DeepSeek's Founder Gave the Most Honest AI Briefing Ever. It Killed Their Fundraise.

There's a leaked transcript making the rounds that does something rare in the AI industry: it tells the truth.

Liang Wenfeng, the founder of DeepSeek, sat down with investors on May 20 for a four-hour meeting. Someone recorded it. Someone transcribed it. Someone translated it. And then it hit the internet last week — right as DeepSeek was trying to close its second funding round at a $74 billion valuation.

The round is now on hold.

Chip export controls couldn't stop DeepSeek. Billions in American lobbying couldn't stop DeepSeek. But a 118-point transcript of one man speaking too honestly to a room full of investors? That did the trick.

What the transcript actually says

I read the whole thing. It's not a PR document. It's not a press release with "we're excited to announce" energy. It reads like a person who genuinely does not care about saying the right thing — which is probably why investors got nervous.

Here are the parts that matter.

The compute gap is real, and it's brutal. Liang told investors that to train a frontier model comparable to what US labs are building, he would need 200,000 Huawei 950 chips. Huawei gave him 16,000. That's 8% of what he asked for. He described Huawei's core problem as "insufficient production capacity" and said he expects the chip crunch to last at least three years.

Let me put that in perspective. US labs like xAI are reportedly running clusters of 200,000+ Nvidia GPUs. Meta is building a data center so large it needed its own nuclear plant. OpenAI's compute spend is estimated in the tens of billions annually. DeepSeek got sixteen thousand domestic chips. The gap isn't a few percentage points. It's an order of magnitude.

The talent gap doesn't exist. This is the quote that everyone latched onto: "The main gap between us and the United States lies in resources, while the disparity in personnel is minimal — there is virtually no difference, as we are essentially the same team of people, possibly from China."

That's not trash talk. It's not nationalist posturing. It's a guy who has worked with these people saying: the engineers are the same. The compute is not. DeepSeek proved this already — their models have repeatedly punched far above their weight class, delivering performance that rivals US frontier models at a fraction of the training cost. The reason isn't magic. It's because the people are genuinely good, and when good people don't have unlimited resources, they get creative.

Two billion dollars would be a great year. Liang said that if he could spend $2 billion on compute this year, "it would indicate that our procurement department has achieved outstanding performance." For context, Anthropic raised $61 billion in its last funding round. xAI secured $10 billion in a single raise. OpenAI has reportedly committed over $100 billion to compute infrastructure through Stargate and related deals. DeepSeek's founder considers $2 billion a stretch goal.

And yet, they're competitive. That fact alone should keep every US lab executive awake at night.

The strategy: win with one-twentieth of the compute

This is where the transcript gets interesting from a strategy perspective. Liang isn't delusional about the gap. He knows the math. But his plan isn't to match US spending dollar for dollar. It's the opposite.

"In the future, we need to rewrite this narrative: using only a fraction of their computing power while significantly shortening the timeline to just six months or three months — I believe this is our goal. Moreover, we could even surpass them in certain aspects."

He's describing algorithmic efficiency as a competitive weapon. If you can't outspend your opponent, you out-engineer them. DeepSeek has already done this repeatedly — their Mixture-of-Experts architecture, their training efficiency innovations, their cost-per-token pricing that broke the AI API market in early 2025. Every one of those moves came from a team that couldn't afford to brute-force anything.

The DeepSeek V3 model, released last year, was trained for roughly $5.6 million. That's not a typo. US labs spend more than that on electricity in a week. V3 rivaled or beat models that cost hundreds of millions to train. The next model, the R1 reasoning series, did the same thing to the reasoning model market — OpenAI's o1 was suddenly competing with a free, open-weight alternative that anyone could download.

The question is whether efficiency can keep closing a gap that's getting wider on the hardware side. Liang's own answer is measured: "achieving comprehensive superiority is unrealistic; yet in key areas where trade-offs are necessary, surpassing them may be feasible."

He thinks they can win in specific domains. Not everywhere. Not the full stack. But enough to matter.

Why the fundraise actually paused

The reporting from Bloomberg and Cyber Kendra is clear on the timeline. The transcript leaked, went viral on WeChat, links got pulled, and then investors were told the round was suspended. The valuation target was 480 billion yuan ($74B) on a raise of at least 10 billion yuan ($1.4B). That first round, which closed in June, pulled in Tencent and CATL for about $7 billion at a ~$50 billion valuation.

But here's the thing: the transcript doesn't contain anything that should scare an investor. Liang said the compute gap is real. Everyone already knew that. He said Huawei can't produce enough chips. Everyone already knew that too. He said DeepSeek's edge is efficiency, not raw power. That's been their entire brand since day one.

What actually happened, I suspect, is simpler. The transcript made the fundraise public before DeepSeek wanted it public. It revealed valuation numbers, strategy details, and specific hardware constraints that a company preparing for an IPO would rather disclose in a prospectus than in a viral WeChat thread.

Reuters reported last week that DeepSeek is eyeing Shanghai's STAR Market for a public listing. A leaked investor call that reveals your hardware constraints, your valuation targets, and your founder's philosophical worldview is not ideal preparation for that process. The pause isn't about the content of the transcript. It's about control of the narrative. When you're about to go public, the last thing you want is a four-hour candid recording of your CEO circulating uncontrollably on the Chinese internet.

The philosophy nobody expected

What makes this transcript genuinely unusual isn't the compute numbers. Those were confirmations of what people suspected. It's the worldview that catches you off guard.

Liang spent the first hour of the meeting not talking about revenue projections or market share. He talked about restraint. Specifically, he argued that the path to success in AI is to deliberately limit how much of the market you try to capture.

"AI represents an enormous scale — ultimately potentially accounting for as much as 10% of humanity's GDP. No one can monopolize this endeavor alone — you must share it with others; otherwise, you simply won't survive."

He frames open source not as charity or idealism, but as a survival strategy. If you try to lock up the value of something this big, the market will route around you. The only way to participate in a $10 trillion opportunity is to let others participate too. "Those who attempt to claim exclusive benefits will inevitably be left behind by history."

This is a fundamentally different logic than what you hear from American frontier labs. OpenAI, Anthropic, Google — their thesis is that building the best model creates a moat, and the moat creates value capture. You build the smartest model, you charge premium API prices, you lock in enterprise contracts, and the moat compounds. Liang's thesis is the opposite: there is no moat big enough for a market this large, and the companies that try to build one will be the ones left behind.

It's hard to say he's wrong. DeepSeek's models are free. They're open-weight. Anyone can download them, fine-tune them, and deploy them without asking permission. And yet DeepSeek is valued at $74 billion — up from $50 billion in their first round. You don't get to that number by giving everything away for nothing. You get there by becoming the infrastructure that everyone builds on, and then capturing a small slice of a very, very large pie.

Small slice, enormous pie. That's the thesis. And it's working.

What this means for the rest of us

If you're building with open-source AI models — and if you're reading this, you probably are — the DeepSeek transcript is reassuring in a way that might seem counterintuitive.

The most efficient AI lab in the world just admitted they're operating with 8% of the hardware they need. And they're still shipping models that compete with trillion-dollar companies. That's not a story about Chinese AI catching up to American AI. It's a story about what happens when brilliant people are resource-constrained: they get creative, and creative engineering tends to produce results that money can't buy.

This is the pattern across the entire open-weight ecosystem right now. Qwen, from Alibaba, runs the entire model range from tiny on-device models to frontier-scale MoE architectures — all open. Zhipu's GLM models are released under MIT license and routinely tie or beat GPT-class models on coding benchmarks. Kimi K3 from Moonshot AI, at 2.8 trillion parameters, is open-weight and approaches frontier performance on long-horizon coding tasks. None of these labs have the compute that US frontier labs have. All of them are shipping models that matter.

Liang also confirmed in the transcript that the next generation of DeepSeek models will be in the 150-250 billion parameter range. Not the trillion-parameter monsters that US labs are chasing. Smaller, denser, and — if the pattern holds — cheap to run.

For anyone running models locally, that's the sweet spot. A 150-250B parameter model with good quantization can run on hardware that a small company or even a dedicated individual can afford. We're talking about models that punch at a frontier level but don't require a data center. That's the world CopperRiver is built for — a desktop AI assistant that runs these open-weight models natively, on your machine, without paying API taxes to five different companies.

The AI race isn't just about who has the most GPUs. It's about who can do the most with what they have. And right now, the most honest person in the room just told us exactly where he stands — compute-starved, talent-rich, and planning to win anyway.

If you want to run these open-weight models yourself — GLM, DeepSeek, Qwen, Kimi — CopperRiver gives you a desktop AI assistant that uses them natively. No API taxes. No rate limits. Your models, your machine.