Five years ago I helped write a framework for the World Economic Forum on equitable access to the data economy. It argued that participation in the coming economy would turn on three principles — integrity, inclusivity, interoperability — enabled by three levers: capital, collaboration, compliance. The framework was correct and is now insufficient, in the way that a map of a river is insufficient once the river has become a sea. Data is no longer the asset that decides who participates in the AI Economy. The asset is capability, and capability has just acquired a meter.

I have spent the past several months updating the framework into an equitable access framework for the AI Economy for the Forum and other bodies, and the exercise forced a discipline the discourse mostly evades: saying what equitable access actually is. Here is the definition the paper stakes:

Equitable does not mean identical: differentiated access, proportionate to genuine risk, is compatible with equity — what equity forbids is differentiation by scale, wealth, incumbency, or geography. Meaningful excludes the nominal: access to a capability one cannot afford to run, integrate, or govern is not access. The definition is multi-scalar, because an AI economy can be equitable within a nation and iniquitous between them.

In January 2025, DeepSeek R1 gave the United States what Andreessen called AI’s Sputnik moment — though one would propose that America’s Sputnik was China’s Apollo: the fruit of a decade of investment catalyzed by AlphaGo’s defeat of Lee Sedol in 2016, which was Beijing’s Sputnik. Either way, the shock produced what Sputnik shocks produce. Sputnik gave America NASA and the National Defense Education Act. DeepSeek gave it Stargate and, last November, the Genesis Mission — a Department of Energy initiative explicitly framed in Manhattan Project terms, yoking seventeen national laboratories, the world’s largest trove of federal scientific data, and an ambition to double the productivity of American science within a decade. Twenty-six grand challenges were named in February. An operating platform is due in August. The institutional response is real, and it is Apollo-shaped: infrastructure first, distribution later, if at all. That ordering is precisely what a framework for equitable access exists to contest.

Then June happened, and the question changed registers.

On June 9, Anthropic released one model under two names. Fable 5 carries the strongest safeguards ever applied to a frontier system and is available to anyone. Mythos 5 is the same mind with the locks removed, dispensed to roughly a hundred and fifty vetted organizations doing defensive security work in concert with the American government. I argued in June that this split was AI governance’s first real act — know-your-customer for cognition, the frontier no longer released or withheld but dispensed by tier to counterparties whose identity has been established. What the intervening weeks added is the stumble, and the stumble is more instructive than the launch.

Three days after release, the government applied export controls to a deployed commercial AI model for the first time in history — triggered by a researcher report later judged to have been inflated beyond its significance. The directive barred access by foreign nationals anywhere on earth, including the developer’s own employees. Since no infrastructure existed to verify nationality in real time, the company had to switch off both models for everyone, everywhere, for eighteen days. The most capable system ever offered to the public spent nearly three weeks dark while less capable open-weight models, several of them Chinese, sold into the vacuum. The controls were lifted at the end of June after review found that numerous other models could reproduce the reported behavior.

Read as farce, the episode is a regulatory overreaction walked back under industry pressure. Read as instrumentation — which is how I read it — it is the first live test of the metered regime, and the meter failed in exactly the ways a first meter fails. The instrument was blunt because no calibrated instrument existed. The restriction was unenforceable because the identity layer it presupposed had never been built. And the regime metered only the compliant: a governance apparatus that can halt a cooperative domestic laboratory within seventy-two hours while exerting no leverage whatsoever on capable open models diffusing globally is not yet governance. It is a handicap with paperwork.

What June proved, above all, is that the old principles of the data economy were built for this moment. The updated Equitable Access framework does not replace them; it extends each into the new terrain of the AI Economy. Integrity, which once meant tracing data through its lifecycle, now runs the length of the whole chain — model provenance, training attribution, watermarked outputs, the ethics of distillation, and the verifiable integrity of the safeguards themselves. Interoperability, which once meant ecosystems speaking across boundaries, now governs model portability, agent protocols, honest benchmarks, and — an extension nobody planned — the very grammar of refusal, since a safeguard that declines a request now does so in machine-readable form with automated fallback. Inclusivity, which once meant that the small enterprise could enter the data marketplace as readily as the multinational, is the principle under the heaviest new load: model access, compute access, inference affordability, and now the metered tier, where a single capability is dispensed under multiple regimes to counterparties whose identities have been established. Inclusivity does not forbid the meter. It disciplines it — and the discipline generates equity questions that nobody has answered because nobody has yet been obliged to ask them in public.

Who vets the vetters? The criteria for trusted access are today set privately by one company in consultation with one government. Does metering entrench incumbency? Unlocked capability flows naturally toward the large, the established, the well-lawyered, the allied; the security firm in Nairobi gets the safeguarded tier at best, forever. Is refusal accountable? Fable’s declines are machine-readable — refusal has been productized, given response codes and billing rules — but a refusal legible to an API is not yet legible to an institution, and nobody publishes false-positive rates on blocked research. And can a metered frontier coexist with unmetered diffusion at all, absent international coordination that does not currently exist?

None of this is an argument against the metering. The deeper truth of Fable is generous: near-frontier capability made broadly available with safeguards is a better equity outcome than capability withheld, and by the definition above it is a genuine inclusivity achievement — the small actor got the capability, not the excuse. The argument is that a practice of this consequence — invented under commercial and geopolitical pressure in a single fiscal quarter — must not remain an improvisation. Banking’s know-your-customer regime took decades to acquire published criteria, appeal rights, independent audit. The metering of minds will need the same apparatus in years, and the three principles tell us exactly what it consists of: for inclusivity, published eligibility standards, due process for denial and revocation, and vetting blind to everything except risk; for integrity, independently auditable safeguards and refusal transparency — categories, rates, false positives on legitimate work; for interoperability, verification infrastructure built deliberately with privacy designed in, and refusal signals that hand off cleanly rather than strand the application. Controls, meanwhile, calibrated to a model’s uplift over the open baseline rather than its headline capability.

Beneath all of this the revised equitable access framework keeps its two longer horizons. The floor: income support funded by the automation that necessitates it — and here the metering delivers an unexpected gift, because an instrumented, metered, billed capability layer is incidentally the measurement infrastructure a compute levy or automation dividend has always lacked. The ceiling: AGI, whose timelines the past year has tended to corroborate rather than embarrass. Within eighteen months of DeepSeek, a model arrived whose unlocked variant outmatches all but the most skilled human security experts, and the interval between frontier capability and government response compressed from years to seventy-two hours. When capability of that order generalizes, the metered regime we are improvising now becomes the institutional form all access governance takes. The rehearsal is the performance.

The original Sputnik moment ended with flags on the moon — a triumph whose benefits remained, for most of humanity, symbolic. The Genesis Mission may yet build marvels on the same pattern. What the equitable-access tradition insists, and what I will keep insisting to the Forum and anyone else who will sit still for it, is that this race be scored differently: by whether participation is meaningful for the many and not merely the metered few, by whether the gains reach the displaced, and by whether the rules get written by published principle or by accumulated accident. The principles already exist. They have existed since 2021: integrity, inclusivity, interoperability. June was an accident. We should not need many more of them to choose.

Stay tuned for the full paper and your feedback is welcome as it goes through review.

Leave a comment