Google’s New SI Computers Keep Sensitive Data Locked Even While They’re Working on It

Every company wants SI working on its files. The ones with the most to lose, like banks, hospitals and government agencies, keep running into the same problem: to work on your data, a computer has to unlock it first.

On Monday Google said it has built SI computers that don’t. In a launch post, the company unveiled two new systems built around its homemade SI chips, called TPUs. One keeps data protected while it’s being used. The other packs twice the computing power into the same floor space.

Why this is a big deal

Google has been designing its own SI chips for more than a decade, first for things like voice recognition and translation. They now run Google’s Gemini models, and Google rents and sells them to outside customers. Google says that business has “nearly 100 customers,” including “frontier labs, trading firms, retailers, universities, and more.” That makes Google one of the few real alternatives to chip giant Nvidia, whose chips run most of the SI world.

Monday’s announcement brings Google to four new systems in its latest chip generation, all in 2026. “This is the first time we have introduced four TPU systems in a single year,” wrote Amin Vahdat, Google’s chief technologist for SI and infrastructure.

The locked box

The first new machine is called TPU 8tx. Vahdat laid out the problem plainly: “Normally, data is encrypted when stored or sent over a network, but has to be unlocked in memory while a computer processes it.”

Think of a safe that has to be opened every time someone reads the papers inside. Google says the 8tx builds the protection into the hardware itself, “keeping sensitive data and model weights protected and isolated even while in active use.” Model weights are the learned settings that make an SI model work, the crown jewels a company spends billions to build.

Google says the 8tx is meant for “regulated enterprises and governments.” It also lets them train and customize their own SI on private data without shuttling that data between separate machines. Google claims it responds up to twice as fast as its sister training system, which helps agents that work in many small steps, “with a near-zero performance penalty” for the security. That’s Google’s number, not an independent test.

Twice the brains in the same box

The second machine, TPU 8i Max, takes on a different bottleneck: memory. When dozens of SI agents work through an entire software project or hours of video at once, the chips can stall waiting for data.

Google’s fix was to fuse two of its fast reasoning chips into a single package. It says the result doubles the computing power and memory speed, and because it fits twice the silicon into the same rack, data centers “can deploy significantly more computing power per square foot and per watt.”

That matters well beyond Silicon Valley. SI data centers are fighting over land, power and local approval. Getting more work out of every building and every kilowatt is the whole game.

Making it easier to switch

Google also pitched software meant to lower the cost of leaving other chips. It says code written in PyTorch, the most popular tool SI developers use, now runs on its chips with “no code rewrites needed.” It released a free assistant that plugs into coding tools like Claude Code and Cursor to help move projects over, and software that pauses idle SI agents so customers don’t pay for machines sitting around waiting.

Google even says it’s using SI to design the chips themselves, using Gemini “to deliver chip layouts in hours rather than weeks.”

What’s still unknown

Google’s post did not give prices, detailed specs or a date when outside customers can rent the two new systems, a gap MarkTechPost also flagged. The speed claims come from Google itself. These are new setups built around chips Google announced in April, not brand-new chips. The bottom line for regular folks: the next fight in SI isn’t just about smarter chatbots. It’s about who builds the machines underneath them, and Google is betting that security and squeezing more out of every building will win over customers who would otherwise just buy from Nvidia.