Foundation Model Startups Through the Lens of DeepSeek: Why Disruptive Innovation Happens Independently

Why Disruptive Innovation Rarely Emerges from Within Large Companies

One central thesis was repeatedly emphasized in this conversation with Luan Yu: truly disruptive innovation tends to happen independently — it rarely grows from within a large company. This may sound counterintuitive. After all, large companies command more capital, more compute, and more talent. So why do they keep failing to produce genuine foundation model breakthroughs?

The answer lies in organizational logic. Large companies are naturally inclined to spread risk, run multiple projects in parallel, and chase short-term user growth and traffic metrics. Foundation model (base model) innovation, as an upstream endeavor, requires the exact opposite: concentrated commitment, tolerance for sunk costs, and long-term strategic conviction. These two logics are fundamentally at odds.

This phenomenon has a well-known theoretical parallel in innovation management — Clayton Christensen's framework of "Disruptive Innovation" in The Innovator's Dilemma. He observed that established companies routinely fall into a "resource dependence" trap: because they are accountable to existing customers and shareholders, their resources continuously flow toward currently profitable businesses, not toward foundational innovations that look small, low-margin, and uncertain in the near term. Foundation model R&D is a textbook example of disruptive technology — no clear commercial return path in the early stages, years of high-intensity investment required, and a natural mismatch with quarterly-earnings-anchored decision-making. On top of that, internal innovation initiatives face an "immune system" problem: existing business units, entrenched interest structures, and reporting chains instinctively reject long-horizon commitments that diverge from the core business direction. These structural barriers — not a lack of funding or talent — are the deep reason why large companies struggle to incubate foundation model innovation.

Three Key Conditions Behind DeepSeek's Breakthrough

Luan Yu broke down the prerequisites that enabled DeepSeek^Ad to succeed — conditions that apply equally to other foundation model players like Kimi and Zhipu AI.

First: a sufficiently strong team. This is the baseline for any innovation. Without an exceptional technical team, everything else is moot.

Second: securing enough capital at the right moment. Luan Yu specifically noted that when these companies started out, servers and compute hardware were "not nearly as contested as they are today" — prices were relatively stable or even trending down. In other words, they caught a window before compute costs were bid up by the market.

Third: compute reserves at an absolute scale. There's a widely circulated claim that DeepSeek's compute stockpile is second only to Alibaba and Tencent. Luan Yu's view is that DeepSeek's ability to move faster at critical junctures — around a certain Lunar New Year period, for instance — was directly tied to founder Liang Wenfeng's entrepreneurial resolve, the depth of his commitment, and the compute reserves built up in advance.

Leaders Must "Understand the Technology" and Get Into the Trenches

One point emphasized strongly in this conversation: doing upstream foundation model innovation absolutely requires someone who understands the technology to be embedded within the technical team — engaging in discussion, driving progress, and making decisions collaboratively.

This also explains a curious pattern worth examining: why did teams from quantitative fund backgrounds produce foundation model breakthroughs? Luan Yu's explanation is that leaders from quant funds at minimum possess technical literacy — they're willing to dig in alongside the technical team and push things forward together. This "technically fluent leadership" is something teams built purely around application development and traffic growth simply don't have.

If a company's leadership treats foundation models as just another "project to file a proposal for" — managing it with an app-building, traffic-chasing mindset — it will almost certainly fail in the foundation model space. The depth of a leader's technical understanding directly determines the direction and conviction behind resource allocation.

Quantitative hedge funds are financial institutions that rely on mathematical models, statistical analysis, and algorithmic systems to make investment decisions. Their core teams are typically composed of researchers with backgrounds in mathematics, physics, and computer science. DeepSeek's founder, Liang Wenfeng, is the founder of High-Flyer, a quantitative hedge fund. The operational DNA of such institutions overlaps significantly with AI foundation model development: both depend on large-scale data processing, iterative model optimization, and long-term bets on uncertain returns. Leaders from quant fund backgrounds not only carry technical expertise themselves — they're accustomed to making decisions through the lens of "expected return vs. risk tradeoff" rather than "quarterly KPIs." This makes them naturally better suited to navigate the long-cycle, high-sunk-cost nature of foundation model development. By contrast, leaders from pure internet or consumer product backgrounds are more accustomed to validating decisions through user growth and traffic data, making it harder for them to maintain conviction in foundation model R&D with no near-term payoff in sight.

Strategic Conviction: The Willingness to Bet Everything on One Project

Another brutal reality of foundation model innovation: it demands a resolute, all-in commitment — the kind that accepts sunk costs without flinching.

Luan Yu offered a concrete scenario: imagine a company has tens of billions in capital. The choice it faces is — do you put all of it behind a single training run? Or do you split it up and spread it across multiple bets? The moment you choose to split your focus, your chances of succeeding in foundation models drop dramatically.

This is fundamentally a question of risk tolerance and strategic conviction. Luan Yu offered a sharp insight: risk is a funny thing — the harder you try to spread it, the more likely it is to materialize. But if you commit decisively to a single bet, you might still end up with nothing, yet at least you preserve a real chance of success.

He also acknowledged plainly that there are plenty of companies that poured hundreds of millions — even billions — into training runs and still failed. So decisive commitment is not a sufficient condition for success, but it is a necessary one. Without the conviction to go all in, you can't even buy your way to the starting line. DeepSeek succeeded precisely because its founder combined all three: the capability, the commitment, and the capacity to absorb the risk.

There's a counterintuitive probability dynamic at play here, sometimes called the "paradox of concentrated betting." In conventional risk management thinking, diversification is the standard approach to reducing overall loss. But in winner-takes-all technology competition, spreading resources often means no single direction reaches the critical scale needed for a breakthrough — every path gets abandoned halfway. Foundation model training is especially prone to this: there are significant scale effects across compute investment, data flywheels, and model iteration. Below a certain threshold of investment, you may not even be able to validate whether a technical approach is viable before the money runs out. "Spreading risk" in this context actually inflates the probability of achieving nothing. Effective risk control, in this setting, means doing rigorous feasibility work before committing — and then concentrating resources fully once you decide to enter, rather than keeping your options open through diversification. This aligns closely with the venture capital logic of "betting heavily on a small number of high-conviction positions."

Focused Bets vs. Scattershot Launches: The Divergence Three to Five Years Out

Toward the end of the conversation, Luan Yu compared two startup trajectories over a longer time horizon — a comparison that carries a clear cautionary message.

He observed that DeepSeek has never rushed to ship a wide range of consumer-facing products. Instead, it has stayed relentlessly focused on getting the core technology and core logic right — making the kind of single-point, concentrated bets that most teams shy away from.

Contrast this with another type of company — including some publicly listed ones — that launched a large portfolio of products and applications from day one, each landing somewhere around a "50–60 out of 100." By funneling existing traffic into these apps, they quickly accumulated users and data. Internally, everything looked calm and productive.

But Luan Yu offered a sobering assessment: stretch the timeline out to three to five years, and those scattershot products may have no real market value at all — and could gradually be phased out. The short-term sense of harmony masks a long-term risk: no genuine breakthrough has been made.

In a deep-tech category like foundation models, whether you pursue a focused breakthrough versus whether you have the strategic conviction to stay the course ultimately determines whether a company is a flash in the pan or something that truly stands the test of time. That, perhaps, is the most enduring lesson DeepSeek leaves for the industry to chew on.