Google launches two new TPUs for the agentic era
Google announced last Monday (November 9) two new systems based on the eighth generation of its Tensor Processing Units (TPUs): the TPU 8tx, focused on confidential inference and training, and the TPU 8i Max, designed for massive context windows of up to millions of tokens. Both are scheduled for release in the first half of 2027 and represent the company's clearest strategy yet to position TPUs as a viable alternative to Nvidia accelerators in real production workloads — particularly those involving autonomous agents.
Two systems, not two chips
A point that Google's communications emphasize is the distinction between a "chip" and a "system." Designing a new silicon chip means reconfiguring billions of transistors on a wafer — a two- to three-year process with TSMC's foundry. Designing a system, on the other hand, is about organizing around an existing chip — or two — the memory, the host processor, and the software layer that makes everything work.
The TPU 8tx contains no new chip. It takes Google's TPU 8t training processor (launched in April) and pairs it with AMD's fifth-generation EPYC processors as the host. According to Google, the combination "builds security directly into the hardware layer," keeping sensitive data and model weights protected and isolated even during active use. The goal is to serve regulated sectors — healthcare, finance, government — where transferring proprietary data across devices is not an option and where autonomous agents need to process confidential information without exposing it to the network.
The TPU 8i Max, meanwhile, basically packages two of the TPU 8i inference chips (also from April) into a single module, doubling memory capacity, processing power, and bandwidth. The result is a chip capable of natively handling context windows of millions of tokens — something essential for autonomous agents that need to maintain long-term memory, track multiple simultaneous interactions, and reason over large datasets without losing track.
Both will be available via Google Cloud and, crucially, for deployment directly into customer data centers — meaning Google is no longer just selling compute cycles, but delivering physical supercomputers on demand.
The race for custom silicon
The eighth-generation TPUs already compete directly with Nvidia in the dedicated AI accelerator arena. The TPU 8t promises up to 2.7× better training performance than the previous generation, while the TPU 8i delivers 80% better inference price-performance. The 8tx and 8i Max don't dramatically improve these raw numbers — the innovation is in the use cases.
Google's strategy reflects a broader shift in the AI ecosystem: the race for custom silicon is no longer just about raw speed. Microsoft, Amazon, OpenAI, and Anthropic are all developing their own accelerator lines to reduce dependence on Nvidia. In Google's case, TPUs are already sold externally as part of Google Cloud — and the company claims nearly 100 customers, including frontier AI labs, trading firms, universities, and retailers.
The arrival of autonomous agents in production environments has accelerated a specific demand that traditional benchmarks don't capture: secure inference for sensitive data and the ability to process massive contexts without breaking. It's the same pattern we saw with the rise of LLMs — first the community asked for bigger models, then for cheaper and more efficient ones, and now for ones that are secure and contextually complete.
What to watch
The 2027 availability date is a relevant detail. The AI chip market is moving so fast that, by launching systems based on existing silicon, Google gains time to refine hardware and software integration before delivering products that could already be considered obsolete. It's a calculated risk: instead of betting on a completely new chip that might take two years to come off the drawing board, the company delivers a ready-to-use solution in less than six months.
Google has not disclosed prices — neither the official blog nor the press release mentioned figures. This is consistent with the company's practice for niche TPUs, which often have case-by-case enterprise pricing. The absence of pricing, however, raises a question: is Google using the 8tx and 8i Max to test the market before defining its monetization strategy?
One thing to watch is whether these systems will attract customers from Nvidia's ecosystem or whether regulated-sector buyers will see them as a complementary option rather than a replacement. The agentic era demands new kinds of workloads, and the winners will be the platforms that address them natively — not just those with the highest FLOPS.
Sources: TechCrunch, Vector, Google Cloud Blog
✓ Independent sources cross-checked and verified before publishing
