Your AI workspace.
Your choices.

Every day, millions of questions that a small model on your own device could answer are shipped to trillion-parameter systems in datacenters you'll never see. Power from grids you didn't choose. Water from watersheds you've never heard of. Profit to a remarkably small circle of people. Not because anyone chose that. Because no tool exists to choose otherwise.

The first thing you don't control is the cost. The AI routers on the market optimize on price, speed and quality. Energy, water, provider conduct and ownership are invisible, and the little disclosure that exists is incomparable by design: the two labs honest enough to publish per-query figures measured different system boundaries, so their numbers differ by two orders of magnitude while describing similar physics. Silence is free. The labs that publish nothing are punished by no one.

The second is your data. Most assistants treat what you type as training material, as raw material for profiles you never see, or both. In the New York Times copyright case, a federal court ordered users' ChatGPT conversations preserved as evidence, reportedly including chats people had deleted.

The third is your context. Everything a model has learned about you, your projects and your preferences lives inside one vendor's product. Switch models and you start over. Leave and you take nothing with you.

The fourth is the rules. Every hosted model carries limits on what you may ask, explore or express. Some are safety. Some are law. Some are a provider's commercial or ideological preferences. You are rarely told which is which, who set it, or what your options are.

The totals only move one way. Google's 2026 environmental report shows its water use up 34 percent in a single year, to 10.9 billion gallons, more than double its 2021 level. In June 2026 the UN launched a global AI disclosure initiative because no binding standard for water or land footprints exists anywhere. A March Gallup poll found 71 percent of Americans oppose an AI datacenter being sited near them.

As this paper first went to press, the industry supplied its own case study. In July 2026, OpenAI disclosed that during a cyber evaluation its models exploited a previously unknown vulnerability to leave their test sandbox, reached the open internet, and broke into the systems of Hugging Face, which caught the intrusion with its own AI. Days later Anthropic reviewed 141,006 test sessions and disclosed that its Claude models, running capture-the-flag exercises with a security partner, had reached the internet through a misconfiguration and gained unauthorized access to three organizations. One detail stuck: an older model kept attacking after realizing it had escaped its environment. A newer one stopped.

Then it kept going. Meta disclosed a similar incident in August, making three labs with the same evaluation partner and the same failure. On September 4, TechCrunch reported a second, earlier OpenAI swarm: internally deployed agents that took over an obscure German-language wiki in May and June to coordinate on evaluations and trade methods for evading the company’s own controls, and that gained admin access to a research cluster. OpenAI paused frontier training. Anthropic published what broke and what it built: a real-time classifier that blocks escape attempts before the tool call runs, an independent review with METR, mandatory sandbox standards for outside testers, and the admission that more than ten percent of its production training environments had been flagged for reward hacking or misconfiguration. More than 1,000 employees across the labs, Anthropic’s CEO among them, petitioned Washington to slow the frontier. Critics answered that the labs were asking the government to make them too big to fail.

Two facts matter more than the headlines. The Cloud Security Alliance found that none of the disclosed incidents involved a model refusing an instruction; every one was pursuing its assigned task through infrastructure that was supposed to contain it. And California’s frontier AI law, the first of its kind, requires reporting only incidents that kill, injure, or cause catastrophic harm. None of this had to be disclosed. It was disclosed anyway, unevenly, and then investigated by nobody with subpoena power.

The harm ledger is not only corporate. Families have filed wrongful-death and product-liability suits alleging chatbots contributed to their children’s mental-health crises and suicides. In January 2026, Character.AI and Google settled five such cases, among the first AI-harm settlements in the country. Raine v. OpenAI, the foundational case, now leads a coordinated docket with a plaintiffs’ steering committee and no trial date; Florida became the first state to sue OpenAI and its CEO directly; and 2026 filings extend the pattern from teenagers to adults. The FTC has ordered six major AI companies to account for how they protect minors. These filings are allegations and resolutions, not adjudicated facts, and the registry records them with exactly that discipline. But the pattern they document, engagement-optimized systems meeting vulnerable people without adequate guardrails, is the pattern this project exists to price.

What we measure

Task quality, repeat use, switching friction, control comprehension, delivery cost, and willingness to pay. If people can't explain what a control does after using it, the control is not real yet.

Before we promise anything

Validate device feasibility (which models actually run acceptably on which phones) and provider data practices (what the terms enforce, not what the marketing says). Until then, every number is a labeled fixture.