. Everyone's code is written around it, and switching chips means rewriting. TorchTPU aims to lower that rewrite wall.
Whether Nvidia can hold pricing power in inference depends on sustaining its per-dollar performance lead — and that lead is now being squeezed from multiple directions by Google, AMD, and others simultaneously.
市场有风险，内容仅供研究参考，不构成投资建议。