$180 billion for one answer.
Over the past year, Alphabet’s capital expenditure jumped from $91.4 billion to $180–190 billion — nearly doubled — and management says 2027 will be higher still. Wall Street keeps asking: where is the return on all this money?
Part of Google’s answer is a chip with the internal codename Frozen v2. It runs no training and serves no external customers. Its sole mission: make the Gemini models run faster and more efficiently. Per The Information, Frozen v2 targets 6 to 10 times the tokens-per-watt of Google’s current newest AI chip — and the day the report landed, Alphabet shares rose about 3.3%.
But the significance reaches well past that one number.
It Is Not a Next-Generation TPU
This is the most-misunderstood point about Frozen v2. It is not a TPU replacement — it is a parallel chip line with a completely different job:
-
TPU: the general AI accelerator — matrix compute, training, inference, reinforcement learning, any model at any scale, rented through Google Cloud and soon delivered as hardware to customers. The eighth generation (April 2026) explicitly split into 8t (training) and 8i (inference) architectures, with native PyTorch support.
-
Frozen: a model-specific chip — Gemini’s structure, parameters, and particular computation patterns mapped into hardware. Generality sacrificed for extreme efficiency.
Google’s chip strategy is crystallizing into a clear dual-track system: TPU expands infrastructure generality and commercial reach; Frozen drives the per-unit running cost of the internal flagship models to the floor. From GPU to TPU to Frozen, AI silicon specialization is being pushed to unprecedented depth.
Why Model-Level Specialization, Why Now
The previous era’s chip competition measured floating-point throughput, parameter counts, interconnect bandwidth, cluster size. But large models have entered mass-inference mode, where single-chip peak compute no longer decides commercial efficiency. The realistic questions: within the same power and server budget, how many user requests? How many useful tokens? How fast the response?
Agent scenarios make this acute — one user request can trigger planning, search, tool calls, code execution, result verification, and multiple reflection rounds. What looks like one task generates token consumption far beyond a chatbot’s. Chip efficiency in data movement, memory access, and low-batch inference now moves product gross margins directly. The eighth-generation TPU already reflects the shift: TPU 8i adds 288GB of high-bandwidth memory and 384MB of on-chip SRAM specifically for low-latency inference. Frozen v2 pushes the division of labor down to the model layer — freezing parts of Gemini’s computation patterns into silicon.
It is feasible because Gemini already runs inside nearly every Google core product — Search, Ads, Workspace, Google Cloud. A small energy and latency saving per call, accumulated over tens of billions of requests, becomes a multi-billion-dollar infrastructure cost difference.
The Biggest Risk: Model Iteration Versus Chip Cycles
Frozen v2 targets 2028 deployment. Google must judge two and a half years out which Gemini capabilities will remain stable enough to etch into hardware. That is the fundamental contradiction of model-specific chips: model iteration cycles run in months; advanced chips take years from design to production. If Gemini’s architecture shifts materially before the chip ships, parts of the design face rework — and Google itself has left explicit room for retreat, noting that “not all research projects reach production.”
The problem does not belong to Google alone. OpenAI released its first custom inference chip, Jalapeño, with Broadcom in June; Anthropic has reportedly discussed chip collaboration with Samsung; custom silicon is becoming standard equipment for first-tier model developers. Every one of them will face the same challenge.
The Bigger Industry Signal
The most notable thing about Frozen v2 is not the unverified “6 to 10×” figure — it is the metric Google chose: tokens per unit of energy. That marks the center of AI competition migrating from model capability to inference economics. As model capabilities converge, whoever serves more users at lower cost holds the next era’s advantage. For the industry, three implications follow: vertical integration becomes standard equipment for model vendors — the ability to co-design model, chip, and datacenter will open gaps; inference costs may fall faster than expected, with specialized silicon driving another sharp token-price drop around 2028; and architecture stability becomes a precondition for chip investment — teams that keep reworking their architectures may abandon the custom-silicon route entirely.
This is not a contest over whose chip wins. It is an answer to a more fundamental question: when AI reaches sufficient scale, can every layer of infrastructure — from model architecture to chip design — still evolve independently? Frozen v2’s answer: no. At least for Google, the optimal move is binding them together.
Implications by Role
-
Building large models? Think about the long-run inference-cost curve. If your architecture is not yet stable, custom chips are probably not today’s priority — but sustained investment in inference optimization is, whether quantization, distillation, or hardware adaptation.
-
Building products on models? Follow these specialized chips. Inference costs may fall further and faster than expected; scenarios blocked by cost today — like agents’ token-hungry loops — may become viable in two to three years.
-
Building AI infrastructure? Model-chip coupling is deepening. Whoever helps model teams balance iteration speed against hardware ROI captures the next wave.
Whether Frozen v2 ever reaches mass production is hard to say today. But the direction’s signal is clear enough: a company spending $180 billion a year on AI infrastructure will find every way to make each dollar burn into more tokens. That logic holds at any scale.
