Google Adds Two Specialized TPUs: a Government-Grade 8tx That Pairs With AMD CPUs, and a Dual-Chip 8i Max for Long Context

Google has added two more specialized variants to its tensor processing unit lineup, pushing the eighth-generation TPU family to four distinct chips. The announcement, confirmed by Bloomberg on Monday, is the clearest signal yet that the company intends to compete with Nvidia on breadth as well as price: instead of one general-purpose accelerator, Google is now carving the TPU line into versions tuned for specific regulatory and technical niches.

The first, the TPU 8tx, is built for confidential computing in government agencies and regulated industries such as finance and healthcare. Data remains encrypted while it is processed, so sensitive workloads never exist in plaintext inside the chip's memory. Unusually for Google's stack, the 8tx is designed to work alongside AMD CPUs rather than Google's own Arm-based processors — a cross-vendor pairing that suggests Google would rather win the accelerator slot than the whole server.

The second, the TPU 8i Max, packages two TPU dies into a single module, trading density for more compute and memory per socket. It is aimed at applications that need to hold enormous context: Google pointed to drug discovery, mathematical work and code generation as the target workloads. Both new variants are expected to reach customers in 2027.

"The future belongs to specialization," Amin Vahdat, who leads Google's chip business, said in the announcement. "The value to be gained from specialization remains immense."

The economics behind the push are substantial. Bloomberg Intelligence analyst Mandeep Singh estimates TPUs are roughly 30 percent cheaper than comparable Nvidia GPUs while delivering similar data throughput, and projects TPU-related sales of about $12 billion this year growing to more than $135 billion by 2028 — analyst forecasts, not company guidance. Apple, Meta, Oracle, Anthropic and the trading firm Jump Trading are already TPU customers, and OpenAI has said it plans to buy capacity as well. Jump Trading's CTO Alex Davies told Bloomberg the firm still leans on Nvidia but has deployed thousands of TPUs for workloads where they perform well.

Markets reacted before the opening bell: Nvidia shares fell about 6 percent in premarket trading on Monday, while the big cloud companies — including Google, Microsoft and Amazon — gained around 2 percent. Pre-market moves are volatile and may not hold at the open, but the direction captured the story: investors read specialized TPUs as a direct threat to Nvidia's roughly 90 percent share of AI accelerators.

The cadence is also notable. Google historically released one new TPU every two years; this year alone it has introduced two new chips and, within months, two more variants on the same eighth-generation base. Oracle said last month that its cloud customers would get access to Google TPUs for the first time next year, extending the distribution channel beyond Google Cloud.

The takeaway: the accelerator market is splitting. Google is betting that a cheaper, specialized chip for every workload — encrypted for regulators, doubled-up for long context, paired with whoever's CPU the customer prefers — is a better strategy than trying to out-engineer Nvidia on a single general-purpose part. For buyers, the arrival of a government-grade confidential TPU with AMD CPU support lowers one of the last practical barriers to non-Nvidia infrastructure in regulated sectors.