The EU Just Handed Google's AI Rivals the Keys to Android

Brussels picks the lock on the assistant slot

The most valuable real estate in consumer AI is not a model. It is a button — the long-press of a home key, the wake word an operating system listens for before any app wakes up. On July 16, the European Commission decided that Google no longer gets to keep that real estate to itself.

In two binding specification decisions under the Digital Markets Act, Brussels ordered Google to open Android’s system-level assistant machinery to competing AI assistants and to share the search data that feeds its own ranking systems with rival search engines — including, explicitly, AI chatbots that offer search functions. The first decision means ChatGPT, Claude, Perplexity, or Mistral’s Le Chat can register a wake word, be summoned by voice, and take actions inside other apps on an Android phone, exactly the privileges Google reserves for Gemini today. The second means the click, query, ranking, and view data that trained two decades of Google Search relevance must flow, anonymized and priced by formula, to the companies trying to replace Google Search. Early coverage framed it as an order to share search data and broaden Android feature access; that undersells it. It is an attempt to unbundle distribution from model quality — to make the assistant war a fair fight on hardware Google controls.

The timing was brutal, and clarifying. The same day the decisions landed, The Information reported that Gemini 3.5 Pro — unveiled at I/O in May with a promised June rollout, a keynote I covered when Google fused Gemini into Android and XR — is months behind schedule because its coding performance missed internal targets. Alphabet shares fell as much as 4 percent, erasing roughly $200 billion in market value at the intraday low. In one trading day, Google absorbed a regulatory strike at its distribution moat and a credibility strike at its model roadmap. Each makes the other worse: distribution is precisely the asset that buys a lagging model time, and Brussels just put that asset on a clock.

The stakes deserve to be stated plainly. The default-search slot was the most profitable contractual position in the history of software — Google paid Apple tens of billions a year for it, and the US v. Google remedies fight was largely about it. The assistant slot is its successor. Whoever holds the wake word owns the query before it is typed, the task before it is delegated, the purchase before it is compared. Google understood this early, which is why it retired Google Assistant and made Gemini the default AI on Android this year, completing the transition across the largest operating system footprint on Earth in early 2026. Apple understood it too, which is why it handed Siri’s brain to Gemini rather than ship a worse assistant on time. Meta understood it when it turned Facebook search into an answer engine. The entire industry has converged on the same conclusion: in the agent era, the scarce good is not intelligence, it is placement.

Brussels has now regulated placement directly. That is what makes July 16 different from the parade of European tech fines that American readers have learned to skim past. The Commission did not fine Google. It specified, in engineering detail, how Android must behave — and it did so at the exact moment the assistant layer is being decided. Regulation usually arrives after the market has ossified; this arrived mid-pour, while the concrete is still wet.

Consider the counterfactual the Commission is trying to prevent. In the search era, regulators spent fifteen years litigating a market that had already tipped — by the time the remedies arguments were heard, Google’s share had been above 90 percent in Europe for a decade, and no order short of divestiture could have changed user behavior. The assistant market has not tipped. Gemini’s default position is months old, ChatGPT’s mobile app remains the most downloaded AI product in Europe, and no assistant yet owns the muscle memory that “just Google it” once encoded. The DMA’s specification mechanism exists precisely for this window: instead of proving harm after the fact, the Commission dictates interoperability terms before the network effects close over. Whether one finds that far-sighted or presumptuous, it is a genuinely new kind of intervention — competition law operating at product-cycle speed.

Eleven doors into Android, one meter on the data

Strip away the legal scaffolding and the Android decision is a technical spec. It identifies eleven specific Android feature points that assistants depend on, and the list reads like a teardown of Gemini’s structural advantages. Wake-word registration at the operating system’s audio-detection layer, so a rival assistant can listen for its own “Hey” without draining the battery from userspace. Access to the long-press home gesture and navigation-handle triggers — system-wide entry points no downloaded app can claim today. The ability to act inside third-party apps: book the taxi, draft the reply, pull the flight number out of an inbox mid-call. Suggested replies, screen context, recent-activity awareness. Everything, in short, that separates an operating-system feature from an app icon.

The search-data decision is the quieter half, and possibly the more consequential one. Google must share anonymized ranking, query, click, and view data with third-party search engines and AI chatbots that offer search functionality, under a fair pricing formula and transparent access process, with sharing to begin in January 2027. This is the data flywheel that made Google Search unassailable: twenty-five years of humanity revealing, click by click, what a good answer looks like. Every AI lab building a search product has been approximating that signal expensively — scraping, licensing, buying browser telemetry. Brussels just ordered the original sold at a regulated price. The Android interoperability changes follow on a longer fuse, due with the next major Android release by July 2027, per CNN’s account of the escalating fight between the Commission, Google, and Apple over how assistants will work in Europe.

The enforcement teeth are real, and recently bloodied. DMA noncompliance carries fines of up to 10 percent of global annual turnover, rising to 20 percent for repeat offenses, and the Commission demonstrated in April 2025 — €500 million against Apple, €200 million against Meta — that it will use them. Against Alphabet’s roughly $400 billion in 2025 revenue, the 10 percent ceiling implies about $40 billion of exposure for defiance. EU tech chief Henna Virkkunen framed the goal as market structure, not punishment: emerging alternatives to Google Search and to Gemini, and genuine choice for European users.

Here is the arithmetic that matters, stitched from the week’s numbers. Alphabet’s intraday loss on the Gemini delay — roughly $200 billion — was about five times the maximum first-offense DMA fine. The market, in other words, still prices losing the model race as a categorically bigger risk than losing a regulatory fight. But the two risks are no longer independent. Gemini’s default placement on Android was the insurance policy against model slippage: even if Gemini 3.5 Pro arrives late and merely adequate, three billion devices would keep handing it queries. The Commission’s decisions convert that insurance into a wasting asset with a twelve-month fuse — search data flows from January 2027, system-level assistant access lands by July 2027. A months-late model was survivable with a locked distribution channel. A months-late model with an unlocked channel is how defaults die.

The beneficiary list writes itself. OpenAI, which I argued has bet GPT-5.6 on the agent race, gains a path to system-level presence on the one mobile platform it does not have to negotiate for. Perplexity, which has been buying its way onto phones one OEM deal at a time, gets the OEM layer bypassed by law. Anthropic gets the same doors opened without spending a dollar of its enterprise-focused capital on consumer distribution. Mistral — Europe’s champion — gets a home-field assistant slot that Brussels will be politically invested in seeing used. None of them earned this; Google’s dominance did. That is how ex ante regulation works: it taxes the incumbent’s position, not its conduct.

Notice who is missing from that list: Apple. Cupertino outsourced Siri’s intelligence to Gemini precisely because it could not build a competitive assistant in time, and the CNN reporting makes clear Apple is clashing with Brussels over parallel demands that rival assistants be as prominent on smartphones as Siri and Gemini. If Brussels applies the Android template to the iPhone — and the Commission’s Apple interoperability proceedings suggest it will — the Apple–Google assistant pact I analyzed in June starts to look less like a moat and more like two incumbents sharing a wall that regulators are dismantling brick by brick. The assistant war was shaping up as a two-platform duopoly with one brain; the DMA is trying to force it into an open market with five.

How Google keeps the moat anyway

Every thesis this confident deserves its stress test, and Google’s counterargument is not frivolous. The company’s response, published within hours, argues the decisions risk undermining vital privacy and security guardrails for millions of Europeans — that granting external apps sensitive device permissions invites abuse, and that “anonymized” search data is a fiction at sufficient scale, exposing private queries to unfamiliar companies without meaningful consent. Dismiss the self-interest and a kernel survives: an assistant with wake-word access and cross-app authority is the most privileged software on a phone. A malicious or merely sloppy one is a keylogger with a marketing budget. The Commission built in security assessments before data flows to any specific third party, but the assessor of first instance is Google — which is both the entity best positioned to evaluate risk and the entity with every incentive to find it.

That points to the likeliest failure mode for Brussels: malicious compliance. Apple’s DMA experience is the template. Interoperability arrived wrapped in fee structures, warning screens, and engineering friction that made the concession technically real and commercially inert. Google has the same toolkit. The anonymization “multi-layered method” could strip the shared search data of the session-level richness that makes it valuable, leaving rivals to pay a regulated price for a degraded signal. The pricing formula itself will be litigated. The eleven feature groups will ship behind permission dialogs written by the world’s best dark-pattern engineers, reminding users at each step what they surrender by leaving Gemini. The EU’s browser choice screens, mandated in 2024, moved European browser share by low single digits. Defaults are not laws of physics, but they behave like them: most users never touch the setting.

The geopolitical risk runs the other direction. President Trump has already threatened 100 percent tariffs on countries that impose digital services taxes on American companies, and his administration treats European tech regulation as trade aggression by other means. A DMA order that visibly transfers value from an American champion to — in part — other American companies is an awkward casus belli, but Washington’s objection is to the precedent, not the beneficiaries. If the decisions get absorbed into a tariff negotiation, enforcement could soften into the kind of perpetual “constructive dialogue” that has quietly defanged prior European actions. Brussels holds formal power; whether it holds nerve through a trade war is a different question.

Skeptics can also attack the remedy’s economics from the challenger side. System-level access is necessary for a rival assistant but nowhere near sufficient: running a wake-word service for tens of millions of European users means always-on inference costs that only the best-capitalized labs can absorb, in a year when token economics already dominate every roadmap conversation. OpenAI is burning capital on data centers and the GPT-5.6 rollout; Anthropic is oriented toward enterprise seats, not consumer telemetry; Perplexity’s entire company is worth less than what Google spends on Android in a quarter. The Commission can open eleven doors, but it cannot pay anyone’s cloud bill for walking through them. If the challengers cherry-pick the profitable hooks — search queries, shopping intents — and skip the expensive ambient-assistant plumbing, Europe ends up with fragmented novelty apps rather than a genuine Gemini rival, and the incumbency reasserts itself by attrition.

There is also the sober possibility that the intervention is simply mistimed — that the assistant slot matters less than everyone currently believes. If agents migrate to earbuds, glasses, and cars faster than expected, Android’s wake word becomes a fortification around an emptying city. If model quality gaps widen rather than narrow, users will route around defaults to reach the better brain, as hundreds of millions already do by opening ChatGPT’s app daily on Gemini’s own platform. And the week’s other Google story cuts both ways: a company that cannot ship Gemini 3.5 Pro on schedule may lose the assistant war on merit before any regulator finishes opening the field. Distribution remedies assume the incumbent’s product is good enough that placement decides the outcome. Google’s July suggests that assumption is, for the first time in two decades, genuinely uncertain.

The twelve-month war for the wake word

Where this leads depends on three dates. January 2027, when search data starts flowing and we learn whether anonymization left anything worth buying. July 2027, when rival wake words either work on European phones with the next major Android release or arrive so encumbered that the Commission reopens proceedings. And the first noncompliance decision, if it comes — because the DMA’s credibility now rests on whether specification decisions get enforced with the same appetite as last year’s fines. Watch, too, what OpenAI and Perplexity actually ship. If no rival assistant of consequence launches in Europe within a year of the doors opening, Google’s lawyers will argue, with some justice, that the moat was never distribution at all.

The second-order effects may outrun the first. A regulated price for search data establishes, for the first time, a public market value for the behavioral exhaust that AI products run on — a number that every data-licensing negotiation, every publisher lawsuit, and every sovereign-AI procurement will reference the moment it exists. And a working wake-word interoperability spec becomes a portable artifact: Japan’s smartphone competition act, the UK’s digital markets regime, and the US v. Google remedies court can each import it wholesale rather than drafting from scratch. The Commission is not just regulating Google; it is publishing reference architecture for everyone who wants to. That is the Brussels effect in its modern form — exported not through market size alone, but through the sheer convenience of pre-written rules.

The wider pattern is worth naming. In the past three weeks, the world’s three regulatory blocs have each made their move on AI’s structural layer. China unplugged 345 million AI companions and, at this weekend’s World AI Conference, wrapped its governance model in multilateral packaging. Washington put frontier models on an export leash, regulating who may possess intelligence. And Brussels, true to form, regulated neither the models nor the morality but the market plumbing — the defaults, the data, the distribution. Three governments, three theories of where AI power actually lives. The European theory is the narrowest and, for exactly that reason, the most enforceable: you do not need to evaluate a frontier model to mandate a wake-word API.

For operators, the practical read-through:

• If you build an AI assistant or agent product, start your Android system-integration planning now. System-level access in the EU is a distribution channel that did not exist last month; the teams that ship wake-word-ready assistants in the first quarter of availability will define the category before default inertia reasserts itself.
• If you build search or answer products, model the January 2027 data feed into your roadmap, but price in degradation. Negotiate access early, benchmark the anonymized signal against your existing click data, and treat it as augmentation, not salvation.
• If you are a Google partner or advertiser, assume Gemini’s European query share erodes from its ceiling and diversify assistant-surface strategies accordingly. The EU is roughly a tenth of the global Android base but a far larger share of its high-value queries.
• If you are watching from the US, treat this as the preview. The US v. Google remedies court, state AGs, and every other jurisdiction with a competition statute now have a working specification for assistant interoperability to copy from. Regulatory diffusion runs faster than product cycles.
• For everyone: the lesson of July 16 is that the industry’s quietest assumption — that model quality and distribution advantages compound privately — is now contested law in the world’s second-largest market. Plan for a world where placement is rented, audited, and shared, not owned.

The default died slowly, then all at once, when it was search. The wake word may get less time than that.

In other news

Xi launches a rival AI governance bloc from Shanghai — At his first World AI Conference keynote, Xi Jinping unveiled the World Artificial Intelligence Cooperation Organization, a Shanghai-headquartered intergovernmental body with 29 founding countries including Russia, Pakistan, Indonesia, and Brazil, declaring AI should not be “a solo performance by any single country.” China also pledged 5,000 AI training opportunities for developing countries over five years, positioning itself as the Global South’s AI patron while US export controls tighten.

Moonshot AI ships the largest open-source model ever — Beijing-based Moonshot released Kimi K3, a 2.8-trillion-parameter model with a 1-million-token context window that benchmarks third on GDPval-AA v2 behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and tops multiple coding and automation leaderboards. Full weights land July 27, with MXFP4 quantization making self-hosting practical on Blackwell and MI400 hardware.

SAP buys an 18-month-old startup to anchor Europe’s frontier lab — SAP closed its acquisition of Freiburg-based Prior Labs, committing more than €1 billion over four years to build a European frontier AI research lab around tabular foundation models. Prior Labs’ open-source TabPFN — 3 million downloads and a Nature paper — predicts directly over the structured business data (ledgers, supply chains, churn tables) that language models handle poorly.

FTC opens a front on AI accuracy claims — The US Federal Trade Commission is seeking public comment on a policy statement addressing AI accuracy, signaling that confidently wrong AI outputs marketed as reliable could be treated as deceptive practices. For US model providers, it is the first concrete hint that hallucination rates may become a consumer-protection liability rather than a benchmark footnote.