
Google Just Split Its AI Chip Into Two and Aimed Both Halves at Nvidia
TPU 8t for training, TPU 8i for inference. 80 percent better performance per dollar. A million chips per cluster. Google is no longer waiting for Nvidia to fail.
The AI Post newsroom — delivering AI news at the speed of intelligence.
For ten years, Google's Tensor Processing Unit was the chip everybody talked about and nobody actually competed with. It powered Search, Gmail, YouTube recommendations, and the Gemini family. It was great at Google's stuff. Nvidia owned everything else.
Today at Google Cloud Next, that polite arrangement ended.
Google announced its eighth-generation TPU, and for the first time it is not one chip. It is two. TPU 8t for training. TPU 8i for inference. Both built on the lesson Nvidia has spent five years denying: training and inference are different workloads, and pretending one chip can do both is leaving 80 percent of price-performance on the table.
TPU 8i is the one that matters for the Nvidia comparison. Google claims up to 80 percent better inference performance per dollar versus its previous Ironwood generation. 1,152 chips per pod. Three times the on-chip SRAM, which means more of the KV cache stays resident during long-context inference instead of getting paged out to slower memory. That is the bottleneck on every modern reasoning workload, and Google just attacked it head-on.
TPU 8t, the training chip, keeps the 3D Torus topology and SparseCore for embedding lookups. The headline number Google is leaning into is scale: networks of up to one million TPUs per cluster. Nvidia's largest planned NVL576 systems top out at 576 GPUs. Hyperscaler-scale training is now a Google-only configuration on paper.
Tom's Hardware called the split "two chips for two crucial tasks at incredible scale." Pune Mirror called it a "bold leap against Nvidia." The trade press is, for once, not exaggerating.
The reason Google is doing this now is not technical. It is political.
Friday morning Google announced it is investing up to $40 billion in Anthropic. Anthropic now has $5 billion of fresh Amazon money on top, plus a CoreWeave deal, plus an October IPO target at $800 billion. The deal terms include Google Cloud providing 5 gigawatts of capacity over five years. Anthropic runs on TPU. Anthropic's competitors run on Nvidia. The implicit deal is: every dollar Anthropic spends on Google Cloud silicon is a dollar that did not flow to Jensen Huang.
Now extend that logic.
OpenAI just amended its Microsoft contract this morning. AWS Andy Jassy says OpenAI models are coming to AWS soon. Microsoft is building its own silicon. OpenAI is reportedly working with Qualcomm and MediaTek on smartphone processors. Apple just elevated Johny Srouji to Chief Hardware Officer to double down on in-house chips for everything. The entire industry is racing away from Nvidia exclusivity.
Nvidia still has 81 percent of the AI chip market. Its stock hit a record high Friday. None of that is in question.
What is in question is whether the next decade looks like the last one. Hyperscalers are not paying $30,000 a GPU because they enjoy it. They are paying because there is no credible second source. Google's pitch with TPU 8 is that there is now a credible second source, the second source ships at scale, and the price-performance gap on inference is 80 percent. That is not a margin Nvidia can defend with software lock-in alone.
The other shoe drops with VAST Data, which closed a Series F today at a $30 billion valuation, more than triple its 2023 round. VAST builds the storage layer that feeds these chips. The AI infrastructure market is not just chips. It is chips, networking, storage, software, power, and cooling. Whoever owns the most layers at a defensible price wins. Google now owns four of those six. Nvidia owns one and a half.
Read this together with the Anthropic investment, the OpenAI-Microsoft amendment, and the Apple silicon promotion, and the picture is clear. The single-vendor era of AI compute is ending in real time. Nobody who matters is building their next product on a single supplier anymore.
If you own Nvidia stock, none of this hits the numbers in 2026. The order book is full and the queue is two years deep. If you are looking past 2027, every line in the price-performance comparison just got harder to defend.
Sources: Google Cloud official blog, Tom's Hardware, Pune Mirror, IndexBox, Crypto Briefing, Shashi.co analysis, Frontier Enterprise (VAST Data round), Wikipedia (VAST Data Series F confirmation).