Google’s timing could hardly be better, or colder. Two new custom AI chip systems arrived days after Nvidia CEO Jensen Huang admitted to an affair with a woman suspected of being a Chinese spy. The real story isn’t the gossip, though. It’s that Google’s TPU business is now a serious commercial product aimed straight at Nvidia’s customers.
Google announced Monday two new variants of its eighth-generation Tensor Processing Units. The TPU 8tx is built for governments and regulated industries, using hardware-level security to protect data even while a model is processing it. The TPU 8i Max targets memory-hungry work, such as dozens of AI agents working across a whole codebase. It joins two existing TPU 8i chips in one package, doubling compute and memory bandwidth. Bloomberg reports both should be available in 2027.
Amin Vahdat, Google’s chief technologist for AI and infrastructure, said the designs came partly from customer feedback. Computing is moving from human scale to agent scale, he wrote, and regulated customers want specialized encryption. That’s a smart read of the market. Nvidia’s GPUs are the default, but agent workloads chew through context windows, and confidential computing is a real sticking point for banks, defense, and health care.
Nvidia’s board has launched an independent investigation, and the company says it has no reason to believe national security or company secrets were compromised. Its stock fell about 5.6% Monday. I’d treat that as noise for the long term. Nvidia’s moat is CUDA and its software ecosystem, not its CEO’s reputation.
The TPU numbers are the better signal. Google started building these chips more than a decade ago for voice recognition and translation. They now train and run Gemini, power Google Cloud, and are sold to customers for their own data centers. Alphabet says TPU sales contributed meaningfully to Cloud revenue, which more than doubled year over year to $30.7 billion in the quarter ended Sept. 30. Google also says it has nearly 100 TPU customers, including Apple, Meta, and Oracle, per Bloomberg.
That customer list matters. Meta and Apple are big enough to buy Nvidia at scale, and they’re hedging anyway. Amazon’s Trainium and Microsoft’s Maia pursue the same logic, but Google is furthest along in selling outside its own walls.
The backdrop is rough. A memory shortage continues, and last week’s midterms produced a Democratic Congress and voter-backed limits on data center development. AI stocks slid on fears of slower buildouts. Chips that squeeze more out of each rack could look better in that world.
Don’t expect Nvidia’s share to crater. The 2027 availability means these are a roadmap statement, not a near-term threat. But every regulated customer that qualifies a TPU is one that now has leverage in Nvidia pricing talks. That’s the actual impact.
Elsewhere at Alphabet, Wing said it will invest up to $125 million in a Fort Worth factory making tens of thousands of drones a year. It has completed more than 1 million U.S. deliveries this year and plans to reach nearly 20 markets with Walmart by 2027.
