Google already has its own TPUs, but it’s looking to create a whole new kind of AI chip.
According to a report from The Information, the company is developing a new server chip, informally called “Frozen v2,” that would bake the blueprint of its Gemini AI model directly into the silicon itself. The goal is to serve Gemini to users far more efficiently than is currently possible, according to two people with direct knowledge of the project.
The idea behind Frozen v2 is a fairly radical departure from how Google’s chips have worked so far. Rather than building a general-purpose accelerator that can run whichever model is thrown at it, Google wants to permanently etch part of Gemini’s architecture into the chip itself — hence the “Frozen” name. Engineers are still working out what the chip’s core features will look like and how its components will fit together, and the current plan is to have it ready for deployment as early as 2028.
Google’s motivation here is capacity, not novelty. The company has been dealing with a serious AI compute crunch, one that has reportedly caused friction internally and even forced Google Cloud to turn away business from outside customers because it simply doesn’t have enough chips to go around. Sundar Pichai has said as much publicly, describing the company as “compute constrained in the near term” even as cloud revenue growth accelerates. Frozen v2 is Google’s attempt to squeeze meaningfully more performance out of the compute it already has, rather than simply buying or building more of the same.
Employees working on the project are projecting big gains — Frozen v2 could be 6 to 10 times more efficient than Google’s newest TPUs, measured in tokens served per unit of power. That’s a significant leap for a company that has already spent a decade refining its TPU line, and it points to just how much slack there still is to capture by narrowing a chip’s purpose.
The logic is straightforward once you look at what a general-purpose chip actually has to do. TPUs, like Nvidia’s GPUs, are designed to run many different AI models, which means the chip has to make a lot of on-the-fly decisions every time it processes a query, depending on which model it’s serving and how that model is structured. Frozen v2 sidesteps a chunk of that overhead by having some of Gemini’s decision-making already baked into the hardware. Fewer decisions to make in real time means fewer steps and less data movement, which in turn means faster responses and lower power draw.
There’s an obvious trade-off buried in that design choice, though. A chip this specialized is also a bet — specifically, a bet that Google will keep building Gemini on roughly the same underlying architecture that Frozen v2 is designed around. If Google’s model architecture changes significantly in future Gemini generations, the chip’s usefulness could be limited to whatever generation of Gemini it was built for. Google would still be able to update the chip with new model weights, according to the report, but a deeper architectural shift is a different story.
Google isn’t positioning Frozen v2 as a TPU replacement. The two are meant to coexist, with TPUs continuing to handle the general workhorse duties while Frozen v2 becomes a purpose-built option specifically for accelerating Gemini inference. It’s a bit like Google building a race car for one specific track, alongside its existing fleet of vehicles that can drive on any road.
Google Isn’t The Only One Betting On Model-Specific Silicon
Frozen v2 is a striking validation of an idea that a much smaller startup has been pushing publicly for a few months now. Toronto-based Taalas emerged from stealth earlier this year with what it calls “Hardcore Models” — AI architectures etched directly into custom silicon rather than run as software on general-purpose hardware. Taalas built a hard-wired implementation of Meta’s Llama 3.1 8B model and released it through a demo chatbot called ChatJimmy, which was reportedly hitting 17,000 tokens per second per user, compared to roughly 1,800 tokens per second on typical GPU setups.
The core pitch from Taalas is essentially the same one now surfacing at Google: unify computation and storage on a single chip instead of splitting them across the memory-compute divide that defines most modern AI hardware, and optimize the resulting silicon for one specific model rather than for general flexibility. Taalas frames it as “The Model is The Computer.” Seeing Google — a company with its own decade-plus head start in custom silicon and a TPU fleet that reportedly runs at close to full utilization — pursue a similar architecture suggests this approach is being taken seriously well beyond one startup’s marketing pitch.
The difference, naturally, is scale and stakes. Taalas is building proof-of-concept chips around an 8-billion-parameter open model to demonstrate what’s possible. Google would be doing this for Gemini, the model that underpins its search, cloud, and consumer AI products, running at a scale where even single-digit efficiency gains translate into enormous savings on power and infrastructure. If Frozen v2 delivers anywhere close to the 6-to-10x efficiency gain being projected internally, it would give Google a meaningful edge in serving its own models cheaply at a moment when compute has become the single biggest bottleneck across the entire AI industry.