HONGKONG, China – When DeepSeek unveiled its R1 artificial-intelligence model in early 2025, Wall Street interpreted the breakthrough as a warning: If increasingly capable AI systems could be built and operated with fewer resources, the technology industry might not need quite so many expensive chips.

The reaction was swift. Nvidia lost about $593 billion in market value in a single trading day — at the time the largest one-day decline ever recorded by an American company. The episode became known as the “DeepSeek moment.”

Now another Chinese company is testing that assumption.

Moonshot AI, a three-year-old Beijing start-up, introduced Kimi K3 last week, describing it as a 2.8-trillion-parameter, open-weight model designed for complex reasoning, software development and long-running autonomous tasks. Moonshot says it is the largest open-weight AI system yet disclosed.

The announcement arrived amid an already broad retreat in technology shares. Nvidia fell 2.2 percent on Friday, while Applied Materials, a major supplier of semiconductor-manufacturing equipment, dropped 5.6 percent. But it would be an overstatement to attribute those moves entirely to Moonshot: concerns about technology valuations, corporate earnings, memory-chip policy and the wider market sell-off also contributed.

A Chinese Challenger Nears the Frontier

Moonshot says Kimi K3 approaches the performance of Anthropic’s Claude Fable 5 and surpasses several other leading systems, including Claude Opus 4.8 and OpenAI’s GPT-5.6 Sol, on selected tests.

Independent evaluations offer some support, though not a definitive verdict. Reuters reported that Arena.ai ranked Kimi K3 first on a web-interface development benchmark, while Vals AI placed it second overall behind Fable 5. Artificial Analysis found it broadly competitive with other frontier models on complex, multistep tasks.

Benchmarks, however, measure narrow aspects of performance and can be affected by test selection and configuration. Model size is also not synonymous with intelligence: A system with more parameters is not automatically more capable or reliable.

What appears less disputed is the speed with which Chinese developers are narrowing the performance gap. Moonshot, DeepSeek, Z.ai, MiniMax and Alibaba are releasing increasingly capable models while generally offering greater access and lower prices than the closed systems of their leading American competitors.

Chinese reporting has emphasized Kimi K3’s scale, native visual capabilities and one-million-token context window, which allows the system to process unusually large collections of documents or code in a single session.

The immediate response exceeded Moonshot’s expectations. On Sunday, the company temporarily stopped accepting new consumer subscriptions after requests approached the limits of its computing clusters. Existing paying customers were given priority while Moonshot worked to add capacity.

“Kimi K3 has received far more love than we expected, and our GPUs are feeling it,” the company said in a social-media post quoted by Reuters.

Efficiency Does Not Mean Small

The comparison with DeepSeek is tempting, but incomplete.

DeepSeek’s R1 challenged the assumption that frontier-level performance necessarily required the same enormous training budgets associated with American laboratories. Kimi K3 also incorporates efficiency improvements, but places them inside a vastly larger system.

Kimi K3 uses a sparse “mixture-of-experts” design. Instead of activating all 2.8 trillion parameters for every request, it selects only a fraction of them for each task. That can reduce the number of calculations required to produce an answer.

But sparsity does not eliminate the need to store the entire model. At four bits per parameter, the raw weights alone would occupy roughly 1.4 terabytes, before allowing for context memory, temporary calculations, concurrent users and other operational overhead.

That makes Kimi K3 difficult and expensive to operate outside a data center. Moonshot’s planned release of the model weights on July 27 may give companies greater control over customization and sensitive data, but “open weight” does not mean inexpensive to run.

This distinction could matter greatly for the semiconductor industry. A large deployment would still require clusters of memory-rich AI accelerators, advanced packaging and substantial quantities of high-bandwidth memory. That could support demand for Nvidia systems, SK Hynix memory and the advanced manufacturing services of Taiwan Semiconductor Manufacturing Company.

The subscription pause provides an early piece of real-world evidence: Kimi K3’s popularity is not reducing Moonshot’s computing needs. It is straining them.

The DeepSeek Lesson Reconsidered

The central investment question is therefore not simply whether models are becoming more efficient, but what companies and consumers do with those savings.

Economists and technology investors frequently invoke the Jevons paradox: When a resource becomes cheaper or more efficient to use, total consumption can rise as new applications become economically viable. Lower-cost AI could similarly bring millions of additional users and thousands of companies into the market, ultimately increasing demand for inference chips, memory and electricity.

That does not guarantee that every chip supplier will prosper. Improvements in model architecture, competition among hardware manufacturers and the tailoring of Chinese software to domestic accelerators could redistribute spending away from some American suppliers.

U.S. export restrictions have also made access to Nvidia’s most advanced processors a persistent constraint for Chinese developers. That pressure is encouraging Chinese companies to optimize their models for domestically produced chips, potentially strengthening China’s own semiconductor ecosystem.

Alibaba added to that momentum over the weekend by previewing Qwen3.8-Max, a 2.4-trillion-parameter model that it says will eventually be released with accessible weights. It has not yet published sufficient benchmark results for outsiders to judge those claims. Still, optimism surrounding Chinese AI and domestic hardware helped lift China’s technology-oriented ChiNext index by as much as 3.6 percent on Monday.

A Price War With Bigger Consequences

The more immediate threat may fall not on chipmakers, but on the companies selling access to AI models.

Kimi K3 is listed at roughly $3 per million input tokens and $15 per million output tokens. That places it below some top-tier American systems, though it is not universally cheaper and remains among the more expensive Chinese models. Its real cost advantage will depend on caching, reasoning settings, workload length and the number of model calls required.

If Chinese laboratories can offer comparable capabilities at consistently lower prices — while also allowing customers to download and modify their models — OpenAI and Anthropic could face weaker pricing power. That would complicate the financial story they must present to investors as both companies consider eventual public offerings.

Moonshot is pursuing an offering of its own. Reuters reported that the company is reorganizing its offshore structure for a possible Hong Kong listing and has engaged advisers including Goldman Sachs and China International Capital Corporation. It raised more than $2 billion in May, bringing its historical financing above $5.5 billion, and was valued at about $30 billion in June. The timetable for any listing remains fluid.

Separate reporting has put Moonshot’s annual recurring revenue at approximately $300 million in June. Because this figure has not been publicly audited, it should be treated as a company or investor estimate rather than established revenue.

Technology Under Political Direction

The release coincided with the World Artificial Intelligence Conference in Shanghai, where President Xi Jinping called for wider international access to AI while insisting that the technology remain secure and controllable.

In his address, Xi advocated a governance system combining laws, regulation, ethical safeguards, monitoring and early-warning mechanisms. The message reflected China’s dual ambition: to broaden the international reach of its AI industry while maintaining tight oversight at home.

That political backing may help Chinese developers raise capital, secure computing infrastructure and accelerate overseas expansion. It may also sharpen concerns in Washington over data security, censorship and the strategic implications of widely adopted Chinese models.

The New AI Divide

Kimi K3 does not yet prove that Moonshot has permanently overtaken OpenAI or Anthropic. Its complete weights and technical report have not been released, its benchmark advantages vary by task, and its early capacity problems demonstrate how costly frontier AI remains to operate.

But the model changes the nature of the competition.

The AI race is no longer confined to a small group of American laboratories competing primarily on raw capability. Chinese companies are combining open access, fast development cycles and aggressive pricing with models approaching the technological frontier.

For the American model providers, that threatens margins and valuations. For the chip industry, the implications are more ambiguous — and potentially more favorable.

Kimi K3 may make each individual task more efficient. But if cheaper and more accessible intelligence leads companies to run vastly more tasks, the world may need not fewer AI chips, but considerably more.

Sources: Reuters, Xinhua