Chinese open-source AI models gain traction, threatening U.S. AI giants

Author: Michael Spencer

Compiled by DeepChain TechFlow

Shenchao Overview: After the release of Moonshot AI’s Kimi K3, U.S. companies have rapidly shifted toward Chinese open-source models to cut token costs, undermining the IPO narratives of OpenAI and Anthropic. While U.S. hyperscale cloud providers burn cash building data centers, Chinese models achieve the same—or even better—results with far less computational power. This isn’t just a technological competition—it’s a realignment of AI power.

As you know, the Chinese company Moonshot AI has released a new version of its Kimi model, Kimi K3, driving numerous enterprises to shift toward open-weight AI models. This could be one of the most significant AI events of 2026. Amid the unresolved threat of war in the Strait of Hormuz and a sharp decline in the Nasdaq 100, which is heavily weighted with tech stocks, this development is evolving into a geopolitical and Trump administration dilemma. In both war and AI, there are clearly no straightforward solutions.

Chart: U.S. geopolitics once again triggers market turmoil—a highly unpopular and costly war.

Generative AI models are evolving, but may be slowing revenue growth for AI giants OpenAI and Anthropic. U.S. hyperscale cloud providers are approaching negative free cash flow due to staggering AI capital expenditures and data center investments, forcing you to question whether it’s all worth it—if Chinese models continue growing larger and more efficient, capable of replicating equivalent performance at significantly lower costs. This could evolve into an AI market crisis on Wall Street, even as the semiconductor boom appears to be entering a bear market correction, shortly after South Korea’s HBM leader SK Hynix listed in the U.S. South Korea (KOSPI) has now become a leading indicator.

With China’s DRAM manufacturer CXMT set to go public in Shanghai, this intensifies the high-stakes rivalry in AI between the U.S. and China. DeepSeek’s remarkable funding round, along with planned IPOs by DeepSeek, Moonshot AI, OpenAI, and others targeting 2027, ensures next year will be anything but ordinary. DeepSeek raised approximately $7.4 billion in funding last month in June. We must assume Databricks is also among the beneficiaries of this wave of enterprise AI shifting toward routing and cheaper tokens from open-source models—Databricks has just announced a new funding round at an $188 billion valuation.

Google's Gemini 3.5 Pro suffered a critical delay at the worst possible moment. The launch of SpaceXAI's new model, Grok 4.5, was completely overshadowed by the Kimi K3 moment. The chaos surrounding Anthropic's Fable 5 reignited China's ultimate DeepSeek moment from January 2025. Even OpenAI's own GPT 5.6 Sol release was almost entirely eclipsed.

The Trump administration quickly restricted models at the Mythos level, but faces a serious problem: what if Chinese models can approximate those same capabilities using open-weight models—capabilities once thought to require months to achieve? The Trump administration is showing signs of potentially banning advanced Chinese models and taking aggressive measures to curb China’s AI rise. I thought they wouldn’t regulate the AI industry. This is how global free-market capitalism ends.

Seven major cloud providers, including Microsoft, Amazon, Meta, and Google, lost some of their AI influence and credibility during this cycle. Mismanagement of Mythos-level models, combined with the rise of models like Kimi K3, created a perfect storm, prompting many enterprises to shift toward more rational and significantly more cost-efficient token usage.

Although Open-Router does not represent the global picture, it serves as an interesting data point for the broader macro trend: cheaper open-source models from China appear to be gaining market share.

Illustration: The era of prompts has given way to the realities of routing and token efficiency.

In my view, some of the best open-source weight models from the U.S. are Thinking Machine’s Inkling (released about a week ago) and whatever Reflection may soon launch. NVIDIA’s Nemotron is frequently cited as an alternative to Meta’s previous leadership. The issue is that the U.S. truly lacks leadership in open-source AI. For the U.S., this could be a potential disaster in its leadership across models, tokens, and enterprise adoption of cutting-edge AI.

More notably, Chinese President Xi Jinping’s most significant recent remarks on artificial intelligence were delivered via a keynote speech at the World Artificial Intelligence Conference (WAIC). President Xi attended the opening ceremony of the 2026 WAIC in Shanghai and delivered a keynote address titled “Join Hands to Build a Fair and Equitable Global AI Governance System.” China appears to be more advanced than the United States in AI governance and regulatory leadership, actively fostering global cooperation on this topic.

Alibaba’s own Qwen 3.8 Max (preview version) will also include an open-source component, paired with an aggressive international token pricing strategy. Kimi K3 is not aimed at hobbyist developers but at enterprise customers, to boost ARR before their anticipated IPO in about six months. The competition between the U.S. and China on models—even amid a scarcity of application-layer products—is inflating demand for compute power, while China benefits from cheaper token generation and more abundant energy. At a time when a key objective of the Trump administration appears to be sustaining the stock market’s AI-driven boom for financial elites and the business class, everyone from Anthropic to Moonshot AI is positioning themselves to maximize IPO hype and revenue sales momentum.

According to the Artificial Analysis Intelligence Index, ranked by the smartest models:

The current status of peak model performance is approximately as follows:

• Anthropic – Claude Fable 5
• OpenAI – GPT 5.6 Sol
• Dark Side of the Moon – Kimi K3 (Open Source Weights*)
• SpaceXAI – Grok 4.5
• Zhipu (Z.ai) – GLM 5.2 (Open Source Weights)

These lists and the benchmarks they're trained on are quite artificial and will likely change next week, and certainly next month.

Chart: Is comparing LLMs (Large Language Models) still the right way to measure progress in artificial intelligence? AI Analysis

Chinese open-source weight models have always been cheaper, but in 2026 they are becoming significantly more capable. This means U.S. companies are beginning to adopt them for everyday tasks to curb token budgets spiraling out of control due to pricing at the Fable 5 and Mythos 5 levels. If each new Chinese model has the potential to rank among the top five best models intelligently, this poses a major problem for the U.S. AI industrial complex. The issue of China’s AI advancement relative to the U.S. not only exists but is worsening.

I deeply respect the founders of DeepSeek, Moonshot AI, and Zhipu AI, because you can sense they possess genuine idealism—not just incredible AI talent and business acumen. Despite constraints on compute and funding, their models demonstrate sophisticated innovation and remarkable AI business insight. You don’t feel the same from executives at Meta, Google, or Microsoft—there’s a clear disconnect. Compared to the massive capital expenditures of big tech companies, their performance in generative AI models (and products) looks surprisingly poor.

Over the past week, a viral tweet by Dean Ball, OpenAI’s newly appointed Head of Strategic Futures, on X best illustrates this point. He argued that open-source weight models are essentially减速主义, and pondered how a world dominated by open-source weight models could lead to full-blown AI communism, resulting in a dystopian hellscape.

Chart: In July 2026, panic surrounding open-source AI reached a new high. The rhetoric about China has become tiresome.

Clearly, capital-intensive, closed-source model companies are concerned about these recent developments, filled with anti-Chinese sentiment and defense-related arguments regarding distillation and cybersecurity. Although Anthropic’s Mythos-level models appear impressive on paper, we do not know their most advanced capabilities because they have been restricted from public access. In the U.S., the line between AI future strategists and lobbyists appears quite blurred. As capital expenditures in the corporate AI token price war are quite poor from an ROI perspective.

Chart: Large tech companies' capital expenditure plans look bleak through 2027. — Bank of America

Our understanding of American protectionism is that anything they cannot compete with in the global market, they will inevitably try to ban or restrict domestically. But by 2026, the dominant story will already be enterprises shifting toward open-source weight models—perhaps too late, as this move could force American companies into an open-source AI race. This is the paradox of open-source AI’s compelling cost argument in 2026: if generative AI technology truly delivered reliable ROI, this wouldn’t be an issue—but it’s still an emerging technology, and there are currently few high-quality AI products capable of leveraging these incredible models.

Last year, the U.S. Department of Commerce considered adding several Chinese AI labs to its Entity List; I’m confident this issue is back on the table now, as U.S. companies struggle to compete in a world where token efficiency is rapidly improving, tools and routing have become more critical than ever, token costs are spiraling upward, and decisions about which models to use for which tasks are increasingly complex.

The U.S. government’s restriction on Anthropic, its top closed-source company’s best model, may be a historic mistake that led to this painful situation. The Kimi K3 moment is made even more dramatic by the ironic timing—just months before Anthropic’s own high-profile AI IPO—while Semianalysis’s publication analyzed Anthropic’s incredible operating profits. You must assume that the shift toward open-source weight models and government interference with Mythos has slowed America’s top AI company. Semianalysis believes Anthropic is on track to achieve over $1 billion in GAAP EBITDA by Q3 2026 (approximately a 6% margin), making it one of the first major frontier AI labs to achieve consistent quarterly profitability. All these recent events have made OpenAI, Meta, and SpaceXAI the biggest losers—even though Meta’s Muse Spark 1.1 is not such a poor model.

If the Trump administration restricts Chinese open-source weight models, NVIDIA, Thinking Machines, Meta, and Reflection (which has not yet released its model) could become the biggest beneficiaries for enterprise customers seeking U.S.-based open-source weight token solutions.

The macro AI narrative in mid-2026 becomes more compelling. The customization level of open-source AI is advancing. Token efficiency and routing have become more critical. AI capital expenditures in the market are facing increasing scrutiny.

Open-source weight models: AI hegemony is not AI communism or a nuclear threat; cheaper tokens benefit Jevons' paradox and expand what companies, developers, and consumers can do with AI. Over the past five years, the entire generative AI model training paradigm has been a multi-string distillation hijacking of the world’s language data. Demand for computing power will not slow down because enterprises choose cheaper models—in fact, it will accelerate. This is the central dilemma of the entire macro AI landscape.

If capital and monopolistic capitalism are their moat, then the United States is a nation in trouble. The future of AI will require new kinds of innovation, not just better venture capital systems. Clearly, there is some narrative abuse at play—business consolidation and geopolitics will resolve themselves as usual. As DeepSeek raises more funding and rushes toward an IPO, its stack keeps growing. Their flagship model, DeepSeek-R2, has never been released, raising questions about the future of China’s most advanced models. Liang Wenhong holds 78% to 84% of DeepSeek, and they are also developing their own chips.

A world where both Anthropic and Moonshot AI are constrained by compute capacity, Google has become a laggard in LLMs, and Meta and SpaceX AI remain on the margins. A world where OpenAI’s IPO looks less appealing each quarter. A world where we’re tired of waiting for IPOs from Databricks, Crusoe, Anduril, Stripe, and others—even as Chinese physical AI, AI chip, and storage giants rush to go public. It is reportedly said that China’s chip manufacturer CXMT’s $8.6 billion IPO was oversubscribed by more than 500 times by institutional investors.

I predict that 2027 will see the ChatGPT moment for robotics—now also called physical AI—because Anthropic is reportedly negotiating to acquire Physical Intelligence. Even though AI pioneer Yann LeCun has warned that these humanoid robots will remain extremely incapable now and for a long time to come, Silicon Valley and Wall Street will rush to find new narratives to sustain AI-themed stock markets, and they may once again overhype the potential.

This occurred during a period when the U.S. market severely lacked publicly listed companies focused solely on robotics or robotic software. China, still struggling to recover from its epic real estate collapse, appeared to be leading the IPO race. While Moonshot AI, Zhipu, Minimax, and smaller Chinese AI labs are impressive, what truly captivated me was the macro dynamic between open-source and closed-source models. In many ways, this is a battle between capital and talent.

"In my view, this future is a dystopian nightmare, yet I’ve never met an open-weight model advocate who doesn’t eventually admit things will go this far," said Dean Ball—or ironically, OpenAI.

"In my view, this future is a dystopian nightmare, yet I have never met an advocate of open-weight models who doesn't eventually admit things will go this far," said Dean Ball—or ironically, OpenAI.

As major tech companies prepare to release their earnings reports this week, we’ll gain more clarity. The narrative around capital expenditures and return on investment has never been more severe. Any misstep in earnings execution will be punished by the market; over the past few weeks, valuations of several growth-oriented companies have been revised downward—SpaceX fell 25.5%, while Grok 4.5 once again failed to make an impact (the bar is high). With its expensive acquisition of Cursor, SpaceX AI still appears to be grouped among the AI losers.

Moonshot AI launched Kimi K3 on July 16, a large model with up to 2.8 trillion parameters that performs comparably to top U.S. systems like Claude and GPT in benchmark tests, and plans to fully open its weights by July 27 to enable low-cost self-hosting.

By 2027, we will know how strong Anthropic’s IPO will be, as its fastest ARR growth in software history faces headwinds from macro AI trends and government intervention. Will U.S. protectionism shield it, or will the market choose the winner? This tension will unprecedentedly extend into a robotics race encompassing space, AI, and defense.

If Anthropic truly acquires Physical Intelligence, the robot GPT moment I’ve mentioned multiple times may be just around the corner. Unitree Technologies’ Chinese IPO is already a major milestone for the future of physical AI—it aims to raise RMB 4.2 billion ($619.4 million) and has recently slashed the price of its entry-level quadruped robots by more than 90%.

Due to China’s natural advantages in robotics, this segment of the AI race is an area where the U.S. is desperately trying to catch up, with multiple large-scale VC funding rounds emerging in this field in 2026. The U.S. appears to lead in world models and software-brain startups focused on making robots more practical, while China has a greater number of humanoid robot (robots with human-like form) startups, along with experiments in drones and novel robotic hardware forms.

Unitree Technologies is expected to go public later this month, roughly three months before the key Anthropic IPO. Unitree is rapidly becoming the de facto standard hardware platform for global AI researchers and robotics labs. The humanoid robot race may heat up so quickly that Tesla will be forced to merge with SpaceX in the near future—I believe this could happen by early 2028. If the generative AI boom cools down from its peak, as many analysts anticipate, I expect the robotics race to replace it as the next major U.S. tech narrative.

The eventual competition between China and the U.S. in AI is beneficial for developers, consumers, and businesses worldwide in unlocking greater real-world utility. Whether generative AI models or robotics will deliver tangible value soon is a separate issue from the scale of investment. The accelerated timelines of startups and significantly raised funding rounds are noteworthy, even as the urgency to commercialize AI grows. Competition is intensifying and will become even fiercer by 2027, laying a massive foundation for technological rivalry in the 2030s. China’s abundant energy and America’s dominance in semiconductors keep the situation uncertain.