A post crossed my feed claiming that DeepSeek is now funded directly by Chinese government VC, that Chinese models will keep being better, faster, cheaper, and open, and that the United States is losing because its models are export controlled, expensive, and closed. The framing is punchy. Most of the underlying facts hold up. A few of them do not, and the ones that do not are worth fixing, because the corrected version is actually a stronger argument than the original.
Let me take the claims one at a time, with the numbers.
The state money is real, and the structure is the story
For its whole existence DeepSeek refused outside capital. Liang Wenfeng funded it through his own wealth and the profits of his hedge fund, High-Flyer. Even after the company went global, it turned away government-linked funds, reportedly to keep its decisions out of other people's hands. That posture broke in 2026. By April investors were talking about a 300 million dollar round at a 10 billion dollar valuation. Within weeks the valuation talk had climbed past 20 billion, then toward 45 to 50 billion.
The round closed in mid-June 2026 at more than 7.4 billion dollars. Tencent and the battery maker CATL were the largest outside backers, reportedly putting in 10 billion yuan and 5 billion yuan. NetEase and JD.com each contributed around 3 billion yuan. China's National Artificial Intelligence Investment Fund, a vehicle tied to the third phase of the state semiconductor fund known as Big Fund III, reportedly put in close to 1 billion yuan.
So when the bookmark says government VC funded DeepSeek directly, that is true. But the headline number is not where the action is. The commercial giants supplied most of the cash. The state fund's contribution was modest in dollars. What it bought was not modest at all. The commercial investors accepted a five year lock-up and surrendered their voting rights. The state fund kept direct equity, kept its voting rights, and accepted no lock-up. Read that twice. The companies with the most money agreed to be silent. The fund with the least money kept the vote.
That is not a normal venture deal. That is a strategic-asset designation dressed as a financing round. Beijing's stance shifted from DeepSeek shielding itself to DeepSeek being aligned with the state's call to harden China against American pressure. China has been telling some AI firms not to take US investment without government approval. The broader pattern backs this up: government-linked investors in China went from fewer than ten AI deals a year before 2018 to more than 140 in 2025, roughly a fifteenfold jump.
If you want the honest one-line version, it is this. The money was small and the control was large, and that is exactly how a government secures an asset it considers national infrastructure.
Cheaper and open are not in dispute. Better is
The claim that Chinese models are cheaper is well supported. DeepSeek's V4 reportedly delivers benchmark scores comparable to frontier US models at roughly fifty times lower API cost, using about ten times less computation per token. The company's old claim that it trained V3 for around 6 million dollars is disputed and should be treated as the company's own number, not a fact, but the efficiency strategy behind it is real and visible across Chinese labs. Baidu, Alibaba, and Tencent have leaned into lean, efficient, open architectures rather than chasing raw scale.
Open is also true, with one precise word. These are open-weight models, not fully open-source. The parameters are shared, often under MIT or Apache licenses, but the training data is not openly licensed. The point still stands. DeepSeek, Qwen, GLM, Kimi K2, and others ship weights you can download and run. By independent ranking, Kimi K2 Thinking is arguably the strongest open model in the world, the best model not made by OpenAI, Google, or Anthropic.
Now the word that does not survive contact with the data. Better. Stanford's 2026 AI Index found the performance gap between the best American and best Chinese models had collapsed to 2.7 percent, down from a range of roughly 17 to 32 points in May 2023. That happened while the United States spent about 23 times more on private AI investment, 285.9 billion dollars against 12.4 billion. A 23x spending gap producing a 2.7 percent quality gap is a stunning indictment of the spending. But 2.7 percent is not zero. As of March 2026 the top of the global leaderboard was still a US lab, Anthropic's Claude Opus 4.6 with an Arena score of 1,503, ahead of ByteDance's entry at 1,464. So the accurate statement is not that Chinese models are better. It is that they are at near parity and closing fast while costing a fraction to run. That is the more damaging claim anyway.
The US side of the slogan needs editing
Three words are attached to the American models in the post: export controlled, expensive, not open. Two of those need correction.
- Export controlled conflates chips with models. The controls run on hardware and on investment, on the export of advanced chip technology and on American money going into Chinese AI firms. US models themselves, the APIs and the open weights, are not broadly export controlled to most markets. And the policy is in flux, with late 2025 reporting describing conditional Nvidia chip exports being cleared again.
- Not open is simply out of date. OpenAI released open-weight models in 2025, the gpt-oss family. Meta's Llama line is open-weight. What is true is narrower: the American frontier leaders, the flagship Anthropic, OpenAI, and Google systems, remain closed. Say that and you are right. Say the US has no open models and you are wrong.
- Expensive is fair at the frontier. Anthropic closed a round at a post-money valuation near 965 billion dollars. OpenAI closed a round at an 852 billion dollar valuation. Pair those valuations with the roughly fifty times API cost gap and the contrast is real.
So the corrected American column reads: chips and investment are controlled, the frontier leaders are closed and very expensive, but open US models do exist. Less tidy than the slogan. Closer to true.
The part the cheerleaders skip
Open is not a free pass to win. A large number of companies cannot use Qwen or DeepSeek weights at all because they come from China, for procurement, security, and political reasons. And open does not mean uncensored. DeepSeek's R1-0528 update was flagged for tighter alignment with official Chinese positions. Download the weights, ask about Tibet or Tiananmen, and you meet the state in the output. Open weights and a constrained worldview can ship in the same file. That matters for anyone who thinks distribution alone settles the contest.
Where I land
I read this the way I read most American framing of the AI race, which is to say from outside it. The dominant Anglosphere story is that vast capital and the best single model equal victory. The evidence says the contest has quietly become closed-and-expensive against open-and-cheap, and open-and-cheap is the distribution model that usually wins over a decade. China already leads on several inputs that compound: a large majority of global AI patent filings, a rising share of publications, far higher industrial robot installation rates, and energy infrastructure to actually run all of this. AI talent migration into the United States has dropped sharply since 2017. None of that guarantees the outcome. The US still holds the single best model and far larger pools of capital, and benchmark parity is not the same as deployment dominance or trust.
But the original question, why would the US win with this strategy, deserves a serious answer rather than a dismissal. The honest answer is that pouring 23 times the money into a 2.7 percent edge, while gatekeeping the frontier and betting export controls will hold, is not obviously a winning play. It might buy time. It has not bought a moat. And the most telling fact in the whole story is not the price of a token. It is that the Chinese state put in the smallest check in the room and walked away holding the votes. When a government does that to a private company, it is telling you what it thinks that company is. Not a startup. A strategic weapon.