Language model → Chain of thought → Agent → Continuous learning → Self-iteration singularity → Embodied intelligence.

Author and source: 0x9999in1, ME News

TL;DR

• DeepSeek has completed its first external funding round since its founding, raising over RMB 50 billion, with a pre-money valuation of approximately RMB 367.5 billion (about $54.3 billion). Liang Wenhong led the round with a personal investment of RMB 20 billion, followed by Tencent with RMB 10 billion and CATL with RMB 5 billion. The company’s former principle of “no funding” has now been broken.
• But this isn’t a compromise—it’s solving the problem. Liang Wenheng calls "team stability" the one non-negotiable core interest; the essence of fundraising is using equity options to retain talent and eliminate the biggest risk all at once.
• He provided a rare, clear roadmap: language models → chain of thought → agents → continuous learning → self-iteration singularity → embodied intelligence. We are currently stuck at the level of "continuous learning."
• Computing power remains the Achilles' heel. DeepSeek positions itself as "6 to 18 months behind the U.S., using only one-twentieth the computing power," with its next goal being to reduce this gap to 3 to 6 months.
• Domestic chips are the biggest variable. V4 has clearly been adapted for Huawei Ascend 950, and Liang Wenfeng bets that "within a year, we will prove that the domestic ecosystem is viable," with production capacity being the only bottleneck.
• Open source remains unwavering—the most powerful models are open source and identical to our internal versions. We generate only reasonable profits aimed at recouping costs within ten months. This restraint is, in fact, our moat.

DeepSeek has let go.

For three years, Liang Wenfeng personally established a rule: no fundraising, no IPO, no commercialization. Now, half of that rule has been broken: the first round of external funding has been secured, exceeding 50 billion yuan, with a pre-money valuation of 367.5 billion yuan. Liang Wenfeng personally invested 20 billion yuan.

What was the initial external reaction? "He finally gave in to capital."

Really?

After reading this nearly four-hour, 118-message transcript of Liang Wenfeng’s conversation, the opposite conclusion emerges: this is not submission, but an extremely restrained person using the most reluctant means possible to secure what he cares about most.

Financing is not a shift—it’s about removing the greatest risk all at once.

First, look at the table of investors: Liang Wenheng personally invested 20 billion, making him the absolute largest single investor with full control. Tencent invested 10 billion, CATL’s ecosystem invested 5 billion, NetEase, JD.com, and IDG each invested 3 billion, and the National Artificial Intelligence Industry Investment Fund invested 1 billion.

This list is not arbitrary. Liang Wenheng himself stated that the investors were "carefully selected," with only one criterion: the greatest alignment of interests, the least hostility toward us, and the strongest desire for our success.

Tencent provides cloud and traffic, CATL provides energy and industrial applications, and state-backed entities provide endorsement. There are no pure financial speculators. This is a list of strategic allies, not a fundraising list.

So why borrow?

The answer lies in the phrase he repeatedly emphasized: “Our greatest core interest is maintaining team stability, which can even be considered our only core interest.”

He spoke about it without hesitation. Team stability, in his view, is the greatest risk and the biggest vulnerability. In an era when Silicon Valley giants are offering $100 million annual salaries to poach talent, what reason does a private company, where employees’ stock options cannot be liquidated, have to retain its oldest and most core researchers?

After the financing, he provided his own answer: “This risk has been significantly mitigated by this recent financing, as the option amounts received by everyone are still quite substantial.”

Understood? The primary goal of this funding round isn’t to buy cards or expand—it’s to give the core team a real, tangible payout. It’s about putting the looming question of “Will people leave?” to rest, once and for all.

Money isn't the issue; cards are. The most valuable resource is human trust. Liang Wenheng thought it through thoroughly.

That roadmap was the true highlight of this meeting.

If funding is the news, then the roadmap is the core.

Liang Wenhong describes the evolution of AGI as a staircase, where each step is worthwhile:

Last year's stepping stone was the chain of thought. This year's stepping stone is Agent. And after Agent, he repeatedly emphasizes one word—continuous learning.

His logic is as sharp as a scalpel. Can AI replace employees now? No. Why? Because it can’t “learn the job in two months” like a new hire. But what if it gains the ability to learn continuously? He said: “Then it could replace everyone.”

Above that is the singularity. When a model can continuously learn, it can develop its own next version. AI accelerates AI. At that point, growth will no longer be linear.

Finally, embodied intelligence—entering the real world to do household chores and provide elderly care.

This sequence is intriguing. Liang Wenhong directly dismissed the most popular video generation and world models as "off the main track." He was blunt: video generation is a good business, but "it has nothing to do with intelligence—we won’t pursue it just because it’s commercially promising."

What kind of discipline is this?

While the entire industry chases multimodal flashy features that promise instant monetization and viral attention, DeepSeek says: we only focus on what truly matters in the core of intelligence. This isn’t aloofness—it’s focus: focusing on the fact that there are so few essential things to do, you don’t need to work overtime.

Interestingly, the market has already voted on this assessment. In July of this year, DeepSeek V4 was officially released. The V4-Pro employs a sparse expert architecture with 1.6 trillion parameters, activating only about 490 billion per token; the V4-Flash is even lighter, with 284 billion total parameters and 13 billion activated per token. Both models feature a 1 million token ultra-long context and are fully open-sourced under the MIT license. On the SWE-bench Verified programming benchmark, V4 achieved the highest score of 80.6% among open-source models.

Staying focused on the main goal has indeed borne fruit.

“Two years behind, using only one-twentieth the computing power”—this self-deprecating remark is full of ambition.

Liang Wenhong's assessment of the gap was surprisingly candid.

He said the main gap between DeepSeek and the U.S. lies in resources, not in talent—“it’s essentially the same group of people, possibly Chinese.” Some stayed in China, while others went abroad; it’s not as if only the most talented left.

So what’s the actual gap? He provided a nearly formulaic statement: lagging behind the U.S. by 6 to 18 months, achieving the same results with only one-twentieth of the U.S. computing power.

On the surface, this sounds like surrender, but beneath it lies pride. With just one-twentieth of the computing power, achieving a result only about a year behind—that’s a terrifying level of efficiency.

But his true ambition lies in the next sentence: “In the future, we will rewrite this narrative, reducing the time lag to six months or even three months using just a fraction of the computing power.”

From “one-twentieth the computing power, two years behind,” to “a fraction of the computing power, just three months behind”—this is the true benchmark for DeepSeek’s future development.

Why is computing power so critical? He stated unequivocally: “All the differences we observe—whether in talent, model capabilities, or applications—can be attributed to differences in computing resources.” He believes in scaling, and acknowledges that the only barrier to scaling is simply not having enough GPUs.

So the question arises—where do the cards come from?

Betting on domestic chips was the boldest move of this forum.

This is, in my view, the most significant judgment in the entire transcript, and the one that will require the most time to verify.

Liang Wenfeng is "optimistic" about domestic computing power. He even made a strong statement: "In terms of domestic computing power, NVIDIA is digging its own grave."

His argument has three layers.

At the first layer, CUDA's moat is rapidly eroding—precisely because of AI itself: with AI that can write code, rebuilding ecosystems is now far easier than before.

Layer 2: DeepSeek has already moved away from the NVIDIA ecosystem in practice. He revealed that although NVIDIA GPUs were still used during V3 training, the NVIDIA ecosystem was almost entirely avoided—by first developing a high-level compiler called TileLang, upon which all other work was built.

The third and most critical point: Huawei. He asserts that Huawei’s 950 super-node can completely replace NVIDIA’s GB200 and GB300 in both performance and price, and he repeats the widely circulated statement: “Four Huawei cards equal one NVIDIA card.”

This is not empty talk. Compatibility with Huawei Ascend 950 was clearly established at the launch of V4 in April this year. DeepSeek has publicly stated that the unit token cost of V4-Pro is only expected to drop significantly after the Ascend 950 super nodes begin mass market release in the second half of 2026. Liang Wenheng attributes the sole issue to “insufficient production capacity,” emphasizing that both hardware and the ecosystem are fully functional.

He even provided a verifiable timeline: “Within the next year, we will be able to verify that the ecosystem for domestically produced chips is completely viable.”

This is a statement that time may refute—or elevate. I tend to believe the latter—because under the practical constraint of being unable to purchase NVIDIA’s high-end GPUs, “everyone is forced to turn to domestic solutions,” and demand will forcibly accelerate ecosystem development. Constraints breed innovation—a phenomenon already demonstrated by DeepSeek.

Open source, fair pricing, earn reasonable profits: restraint itself is a strategy.

Returning to that original question: Having broken the rule of "no fundraising," will DeepSeek also abandon its restraint of "no commercialization"?

Looking at the data, the answer is no.

Open source—Liang Wenfeng was unequivocal: “We will open source it; even the strongest models will be open sourced, and the open-source versions will be identical to those we deploy internally—no hidden features. He asked: What are the benefits of keeping it closed source? ‘ByteDance’s models are closed source—what advantages do they offer? I don’t see any.’”

More importantly, his assessment of the relationship between open source and revenue—“Open source has no impact on our business model.”—is that even if you were given everything, others would still face extremely high barriers to using it and driving costs low enough.

In pricing, he is equally restrained, bordering on “counterintuitive to business norms.” DeepSeek’s API pricing aims only for a reasonable profit—enough to recoup the cost of purchasing equipment within ten months. He admits that, to maximize profits, prices could be set much higher, since demand in this range is nearly inelastic—doubling the price would barely affect token consumption. Yet he refuses to do so. When a previous model was priced too high and left the team dissatisfied, he immediately slashed the price to one-quarter of its original level.

What does this restraint look like in today’s market? V4’s official release adopts a peak-and-valley pricing model, doubling prices during peak hours and setting them extremely low during off-peak periods. Meanwhile, in the open-source model camp, it has created a price gap with closed-source flagship models measured in multiples. Some estimates describe this disparity as ranging between one and two orders of magnitude.

Liang Wenheng summarized it with one statement, which I believe is the key point of the entire conversation:

Those who take more will be defeated by those who take less. If your vision is to take more, you’ve already lost.

This is not moral posturing. It is a strategy he has repeatedly validated: restraint increases the probability of achieving AGI; the more restrained, the more likely it is to succeed.

Conclusion: Where can a company that doesn’t want to become the “next ByteDance” go?

So, where exactly does DeepSeek’s next path lead?

Putting together these 118 pieces, the picture becomes clear.

It doesn’t aim to become a super app or compete with Tencent or ByteDance for market share. It is clearly focused on monetization—“The government won’t give us a single cent”—and in the worst-case scenario, “Fully selling APIs would be enough to sustain a publicly traded company.” Its current business priority is general-purpose agents, especially coding agents, with other vertical agents receiving lower priority. It has even indicated that if it can generate hundreds of millions of dollars in B2B revenue this year, combined with C-end users, “the company is close to profitability.”

The market is also imagining even greater potential for it. Just as the funding round settled, rumors emerged that DeepSeek could list on the mainland as early as next year, with its valuation continuing to rise. The one who once said “no listing” is now being repeatedly pursued by capital eager to invest.

But are you asking whether Liang Wenheng cares about these things?

I guess he doesn’t care much. He’s made it clear: the real goal is to have AGI help him iterate to the next version of the model, and once it’s embodied, have robots build the next generation of robots. As for shares, massive profits, or market cap—he said, “AI is such a huge field; even a small slice of it brings enormous returns.”

A group of people who call themselves "very ordinary," willing to take less, are chasing the world's cutting edge—just over a year behind—using only one-twentieth of the computing power.

Now, they have secured 50 billion, won over public trust, bet on domestic chips, and laid out their timeline on the table.

The coming year is the validation period. Whether the domestic ecosystem can succeed, whether continuous learning can break through, and whether the “three-month” narrative can hold up—answers lie not in PowerPoint slides, but in Ascend’s production ramp-up and V5’s training logs.

Those who exercise restraint often have the last laugh. This time, I’m willing to wait and see.

Source:

• Sina Finance: "Transcript of the DeepSeek Funding Meeting: Liang Wenhong Breaks Down the Technology and Business Logic of Large Models," July 23, 2026
• East Money Web: "DeepSeek Secures First Round of Funding Amounting to Approximately RMB 51 Billion, with Investments from Tencent, CATL, and Others," July 2026
• Dahe Daily · Dahe Finance Square: "Lei Wenfeng's 4-Hour Investment Talk Leaked: Why DeepSeek Doesn't Pursue 'More Profitable' Opportunities," July 2026
• Sina Finance (Caixin report): "DeepSeek-V4 Explicitly Supports Huawei Ascend 950 Chip," April 24, 2026
• Caixin: "DeepSeek: V4-Pro Usage Costs Expected to Drop Significantly After Bulk Launch of Ascend 950 Super Nodes in the Second Half of the Year," July 23, 2026
• EET-China: "DeepSeek V4 Ignites a National Computing Power Boom: Huawei Ascend 950 Snatched Up by Major Companies," April 2026
• meshlaunch: "DeepSeek V4 Full Release (July 2026): Pricing, Benchmarks & Migration," July 20, 2026
• n1n.ai Explore: "DeepSeek V4 GA and Open-Source: Analyzing the LLM Price Gap," July 21, 2026