How does Google TPU expand the AI market? What impact does it have on NVIDIA?

The emergence of Google's TPU is not essentially "cutting the pie of AI infrastructure," but rather dividing the cake while also making it bigger.

First, even without TPU, Google's AI CapEx wouldn't be less.
As a tech giant, Google must continuously invest in AI infrastructure for businesses like search reranking, YouTube recommendations, and Gemini. Without TPU, it would still need to purchase NVIDIA GPUs on a large scale. After TPU's launch, the gross margin from self-developed chips can be retained within the group, making the financials look better, and management becomes more willing to increase long-term investments. Therefore, TPU is more like adding a new form of AI infrastructure on top of the existing foundation, rather than shrinking the original market.

Second, self-developed chips improve "computing power cost-effectiveness," directly increasing total computing power demand.
With TPU, Google can obtain more usable computing power under the same capital expenditure, with lower marginal costs for inference. This gives it the confidence to widely deploy large models across more products: search, maps, Gmail, Docs, the Android ecosystem... all can more aggressively leverage large models. The result is that overall AI usage is pushed higher, total computing power demand expands, and the entire "AI infrastructure market" (data centers, power, cooling, networking, storage, etc.) grows alongside it.

Third, TPU has triggered an "AI infrastructure arms race" among cloud providers.
When Google introduced TPU as a self-developed accelerator card, other cloud providers couldn't sit still: AWS developed Trainium/Inferentia, Microsoft created Athena/Maia, and domestic cloud players are also pushing their own NPUs. Those who don't invest in AI infrastructure risk being squeezed out of the cloud business. The result isn't "with TPU, everyone buys fewer GPUs," but rather: to maintain cloud competitiveness, all providers must keep pouring money into AI infrastructure, further driving up the industry's total CapEx.

Fourth, more "platforms" mean more developers and more demand.
TPU isn't just a chip—it's connected to a full suite of services including XLA, JAX, TF, and GCP. NVIDIA has the CUDA ecosystem, while self-developed chip factions are forming their own platforms. More platforms mean more entry points, lowering the barrier to starting AI projects: some teams are accustomed to CUDA + GPU, others directly use TPU services on GCP, and some build on local NPUs. Long-term, the number of successfully launched AI projects will only increase, further expanding the overall demand pool.

For NVIDIA: short-term it's "dividing the cake," long-term it's "forcing evolution."
In the short-to-medium term, TPU will undoubtedly take away some orders that might have gone to NVIDIA—this is real competitive pressure. But with AI's total demand expanding, NVDA still has a chance to capture the largest share of the general market. More crucially, the rise of self-developed chips is forcing NVIDIA to evolve from "selling chips" to "selling platforms"—as Jensen Huang himself said, all current investments are aimed at expanding the CUDA ecosystem. If NVIDIA can truly turn CUDA + various AI software stacks into a foundational standard like "Windows + DirectX" for the AI era, then even if pure hardware margins are partially diluted, its moat will actually grow stronger due to "platformization."