The United States still produces some of the world’s most powerful artificial intelligence systems. Price is becoming the awkward part.
Chinese developers including DeepSeek, Alibaba and Moonshot AI have pushed capable open-weight models into the global market at prices that American businesses find difficult to ignore. Developers can download many of these models, modify them and run them on their own infrastructure rather than paying repeatedly for access to a closed platform.
That has left a small group of US startups chasing a difficult target: build American open-weight AI models that are powerful, affordable and genuinely useful outside a research demo.
Arcee AI Takes a $20 Million Swing
Arcee AI has become one of the more visible companies in that race.
The Silicon Valley startup reportedly spent around $20 million developing Trinity Large, an open-weight foundation model trained using 2,048 Nvidia Blackwell B300 GPUs. That budget is modest by frontier AI standards, where individual training runs can cost hundreds of millions of dollars before post-training, staffing and infrastructure enter the calculation.
Trinity Large uses a sparse mixture-of-experts architecture. The full model contains roughly 398 billion parameters, but only about 13 billion activate for each token. This design can cut inference demands while preserving the capacity of a much larger model. Arcee says it trained the system on approximately 17 trillion tokens.
The company released the model weights so organisations can inspect, fine-tune and deploy the technology on their own terms.
That is the pitch, anyway. An American-built model without the recurring cost and dependency that comes with relying entirely on a proprietary API.
Chinese AI Models Have Already Won on Price
American AI companies are not suddenly interested in open models because the idea is new. They are reacting because Chinese developers have made the economics hard to dismiss.
Models from DeepSeek, Alibaba’s Qwen family and Moonshot AI’s Kimi line have gained attention by offering strong reasoning, coding and language performance without the premium pricing attached to many leading US systems.
For startups processing millions of tokens a day, a small difference in API pricing quickly becomes a large operating expense. A cheaper model does not need to win every benchmark. It only needs to handle the company’s actual workload well enough.
That is why Chinese models have started appearing inside American businesses and research environments, even as Washington continues tightening restrictions around Chinese technology.
The uncomfortable reality is simple: companies usually talk about national AI leadership until the infrastructure bill arrives.
Open-Weight AI Gives Companies More Control
Closed AI platforms remain convenient. A company connects to an API, sends a prompt and receives an answer without maintaining expensive computing infrastructure. The trade-off is dependency.
Providers control the pricing, usage policies, model updates and availability. A model can change without warning. Features may disappear. Costs can rise once a product reaches scale.
Open-weight systems give enterprises another route. They can download the model parameters, run the technology inside their own environment and fine-tune it using specialised data.
This matters for banks, government agencies, healthcare providers and companies handling sensitive intellectual property. Keeping prompts and data inside controlled infrastructure can be more attractive than sending everything to an outside service.
Open-weight does not always mean fully open-source. Developers may release model weights while withholding training datasets, code or detailed information about how the system was created. Even with those limitations, the deployment freedom remains valuable.
Poolside and Reflection AI Join the Push
Arcee is not working alone.
Poolside, Reflection AI and Thinking Machines Lab are among the American companies trying to develop more competitive open or customisable AI systems. Each is approaching the problem differently, but the pressure comes from the same place: US developers need alternatives that can compete on cost as well as raw intelligence.
Poolside has focused heavily on coding models and software-development agents. Nvidia now offers access to one of its efficient mixture-of-experts coding models through the NVIDIA NIM platform, showing how chipmakers are helping bring open and specialised systems closer to enterprise deployment.
Building the model is only part of the job. These companies also need reliable inference infrastructure, developer tools, enterprise support and enough financing to survive while larger rivals spend aggressively.
Investors Still Prefer Closed AI Platforms
The funding problem has not disappeared. Venture investors have poured enormous sums into closed-model companies such as OpenAI and Anthropic. Those businesses can charge customers for every API call, subscription or enterprise seat.
Open-weight models are harder to monetise in the same way. Once the weights become available, users can host the system themselves. Competitors can fine-tune it. Cloud providers can offer access without sending every dollar back to the original developer. That scares some investors.
It also creates a strange gap in the American AI market. Companies say open models matter for national competitiveness, yet the easiest investment returns still appear to sit inside proprietary platforms. China has been much more aggressive about treating widely available models as a way to build influence, developer adoption and technical standards.
Nvidia Has a Reason to Back Open Models
Nvidia sits in a different position. The company can benefit whether customers choose a closed model, an open model or something built internally. All of them require computing power.
Nvidia supplied the Blackwell hardware used for Arcee’s Trinity Large training run and has promoted the project as an example of how smaller companies can build frontier-level open systems. The chipmaker says Arcee combined Blackwell Ultra GPUs with Nvidia networking and inference software to reduce token costs during deployment.
Nvidia has also released its own models and datasets while contributing to open-source projects across the AI ecosystem. The company describes open platforms as a way to expand adoption and give enterprises more control over deployment. More models usually mean more GPU demand. Nvidia does not need one AI laboratory to win. It needs everyone to keep building.
Washington Now Sees Open AI as Strategic Infrastructure
The debate has moved beyond startup competition. US policymakers increasingly see open and affordable AI as part of a broader national-security question. If developers around the world build their products using Chinese foundation models, those systems could shape technical standards, research habits and future AI infrastructure.
Banning access would not solve the underlying problem. Developers choose these models because they work and because they are cheap. The more durable response is to produce American alternatives that developers actually want to use.
The US Department of Energy is expanding work around AI for scientific discovery through its Genesis Mission, combining national laboratory data, supercomputing resources and private-sector technology. The department has described AI leadership as central to energy research, national security and critical technologies. Arcee has also been linked to the department’s scientific AI efforts, giving the small startup a potential role well beyond commercial chatbots.
Cheap AI Could Matter More Than the Smartest AI
The US remains strong at the top of the AI market. OpenAI, Anthropic and Google continue building highly capable proprietary systems backed by huge amounts of capital and computing infrastructure. The global market may not run entirely on the smartest available model.
Most businesses do not need a system that tops every benchmark. They need something reliable enough to summarise documents, write code, automate support requests or analyse company data without destroying the technology budget. Chinese developers understood that early. They treated efficiency and accessibility as core product features rather than technical compromises.
American open-weight AI models now have to catch up on that ground. Arcee’s Trinity Large does not settle the race. Neither will one release from Poolside or Reflection AI. But the direction is becoming clear. The next stage of AI competition will not only measure who can build the most powerful model. It will measure who can make intelligence cheap enough for everyone else to use.
Sources
- The Wall Street Journal – The Race to Build an American Alternative to Cheap AI From China
- Arcee AI – Trinity Large: An Open 400B Sparse MoE Model
- Hugging Face – Arcee AI Trinity Large Base
- NVIDIA – Arcee AI Trains Frontier Open Model on Nvidia Blackwell
- US Department of Energy – Artificial Intelligence
