Most AI companies have to queue up to buy GPUs from Nvidia. Google holds an extra card: an AI chip it designs itself, the TPU. That makes it the least Nvidia-dependent of the big AI players. For the full picture of Google first, start with What kind of company is Google.
This piece answers three questions: where the TPU saves Google money, how deep the moat is, and which hurdles it still can’t clear.
The TPU is Google’s own AI chip
TPU is short for Tensor Processing Unit. It is a specialized chip (an ASIC) that Google built for its own AI workloads. It isn’t sold at retail. In the past, Google mainly used it in-house and rented it out through Google Cloud. Starting in Q2 2026, Google began shipping TPU systems directly into a few customers’ own data centers and recognized revenue from this for the first time. The CFO said most of this revenue will land in 2027.
The seventh generation, Ironwood, is now generally available to cloud customers. The eighth generation splits into TPU 8t for training and TPU 8i for inference, announced in April 2026. In September, Broadcom said it had shipped TPU 8i to Google. Chip shipments and Google Cloud availability to customers are separate milestones.
For how the TPU differs from the GPU, read our AI accelerator chip breakdown and the dedicated TPU piece.
Where making your own chips saves money
The most direct evidence comes from Google’s earnings: in 2025, it cut Gemini’s per-unit serving cost by about 78%, thanks to model optimization plus better efficiency and utilization of its own hardware.
The savings come in two layers. First, no middleman margin: Google designs its own chips, so it doesn’t buy GPUs at Nvidia’s prices. Second, a custom fit: the TPU is designed for Google’s own work, so it runs that work well. That is also why the TPU saves the most on Google’s huge, schedulable in-house jobs. It isn’t necessarily the cheapest for every kind of compute sold to others.
Behind the end-to-end story sits a long supply chain
“In-house” doesn’t mean Google does everything itself. Google leads the architecture and splits design work among several partners:
| Partner | Role | Source |
|---|---|---|
| Broadcom | Long-term agreement to develop and supply future TPU generations; rack component supply through 2031 at the longest | Official (Broadcom filing) |
| Marvell | Signed in July 2026 to develop custom chips for the TPU ecosystem, including inference accelerators; Google also received warrants that vest in stages subject to conditions | Official (Marvell filing) |
| MediaTek | TPU 8t (training chip) | Mentioned by Broadcom’s CEO on its September 2026 earnings call |
TSMC is the one that actually makes the chips. The TPU also needs large amounts of HBM (high-bandwidth memory) and advanced packaging (CoWoS). Like TSMC’s advanced process, these are bottlenecks the whole industry is fighting over, and all fall under U.S. export controls. Google has more control than rivals that can only buy GPUs, but it hasn’t left the global supply chain. For how the whole chain works, see The end-to-end AI hardware supply chain.
Even its rivals are renting its chips
The TPU’s strength shows most clearly in who else uses it:
- Anthropic: officially announced access to up to 1 million TPUs (more than 1 GW in 2026), plus about 3.5 GW of next-generation TPU compute through Broadcom starting in 2027. According to the IPO filing Reuters reviewed, Anthropic has committed to spend at least $111.1 billion with Google from April 2026 to July 2033, and must pay the difference if it uses less.
- Meta: according to The Information (as reported by Reuters), Meta signed a multibillion-dollar TPU rental deal in February 2026. Talks about buying TPUs for Meta’s own data centers are unconfirmed.
Google invests in Anthropic while also renting it compute. We unpack that triangle in Why Google invests in its rival Anthropic. That rivals trust the TPU to train their next models is itself an endorsement of its cost-performance.
Chips need power—where does it come from
More compute means more electricity, so energy has become another front for Google. One main line is nuclear:
- An agreement with Kairos Power plans up to 500 MW of small modular reactors (SMRs) between 2030 and 2035, with the first targeted to come online in 2030.
- In October 2026, Google signed a 20-year power purchase agreement with Constellation, adding 890 MW of new capacity by uprating existing nuclear plants. Google says it has now helped bring more than 1.5 GW of new nuclear capacity to the U.S.
Penchan’s summary
The TPU’s value isn’t being “the cheapest.” It is having “the most control.” While others scramble for Nvidia’s supply, Google has a chip line it can schedule itself.
But this moat has limits. When capacity runs short, Google still rents Nvidia GPUs from outside. According to CNBC and other outlets citing a SpaceX filing, starting in October 2026 it pays about $920 million a month to rent about 110,000 GPUs from SpaceX as bridge capacity. The chips themselves still depend on TSMC, HBM and advanced packaging. Google is closer to chip independence than anyone, but it isn’t there yet.
Further reading: What kind of company is Google, The end-to-end AI hardware supply chain, Why Google invests in its rival Anthropic.