News · 2026-07-21

Google's two opposite bets: a Gemini-specialized chip and an EU order to open Android AI

Google is making two opposite lock-in bets at once. It is reportedly designing a server chip, informally called Frozen v2, that would bake Gemini's architecture into silicon for a large efficiency gain, betting Gemini's internals stay stable. At the same time, the European Commission has ordered Google to open eleven AI-relevant Android capabilities to rival assistants on parity terms, forcing it to unlock the very features that make Gemini sticky on phones. One bet assumes stability; the other is imposed change.

Key facts

• The Information reports Google is working on "Frozen v2," targeting deployment as early as 2028, with an internal projection of six-to-ten times more tokens served per unit of power than its newest AI chips. The Information
• Google has not announced the chip; a spokesperson said only that its teams research hardware-software co-design and not every project ships.
• The European Commission adopted binding Android AI-interoperability measures on July 16. Commission announcement
• Most measures are due in Android 18 by August 1, 2027; concurrent wake-word detection follows in Android 19 by August 1, 2028.

Take the chip first. General-purpose AI accelerators like Google's TPUs are flexible: they can run many models, which is convenient but wastes energy on runtime generality. Frozen v2, as reported, makes a deliberate trade: freeze the model's architecture into the hardware to cut repeated decisions and data movement, but keep the weights replaceable so the chip is not obsolete after every training run. An earlier design that would have frozen the weights too was reportedly abandoned because its useful life would be too short. The reported six-to-ten-times figure is more output per watt, an internal engineering projection, not a shipped-chip benchmark, cost, or quality claim, and Reuters reports the design is still being finalized.

The real-world comparator sharpens the tradeoff rather than settling it. Taalas's HC1 demonstrator hardwires an entire Llama 3.1 8B model into silicon and claims very high per-user speed, but its own materials call it a single-model technology demonstrator. Frozen v2, if the reporting holds, would be less brittle because weights remain replaceable, but still brittle if Gemini's architecture changes. For context, Google's publicly announced TPU 8i takes the opposite, more flexible path, adding on-chip memory and a serving-oriented network while keeping broad framework support, which is why the reported 6-to-10x figure should be read as an unusually aggressive, unvalidated target.

Now the EU order, which is concrete and binding where the chip is speculative. The Commission's measures require free, equally effective interoperability across eleven feature areas: invocation through long-press and custom wake words; context such as opted-in app data and live sensor inputs; actions like screen automation and system integration; and resources including access to system-level on-device models such as Gemini Nano. The competitive change is that a rival assistant can become an actual device agent rather than a foreground app. And it is parity, not just an API: equal effectiveness on setup, data-transmission speed, and even energy consumption, and Google cannot condition access on a rival holding the default-assistant role.

Why it matters: the EU is not merely opening a voice-assistant slot; it is requiring parity around the contextual, on-device, and background capabilities that make an AI assistant sticky in the first place. That is a deeper intervention than a default-choice screen. Google's strongest counter-case, from policy chief Kent Walker, is security governance: he argues the order gives external apps sensitive permissions without the device-maker validation Google says protects users, and that the parallel search-data-sharing order risks inadequate anonymization. That is Google's stated position, not an independently established finding, and the measures do preserve consent, integrity controls, and certification for five defined restricted features.

The honest caveats: Frozen v2 is reporting from unnamed sources, not a Google-confirmed product, and the EU's specification decision guides compliance rather than itself declaring Google non-compliant. Analyst John Gruber's sharper reading is that parity could turn EU Android into a desktop-like environment of multiple assistants listening for wake words, and that Google might limit Gemini's EU integration rather than build the APIs, though that "unfettered access" framing overstates the legal text's safeguards. The combined beat is the tension: Google wants to make Gemini cheaper by assuming its internals stay fixed, while Brussels makes Gemini's Android advantage less exclusive by forcing the surrounding capabilities open.

Primary source, verified: read the paper →

Ground Truth. (2026, July 21). Google's two opposite bets: a Gemini-specialized chip and an EU order to open Android AI. Ground Truth. https://groundtruth.day/news/google-frozen-v2-chip-and-eu-android-ai-mandate.html

@misc{groundtruth:google-frozen-v2-chip-and-eu-android-ai-mandate,
  title  = {Google's two opposite bets: a Gemini-specialized chip and an EU order to open Android AI},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/google-frozen-v2-chip-and-eu-android-ai-mandate.html}
}

Topics: industry · google · hardware · regulation · eu · android · inference