Google is working on a new AI chip designed to make Gemini more efficient

Google's stock jumped 3% on a chip that doesn't exist yet. That tells you everything about where the AI arms race actually stands. The market isn't rewarding breakthroughs. It's rewarding the promise of escape.

Frozen v2 — the internal codename alone signals how early this thing is — targets a 2028 release. Six to ten times more tokens per watt than Google's current TPU fleet. If true, that rewrites the economics of serving Gemini at scale. But 2028 is three generations of model evolution away. The architecture Gemini runs on today won't resemble what ships in 2028. Designing silicon for a moving target is the oldest trap in computing. Google knows this. They've fallen into it before.

The company's non-denial denial to TechCrunch was textbook. "Constantly researching and experimenting" translates to: we have a tape-out schedule and a power budget, but yield is still a prayer. Every hyperscaler says they co-design hardware and software. Few survive the collision between compiler maturity and process node delays. Google's TPU lineage proves they can execute. It also proves execution takes longer than roadmaps admit.

Investors cheered anyway. They cheered because Alphabet committed $180-190 billion to AI infrastructure this year alone. That number demands a return narrative. Custom silicon is the only story that scales. Buying Nvidia at margin forever doesn't pencil out. The market knows it. Google knows it. Jensen Huang knows it — which is why Nvidia now sells systems, not just GPUs, desperate to lock in the integration layer before customers flee.

OpenAI's Jalapeño. Anthropic's Samsung talks. Microsoft's Maia. Amazon's Trainium. Meta's MTIA. The stampede is real. Every frontier lab has concluded that Nvidia's moat — CUDA, NVLink, driver maturity — is wide but shallow. The moment a competitor delivers 80% of the performance at 30% of the power draw, the lock-in dissolves. Power is the new bottleneck. Data center build-outs hit utility limits before they hit capital limits. Tokens per watt is the only metric that converts directly to margin.

Google's advantage is verticality. They own the model, the framework, the compiler, the data center, the network, the cooling, and now the transistor layout. That stack integration lets them optimize across boundaries Nvidia can't see. They can strip instructions Gemini never uses. They can widen memory paths exactly where attention heads bottleneck. They can trade die area for sparse compute density because they control the sparsity pattern. Nvidia builds for everyone. Google builds for Gemini. Specialization wins when volume justifies it.

Volume justifies it. Gemini serves billions of queries daily across Search, Workspace, Cloud, and the API surface. A 6x efficiency gain across that footprint pays for the tape-out in weeks. The math is brutal and beautiful.

But the risk is equally brutal. Process node delays. Yield excursions. Package thermals. Compiler bugs that only manifest at 100k-chip scale. The graveyard of custom AI silicon is deep — Graphcore, Habana, Wave Computing, half of Intel's Xeon-HPC line. Google's own TPUv4 slipped a year. v5 slipped more. Frozen v2 is v6 by another name. The naming reset signals a microarchitecture break. Those breaks are where schedules die.

The 3% pop is rational speculation. The $190B spend is rational desperation. The 2028 target is rational hope. None of it is certainty. The only certainty is that the industry has stopped believing in general-purpose acceleration. The history of computing is the history of specialization eating generality. GPUs ate CPUs for graphics. TPUs ate GPUs for training. Inference ASICs will eat TPUs for serving. Frozen v2 is Google's bet on that timeline.

Three years is forever. Three years is tomorrow. Both are true. The editorial line: watch the tape-out, not the press release. Watch the power density at hot chassis. Watch whether Gemini's 2026 architecture still maps to Frozen's 2028 datapath. The chip that matters isn't the one they announce. It's the one that yields.