No business model--DeepSeek's strongest "moat"

```

On the upcoming January 27th—marking the first anniversary of “DeepSeek Moment”—the global AI community is eagerly anticipating another big move from DeepSeek.

As market speculation heats up, well-known tech commentator Kevin Xu published a lengthy analysis of DeepSeek’s business model and moats, drawing considerable attention within the AI community.

He believes DeepSeek’s strongest moat is its unique model of zero external funding and no commercialization pressure. While global AI giants are all caught up in capital’s demand to make money, DeepSeek is the only one that can run freely for its AGI dream, regardless of costs or external influences—a true “free spirit.”

Kevin S. Xu is an independent observer focused on the intersection of Chinese-American tech and capital, as well as a well-known technology commentator. His ChinaTalk podcast and newsletter are influential in specialized circles, and he excels at analyzing the logic of technological evolution from angles such as capital flow, organizational behavior, and geopolitics.

Here’s a summary of the key points in the article:

Cooled market expectations: While everyone is waiting for DeepSeek’s new model, the author bluntly says “don’t expect it to shock the world like last year.” The market has already been saturated with “open-source models.” Though DeepSeek was a pioneer, it’s no longer the only or even the most open player (for example, it hasn’t open-sourced its datasets yet).
The only “zero-funding” outlier: The current AI field is a “money pit”—even Musk couldn’t hold out and recently raised a massive $20 billion for xAI. But DeepSeek’s founder, Liang Wenfeng, still insists on “zero external funding,” making it unique among top laboratories; he values control far more than money.
Super “money printer” backing: Liang Wenfeng can be so resolute because his quant fund, “High-Fly Quant,” is extremely profitable. Last year, High-Fly achieved a 53% return, earning over $700 million (~5 billion RMB) in profits. Liang directly used this to buy GPUs and hire talent, fueling “new dreams” with “old money,” allowing him to ignore investors’ expectations entirely.
Rejected by VCs was a blessing: DeepSeek actually sought funding when it was founded in 2023, but domestic VCs were short-sighted and didn’t invest—turning setback into opportunity. The author suggests that once you take VC money, you’re saddled with commercialization KPIs, distorting your actions. Now, DeepSeek only answers to technology, not financials.
“Too much money” means trouble: A deep insight—having too much money breeds “big-company syndrome.” The author cites a sarcastic example: some cash-flush labs (like Thinking Machines) even have custom barbell plates with their logo in the office gym. Such ostentation comes with “paper fortune” from options and departmental “power struggles” over computing resources. Since DeepSeek has no outside valuation, its team is extremely flat, with no internal resource strife.
Compute power isn’t a panacea: The article quotes former OpenAI expert Ilya Sutskever—truly disruptive innovation rarely requires infinite compute. The Transformer architecture was run on just 8–64 GPUs. Too much money can make teams lazy, assuming that scaling compute solves everything, thus losing “research taste.”
Investor’s ultimate paradox: Lastly, from an investor’s view the author sighs—as an investor, he’d love to be involved with DeepSeek; but he is well aware that once DeepSeek accepts external investment, its pure essence will vanish.

The original text, as translated by AI, is as follows:

No Business Model: DeepSeek’s Long-Term Advantage

By: Kevin S. Xu

As the first anniversary of the “DeepSeek Moment” (January 27th) approaches, anticipation for DeepSeek’s release of a stronger model before the Lunar New Year (February 17th) is mounting.

However, inflated expectations are almost destined to bring disappointment. Although I’m eager for the new model and curious about what DeepSeek is building, I do not expect it to shock the market as it did a year ago. The entire AI industry, and the broader market, has become accustomed to seeing a new “open-weight” model released every month or two, especially from Chinese labs. Their models, while maybe not the most cutting-edge, often perform robustly. Thanks to being open and free, these models are spreading in almost untraceable ways—a general virtue (or flaw) of open source. As Nvidia CEO Jensen Huang likes to say: measured in tokens generated, the world’s largest AI is still OpenAI, but the second largest is open-source AI.

Perhaps DeepSeek fired the first shot last year by releasing V3 and R1 models under the MIT License, but other labs quickly followed suit—including Chinese competitor Alibaba and its Qwen model series, as well as OpenAI with its gpt-oss model. (Both use the Apache 2.0 license, another common permissive open-source license.) As I’ve noted before, if not for DeepSeek, OpenAI might never have felt pressured to fulfill its open-source roots, even superficially, like releasing gpt-oss. While DeepSeek continues to open-source its model weights and various tools and libraries, it’s no longer the most open lab—the training datasets and main codebase have not been released.

According to Artificial Analysis, today’s top three most open models globally come from NVIDIA (Nemotron 3), Allen Institute (Olmo 3), and Mohammed bin Zayed University of AI (MBZUAI, K2-V2).

Thus, DeepSeek’s models are no longer the most capable among open models, nor the cheapest, or even the most open. So, is there still a reason for us to pay extraordinary attention to this once world-shocking lab?

Yes, I think so. But not because of its models or technological breakthroughs—instead, it’s because of its internal incentives and business model. More precisely, we should care about DeepSeek because it has no business model—a unique and counterintuitive trait among top labs in China, the US, and elsewhere, and not a bug.1. Self-Financing to Sustain the AGI Dream

Even weeks before DeepSeek’s “big moment” when it became a household name, I wrote about DeepSeek (and discussed it on the ChinaTalk podcast), noting this “no business model advantage.” Remarkably, this advantage has lasted stubbornly as the entire AI world has been drowned in capital.

Liang Wenfeng has not raised any outside money. The lab does generate revenue from its API services, but the pricing keeps dropping. Liang is satisfied funding DeepSeek’s R&D with profits from his quant fund, High-Fly Quant, which originally incubated DeepSeek. To clarify, when DeepSeek was founded a few months after ChatGPT launched in 2023, Liang did attempt to raise VC money in China. But his “AGI-pilled” idealism, absence of a business plan, and the notoriously short-sighted and risk-averse Chinese VCs led to failure—turning out to be a blessing in disguise.

As 2025 approaches, and to not miss any potential progress, more and more capital keeps flooding in to fund AI advances. Every lab with a whiff of traction raises more cash, and with it, expectations of future commercial pay-off. A crop of “new AI labs” led by star researchers and backed by venture capital has sprung up, such as Thinking Machines Lab (Mira Murati), SSI (Ilya Sutskever), and AMI Labs (Yann LeCun). Even xAI, the closest analogy to DeepSeek (because it too lacked commercialization urgency), has recently bowed to outside capital—Elon Musk’s lab just raised $20 billion in Series E funding, through a mix of equity and debt.

Even one of the world’s richest people can’t say “no” to more money! Meanwhile, Liang Wenfeng keeps funding his mysterious shop and his AGI dreams out of pocket. Certainly, last year his quant fund’s stellar performance, making over $700 million at a 53% ROI, helps a lot. Presumably, most of those profits go toward buying more GPUs (dodging US export controls as much as possible) and hiring talent, to keep DeepSeek’s research roadmap moving. Still, as DeepSeek became globally recognized, raising $1–2 billion to supercharge its pace would be trivial. You can love it or hate it—everyone knows its name.

Yet, saying “no” to outside capital lets you control your own destiny. If your self-defined destiny is “to make AGI a reality… to use curiosity to unlock the mystery of AGI… to answer essential questions with long-term thinking”—DeepSeek’s slogan on HuggingFace—then trading less capital for total control is a worthy bargain. Of course, you can use “corporate governance innovations” to achieve similar control. Thinking Machines Lab did that: the founder’s voting power counts for one vote more than the rest of the board, giving her actual final say. But no matter how “innovative” you get at the board level, if you accept venture capital, large-scale commercialization for VC-sized returns will inevitably be expected.

DeepSeek has none of those expectations. So, it has no business model, and needs none. The road to AGI needs compute, talent, and a healthy measure of good research taste. No one ever said it requires a business model.2. Lots of Money, Lots of Trouble; No Money, No Worry

Of course, the usual rationale for raising more money is to buy more compute for research. Yet, it’s not obvious that more compute is always required for good research output.

This insight isn’t unique to DeepSeek’s self-financing structure. AI research legend Ilya Sutskever agrees. In the Dwarkesh podcast, he explains:“Compute scale is now so large that it’s not obvious you need so much extra compute to prove a point. For example, AlexNet used two GPUs—total compute. Transformer used 8 to 64 GPUs. No 2017 Transformer paper ran on more than 64 GPUs, which today is like two cards. ResNet too. You could argue o1 reasoning models aren’t the most compute-intensive in the world.

So for research, you absolutely need some compute, but it’s far from obvious you need the maximum possible amount for good research.”

On the other hand, by eliminating strings attached to outside capital, and the “must buy more compute just because you can,” two organizational benefits emerge. DeepSeek has both.

First, even with limited resources, there is no internal resource competition. No bureaucracy, power struggles, or horse-trading over whether GPUs should support a new product launch, expand existing inference demand, or back a new research idea. In a small lab without external money or business models, good research taste and new ideas get maximum support, even if the total compute available is limited.

Second, less jealousy and hierarchy based on pay, perks, or GPU allocation—often deeply toxic but rooted in human nature. Your organization instantly becomes flat, and stays flat. Outside capital brings not just business model expectations, but also valuation, options, and all the flashy frills of a high-valuation AI lab, even if such valuation is meaningless early on. Paper wealth from magical valuations breeds pride, envy, and easy poaching by deeper-pocketed rivals. These frills aren’t just about personal wealth; they’re tied to the lab’s brand, reputation, and ability to attract talent. Sometimes, they get ridiculous—like the branded barbell plates in Thinking Machines’ office gym:

Does DeepSeek have its own gym? I don’t know. Maybe. But I’m pretty sure it doesn’t have branded barbell plates. Does Ilya Sutskever wish he could run SSI with self-funding, like Liang Wenfeng (it’s raised $3 billion so far)? Maybe. He may have great research taste and instincts, but he can’t stop poaching—even his co-founder CEO was poached by Meta’s better-funded Zuckerberg. That’s not to say DeepSeek is immune to poaching: one star researcher, Luo Fuli, now leads AI work at Xiaomi, which has deeper pockets.

In an industry drowned in money—and the drama and scheming that always come with money—DeepSeek’s lack of a business model caused by zero outside capital is the source of its only durable advantage: maximal internal team alignment for AGI research, and nothing else.

I’m not even sure I believe in AGI. But as an investor and capitalist, I’d love to be on DeepSeek’s cap table. That said, if DeepSeek ever let me or any other outside investor in, the very essence that makes DeepSeek what it is would be utterly destroyed.

This article is from WeChat public account “Hard AI”. For more AI frontier news, visit here

Risk Warning and DisclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice, nor does it take into account the individual investment objectives, financial situations, or needs of specific users. Users should consider whether any opinions, views, or conclusions in this article fit their specific situation. Please invest at your own risk.

```