01 / The diffusion race
Who is the world
building on?
Most countries will never train frontier models. Nearly all will deploy them. Explore three signals of where adoption is heading.
Choose a signal to explore.
Figures reproduced from the report.
The developer base is shifting.
Share of new Hugging Face fine-tunes and adaptations of open models
No intermediate values are implied. Source: The ATOM Report, as cited in the report.Source notes ↗
Open models are running real products.
Chinese share of open-model requests on OpenRouter
of token volume on Vercel AI Gateway was handled by open-weight models of all origins in August 2026.
Final bar shows the 70% lower bound. Sources: ATOM and Vercel, as cited in the report.Source notes ↗
The capability gap is measured in months.
The most capable open models versus leading closed systems
of average lag
The report argues that open and closed models are complements. Wider access allows developers to evaluate, adapt, and deploy capable models on infrastructure they control.
Source: Epoch AI, as cited in the report.Source notes ↗
02 / The American answer
Three lines of effort.
One open ecosystem.
Choose a priority. Explore the actions that turn American AI leadership into adoption.
Build the AI ecosystem America can win.
Expand access to compute, capital, and shared infrastructure so researchers, startups, and government can inspect, adapt, and trust American models.
1aResource the Genesis Mission
Fully resource the Genesis Mission so that it can benefit open model ecosystem development. DOE holds the purchasing authority and has launched the open models program; however, it lacks the dedicated resources to run either at scale. Alternatively, an open-model mission can be assigned to NNSA and funded through its existing authorization and budgetary processes, provided any designated laboratory or laboratory consortium receives an explicit mandate and dedicated funding together.
1bFinance compute for open-model research
The Department of Energy should explore pathways to finance, or guarantee financing, for large-scale compute infrastructure dedicated to open-model research. DOE’s existing credit and loan programs are limited to energy supply projects, and no federal agency has a dedicated financing authority for open model compute infrastructure.
2Streamline access to compute
Streamline access to Genesis Mission compute. Compute access should require a single justification, with consolidated process steps and rolling applications.
3Fund fully open-source science models
Fund fully open-source models for scientific fields critical to Americans and national security missions. Models that release training, data documentation, evaluation results, and reproducibility materials should be funded through academic, nonprofit, and national laboratory programs, with full release as a condition of support. Give priority to research where a model’s output is a claim others must be able to reproduce, and to missions where an agency must know how a model was trained.
Turn leadership into adoption.
Put American AI to work in the agencies, schools, hospitals, and small businesses where American life happens. Build confidence through evidence of what works.
1Fund justified-confidence tooling
Fund justified-confidence tooling as a capability distinct from safety evaluation. Institutions need a way to establish that a model works for their purpose and keeps working. NSCAI identified this requirement in 2021; the tooling was never built.
2Share deployments that work
Establish and fund an extension capability built around aggregating and sharing working deployments rather than only delivering advice. The product is a transferable record of results and, where it can be assembled, a working package the next institution can run.
3Measure cost, speed, and mission outcomes
Measure open-model deployments on already-existing agency scorecards. Add cost per task, time to complete, and mission outcome for open-model deployments to existing portfolio scorecards and FITARA reporting, and publish the results so the record is available to the next agency and to the extension network.
Govern AI without making it inaccessible.
Address risk according to a model’s capability, how it is deployed, and its impacts. Give institutions the standards, expertise, and reporting infrastructure to do that well.
1aEvaluate risk beyond openness
Avoid any policy approach that treats openness as a proxy for risk. Hold open and closed models to the same evaluation and provenance standards when used for similarly sensitive tasks. Requirements should be proportionate to capability and deployment context. Provide shared testing resources and technical support so smaller developers can meet these standards without being priced out.
1bFund independent research on real risks
Fund independent, published research on the actual, not suspected, risks of open models. Research should be supported through NIST, the national laboratories, or NSF into how open models fail in deployment, how their safeguards hold up under attack, and what marginal capability a release adds beyond what is already available. Rules keyed to openness should rest on that evidence or not be written.
2Equip sector regulators
Direct sector regulators to develop frameworks for identifying highly consequential AI use cases in their own domains, using SCSP’s Framework as a model, in place of rules keyed to release category, and fund the technical expertise inside those regulators that applying them requires.
3Build a national risk and harm reporting capability
Establish a standardized capability for reporting, adjudicating, and disseminating AI risks and harms, with common definitions, a national database, and the staff to run it.
03 / Inside the report
The full picture,
on your terms.
Read the argument, understand the terminology, or go straight to the source.
Executive summary
The United States can lead at the technological frontier and still lose the most consequential part of the AI competition, if developers, firms, and governments at home and abroad build on another country’s models and infrastructure. Open models will help determine which technology ecosystem becomes the global default. Building on decades of open-source software development, openness in AI enables others to inspect, adapt, and improve shared technology, broadening participation and accelerating innovation. Openness spans the spectrum, from publicly available model weights to broader access to code, training data, and methods, with licenses determining how these resources can be used and shared. Open-weight models, systems whose learned parameters (i.e., the billions or even trillions of numerical settings a model adjusts during training to shape how it responds) are publicly released, allow users to download, evaluate, customize, and deploy them on infrastructure they control. History shows invention alone does not secure lasting leadership. For example, although the United States led the world in wireless research, Huawei won the diffusion race and connected much of the globe by out-competing on cost, financing, and speed of deployment.
China is implementing a similar strategy to AI. Its developers are using capable, low-cost, downloadable models to expand their footprint across the global AI ecosystem. The advantage does not come from openness itself, but the adoption it enables: lower costs for initial adoption as well as for switching models, the ability to run a model on one’s own hardware, customization, a larger base of derivative models and tools, and a growing community with reasons to develop further within the same technical ecosystem. What matters about these models is not how they score on benchmarks, the standardized tests the industry uses to assess a model’s capabilities. It is that China’s open models are becoming the foundation on which developers are building precisely because they are open.
In Washington, meanwhile, open models are discussed as though all of them are Chinese models, and Chinese models are discussed as though they all are unsafe. Neither follows from the other.
While there are trust and security concerns regarding Chinese open models, the American response should be to strengthen its own open model ecosystem alongside its proprietary frontier laboratories. The two ecosystems are structurally different. Leading U.S. frontier laboratories monetize access through subscriptions, enterprise products, and managed platforms; that commercial engine, rather than large public investment, funds the next generation of models. But a strong American strategy should also support open model developers and the researchers, startups, and businesses that build on their work. The United States needs a strategy that builds on its distinctive strengths: invest in a thriving open AI ecosystem by expanding access to compute, capital, and shared infrastructure; accelerate adoption of American AI across society; and govern risks based on what models can do, how they are used, and their impacts and outcomes, not simply whether their weights are public.
What do we mean by “open”?
Openness spans a spectrum: from publicly available model weights to broader access to code, training data, and methods. Licenses determine how those resources can be used and shared.
Open-weight models make their learned parameters available so users can download, evaluate, customize, and deploy the model on infrastructure they control. That does not automatically mean the training data, code, or full development process is open.
The report also recommends funding fully open-source models with training and data documentation, evaluation results, and reproducibility materials in scientific and national security fields.
Sources & methodology
The interactive figures reproduce indicators cited in the report and the supplied report design:
- The ATOM Report (2026): Hugging Face fine-tunes and adaptations; OpenRouter open-model requests. Chinese laboratories collectively accounted for roughly 70% of new derivative models by February 2026.
- Vercel AI Gateway (September 2026): open-weight models handled 56% of token volume in August 2026.
- Epoch AI (2026): an average capability lag of roughly four months between the most capable open models and leading closed systems.
These measures describe different platforms, populations, and periods. They should not be combined into a single market-share estimate. Bars show only the stated observations; no missing months are interpolated.
About this work
On September 21, 2026, SCSP convened industry, academia, and government for a half-day workshop, The Open Frontier: Winning the AI Diffusion Race, under the Chatham House Rule. SCSP staff prepared the report and recommendations, drawing on many, though not all, of the insights from that discussion.
SCSP is a non-partisan, non-profit initiative focused on strengthening America’s long-term competitiveness as AI and other emerging technologies reshape national security, the economy, and society.
The open frontier / SCSP