Moonshot and the DeepSeek 2 Moment: What Changes in the AI Race

The world's largest open-source model was born in Beijing

A two-year-old Chinese startup has just put more pressure on the trillion-dollar valuations of American Big Techs. Moonshot, founded in 2023 by alumni of Tsinghua University, launched the Kimi K3, a language model with 2.8 trillion parameters. It is the largest open-source model ever created and will be available for free download starting on the 27th.

To put it into perspective: the larger the number of parameters, the greater the number of "neurons" in the model and, theoretically, the greater its ability to process information and store knowledge. Anthropic, the owner of Claude, does not officially disclose this data, but researchers estimate that Claude Opus 4.8, released in May, has just over 1.5 trillion parameters. The Kimi K3 has nearly double that.

According to results presented by Moonshot itself, the K3 outperformed both Opus 4.8 and OpenAI's GPT 5.5 in most programming and agent benchmarks. It still lags slightly behind the two most advanced models from American companies, Fable 5 (Anthropic) and GPT-5.6 Sol (OpenAI), but the difference is marginal. In a blind user preference test for web interface engineering, the Kimi K3 came in first place, ahead of Fable.

Why this matters: pressure on the business model of Big Techs

The episode is already being referred to as the "DeepSeek 2 moment," in reference to the shock that the Chinese DeepSeek caused in the markets early last year when it presented a frontier model developed at a fraction of the cost of its American competitors. The logic is repeating itself now with Moonshot, but on an even more significant scale.

The central point is not just technical performance. It's the economy. As we detailed when analyzing the impacts of AI on the technology sector, the revenue model of American Big Techs in the AI segment relies on charging for tokens, the units of information processed by the tools. If an open-source model delivers equivalent results without licensing costs, the value proposition changes drastically.

The numbers illustrate the problem. The average cost of Chinese AI services is about one-tenth of what American competitors charge. To perform a standardized task unit using Kimi 2.6 or DeepSeek's V4 Flash, the price ranges from $0.02 to $0.33. The same task on Claude Fable 5 costs $2.75, according to data from Artificial Analysis.

Moonshot: from reference to Pink Floyd to a valuation of $31.5 billion

Moonshot was founded by Yang Zhilin, who is 33 years old and has experience at Google and Meta. The startup was originally named "Dark Side of the Moon" in Chinese, referencing the Pink Floyd album. The meeting rooms are named after Radiohead, Led Zeppelin, and Nirvana. The subscription plans for Kimi follow classical music tempo indications: Adagio, Andante, and Moderato.

Despite the casual tone, the numbers are serious. The company received investments from Alibaba and Tencent, reached a valuation of $31.5 billion, and achieved an annualized revenue of $200 million in April. All this in just over two years of operation.

For those following the financial markets, the parallel with DeepSeek's trajectory is inevitable. Both emerged from cutting-edge Chinese academic ecosystems, received capital from large conglomerates, and managed to compete on equal footing with companies that spent tens of billions of dollars on infrastructure.

Market reaction and the debate over data centers

The impact on the markets was immediate. The Philadelphia SE semiconductor index fell 1.9% on the day of the announcement and has accumulated a 10% drop for the week. Still, the index maintains a 62% increase year-to-date, compared to 9% for the S&P 500.

Nvidia has lost its position as the most valuable company in the world. With a market cap of $4.8 trillion, it has fallen behind Apple, which is up 22% this year and trading at an all-time high of $4.9 trillion. The Round Hill Magnificent Seven ETF, which includes the top Big Tech companies, has declined by 1.9%.

David Sacks, a technology investor and advisor to the White House on the topic, stated in a post on X that the United States is "on track to lose the AI race." According to him, while China is advancing, the U.S. is complicating matters with bans on new data centers, a buildup of state laws, and pressure to create federal agencies for pre-approval of models.

Analysts at Bank of America reinforced this point. Despite the ongoing hardware and computational capacity restrictions imposed by the U.S. on China, the K3 demonstrates that scaling parameters combined with architectural innovation can provide significant leaps. In other words, Nvidia's export restrictions on chips to China may be circumvented by software engineering.

What This Means for Technology Investors

The debate is not just geopolitical. For investors, the advancement of Moonshot raises a practical question: to what extent are the valuations of American Big Tech companies supported by the premise that their models will remain superior?

If Chinese open-source models deliver 90-95% of the performance at 10% of the cost, the willingness of companies and developers to pay for premium tokens may decrease. This would directly pressure the margins of companies like OpenAI, Anthropic, and Google, which are betting their revenue projections on the monetization of generative AI.

For those following the evolution of artificial intelligence, the pattern is clear: the competitive advantage in AI is becoming more ephemeral. The investment race in data centers, which moves hundreds of billions of dollars globally, may face a reality check if algorithmic efficiency continues to offset computational brute force.

The "DeepSeek 2 moment" may not cause the same market shock as the first. But it reinforces a trend that investors would do well not to ignore: cutting-edge AI is commoditizing faster than anyone expected.