In January 2025, DeepSeek's R1 model wiped nearly a trillion dollars off U.S. markets overnight. Nvidia alone cratered. Headlines screamed about the end of American AI dominance. Sixteen months later, DeepSeek is reportedly raising its first-ever external funding round at a $10 billion valuation, and the reaction has been... a collective shrug. That muted response tells us more about the state of AI funding than any earnings report could.

The numbers that used to be shocking

DeepSeek is in talks to raise at least $300 million at a valuation of $10 billion or more, according to The Information. This would be the company's first outside capital. Until now, DeepSeek has been funded entirely by its parent company, High-Flyer Capital Management, a Chinese quantitative hedge fund co-founded by Liang Wenfeng. The company had previously turned down approaches from major Chinese VCs and tech giants. A $10 billion valuation for a Chinese AI lab that shook global markets just over a year ago should be a major story. Instead, it landed as a footnote in a week already saturated with AI funding news. The reason is simple: the numbers around it have gotten so large that $10 billion barely registers. Consider the context. In Q1 2026 alone, global startup funding hit $297 billion, with roughly $242 billion, about 80% of the total, flowing into AI companies. That single quarter outpaced every full year of global VC activity prior to 2019. OpenAI closed a $122 billion round at an $852 billion valuation. Anthropic raised $30 billion at $380 billion. SpaceX absorbed xAI in a merger that valued the combined entity at $1.25 trillion. Against that backdrop, DeepSeek's $10 billion is a rounding error.

Anchoring and the normalization of absurd figures

There's a well-documented cognitive bias at work here: anchoring. When we encounter a number, it becomes our reference point for evaluating subsequent numbers. Tversky and Kahneman described this in their foundational 1974 paper on heuristics and biases, and it explains a lot about how AI valuations feel right now. Once you've processed the idea that OpenAI is worth $852 billion, or that investors poured $242 billion into AI startups in a single quarter, your brain recalibrates. The anchor shifts. Suddenly $10 billion for a company that employs around 200 people and was valued at roughly $3.4 billion just last year doesn't trigger the alarm bells it should. This is how normalization works. The numbers escalate, each one making the next seem more reasonable by comparison. A year ago, a $10 billion valuation for a two-year-old Chinese AI startup would have dominated tech headlines for a week. Now it competes for attention with rounds that are 10x or 80x its size.

But is the valuation actually crazy?

Here's where it gets interesting: when you look past the headline number and focus on what DeepSeek has actually built, $10 billion might be the most rational valuation in the current AI landscape. DeepSeek V3, released in late 2024, is a 671-billion-parameter Mixture-of-Experts model that activates only 37 billion parameters per token. Its training reportedly required just 2.788 million H800 GPU hours, translating to roughly $5.6 million in compute costs for the final training run. For comparison, GPT-4's training is estimated at $50 to $100 million. The technical innovations are real. Multi-head latent attention reduces memory usage to 5-13% of previous methods. An auxiliary-loss-free load balancing strategy and multi-token prediction training objective push performance without the typical overhead. DeepSeek R1, their reasoning model, achieved competitive results with top closed-source models while relying on reinforcement learning with verifiable rewards rather than expensive human-labeled chain-of-thought data. These aren't incremental improvements. DeepSeek demonstrated that you can build frontier-competitive models at a fraction of the cost, and then released the weights for anyone to use. That combination of technical innovation and openness is genuinely rare.

The funding paradox

DeepSeek's decision to raise outside capital is itself noteworthy. Liang Wenfeng had long positioned the company as self-sufficient, funded by High-Flyer's profits. The hedge fund posted 56.6% returns in 2025, ranking second among China's large quant funds. With that kind of performance and over $10 billion in assets under management, the financial backing was never really in question. So why raise now? The likely answer is scale. Building and operating frontier AI models is getting more expensive, not less. Even DeepSeek's efficiency advantages can't fully offset the capital requirements of staying competitive as the field accelerates into reasoning systems and agentic applications. The shift from self-funded research project to externally capitalized company signals that DeepSeek sees an opportunity, or a necessity, to grow beyond what a hedge fund's side project can sustain.

What the number actually tells us about the market

The real story isn't whether $10 billion is too high or too low for DeepSeek. It's what the market's non-reaction reveals about where we are in the AI funding cycle. We've entered a phase where capital concentration is extreme. CB Insights reports that mega-rounds of $100 million or more accounted for 94% of total AI funding in Q1 2026. The average deal size hit $160 million, more than four times the 2025 full-year average. OpenAI's single $122 billion raise accounted for over half of all AI funding in the quarter. This is a market that is becoming increasingly top-heavy, with fewer, larger bets. DeepSeek's $300 million raise, substantial by any normal standard, is small enough to get lost in the noise. The question worth asking is whether this concentration of capital actually maps to a concentration of capability. OpenAI at $852 billion is roughly 85 times DeepSeek's proposed valuation. Is OpenAI 85 times more capable? More impactful? More likely to capture value? DeepSeek's track record suggests the relationship between spending and output is far less linear than these valuations imply.

The open-weight wildcard

DeepSeek's open-weight approach adds another dimension. Most of the companies commanding stratospheric valuations, OpenAI, Anthropic, xAI, keep their model weights proprietary. DeepSeek releases theirs. This creates an unusual dynamic: a well-funded company systematically undermining the moats that justify its competitors' valuations. With $300 million in fresh capital and a demonstrated ability to produce frontier models efficiently, DeepSeek could accelerate the commoditization of AI capabilities. Every open-weight release puts pressure on closed providers to justify their pricing and, by extension, their valuations. Whether you view this as a competitive threat or a public good depends on your perspective. But from a market structure standpoint, it's one of the most consequential dynamics in AI right now, and it's happening at a company valued at a fraction of its peers.

When numbers lose meaning

At some point, valuations in AI stopped being predictions about future cash flows and started being statements of belief. A $10 billion bet on DeepSeek is a belief that efficient, open AI development can sustain a business. An $852 billion bet on OpenAI is a belief that the company will capture enough value from artificial general intelligence to justify being worth more than all but ten public companies on earth. Both of these can't be right in the same way. If DeepSeek's approach works, if you really can build competitive models for a fraction of the cost and release the weights, then the value capture assumptions baked into OpenAI's and Anthropic's valuations need serious scrutiny. If the closed, capital-intensive approach wins, then DeepSeek's efficiency advantage matters less than its ability to scale. The fact that the market is simultaneously pricing both narratives as viable, without apparent contradiction, suggests we've moved past the point where these numbers are meant to be taken literally. They're not valuations. They're vibes. And $10 billion, for a company that briefly broke the market, is apparently not enough vibes to make anyone look up from their screens.

References