Below the Ice — The Public Option for AI Infrastructure

This is the print twin of tonight's Below the Ice — our evening deep-dive, one topic told properly. Prefer it in your ears while you wind down? Listen to today's episode.

What if AI infrastructure worked like a highway — built with public funds, maintained as a shared resource, open to whoever needs it? That's the question behind tonight's topic. At the AI Action Summit in Paris in February 2025, a group of researchers, governments, and foundations launched Current AI, a non-profit coalition with $400 million already committed and a deceptively simple goal: build a public option for AI. This week, Simon Willison flagged their Open Source AI Gap Map, and it's worth sitting with. Not because the money is large — relative to what the private labs spend, it isn't — but because the map itself is an honest accounting of a real problem most people in the industry prefer not to name out loud.

What it is

Current AI describes itself as "a global partnership building a public option for AI." The Open Source AI Gap Map is their attempt to take stock of where the AI stack is closed — where proprietary infrastructure has no open alternative — and make those gaps visible enough to fund against.

The phrase "public option" is borrowed deliberately from healthcare policy. A public option isn't a government takeover; it's a baseline alternative that any organization can use if the private market's offerings are too expensive, too closed, or too extractive. Think of it the way a public library relates to a bookstore: you don't have to use it, but it's there, and that fact alone changes the market.

How it actually works

Think of the AI stack as having about five floors: hardware and compute, data pipelines, foundational models, safety and alignment infrastructure, and the application layer. Today, floors two through four are almost entirely controlled by a small number of private companies — and floor one, chip supply, is controlled by even fewer.

The Gap Map is a structured inventory of that stack. For each layer, it asks: is there an open alternative that meets production-grade requirements? Can a government, a hospital, a university, or a startup in a country without a major AI lab actually use something that isn't dependent on one of three or four companies' goodwill and pricing decisions?

The answer, in most layers right now, is "not yet." Open models exist — but they typically lag commercial frontier models by months or years, are rarely optimized for low-resource languages, and still depend on proprietary cloud compute to train and run at scale. The Gap Map's job is to make that shortfall precise enough to fund against.

Current AI's approach is not to build everything itself. It's to coordinate: identify the gaps, secure funding from governments and foundations, and direct that capital toward open research teams and infrastructure projects that can fill specific holes in the public stack. Think of it less as a company and more as a venture-allocation map for public AI investment.

Why it matters now

There is a version of the next ten years where the entire AI stack — models, inference hardware, the data pipelines that run everything — is owned and operated by three or four companies, all headquartered in one country. Every government, every hospital, every university that wants capable AI does so via API, on terms that can change at any time, with no recourse and no alternative.

That is not a hypothetical. It is the trajectory we are already on. And the reason it matters now rather than in five years is that infrastructure tends to ossify. The window where an alternative is still buildable — before the lock-in is complete, before the network effects make switching impossible — is open, but probably not for long.

The $400 million committed to Current AI sounds large until you know that a single frontier model training run costs several hundred million dollars, and the labs do several per year. So the real argument for the Gap Map approach isn't budget scale. It's target selection: if you can't match the private labs on raw spend, the only play is to identify the specific infrastructure pieces where open alternatives are feasible and disproportionately valuable — where plugging one gap frees up everything above it.

For builders specifically: the public option changes your risk calculus. Today, if you build on a proprietary API, you are betting that the price, terms, and model behavior you rely on today will remain acceptable. A credible public alternative in even one or two layers of the stack gives you leverage that doesn't yet exist.

What is overhyped

The honest caveat: $400 million over a non-profit coalition is a start, not a plan.

The announcement math is tricky. "Committed capital" in non-profit contexts often means pledges from governments and foundations that disburse slowly, with conditions, over multi-year timelines. The AI labs will spend that much on compute alone in a quarter. The Gap Map is a prioritization tool, not a magic wand — it can tell you where the holes are, but filling them requires sustained engineering capacity, not just a well-lit inventory.

There's also a governance question nobody in the open-AI-infrastructure space has answered cleanly: who decides what "the public option" actually builds, and in whose interest? Open infrastructure built to serve Western researchers and a handful of well-funded governments is not the same as AI infrastructure that genuinely works for a Swahili-speaking doctor with a 2G connection. The Paris summit participants were mostly governments and foundations from wealthy countries. That's a starting point, not a representative sample.

What to watch

Three concrete things, the way we close every dive.

1. Whether governments actually use it. Public option thinking works in the countries where the public option is good enough to be the default for a meaningful share of the population. If Current AI's funded alternatives are only used by researchers, the leverage argument never lands. Watch for the first major government ministry — health, education, justice — to deploy a public-stack AI system at operational scale.

2. Whether the Gap Map gets specific enough to fund. An inventory of vague deficiencies doesn't move capital. Watch for the map to publish concrete engineering targets: inference stack benchmarks, cost-per-token thresholds, language coverage milestones. That's the sign it's becoming a build plan rather than a policy statement.

3. Whether the coalition holds together past the launch energy. Multi-government, multi-foundation coalitions tend to dissolve when the original convening moment passes. The AI Action Summit had real energy. Whether that energy survives a year of slow procurement cycles, diverging national priorities, and private labs writing compelling checks to the same talent pool is the real story.

Tonight's headline is real and the project is worth watching. The harder story is the one underneath: we are in a narrow window where public AI infrastructure is still feasible to build, and whether that window stays open depends on whether the map becomes a build plan before the gap closes on its own — with a single provider inside it.

---

That's tonight's Below the Ice. The full episode — same topic, out loud, while you wind down — is up now: listen to today's episode. More deep-dives at penguinalley.com.

Sources: Open Source AI Gap Map — Current AI · Current AI — Simon Willison's annotation · Current AI org · Introducing the Gap Map — Current AI blog