Google Adds TPU 8tx and 8i Max, Betting on Specialized AI Chips

Security and memory

The TPU 8tx targets governments and regulated industries that want to run AI models and agents on confidential data. Data is usually encrypted in storage and in transit, but must be unlocked in memory while being processed. Google says the 8tx builds protection into the hardware, keeping data and model weights isolated even during use. It uses the same silicon as the 8t, is paired with fifth-generation AMD EPYC processors, and offers up to twice lower latency than the 8t with almost no performance penalty, according to the company.

The TPU 8i Max tackles what Google calls the "memory wall": when many AI agents work across a whole codebase or hours of video, memory bandwidth becomes the bottleneck. Google places two TPU 8i chips in one package, which it says doubles compute, on-chip SRAM and memory bandwidth versus the standard 8i, and fits twice the silicon in the same rack space. Vahdat pointed to long-context uses such as drug discovery, complex math and coding.

Reconfigured, not brand new

Vahdat said both variants are reconfigurations of existing silicon, adapted to customer requests. A new chip takes years to design and build, while the system around it, including networking, memory and packaging, can be adjusted in months. His thesis is that "the future is going to revolve around specialization".

The market is growing quickly. Bloomberg Intelligence expects TPU sales, including related hardware and software, to top $135 billion by 2028, up from an estimated $12 billion this year. Customers include Apple, Meta and Oracle, and Anthropic has the option to buy up to $150 billion in TPUs between 2028 and 2031, according to a regulatory filing. OpenAI is also reported to plan purchases.

Obstacles remain. Google has said all its available TPUs are already committed to customers, memory and manufacturing capacity are tight, and the company is reportedly widening its supplier base beyond Broadcom, including a deal with Marvell. Rivals are emerging too, among them startup Tenspresso, founded by two former TPU engineers, though its chip is at least a year from a finished design.