The $190 Billion Discoverability Problem
The U.S. government spends roughly $190 billion annually on R&D, and federal policy increasingly mandates that the results be shared publicly. NIH and NSF have strengthened open-access requirements. Researchers file data management plans. Repositories fill up.
But here's the thing: publicly available is not the same as usable.
Scientific datasets, models, and code end up scattered across incompatible systems, stored in legacy formats, released without the metadata or provenance needed to actually do anything with them. It's open in the legal sense, closed in the practical sense. And if humans struggle to reuse this stuff, AI systems don't stand a chance.
Project ProtoStar is a new consortium trying to fix this—not by yelling at researchers to do better, but by building infrastructure, setting standards, and running challenges that prove the value of AI-ready science.
Who's In and What They're Committing To
The founding partners span universities, nonprofits, and infrastructure:
- University coalition via Open Forum for AI (OFAI): Carnegie Mellon, Purdue, Georgia Tech, Arizona State, George Washington University
- Infrastructure: Hugging Face
- Community coordination: Open Research Community Accelerator (ORCA)
- Funding/support: Astera Institute
Participating organizations sign a memorandum of understanding committing to release suitable scientific artifacts—datasets, models, code—in AI-ready formats on open infrastructure. CMU's Open Source Programs Office is already encouraging researchers to share on Hugging Face while still supporting domain-specific repositories where appropriate.
This isn't a replacement for discipline-specific repos like GenBank or arXiv. It's a parallel track optimized for machine actionability.
The Numbers Behind the Push
Hugging Face hosts more than 17,000 public datasets across core scientific domains (genomics, protein design, chemistry, climate, Earth observation, astronomy). Expand to medicine and math and you're approaching 29,000 datasets, with a comparable number of models.
The pace is accelerating. Researchers shared more than three times as many new science datasets on Hugging Face in the first nine months of 2026 as in the same period of 2025—and that's after excluding the biggest single uploader.
When institutions do this right, the usage is real. NIH's National Center for Biotechnology Information maintains 18 public repos on Hugging Face with more than 11 million collective downloads. NASA IMPACT's 27 repositories have clocked approximately 770,000 downloads as of late October 2026.
Documentation matters too: 90% of dataset downloads on Hugging Face come from fully documented datasets, according to research cited in the announcement. This isn't a "build it and they will come" story—it's a "document it thoroughly or no one can use it" story.
The Open Science Challenges
Making data AI-ready is step one. Proving that readiness translates to actual scientific value is step two.
ProtoStar's core mechanism is a series of open challenges, each hosted by a participating institution and built around scientific resources that institution has released openly. The challenges have dual goals:
- Demonstrate reuse: Produce concrete examples of people building on publicly funded artifacts, ideally combining resources across institutions or disciplines
- Learn what enables reuse: Identify technical barriers, documentation gaps, interoperability issues, and what actually works
The idea is to build an evidence base for better research infrastructure, smarter funding decisions, and more effective public policy.
First Up: CMU's Climate Risk Intelligence Challenge
The inaugural challenge runs November 2–13, 2026, hosted by Carnegie Mellon.
Participants get historical weather data from 210 U.S. cities and are asked to predict 2019–2026 conditions using only data through 2018. The twist: these aren't hypothetical future scenarios—they're predictions that can be evaluated against actual observed weather.
The targets are four indicators relevant to urban planning:
- Average annual temperature
- Total annual precipitation
- Hottest day of the year
- Cooling degree days (a proxy for AC demand)
Predictions get compared against simple statistical baselines—historical averages, linear trends—to see if AI approaches offer meaningful accuracy improvements.
There's also a policy component: teams propose a short climate adaptation idea informed by what they learned from the data. Connect heat waves to energy infrastructure. Think about precipitation variability and stormwater systems. Make the science actionable.
Submissions are Hugging Face datasets with forecasts and policy proposals. A public leaderboard tracks forecasting performance. Final results get published in early 2027 once complete 2026 weather records are available.
Future challenges from other institutions will follow sequentially, each highlighting different scientific resources and reuse opportunities.
Why This Might Actually Work
A few things make ProtoStar more credible than the average "let's fix science" announcement:
Infrastructure that already exists: Hugging Face isn't building a new platform from scratch. The Hub is production infrastructure with 29,000 science datasets already living on it. Researchers know how to use it. The tooling for versioning, documentation, and access control is mature.
Institutional skin in the game: When CMU's Open Source Programs Office tells researchers to upload to Hugging Face, that's not a suggestion—it's institutional policy backed by the university's administration. These MoUs have teeth.
Learning-focused design: The challenges aren't just competitions with prize money. They're structured to surface what breaks, what's missing, where the friction lives. That feedback loop back to infrastructure and policy is the whole point.
Examples already in the wild: HuggingFaceBio's Carbon-A and Carbon-X projects show what's possible when you build on openly available genomic data. These aren't speculative demos—they're production research tools built on public resources.
The Harder Questions
A few tensions ProtoStar will need to navigate:
Domain repositories vs. generalist platforms: Biologists trust GenBank. Physicists trust arXiv. How does Hugging Face coexist with—rather than compete with—these specialized systems? The answer seems to be "both/and," but the workflows for maintaining artifacts in multiple places need to be near-zero friction or they won't happen.
Incentive misalignment: Academic careers are built on citations and publications, not on meticulously documented datasets. Until funding agencies and tenure committees value data release as a first-class contribution, researcher incentives remain skewed. ProtoStar can make it easier to share well, but it can't fix the underlying incentive structure.
The private sector free-rider problem: If public institutions invest in making data AI-ready and companies simply hoover it up for commercial products without contributing back, does that advance the public interest or just subsidize private R&D? The open licenses make this legal, but the optics and sustainability get thorny fast.
Evaluation and what counts as success: How do you measure whether ProtoStar worked? Number of datasets released? Downloads? Citations? Actual scientific breakthroughs that wouldn't have happened otherwise? The challenge design is a good start, but the long-term impact metrics are fuzzier.
What to Watch
If you care about open science infrastructure, a few things to track:
- Participation in the CMU climate challenge: Does it draw dozens of teams or hundreds? Who shows up—students, researchers, companies, hobbyists?
- Future challenge designs: What domains do subsequent institutions pick? How do they structure evaluation?
- Policy ripple effects: Do other universities adopt similar institutional commitments? Do funding agencies start requiring Hugging Face uploads as part of data management plans?
- Sustainability model: Who pays for hosting and bandwidth at scale? Hugging Face isn't a charity, and this kind of infrastructure requires long-term resourcing.
The CMU climate challenge launches November 2. Details are available at the Project ProtoStar page on Hugging Face.
Research organizations interested in joining can reach out to protostar@orcaopen.org.
The Meta-Point
Project ProtoStar is interesting less for any single technical innovation and more for what it represents: a coalition of institutions with resources and credibility willing to put weight behind the idea that publicly funded science should be practically open, not just legally open.
The $190 billion number is the hook, but the real story is whether this shifts norms around what "releasing research" actually means in 2026 and beyond. If ProtoStar succeeds, "we published the paper" becomes insufficient. "We published the paper and the AI-ready dataset with full documentation and provenance" becomes the baseline.
That would be a meaningful shift. And it would make all that public funding considerably more useful.