Google plans to deploy its next-generation Frozen V2 AI chips by 2028, a move the company says will deliver six to ten times the efficiency of its current TPUs. The timeline and performance targets escalate the already intense competition in AI hardware.
Why the efficiency jump matters
The 6-10x efficiency gain over today's Tensor Processing Units isn't a small step. It's a leap that could reshape how Google runs its AI workloads, from training massive models to serving real-time predictions. Current TPUs already power much of Google's AI infrastructure, but the company is betting that a new chip design can cut energy use and speed up operations dramatically.
Google hasn't disclosed technical details about Frozen V2's architecture or how it achieves those gains. The company described the deployment as part of an escalating AI hardware arms race, suggesting the chip is a direct response to pressure from competitors racing to build faster, more efficient processors.
What the 2028 deadline means
Setting a target four years out gives Google time to design, test, and manufacture the chips at scale. It also signals that the company expects the current pace of AI hardware improvement to continue — and that it's willing to commit to a long-term roadmap. The 2028 date is ambitious given the complexity of bringing a new chip from concept to production, but Google has a track record of iterating on its TPU line.
The company's previous TPU generations have rolled out roughly every two to three years. Frozen V2 would mark a longer gap, which could reflect the difficulty of achieving such a large efficiency jump or a strategic decision to wait for a manufacturing process that can deliver it.
The broader hardware arms race
Google's announcement comes as demand for AI computing power surges across the industry. Every major tech company is investing in custom chips, and startups are racing to build alternatives to the dominant GPU architecture. The phrase “AI hardware arms race” captures the stakes: whoever builds the most efficient chip can train bigger models faster and at lower cost.
Frozen V2's efficiency gains could give Google an edge in running its own AI services, from search to cloud offerings. But the company also sells access to its TPUs through Google Cloud, so the chip could become a product for external customers as well. Google hasn't said whether Frozen V2 will be available to third parties.
The 2028 target leaves room for competitors to respond. Several companies have announced AI chips with similar efficiency claims, though none have shipped at scale yet. The next few years will test whether Google can deliver on its promise — and whether the arms race accelerates further.
