Public AI is the bet that artificial intelligence can be built as public infrastructure, the way roads and libraries were, rather than rented from a handful of companies. This month the bet got real money, working prototypes, and the UN's first all-member AI governance dialogue. Three developments mattered most.

In Geneva this month, the nonprofit Current AI switched on Alpha Chat, an open-source chatbot assembled in seven weeks by a coalition of 10 organizations including Hugging Face, Mozilla, and the MIT Media Lab. Each partner brought one piece of the stack: a language model, safety tooling, computing power. The coalition assembled from what already existed rather than training a new foundation model from scratch. (TechCrunch, Jul 19)

Why it matters: Assembly is now the movement's method. The next stage, a project called AI Potluck, assessed 24,626 open-source projects from foundation models to inference backends. Its 12-month sprint, launched this month, promises to "integrate the best tools and products and ship feature-by-feature until there is a viable alternative to proprietary AI that isn't owned by any one company or country." (aipotluck.org, July)

By the numbers: Portugal ran the same play at national scale. The government unveiled Amalia, described by its creators as the first open large language model developed in European Portuguese, in early July: a 9-billion-parameter core built on EuroLLM-9B and the earlier Portuguese model GlorIA, with public investment rising to 7 million euros by 2027. The reason the price stayed low: "Adapting an existing foundation model costs an order of magnitude less than training one end to end, an operation that runs into tens or even hundreds of millions for state-of-the-art models." (ActuIA, Jul 3)

Brussels is backing the opposite theory with public compute. In June the European Commission selected EUROPA, a consortium led by the Italian AI company Domyn with Germany's Fraunhofer-Gesellschaft, to develop an open-source model of more than 400 billion parameters covering all 24 official EU languages, and granted it "up to 2.5 percent of the total EuroHPC computing capacity" for a year. (heise online, Jun 19)

Between the lines: Two theories of public AI now enjoy public backing. Current AI CEO Ayah Bdeir rejects the premise behind the bigger bet: "Scale is not always the measure. That is the Big Tech paradigm." Her organization's backers, she says, are "not investors, they're funders": $400 million committed, including $100 million from the French government, against a five-year target of $2.5 billion. (The Next Web, Jul 20)

What's next: Both bets run on public clocks. AI Potluck has 12 months of sprint; EUROPA has one year of granted supercomputer time. By next summer, public AI will know which theory produces working software people use.

For eight months the Swiss digital agency Liip ran Apertus, Switzerland's public language model, behind its assistant for the city of Zurich and logged more than 39,000 production conversations against OpenAI's GPT-4o-mini. Developer Josef Kruckenberg wrote: "The question is not 'sovereign AI or quality' but 'how much quality gap are you willing to accept for full sovereignty?'" (Liip, Jul 9)

By the numbers: In Liip's own rating of that deployed traffic, one assistant and one evaluation scheme rather than a public benchmark, the two models tie on answers judged acceptable: 82% for Apertus against 81% for GPT-4o-mini. On the project's stricter quality measure the gap opens, 55% against 72%. The direction favors patience, though. The same analysis found Apertus climbing to 63% by June, "the best month on record," without a single model update, and version 1.5 is due with tool calling, reasoning, and better regional-language coverage.

The big picture: Latin America makes the same argument from identity rather than telemetry. LatamGPT, the 70-billion-parameter regional model coordinated by Chile's CENIA and trained on data from 20 countries, launched in February with President Gabriel Boric declaring: "Here we're defending our identity and our right to exist." CENIA director Alvaro Soto said: "No matter how powerful large models are, they cannot cover all aspects relevant to our reality." (TechPolicy.Press, Mar 2)

Yes, but: The market is arriving at the same place for its own reasons. Chinese open-weight models took 41% of Hugging Face downloads this spring, and Hugging Face CEO Clem Delangue expects "most of the production workloads" to run on private or open-source models within a few years, while Anthropic CEO Dario Amodei maintains the standing objection that open weights, "once they are released, they become difficult to control." (TechCrunch, Jul 14) Independent technologist Ben Werdmuller's reading: "Open almost always wins when it comes to infrastructure adoption." His post cites venture investor Martin Casado's estimate of "an 80% chance that any given startup is using Chinese models." (Werd I/O, Jul 20)

Apple removed the 512-gigabyte memory option from the Mac Studio, one of the few consumer workstations able to hold some very large open models in unified memory, without public explanation; industry reporting ties the shortage behind it to chip suppliers prioritizing datacenter demand. Jonathan Fulton, an independent analyst who tracks the self-hosting market, published the numbers behind the squeeze in early July: "The boom that made open weights frontier-class is the same boom that made them impossible to run at home." (Jonathan's Musings, Jul 5)

By the numbers: Industry forecasts he compiles project datacenters consuming roughly 70% of the world's high-end memory-chip output this year, and some consumer RAM configurations have climbed 300-600% off their 2024-2025 cyclical lows. Some enterprise GPU systems reportedly carry lead times of 30 to 52 weeks with non-refundable deposits.

The big picture: Institutions are building what individuals can no longer buy. Jan Hajic, who coordinates the 20-organization OpenEuroLLM project from Charles University, secured time on four EuroHPC supercomputers and still reported in February that "significant challenges, especially in securing more compute for creating the final models, still remain." (Tubingen AI Center, Feb 26) In the United States, New York's Empire AI has put $500 million over 10 years behind shared academic supercomputers. "With Empire AI, a researcher in New York can now build their own," said Venu Govindaraju, senior vice president for research at the University at Buffalo, while UC San Diego physicist Michael Norman explained what the program replaces: "We just don't have the money." (Science, Dec 3)

What's next: The federal layer lags behind the states. NAIRR, the US National AI Research Resource pilot, offers about 3.77 exaFLOPS, roughly 5,000 H100-class processors, a capacity researchers Sarosh Nagar and David Eaves judged insufficient for national research needs last August. (Lawfare, Aug 7) Hanna Hajishirzi of the Allen Institute for AI has recommended $2.6 billion over six years to expand it, per the same Science report. Public AI spent the month proving it can assemble models, money, and coalitions. The hardware underneath is still priced by someone else.