Open Science Needs Open Compute

Why independent science will depend on compute, not just funding, and why providing it is the best bet we can currently make

By Allison Duettmann, CEO of Foresight Institute

Every scientific era has a defining instrument: the telescope, the microscope, the particle accelerator, the gene sequencer. Our era’s defining instrument is compute. No astronomer owned the Palomar telescope, and no physicist owned a particle accelerator either. But those instruments belonged to universities, governments, and philanthropies. The compute that matters for frontier science today does not: five companies hold an estimated 71% of the world’s AI compute, while the entire US public research compute program runs on $30 million a year, a ratio of about 1 to 25,000 against private capex.

Foresight’s founding bet back in 1986 was that it is not enough to ask whether and when transformative technologies will arrive. We need to ask who controls them and how. These questions are still with us across bio, nano, and neuro sciences but nowhere are they perhaps more pressing right than with AI. Depending on how AI is built and accessed today, it could compress decades of science into years, or lock the lab door behind itself.

Most of the debate about open AI so far has been about model weights. But open models and open compute are two halves of one thing: weights you can download are inert without hardware you control to run them on, and owned hardware is idle without capable models you’re allowed to run. Open science needs both, and the hardware half is currently lagging behind.

In 2026, Foresight opened the doors to its first AI Nodes, one in Berlin and one in San Francisco: physical hubs providing on-site compute and funding for independent AI safety and AI science projects. The Nodes are a small prototype of what open compute for research can look like. What could a future look like if more funders and builders stood up independent compute collectives like ours across geographies, dedicated to uplift independent AI science and safety research?

TL;DR

This article argues four things: 1) open compute is necessary for open science, and AI centralization makes it urgent. 2) the open AI ecosystem is better than most think, and is ready to carry serious science and safety work. 3) what the ecosystem sorely lacks is infrastructure: sovereign and secure compute. 4) nothing is stopping us from providing it, especially now, with the wealth that AI windfalls are strating to unlock.

AI centralization

I see four critical centralization bottlenecks hobbling science in the AI era: institutions, geography, research agendas, and compute itself.

Institutions

Last year, only 4 of 102 notable models released their training code, and industry produced over 90% of these models, while only 2% came from academia. That this can come with access issues became clear in June when Anthropic was ordered by the Commerce Department’s export-control directive to restrict two new models. Mythos 5 had never been publicly released at all and Fable 5 went entirely offline for about three weeks while Anthropic debated this takedown with the government.

It looks like, at least partially, the govt and Anthropic had real concerns about real risks. But when Fable went down, every project that was built on it lost access, without notice. This is a structural property of closed models. When openness at the frontier becomes unreliable for the ecosystem as a whole, it becomes a safety issue. Open weights are what let outside groups red-team AI models, reproduce evaluations, do interpretability work without relying on permission, and, as we saw with the Hugging Face incident, defend systems from other AIs. More eyes on the ball find more flaws, and flaws keep being found long after models are deployed in the wild.

Geography

The US hosts ~75% of global AI supercomputer performance. The entire EU hosts under 5%. The talent, though, is global even if the infrastructure isn’t: 38% of the world’s top AI researchers were educated in China, versus 24% in the US, yet the US employs 59% of them. I’m pretty surprised, every time I go back to Germany, by how much talent is sitting there underused.

In 2025, Bay Area companies captured $126 billion of the $211 billion in global AI venture funding; 60%, on 22% of deals. OpenAI and Anthropic between them took 14% of all global venture investment, across every sector.

No particular government needs to be hostile for geographic centralization to be a problem. One jurisdiction means one regulatory regime, one political cycle, one epistemic view, and one memetic climate. Distributed research capacity is a hedge against the blind spots of any single community. Well-intentioned decisions in one capital or one lab shouldn’t determine whether a whole field can independently evaluate or build on the most consequential technology in history.

Research agendas

Then there’s intellectual consolidation. When a handful of institutions fund the majority of the work, the field converges around their priorities. The share of US AI PhDs heading into industry went from 21% in 2004 to about 70% in 2020. With frontier labs now paying multiples of a professor’s salary, academia stops being an attractive choice for many researchers. And past a certain scale, the work simply cannot be done on a university cluster.

Clearly, industry’s influence on AI research is growing. But is that a problem? The scientific output of the frontier labs includes some of the best work of the decade: DeepMind’s AlphaFold won a Nobel Prize and, open-sourced, turned up in 35,000 papers, with a structure database used by researchers in 190+ countries. Anthropic’s launched the Claude Science program. OpenAI is giving free frontier-model access to 10,000 researchers and scaling to 100,000.

This is great, and I want to see more of it. But frontier labs can’t be the only game in town. Science has historically been a more or less open-source process. And there is a real risk of intellectual narrowing here: industry-funded AI research is measurably less thematically diverse and more influential than academic or independent research.

Thinking and compute are becoming more and more intertwined. Whoever controls access to the latest AI models gets to pick, in some ways, who and what gets thought about. If the only way to do good thinking is sitting inside a frontier lab, we significantly reduce research agenda diversity.

Philanthropic AI safety funding was around $100-150 million a year in 2025, three orders of magnitude below private AI investment. The entire global brain-computer interface field raises about $1 billion a year, and longevity biotech is similar. In the coming compute-dominated era, scientific fields that don’t map onto a product roadmap will starve, comparatively, if left to their own devices. A healthy scientific ecosystem needs many bets placed by and on lots of different people, and right now nearly all the resources sit with five companies.

Compute

Five companies hold 71% of the world’s compute against a $30 million a year US public AI program. And the gap is widening: hyperscaler spending plans reach $745 billion this year, and frontier training costs go up about 2.4x every single year, with runs expected to reach billions of dollars by 2027. Concentration on this scale comes with risks I don’t think we take seriously enough.

For one, there is the risk of exposure. Do you know anything about the legal structure governing the AI you depend on? Most of us don’t really. Every prompt, including trade secrets, patient data, unpublished scientific results and confidential drafts, runs on hardware owned by someone else, logged in jurisdictions where governments can demand disclosure without your knowledge. And they frequently do: hundreds of thousands of National Security Letters have gone to companies before, demands for user data, usually with gag orders attached. The security community spent a decade learning to secure the data, verify the hardware, and decentralize the protocols. But now, our most sensitive thinking runs mostly on rented hardware in reach of a handful of governments whose internal workings we can’t always trust.

Then, there is price risk. After two years of falling GPU prices, the market has turned. In five months, H100 rental rates rose about 40%, and kept climbing. Dwarkesh Patel argues this is the start of something much larger: compute supply grows around 3x a year, while labs aim for 10x in revenue. When the cognitive work of models becomes economically valuable, the price of a GPU will get benchmarked against the human labor it replaces. On that logic, an H100 “should” rent for something like $250k a year. A sum frontier labs can pay, while academic groups and independent safety researchers mostly cannot. Nobody needs to decide to exclude independent researchers when the market does it automatically. Ajeya Cotra’s 80,000 Hours interview adds the uncomfortable timing: if there’s a crunch window, say twelve months between automated AI R&D and truly dangerous systems, that’s exactly when labs may redirect every marginal GPU into the recursive loop, and exactly when independent evaluation and alignment work is needed.

Finally, there is the risk of access. In his essay “Cut Off”, Anton Leicht sketches where the current trajectory seems to lead: worries about security, distillation, and government oversight all converging on tiered access, where security agencies are first, trusted domestic firms second, and everyone else gets access through restricted product layers, far behind the frontier. We got a small dose of this in June. And nothing says it stops at models: a government that can restrict a model can restrict a datacenter. Nothing guarantees frontier access, or access to compute itself unless you own it.

A scientific community that runs on rented compute runs into exposure, price and access risk all at once. That is the arrangement almost every independent AI researcher in the world is currently in. So ‘cheap’ rented compute comes at a real cost.

Why the open ecosystem is better than you think

Fortunately, there has been a recent notable development: in the last eighteen months, open models have become good enough that a big part of AI science no longer requires access to the labs. For the first time, running serious research on hardware you own is actually viable. Here’s why:

Starting with capability, according to at least some benchmarks, the best open models aren’t that behind anymore. GLM-5.2 currently beats GPT-5.5 on a benchmark called SWE-bench Pro (though Claude Opus 4.8 still leads on most coding benchmarks) at roughly a sixth of the price. Kimi K3’s official numbers put it ahead of every closed model on BrowseComp agentic browsing, 91.2, against GPT-5.5’s 84.4, at about 60% of flagship API prices. OpenAI now ships open weights, too: gpt-oss-120b almost matches its o4-mini on core reasoning and runs on a single GPU. And exciting new players are entering the open frontier: Thinking Machines released Inkling, a 975B-parameter multimodal model with open weights this July, alongside a platform built around user control and an explicit safety case for open weights.

Costs are falling too. The price of inference for a fixed capability level has fallen somewhere between 9x and 900x per year, depending on the task. Getting GPT-4-level performance on PhD-level science questions has gotten 40x cheaper every year, and for research, small open models often win. A fine-tuned 27B Gemma beat Claude’s Sonnet 4 by 60% on a specialized clinical task at 10-100x lower cost. Measured in cost per discovery instead of cost per token, open models can come out ahead. And capability that is released at the closed frontier often becomes runnable on a single consumer GPU within six to twelve months. If the timelines are ‘long enough’ to allow for open AI efforts to build up on each other’s wins, there’s a real chance open models end up a viable competitor.

But there’s a catch: a shrinking lag may stop mattering if capability jumps become larger. If frontier models start meaningfully accelerating frontier research, aka ‘the recursive loop’, then being a few months ahead could compound to a capability difference that equals being years ahead before. Suppose even a small lead becomes decisive. That makes the case for independent research capacity stronger, not weaker. A world where one or two actors hold a compounding lead is exactly where everyone else needs standing infrastructure to evaluate, verify, and check their claims. You don’t necessarily need to match the leader to audit the leader, but you do need capable models and you do need compute.

And open ecosystems compound too, just through breadth rather than depth. Every open release becomes a floor that thousands of independent teams build on, and their improvements are fed back into the commons. Linux started years behind proprietary Unix and now runs nearly every supercomputer and cloud VM on Earth. The open release of AlphaFold generated OpenFold, and suddenly we had an entire ecosystem of structural-biology tools. After Meta closed down its ESM protein-modeling team, the weights were public, so the research outlived the project; the team went on to found EvolutionaryScale, and kept building. If one lab compounds its lead, an open ecosystem compounds its floor, and over time that’s hard to beat.

It’s not just science that can benefit from openness but safety and security, too: When a fully autonomous AI cyberattack hit Hugging Face, and a frontier API model refused to analyze the malicious payloads, the responders switched to an open-weight model on their own hardware that processed 17,000+ attack logs and reconstructed the breach in hours, with no attacker data ever leaving their environment. Their post-mortem points to the central issue: defenders in attacks like this face an asymmetry problem, as attackers run unrestricted models while defenders get guardrails and account flags.

Their advice: ensure you have a vetted and capable model you can run on your own infrastructure, before an incident. That is very much in line with Vitalik’s d/acc philosophy of differential and decentralized acceleration of the tools and enabling the infrastructure that help defenders specifically.

Open vs. closed: why d/acc

Foresight has been in the openness business for a while; our co-founder Christine Peterson coined the term “open source software” in 1998. Open source won its first era because it was distributed and inspectable. Whether it can compete in the AI era is genuinely still open.

I already touched on how centralized AI can go wrong: a few actors end up with visibility into everyone’s thinking and a veto over everyone’s access. Any blind spots they have get embedded into systems everyone must use, and the whole arrangement has single points of failure. But open AI comes with its own failure modes. The obvious one is that if powerful AI is open to everyone, it’s also open for misuse by anyone. A more subtle one is that maintaining an open ecosystem is a continuous battle: a well-funded player that’s closed will always try to absorb or out-compete an open ecosystem unless it’s actively defended. And then there’s the issue of collective action: a thousand independent actors will have a much harder time coordinating to add safeguards or agreeing to provide some benefit to humanity. These are real tradeoffs between open and closed paradigms.

Which is why the d/acc paradigm (decentralized, differential, defensive acceleration) matters: it takes both sides seriously. Don’t open everything indiscriminately, don’t centralize everything defensively. Instead, differentially accelerate the technologies that favor defense, privacy, and distributed power, with security built in from the start. This is the frame Foresight’s grant-making runs on, and it’s the design brief for our AI Nodes.

Our AI Nodes: funding, community, and most importantly compute

In April 2026, we launched the first two AI Nodes in San Francisco and Berlin. Each site provides three things: funding for open AI science and safety research, a physical hub where this work can happen, and a local compute cluster, so researchers working on privacy-preserving and secure AI aren’t dependent on hyperscaler infrastructure.

We deliberately designed the Nodes to map onto the four bottlenecks above:

1. Institutions: we are happy to support independent projects even if unaffiliated with popular organizations with the expectation that results are shared openly.

2. Geography: the Berlin Node is one of few independent homes for open AI safety research in continental Europe, an alternative jurisdiction for a field that badly needs one.

3. Research agendas: our grants fund scientific fields outside the current commercial research agenda.

4. Compute: every Node runs its own cluster, private and local by default, providing researchers with more sovereign access to compute.

Here is what that looks like in practice:

The grants, roughly $3M a year, go to the underfunded end of the science portfolio: AI for secure AI (automated red-teaming, formal verification, hardening the digital commons), private AI (cryptographic ML, confidential compute, open hardware that closes off supply-chain backdoors), decentralized and cooperative AI (local agents that represent their users, multi-agent coordination, cryptographic protocols), AI for better epistemics (forecasting, collective sense-making), and AI for science across longevity bio, frontier neuro, and molecular nano. Almost none of this work sits on a commercial product roadmap but all of it depends on compute.

The physical hubs exist because a project needs so much more than just funding: people need a community, introductions, help hiring. We started in the Bay Area because it’s close to the AI talent and it’s our home turf, and in Berlin because there’s a lot of appetite for privacy-preserving work but still too little community focused on the existential AI part. A Node is basically a social clubhouse with deep work space, coworking days, happy hours and salons (recent guests include Robin Hanson of George Mason, Seb Krier of Google DeepMind, and Isaak Freeman of Capable). Each Node also anchors an annual technical workshop. The Secure & Sovereign AI Workshop ran in Berlin in July: around 60 curated researchers, engineers, cryptographers and funders, a $10k prize for the top project, and the invitation to stick around to sprint together for a few weeks. The AI for Frontier & Meta Science Workshop follows in San Francisco on September 29-30.

As for the compute onsite: each cluster can run multiple frontier open models at once, supports small-scale training and is freely provided to Node projects. Our Node clusters are extremely humble compared to a hyperscaler datacenter, and that’s ok. They don’t need to train large frontier models yet. A model a few months off the frontier, running on hardware we own, can often help keep independent evaluation, safety and science research somewhat alive.

Our community compute model is certainly a step towards sovereignty when compared to Europe’s ‘sovereign AI’ efforts that now try to stand up govt-run national efforts vis-à-vis US dependency. But the true logical endpoint of the ‘sovereignty’ spectrum would be a user-owned compute node at one’s house. Given how fast frontier capability reaches consumer hardware, this is less sci-fi than it sounds. The more local one can go, in ownership, in jurisdiction, in key control, the better - at least in theory. The practical benefit of a community cluster is, of course, that we can put idle compute time to use across the community. Different use cases will require different compute setups.

Either way, if AI goes as far as many of us expect, most value creation shifts from labor to capital, and the capital that matters now is compute: the one input one can accumulate and directly convert into cognitive work at scale. Whoever owns the compute owns the hardware doing the work; everyone else rents access to the future. We want everyone to own a share of the singularity, and in a compute-as-capital world, that means diversifying compute ownership itself: a box at your house, a cluster your community governs, a Node your neighborhood runs. Broad compute ownership is to the intelligence age roughly what broad land and home ownership were to earlier ones: the difference between a society of stakeholders and a society of tenants.

From Node to network: A note to funders

Two AI Node buildings with a little bit of compute don’t fix the structural problems in this article. But a network might. A group in South Africa is already getting an independent AI Node going. Groups in London, Paris, Boston and China have expressed serious interest in launching their own. Imagine independent compute collectives for non-profit safety and science work, spread across places, ideally each with a research community and a grant program attached to grow dedicated, cross-jurisdictional research cohorts. Universities, philanthropies, and regional science funders could each stand one up. So could the labs themselves, at arm’s length. And for the user-owned end of the spectrum, we’re launching an RFP to build the compute node for your house.

A network is more than the sum of its nodes: ideally, every new AI Node joining the growing network would keep its own governance and character, but where aligned, the Nodes might share compute, talent, and research agendas across jurisdictions. That’s the resilience the field is missing: if one hub shuts down or gets captured, the work can move elsewhere.

Directing even a small slice of the AI boom’s windfalls into this independent layer is cheap insurance for everyone. That includes the labs: their safety claims become more credible when outside infrastructure exists to verify them, and the value of their open releases increases when there is a commons equipped to build on them. So if you’re an AI-oriented funder or builder thinking along these lines, let’s compare notes so we can lower the barrier to entry for projects seeking to provide AI safety and science-critical compute.

If you want to join a Node, apply at foresight.pub/ai-nodes. If you want to launch one in your city, fund someone to do so, or just chat about shared, independent compute infra, let me know at allison [at] foresight.org. The next few years may decide whether the infrastructure gets built while it is still possible.

Note: Special thanks to science writer Linda Petrini for the editorial rewrite and to the Foresight team for feedback and support. This article was developed with some AI assistance.