中

Google TPUv8 Officially Launches: Meta and Anthropic Sign Big Deals, Inference Costs Drop 40%

Cloud Computing, Google, AI Chip, TPU —

Honestly, I did not see this coming so soon.

On April 22, Google Cloud Next 2026 kicked off in Las Vegas. Google officially unveiled the TPUv8 series chips—and I mean officially launched, not another “coming soon” slide. According to data shared at the event, TPUv8 delivers 2-3x performance improvement at the same power consumption, directly cutting inference costs by 40%.

What surprised me even more: Meta and Anthropic have already become the first major customers. These are not LOIs—they are real contracts. What does this mean? It means Google TPU has passed validation from the world top AI companies. It is no longer just an internal Google system.

Broadcom and MediaTek Dual-Chip Strategy

There is an interesting design choice in the TPUv8 series: the dual-chip strategy. According to supply chain sources, TPUv8 is actually manufactured by two companies—Broadcom and MediaTek—with a clear division of labor:

  • Broadcom version: Focused on training scenarios, for companies needing large-scale pretraining
  • MediaTek version: Focused on inference scenarios, for companies needing cost-effective deployment

This makes me think of a phrase—specialization. Previously, everyone tried to build one chip to do everything, but the reality is that training and inference have wildly different requirements. Google seems to have figured this out this time.

Nvidia Faces Real Competition

To be honest, every time someone said “Google is challenging Nvidia” over the past few years, I thought it was hot air. But this time feels different.

TPUv8 key advantage is not explosive performance—it is cost performance. A 40% reduction in inference costs translates to real money for large cloud providers. The electricity savings alone could fund another batch of chips.

Plus, Google has its own cloud service, which means TPUv8 is not just selling chips—it is selling a complete solution package. For companies that do not want to be “held hostage” by Nvidia but need large-scale computing power, this is extremely attractive.

My Take

This move from Google is serious. But Nvidia is not sleeping either—Blackwell ecosystem is incredibly mature, and CUDA moat is not something you can fill in a day.

That said, competition is good for the entire AI industry. Monopoly ultimately hurts users and developers.

Do you think Google can actually dent Nvidia dominance this time?

Related Posts