Key Points

  • Google unveiled two new TPU systems on Monday: the TPU 8tx for confidential computing and the TPU 8i Max for long-context AI work.
  • Both are new configurations of existing eighth-generation silicon, not new chips, and are due in 2027.
  • The launch takes Google’s eighth-generation TPU lineup to four systems in a single year, a first for the company.
  • The 8tx works alongside AMD processors and keeps data encrypted while it is being processed, aimed at governments and regulated industries.
  • The 8i Max combines a pair of TPU 8i chips in one package, giving it twice the processing power and memory throughput for work such as coding, maths and drug discovery.
  • Google says nearly 100 outside customers now use its TPUs, and analysts forecast its chip sales could pass $135 billion by 2028.

Google on Monday unveiled two new AI chip systems, the TPU 8tx and TPU 8i Max, completing a four-strong eighth-generation lineup aimed squarely at Nvidia’s grip on the market for AI hardware.

What happened

Google announced the two systems in a blog post on Monday, 9 November, written by Amin Vahdat, its chief technologist for AI infrastructure. Bloomberg News reported the launch the same day. The systems will reach customers in 2027, according to Vahdat.

The two additions are not new chips. They are fresh configurations of the silicon Google introduced in April, when it launched the TPU 8t for large-scale training and the TPU 8i for real-time reasoning at its Cloud Next event. Vahdat told Bloomberg that the future of AI hardware lies in specialisation, with each system tuned to a different job.

The TPU 8tx is built for secure, confidential inference. It runs on the same silicon as the TPU 8t and is paired with fifth-generation AMD EPYC processor hosts. Encryption normally protects data at rest and in transit, yet the data must be decrypted before a processor can work on it — the gap the 8tx is built to close. It locks data down in the hardware, so private information and model weights remain shielded throughout processing. Google says it cuts latency by up to half compared with the TPU 8t, with almost no loss of performance, and it can also handle small-scale training. That means a bank, a hospital or a government agency could fine-tune a model on its own private data and run AI agents on it without moving that data between separate systems.

The TPU 8i Max attacks a different problem: memory. Memory bandwidth becomes the choke point once scores of AI agents chew through an entire codebase or hours of video in parallel. The fix is a dual-chip package holding two TPU 8i dies side by side. That gives the Max twice the processing power, twice the fast on-chip memory and twice the memory throughput of one 8i chip, so operators squeeze much more compute from the same rack space.

Google also used the announcement to open its chip platform further to developers. Last month it released TorchTPU as open source, letting ordinary PyTorch code and Hugging Face models execute on TPUs with no changes, and it released new tools that plug into coding assistants such as Claude Code and Cursor. Another new tool, Agent Compute Engine, freezes agent sessions that go idle and moves their working memory off the chip within milliseconds, trimming the cost of agents that spend long stretches waiting.

Why Google’s new TPU systems matter

Nvidia’s graphics processors dominate AI workloads around the world, and Google’s home-grown chips, once reserved for its own use, have become the most serious challenger to that position. The scale of the shift is striking: Bloomberg Intelligence estimates Google’s TPU-related sales will grow from about $12 billion this year to more than $135 billion by 2028, and its analysts say the chips cost roughly 30 percent less than Nvidia’s graphics processors while delivering competitive performance. Both figures are analyst estimates, not confirmed results.

The customer list tells the same story. Google says nearly 100 outside organisations now run workloads on its TPUs, including frontier AI labs, trading firms, retailers and universities. Bloomberg’s reporting names Apple, Meta, Oracle and OpenAI among them, alongside Anthropic, which holds an option to buy up to $150 billion of the chips between 2028 and 2031, according to a regulatory filing. Oracle signed a deal last month to offer TPUs to its own cloud customers from next year.

Demand is already outrunning supply. Bloomberg reported that all of Google’s available TPUs are already spoken for by customers, and that the company is exploring supply deals beyond its long-time manufacturing partner Broadcom, including with Marvell Technology. The two new systems are part of that answer: by repackaging existing silicon into workload-specific versions, Google can serve new customers in months rather than waiting years for a new chip design.

The 8tx also opens a market Nvidia’s general-purpose chips have not fully reached. Governments and regulated industries have held back from cloud AI because their most sensitive data would be exposed while it was processed. Hardware-level encryption removes that obstacle, and it arrives as AI agents move from experiments into production inside banks, hospitals and public agencies.

Investors welcomed the news. Alphabet shares rose 2.4 percent on Monday, according to market reports. LG made its EXAONE 2.5 AI model API free for developers on Monday, another sign of how fast the AI hardware and software race is moving. Anthropic, one of Google’s biggest TPU customers, has struck a $3.4 billion deal for gas generators to power its own AI expansion.

What happens next

Neither new system has a price or a firm release date yet. Google has not published detailed specifications for the 8tx either, including its core count or memory size. The company broadcast the launch at 9 a.m. Pacific Time on Monday, and the first test will be whether cloud customers sign up for the two systems when they arrive in 2027.

The wider contest is only heating up. Nvidia has launched an independent investigation after reports involving its chief executive Jensen Huang, while Google presses ahead with its own silicon. If the 8tx and 8i Max find buyers quickly, Google will have turned a chip business that was once an internal tool into a genuine alternative for the customers now queuing for Nvidia hardware — and the next year will show whether that challenge can hold.

Sources: Google, Reuters, Bloomberg, The Circuitry. 9 November 2026.