Google Challenges Nvidia With Two New AI Chips

Building high-end artificial intelligence has almost always meant paying whatever Nvidia charges for its hardware. Google is working to break that dependency. The company just expanded its custom silicon family with two purpose-built processors: the TPU 8tx and the TPU 8i Max.

Rather than offering generic workhorse chips, Google is focusing on two specific pain points: handling strictly regulated private data and keeping up with massive, memory-heavy AI prompts.

Locking Down Data with the TPU 8tx

Banks, hospitals, and security-focused platforms often avoid public clouds because they cannot risk exposing sensitive customer data or proprietary code to shared hardware.

The TPU 8tx solves this through confidential computing. Paired with AMD EPYC processors, it uses hardware-level encryption to keep information scrambled even while it is actively being processed in memory. Because the data is never exposed in plaintext during computation, companies can run sensitive AI models without worrying about rogue access or memory leaks.

Cracking the Memory Bottleneck with the TPU 8i Max

Modern AI agents and reasoning tools often need to process hundreds of pages, dense codebases, or complex documents in a single prompt. Doing that quickly takes enormous memory bandwidth, which usually creates severe bottlenecks for standard chips.

Google tackled this with a clever hardware trick in the TPU 8i Max: it combined two chips onto a single module. This design doubles memory bandwidth and compute density without taking up extra space on the server rack. As a result, long-context models run noticeably faster without demanding more physical room in the data center.

Why Custom Chips Matter

Relying entirely on outside vendors leaves tech platforms exposed to price surges and supply shortages. By designing its own silicon, Google bypasses third-party markups, trims operational expenses by roughly 30%, and optimizes the hardware specifically for the software running on it.

Nvidia still dominates the AI hardware market, but Google’s custom strategy shows where the industry is heading. Winning today isn't just about packing raw, brute-force power into a server; it's about deploying chips built directly for the specific job at hand.