Author: Michael Spencer

Translation: Shen Chao TechFlow

Shen Chao Guide: After the release of Kimi K3 by Moonshot AI, American companies have turned to Chinese open-source models to cut token costs, shaking the IPO narrative for OpenAI and Anthropic. While American hyperscale cloud providers spend exorbitantly on data centers, Chinese models achieve the same or even better results with significantly less computing power—this is not a technological competition, but a reshuffling of AI power.

Geopolitics and AI Confrontation 🔥

As you know, the Chinese company Moonshot AI has released a new version of its Kimi model, Kimi K3, prompting a large number of enterprises to shift towards open-source weight models for AI. This could be one of the most significant AI events of 2026. Amidst the unresolved Iran war in the Strait of Hormuz and the dramatic decline of the tech-heavy Nasdaq 100 index, this is evolving into a geopolitical dilemma with the Trump administration. In war and the realm of AI, there is clearly no clear-cut solution.

Image: American geopolitics reignites market turmoil. An extremely unpopular and costly war.

Generative AI models are evolving, but they may be dragging down revenue growth for AI giants OpenAI and Anthropic. American hyperscale cloud providers are nearing negative free cash flow due to astonishing capital expenditures and investments in data centers, and one must question whether all of this is worth it—if Chinese models continue to grow larger and more efficient while greatly reducing costs that can replicate comparable performance. This could lead to a crisis for AI stocks, even as the semiconductor boom seems to have entered a bear market correction following the listing of Korean HBM leader SK Hynix in the US. South Korea (KOSPI) has now become a leading indicator.

With Chinese DRAM manufacturer CXMT set to go public in Shanghai, this presents a high-pressure situation for the US-China showdown in the future of AI. DeepSeek's impressive funding rounds, as well as plans by DeepSeek, Moonshot, OpenAI, and others to IPO in 2027, ensure that next year will be extraordinary. DeepSeek raised approximately $7.4 billion in funds last month. We must assume that Databricks is also among the beneficiaries of this corporate AI shift towards routing and the cheaper tokens of open-source models—Databricks has announced a new funding round, valued at $188 billion.

Google's Gemini 3.5 Pro faced critical delays at the worst possible time. The launch of SpaceXAI's new model Grok 4.5 was completely overshadowed by the Kimi K3 moment. The confusion surrounding Anthropic's Fable 5 brings China back to the ultimate return of the DeepSeek moment of January 2025. OpenAI's own GPT 5.6 Sol release was also nearly entirely eclipsed.

The Trump administration quickly restricted Mythos-level models but faces a serious problem: What should they do if Chinese models can approach the same capabilities using open-source weight models—capabilities that were once thought to take months to achieve? The Trump administration has begun to show signs of potentially banning cutting-edge Chinese models and taking drastic actions to contain China's AI rise. I thought they would not regulate the AI industry. This is global free-market capitalism.

A Brave New World of Token Efficiency Looks Very Chinese

The seven major hyperscale cloud providers—including Microsoft, Amazon, Meta, and Google—have lost some AI discourse power and credibility during this cycle. The government's mishandling of Mythos-level models, combined with the rise of models like Kimi K3, has created a perfect storm, forcing many corporations to shift toward a much more rational and cost-efficient use of tokens.

While Open-Router does not represent the overall situation, it serves as an interesting data sampling point of the overarching macro trend: cheaper open-source weight models from China appear to be winning market share.

Image: The era of hints has given way to the reality of routing and token efficiency.

In my view, some of the best open-source weight models from the US are Thinking Machine's Inkling (released about a week ago) and any product Reflection might launch soon. Nvidia's Nemotron is often cited as a substitute for Meta's previously leading position. The issue is that the US lacks true leadership in the open-source AI space. For the US, this is a potential disaster concerning its edge AI leadership in models, tokens, and enterprise adoption.

Curiously, China's President Xi Jinping's most significant recent remarks on artificial intelligence were made in 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 titled "Working Together to Build a Fair and Just Global AI Governance System." China appears to be more advanced than the US in AI governance and regulatory leadership, actively building global cooperation around this theme.

Alibaba's own Qwen 3.8 Max (preview version) will also have an open-source weight component, accompanied by an aggressive international token pricing plan. Kimi K3 is not aimed at amateur developers, but rather at enterprise customers, to enhance ARR before they also IPO in about six months. The competition between the US and China in models is exaggerating the demand for computing power, even though there are currently few application-layer products, as China offers both cheaper token generation and more abundant energy. Meanwhile, the Trump administration's crucial task seems to be maintaining stock market AI prosperity for financial elites and the business class. At the same time, everyone from Anthropic to Moonshot is adjusting their positions to maximize IPO hype and revenue momentum.

Does Model Ranking Still Matter?

If we rank by the Artificial Analysis Intelligence Index and by the smartest models, the rankings are as follows:

The current state of peak model performance is roughly as follows:

• Anthropic – Claude Fable 5
• OpenAI – GPT 5.6 Sol
• Moonshot AI – Kimi K3 (open-source weights*)
• SpaceXAI – Grok 4.5
• Z.ai – GLM 5.2 (open-source weights)

These lists and their training benchmarks are quite arbitrary and are likely to change next week, definitely next month.

Image: Should comparing LLMs (language learning models) still be the right way to measure AI progress? AI Analysis

"AI Disconnection" is Worsening 🔎

Chinese open-source weight models have always been cheaper, but as of 2026, they are becoming significantly more capable. This means American companies are starting to adopt them for everyday tasks to mitigate runaway token budgets due to the Fable 5 and Mythos 5 level pricing. If every new Chinese model has the potential to enter the top five best models in intelligence ranking, this is a major issue for the American AI industrial complex. The problem of China in AI compared to the US not only exists but is worsening.

I have great respect for the founders of DeepSeek, Moonshot AI, and Z.ai, as you can feel their genuine idealism, not just incredible AI talent and business acumen. Under their constraints of computing power and funding, their model capabilities display sophisticated innovation and incredible AI business acuity. You do not feel the same from executives at Meta, Google, or Microsoft; there is a noticeable disconnect. Relative to the capital expenditures of big tech companies, this is starting to look really bad. Big tech's execution on generative AI models (and products) is surprisingly poor.

Last week, OpenAI's new strategic future head, Dean Ball, made perhaps the most explanatory tweet on X. He noted that open-source weight models are inherently about slowing down and considered how a world dominated by open-source weight models might lead to a thorough AI communism, and consequently, a dystopian hellscape.

Image: In July 2026, the panic surrounding open-source artificial intelligence reached a new height. Commentaries about China have become tiresome.

Clearly, capital-intensive closed-source model companies are concerned about these recent developments, filled with anti-China sentiments and national defense arguments regarding distillation and cybersecurity issues. While Anthropic's Mythos-level models are impressive on paper, we do not know their cutting-edge capabilities because they have been restricted from public release. In the US, the boundaries between AI future strategists and lobbyists appear quite blurred. As capital expenditures in the corporate AI token price war look rather poor from an ROI perspective.

Image: The capital expenditure plans of major tech companies look bleak before 2027. — Bank of America

Our understanding of American protectionism is that they will try to ban and restrict things in the global market that they cannot compete with. But the corporate shift to open-source weight models has become the dominant story of 2026, which could already be too late; in doing so, they may be setting up a race for open-source AI for American companies. This is the peculiar dilemma of open-source AI regarding costs in 2026. If generative AI technology indeed delivers reliable ROI, this will not be an issue, but this is an emerging technology, and there are currently few excellent AI products to leverage these incredible models.

The US Department of Commerce considered adding several Chinese AI labs to its "Entity List" last year, and I am sure this has come back to the table, as American companies struggle to compete in a world where token efficiency is rapidly increasing, tools and routing are more important than ever, token costs are spiraling upward, and deciding which models to use for which tasks is challenging.

The US government's restriction on the best models from its leading closed-source company, Anthropic, may prove to be a historic mistake that has allowed this painful situation to arise. The Kimi K3 moment is made more dramatic by the comedy of errors just months away from Anthropic's own high-profile AI IPO, as Semianalysis analyzed Anthropic's incredible operational profit. One must assume that the shift to open-source weight models and government hijacking of Mythos have slowed down America's top AI firms. Semianalysis predicts that Anthropic is expected to reach over $1 billion in GAAP EBIT (approximately 6% profit margin) by the third quarter of 2026, making it one of the first major edge AI labs to achieve sustained quarterly profitability. All of these recent events have made OpenAI, Meta, and SpaceXAI the biggest losers. Even if Meta's Muse Spark 1.1 is not such a bad model.

If the Trump administration restricts Chinese open-source weight models, Nvidia, Thinking Machines, Meta, and Reflection (which has yet to release a model) might emerge as the big winners of American-based open-source weight token solutions for enterprise customers.

The macro AI narrative in mid-2026 is becoming more interesting. The level of customization in open-source AI is upgrading. Token efficiency and routing have become more important. AI capital expenditures in the market are receiving increasing scrutiny.

Many Stories of DeepSeek Moments

The hegemony of open-source weight model AI is not AI communism or a nuclear threat; cheaper tokens benefit the Jevons Paradox and what companies, developers, and consumers can do with AI. For the past five years, the entire generative AI model training paradigm has been a multivariate distillation hijack of language data based on the world. The demand for computing power will not slow down as enterprises choose cheaper models; in fact, it will accelerate. This is the entire macro AI situation's dilemma.

If capital and monopoly capitalism are their moats, the US is a country in distress. The future of AI will require a new type of innovation, not just a better venture capital system. There is clearly some narrative abuse; business integration and geopolitics will resolve themselves, as always. As DeepSeek raises more funds and rushes toward IPO, the accumulation of DeepSeek moments continues. Their flagship model DeepSeek-R2 has never been released, raising questions about the future cutting-edge models from China. Liang Wenfeng holds 78% to 84% of DeepSeek's shares; they are also developing their own chips.

The Roaring 20s IPO Race

A world where both Anthropic and Moonshot are constrained by computing capacity, Google has become an LLM laggard, and Meta and SpaceXAI remain on the margins. An OpenAI IPO looks less and less appealing each quarter. A world where we are tired of waiting for IPOs from Databricks, Crusoe, Anduril, Stripe, and others, even as Chinese physical AI, AI chip, and storage giants are eager to go public. Reportedly, Chinese chip manufacturer CXMT's $8.6 billion IPO was oversubscribed by more than 500 times by institutional investors.

I predict a ChatGPT moment in the robotics field in 2027 (now also referred to as physical AI), as rumors suggest that Anthropic is negotiating the acquisition of Physical Intelligence. Even if AI pioneer Yann LeCun warns that these humanoid robots will be very incompetent now and for a long time to come, Silicon Valley and Wall Street will be eager to find new narratives to support AI concept stocks, and they may exaggerate once again.

This is happening at a time when the US market severely lacks publicly listed companies in pure robotics or robotics software. China, which is struggling to recover from an epic real estate crash, seems to be leading the IPO race. While Moonshot AI, Z.ai, Minimax, or smaller Chinese AI labs are impressive, what truly fascinates me is the macro momentum of the open-source versus closed-source battle. In many ways, this is a contest of capital versus talent.

>"Such a future appears to me as a dystopian hellscape, but I have never encountered an advocate for open weight models who does not ultimately concede that things are headed in that direction," Dean Ball ironically states as a representative of OpenAI.

"Such a future appears to me as a dystopian hellscape, but I have never encountered an advocate for open weight models who does not ultimately concede that things are headed in that direction," Dean Ball ironically states as a representative of OpenAI.

As earnings reports from big tech companies are set to be released this week, we will get more answers. The story of capital expenditures versus ROI has never been so severe. Any failure in execution on earnings reports will incur market punishment, and in recent weeks, valuations for some growth companies have been corrected, with SpaceX down 25.5%, and Grok 4.5 again failing to make an impact (expectations are high). With the high-priced acquisition of Cursor, SpaceX AI seems to remain among the AI losers.

Moonshot AI launched Kimi K3 on July 16, a large model with up to 2.8 trillion parameters, which in benchmark tests matches top American systems like Claude and GPT and plans to fully open weights for low-cost self-hosting by July 27.

By 2027, we will know how strong Anthropic's IPO will be, as its software technology has the fastest ARR growth in history, currently challenged by macro AI headwinds and government intervention. Will American protectionism safeguard it, or will the market choose the winners? This tension will extend unprecedentedly into a robotic race that encompasses space, AI, and defense.

The GPT Moment for Robotics is Approaching

If Anthropic does indeed acquire Physical Intelligence, the robotic GPT moment I have mentioned several times may be just around the corner. Yushu Technology's Chinese IPO has already become a significant milestone for the future of physical AI. It will raise 4.2 billion yuan (619.4 million USD) while significantly lowering the price of entry-level quadrupedal robots by over 90% recently.

Due to China's inherent advantages in robotics, this part of the AI race is what the US is frantically trying to catch up with, as several large-scale VC funding rounds have occurred in this sector in 2026. The US seems to be leading in the world of model and software brain startups dedicated to making robots more practical, while China has more humanoid robot (robots with human form) startups as well as experiments involving drones and new forms of robotic hardware.

Yushu Technology should be listed later this month, approximately three months before the critical Anthropic IPO. Yushu Technology is rapidly becoming the de facto standard hardware platform for global AI researchers and robotics labs, and the humanoid robot race may heat up so quickly that Tesla may have to merge with SpaceX in the near future. I believe this could happen in early 2028. If the generative AI craze subsides from its peak (as many analysts expect), I believe we can anticipate that the robotics race will take its place as the next major narrative in American technology.

Ultimately, the competition between China and the US in the AI field will be beneficial for global developers, consumers, and enterprises to achieve more real-world utility. Whether generative AI models or robotics will quickly deliver practical value is another question compared to the scale of investment. The accelerated timelines for startups and significantly heightened funding rounds are noteworthy, even as the urgency to accelerate AI commercial applications is also increasing. Competition is heating up, and 2027 is set to become even more intense, laying a huge foundation for the technology competition of the 2030s. China's ample energy and America's semiconductor dominance keep the situation in suspense.