DeepSeek told prospective investors on July 25 to halt a second funding round that would have valued the Chinese AI lab at $71 billion—triggered by a leaked transcript of founder Liang Wenfeng’s private investor briefing that spread rapidly across Chinese social media. The 118-point document, drawn from a roughly four-hour meeting organized by Tencent Technology, contains Liang’s unguarded assessment that China trails the US in AI by 6 to 18 months and that the decisive gap is computing resources, not talent. DeepSeek has not confirmed the document’s authenticity. It doesn’t need to—the fundraising confirmed enough.
How a Private Investor Call Froze DeepSeek’s $71 Billion Round #
DeepSeek had only begun preliminary talks for its second round in mid-July 2026, weeks after closing a $7 billion first round in June—one of the largest Chinese startup financings on record. Tencent, CATL, and the state-backed National AI Industry Investment Fund anchored that first round at a $52 billion valuation. Liang himself put in roughly $3 billion of personal funds. The second round would have stepped that up 37% to $71 billion pre-money. Then Liang’s private remarks leaked, and he verbally told would-be backers to stand down.
According to Fortune’s reporting on Bloomberg’s scoop, the trigger wasn’t a business problem—it was a privacy violation. Liang grew furious that a confidential four-hour investor briefing had been turned into a 118-point document circulated on Chinese tech platforms. The is partly a rebuke to whoever leaked it. It’s also a signal that what he said in that room was sensitive enough that he’d rather kill a $71 billion financing than let it keep circulating unaddressed.
What DeepSeek’s Leaked Transcript Says About China’s AI Compute Gap #
The leaked document’s most explosive section covers compute. Liang stated plainly that China’s AI gap with the US comes down to computing resources. His figures: US frontier models have roughly 800 billion active parameters; China’s most advanced models have tens of billions—about a 10x gap in model scale. DeepSeek, for its part, runs on roughly one-twentieth of what American AI labs spend on compute. The quote attributed to him at point #55 reads: “The biggest gap between us and the US is in resources…The bulk is compute.”
On chips, the transcript reveals DeepSeek has deployed 16,000 Huawei Ascend 950 units, which Liang claims can replace Nvidia’s GB200/GB300 in a super-node configuration. However, he also gives the ratio: four Huawei Ascend 950 chips approximate one Nvidia B-series chip in performance. That means DeepSeek’s 16,000 Huawei chips represent roughly 4,000 Nvidia-equivalent units. As HelloChinaTech’s transcript analysis notes, Liang set a one-year deadline—by mid-2027—to validate the domestic chip ecosystem through real-world deployment. If that deadline slips, the compute ceiling holds.
Related:[China’s Open-Weight AI Is Winning. OpenAI Is Scared.]
What This Changes for Developers (Almost Nothing—For Now) #
The fundraising changes nothing about DeepSeek’s API pricing or open-source model strategy. The company’s V4 Pro API runs at roughly $2.17 per million tokens blended—20x to 50x cheaper than OpenAI’s GPT-o series. Liang said in the transcript that they “won’t open-source a weaker model,” meaning released models equal deployed models. That commitment is unchanged. According to The Next Web, DeepSeek may restart the second round at a later date—this is a , not a collapse.
The real concern for developers isn’t whether DeepSeek shuts down tomorrow—it won’t. The concern is model capability trajectory. Liang’s own transcript shows the absolute scale gap (10x parameters) and compute gap (20x resources) are not closing fast. Export restrictions are working as intended. DeepSeek’s efficiency innovations are genuinely impressive, but they’re compensating for a resource deficit, not eliminating it. Over the next 12-18 months, watch whether DeepSeek’s model releases keep pace with GPT and Claude updates, or start falling behind on benchmarks that matter for production workloads.
Key Takeaways #
- DeepSeek d a $71 billion second funding round on July 25 after founder Liang Wenfeng’s private investor transcript leaked on Chinese social media
- The transcript reveals China trails the US by 6-18 months in AI, with the gap rooted in compute resources, not talent—a direct validation that US export restrictions are biting
- DeepSeek’s 16,000 Huawei Ascend 950 chips perform at a 4:1 ratio to Nvidia B-series; the company targets domestic chip validation by mid-2027
- For developers: API pricing, open-source model releases, and stated strategy are unchanged—the fundraising is a response to a privacy breach, not a business crisis
- The actual risk to watch is model capability stagnation relative to US labs, not service disruption in the near term