The Gap Closed: DeepSeek, Open Weights, and the End of the Moat
DeepSeek-R1 arrived in January 2025 with open weights and frontier-matching reasoning. The closed-model moat is thinner than anyone admitted.
On January 20, 2025, a Chinese AI research lab most people outside the field had never heard of dropped a model under an MIT license that matched the performance of OpenAI’s o1 on serious reasoning benchmarks. Not approximately. Not “close enough.” Matched it, on AIME, MATH-500, and SWE-bench Verified, according to DeepSeek’s own published results.
The model was DeepSeek-R1. The lab was DeepSeek, a Hangzhou-based company founded by Liang Wenfeng, who also founded the quantitative trading firm High-Flyer. The license was MIT — use it commercially, modify it, redistribute it, do whatever you want with it.
The tech industry had a minor crisis. NVIDIA’s stock dropped significantly on the news, because DeepSeek’s training efficiency claims implied you might not need as many H100s as everyone assumed. The crisis was partly overblown in the short term — the training efficiency gains were real but did not immediately obsolete anyone’s infrastructure plans. But the deeper implication is lasting, and I think it has not fully settled into the industry’s mental models yet.
The closed-model moat is genuinely thinner than it appeared.
What DeepSeek-R1 Actually Is
DeepSeek-R1 is a 671 billion parameter model with a mixture-of-experts architecture — 37 billion parameters are active per token, which is why it runs more efficiently than its total parameter count suggests. It is a reasoning model, meaning it generates extended chain-of-thought before producing a final answer, similar to OpenAI’s o1 family.
The training approach is what makes it technically interesting. DeepSeek used reinforcement learning — specifically an algorithm called Group Relative Policy Optimization (GRPO) — as the primary driver of reasoning capability, after an initial cold start with supervised fine-tuning on chain-of-thought examples. The model learned to self-verify, correct errors mid-reasoning, and develop reasoning strategies through RL rather than through manually curated reasoning data. The precursor R1-Zero, trained purely with RL from scratch, showed that advanced reasoning behaviors can emerge from RL without any human-labeled chain-of-thought data — a result that was genuinely surprising to many researchers.
The weights are publicly available on Hugging Face under MIT license. You can download them and run them today, on your own infrastructure, without any API call, without any license fee, without any usage tracking.
Liang Wenfeng’s Framing
In an interview originally conducted in July 2024 and widely circulated after R1’s release, Liang Wenfeng said something I keep returning to: “We’re done following. It’s time to lead.” He also said that DeepSeek “never intended to be a disruptor; it just happened by accident” — that they were simply following their own research agenda, pricing based on their actual costs with a modest profit margin, not trying to undercut anyone.
His description of DeepSeek’s purpose was unusually direct: “Because we believe the most important thing right now is to participate in global innovation… Our goal isn’t quick profits but advancing the technological frontier to drive ecosystem growth.”
This reads differently from how Silicon Valley AI labs typically speak. There is no “safety first” language, no “democratizing AI” marketing speak, no fundraising narrative. It reads like a research lab that stumbled into a product because the research was good.
Whether or not you take Liang at face value on the accident framing, the output is real: a frontier-class reasoning model, open weights, MIT licensed, from a team that was not on most people’s radar twelve months earlier.
Yann LeCun Has Been Saying This for Years
Yann LeCun, Meta’s Chief AI Scientist, has been the most prominent and persistent advocate for open-source AI in the industry. In 2023, he stated flatly: “Open source AI models will soon become unbeatable. Period.” He has argued consistently that the future of AI is open — that the dynamics of open-source software, where community contributions compound over time, will ultimately outpace closed systems.
LeCun’s position is principled but also strategic: Meta releases its frontier models (the Llama family) openly, which is a competitive move against OpenAI and Google as much as an ideological stance. If the underlying base models become commodities — freely available, widely fine-tuned, integrated into every framework — then the companies that depend on proprietary base models as their moat are vulnerable.
DeepSeek-R1 is the strongest single piece of evidence for LeCun’s thesis. A Chinese research lab, with reportedly far fewer resources than the leading US labs and operating under export restrictions that limited access to the most advanced NVIDIA chips, produced a frontier-class reasoning model and gave it away. The “advantages” that were supposed to create durable moats — massive GPU clusters, proprietary training data, huge research teams — did not prevent this from happening.
What the Moat Actually Consists Of
I want to be careful here not to overcorrect. The closed-model labs are not irrelevant. OpenAI, Anthropic, and Google are still producing the absolute frontier models, and for many enterprise use cases, that frontier matters. But let me be precise about what the moat actually is and is not.
The moat that exists:
- Proprietary training data — closed labs have negotiated data licensing deals, RLHF pipelines built from millions of user interactions, and safety fine-tuning datasets that are not public. This matters for model quality and alignment.
- Research iteration speed — frontier labs are running continuous experiments at a pace that open-source replication lags by months. The gap at the absolute cutting edge exists.
- Infrastructure for serving — OpenAI, Anthropic, and Google have the engineering to serve hundreds of millions of queries daily with sub-second latency. Running the weights yourself at that scale is non-trivial.
- Trust and compliance — for regulated enterprise deployments, a vendor relationship with auditable SLAs and compliance certifications matters. Open weights alone do not give you this.
The moat that does not exist:
- Architecture monopoly — transformer architecture is public. The RL training techniques DeepSeek used are described in published research. The ideas are not secret.
- Capability monopoly — DeepSeek-R1 proved that frontier-class reasoning is not limited to the three or four labs with the largest GPU clusters.
- Switching cost moat — if you are using a model through an API, switching costs are low. Models are not applications. The stickiness in AI is in the application layer, not the model layer.
The Implication for Builders
I have spent a lot of time thinking about what open weights mean for what I am building with Zero. The honest answer: they are mostly good for the ecosystem.
If serious reasoning capability is freely available under an MIT license, then the limiting factor is not the model. The limiting factor is domain knowledge, user experience, data, and integration. That is where builders should be competing. That is a competition where a small, focused team can beat a large, unfocused one.
Open weights also change the security and privacy calculus for enterprise AI. Running DeepSeek-R1 locally — on your own hardware, with your own data, with no API calls leaving your environment — is now a real option for organizations that have serious data governance requirements. That unlocks enterprise use cases that API-dependent solutions cannot serve.
The concern I hear is about geopolitics — a Chinese lab releasing a powerful open model raises questions about what is in the training data, what backdoors might exist, whether using it creates risks. These are legitimate concerns that I would not dismiss, especially for sensitive applications. But they are separate from the technical and economic implications of the capability gap closing.
The gap closed. A team most people had not heard of, in a country facing export controls on advanced chips, produced a frontier reasoning model and gave it away. That happened. Whatever you think about DeepSeek’s motives or governance, the implication for the structure of the AI industry is real: the assumption that capability requires closed systems and massive resources has been falsified.
Yann LeCun has been saying open source will win for years. He might be right, but not because of ideology. He might be right because, it turns out, the best ideas spread, and the best models get built by people who are genuinely curious, not just capitalized.
