Google Adds Two Specialized AI Chip Systems to Press Its Challenge to Nvidia

Google on Monday added two systems to its tensor processing unit line, its in-house AI chips, widening its direct challenge to Nvidia.

The new systems are the TPU 8tx and the TPU 8i Max. They join the TPU 8t and the TPU 8i in the eighth-generation lineup.

Google said this is the first time it has introduced four TPU systems in a single year. The first two arrived in April at Google Cloud Next.

Neither addition is a new chip. Google built each as a new system around silicon it already had.

Amin Vahdat told Bloomberg the variants are reconfigured versions. He said they answer workloads customers have been asking for.

A system that keeps data protected while it runs

TPU 8tx is built for secure and confidential inference. It is aimed at governments and companies in regulated industries.

Data is normally unlocked in memory while a computer processes it. TPU 8tx keeps sensitive data and model weights protected during that step.

The system uses the same silicon as TPU 8t. It pairs that silicon with 5th Generation AMD EPYC CPU hosts.

Google said it delivers up to 2x lower latency than TPU 8t. It put the penalty for the added protection at near zero.

The same system can also fine-tune models on private data. Teams need not move sensitive data between separate clusters to do it.

Two chips in one package, aimed at the memory wall

TPU 8i Max attacks a different bottleneck: memory. When many AI agents work across a whole codebase at once, memory bandwidth runs out first.

Google combined two TPU 8i chips onto a single package. It doubles TPU 8i's compute and its on-chip SRAM.

It also doubles the bandwidth of its high-bandwidth memory, or HBM. Twice the silicon fits within the existing rack footprint.

Google said that means more computing power per square foot. It also means more power per watt in the same space.

The target work needs very long context windows. Bloomberg linked the system to drug discovery, mathematics and coding work.

Specialization is the strategy, in Google's own telling

"The future is going to be about specialization," Vahdat said in a Bloomberg interview. He is Google's chief technologist for AI infrastructure.

His argument is that specialized versions deliver more value. One chip built for the whole market is the alternative he rejected.

Both new systems are due in 2027, Vahdat told Bloomberg. Google's own announcement listed no availability date and no price.

The chip business is already moving Cloud's numbers

Google Cloud revenue rose 103% to $30.71 billion in the third quarter. Alphabet reported the quarter, ended September 30, on October 28.

TPU system sales were a meaningful contributor to that growth, Alphabet said. Reuters reported the same point in its account of Monday's launch.

Google said TPUs now serve nearly 100 customers. That count includes frontier labs, trading firms, retailers and universities.

Bloomberg named Apple, Meta Platforms and Oracle among TPU customers. The chips were once reserved for Google's internal use, Reuters reported.

Bloomberg Intelligence estimates TPU sales will top $135 billion by 2028. It put sales this year at an estimated $12 billion.

Those figures include related hardware and software. They exclude capacity rented through Google Cloud.

Alphabet shares were up about 3% in premarket trading on Monday. The reading is as of 8:11 a.m. ET, at $293.26 a share.

That is a premarket indication, not a settled reaction. Monday's closing price is the first settled verdict.

Through Friday's close, the stock had fallen 9% this year, Reuters reported. Friday's close itself was $284.78 a share.

What remains unknown

What remains unknown starts with price. Google published no price for either new system.

Exact 2027 availability is also unknown. Vahdat gave the year to Bloomberg, not a date.

No independent benchmark of either system has been published. Every performance claim so far is Google's own.

Whether Nvidia answers with pricing or new inference products is unknown. Demand for its processors still exceeds supply, Bloomberg reported.

The 2027 launch is the first real test

What to watch: the exact 2027 launch dates and prices Google has not yet given. Those terms decide how directly the systems compete with Nvidia.

Customer orders are the second test. Named buyers beyond Apple, Meta and Oracle would show the sales push is broadening.

Independent benchmarks are the third. They will test Google's latency and memory claims against Nvidia's current products.