Google to Release New AI Chips, Challenging Nvidia | Bloomberg Tech 4/20/2026
Why It Matters
The new inference‑focused TPU gives Google a hardware foothold against Nvidia, potentially reshaping AI‑chip pricing and accelerating the rollout of generative AI services.
Key Takeaways
- •Google to unveil next‑gen TPU inference chip at Las Vegas event.
- •New chip separates training and inference, targeting growing inference demand.
- •Major customers like Anthropic, Meta, and Celer signed multi‑year TPU deals.
- •Supply constraints with Marvell highlight potential bottlenecks for scaling.
- •Chip launch could intensify competition with Nvidia’s latest inference GPUs.
Summary
Google is set to announce a new generation of Tensor Processing Units (TPUs) at its Las Vegas AI infrastructure event this week, marking the first dedicated inference‑only accelerator from the company.
The chip splits training and inference workloads, a move analysts say reflects the surge in demand for low‑latency model serving. Google has already secured multi‑year agreements with frontier AI labs such as Anthropic, Meta and Celer, and is working with Marvell to source the silicon, though supply‑chain constraints remain a concern.
Bloomberg’s AI infrastructure reporter quoted Google insiders that the new TPU will be tuned using data from Google’s own Gemini models, aiming for higher efficiency than competing Nvidia GPUs. Nvidia, meanwhile, has rolled out a fast inference chip for its Grok model, underscoring the intensifying rivalry.
If successful, the launch could erode Nvidia’s dominance in the inference market, boost Google’s hardware revenue, and give investors a clearer view of the $650 billion AI‑infrastructure spend timeline. The chip’s availability will also influence pricing and capacity decisions for hyperscalers and enterprise AI adopters.
Comments
Want to join the conversation?
Loading comments...