Open weights vs. closed borders

news23072026

For many, the global artificial intelligence market seems like a territory of an unshakable status quo. The rules of the game were previously dictated by a narrow circle of American corporations, and the industry itself developed along an extensive path: more parameters in the model, more transistors on a die, and more advanced process technologies from Taiwanese silicon manufacturers. Access to cutting-edge technologies was measured by the subscription cost for closed APIs and the volume of purchases of scarce graphics cards, creating an illusion of comfortable Western technological leadership.

However, the foundation of this monopoly was fragile logistics—unhindered access to Western hardware and software stacks. When geopolitical contradictions escalated into strict export controls, the familiar model cracked. Finding itself behind an artificially erected digital barrier, the Chinese tech sector faced an existential threat of its data centers grinding to a halt. Instead of resigning itself to the role of a perpetual follower or trying to break through the wall of sanction policy, Beijing initiated a radical paradigm shift.

What began as a forced measure for survival swiftly transformed into a systemic engineering revolution. The Chinese industry made a conceptual breakthrough, shifting the competition from the plane of microelectronics—where TSMC’s dominance seemed absolute—to the plane of system architecture. While Western engineers were squeezing the last percentages of performance out of individual GPUs, their colleagues in China began designing computing clusters as unified living organisms.

Hardware Independence: The End of the Nvidia Era

DeepSeek CEO Liang Wenfeng’s statement about erasing the boundaries between Chinese and American AI is based not on software optimizations, but on a radical rethinking of the physics of computation itself. Trapped behind a digital iron curtain due to US export restrictions, Beijing had to abandon attempts to catch up with TSMC in the nanometer race. Instead, Chinese engineers executed a flanking maneuver, shifting the competition from the level of a single transistor to the level of the entire data center architecture.

Systemic Leap: The Atlas 950 SuperPoD Architecture

Instead of purchasing thousands of discrete video cards, Chinese developers switched to monolithic compute nodes. Huawei’s flagship supercomputer, the Atlas 950, represents a SuperPoD architecture—a single logical organism uniting an array of 1,024 Ascend 950 neural accelerators.

The key innovation was the abandonment of the traditional PCIe bus in favor of HCCS (Huawei Cache Coherent System) technology. Thanks to ultra-low latency data exchange, the cluster functions as a single giant processor with a shared memory address space of 256 TB. This allows scaling the system up to 8,192 chips in a single domain, with a theoretical supercluster limit of half a million cores.

For corporate clients whose old data centers are not adapted for aggressive liquid cooling, the Atlas 850E node with an air-based heat dissipation system was released. The aggregate power of such systems reaches 8 ExaFLOPS for model training (FP8) and 16 ExaFLOPS for inference (FP4), with an internal interconnect bandwidth claimed at 16.3 petabytes per second.

The Math of Survival: Token Economics vs. Nvidia’s Price List

US export sanctions became an economic gift for Chinese developers. Deprived of access to Hopper and Blackwell GPUs, they were forced to create infrastructure where the main efficiency criterion became the cost of generating a single token, rather than the peak crystal performance in TFLOPS.

The result was devastating for American venture capital. Chinese open-weight models crashed market prices:

DeepSeek V4 Flash: $0.09 input / $0.18 output per million tokens.
Kimi K3: $3.00 input / $15.00 output.
Western closed analogs (OpenAI, Anthropic): $5–7 input / $30–60 output.

The cost gap is dozens of times. Consequently, business began mass migration: by July 2026, the share of Chinese providers on the OpenRouter wholesale platform reached 58%, leaving American giants with only 42% of the traffic.

Hardware is useless without drivers. Understanding that Nvidia’s proprietary CUDA stack is the main anchor holding developers, the Chinese IT sector simultaneously attacked this monopoly:

Alibaba brought to market a bundle of Hanguang accelerators and Panjiu servers with its own high-speed data transfer protocol.
ZTE initiated the creation of the Matrix alliance—an open interface architecture designed to forever eliminate the risk of vendor lock-in.
Biren and MetaX presented mature versions of their own software SDKs, acting as direct CUDA analogs. These stacks allow transferring neural networks trained in the West to Chinese hardware with practically a single click of compilation.
Sugon presented the Dengfeng supercluster (Sugon 8000), which won the WAIC 2026 prize for revolutionary component mounting density and new immersion cooling schemes.

A symbol of complete autonomy was the switch to its own HiBL 1.0 memory in Ascend 950PR accelerators. Even if it currently lags behind classic HBM standards in peak bandwidth, this step guarantees China’s immunity to Western microelectronics shortages.

Theoretical announcements have already turned into industrial routine. The DeepSeek V4 model is fully rewritten for the Ascend architecture and runs on Huawei’s own hardware at real business scale.

A new reality of the global AI market is forming. The American development model hits the physical limit of silicon and the cosmic cost of infrastructure. The Chinese model offers the world modular, infinitely scalable, and affordable systems. The era of Jensen Huang’s unconditional dictate is over: it has become mathematically unprofitable for global business to pay a fifty-fold markup for a brand when a competing system delivers comparable results on open-source algorithms.

Closed APIs vs. Open Weights: The Hugging Face Breach

The investigation into the July 2026 hack on the Hugging Face platform exposed a deep systemic crisis within the Western tech industry. For the first time in history, the largest platform for AI developers was attacked by a fully autonomous algorithm acting without human participation. However, the true watershed was not the breach itself, but the helplessness of American commercial giants in the face of a threat that an open Chinese model managed to mitigate.

The Dead End of Proprietary APIs: Why OpenAI Tools Were Useless

During the investigation (DFIR), the Hugging Face security team faced an insoluble operational contradiction. Analyzing the complex attack chain required feeding raw invasion artifacts into an LLM: RAM dumps, fragments of stolen source code, and service access tokens.

Using OpenAI or Anthropic cloud interfaces in this situation would have meant capitulation. Transferring such sensitive data to external providers via public APIs created an unacceptable risk; the company would effectively be broadcasting the details of its own vulnerability to a third party’s servers. Furthermore, it threatened catastrophic legal consequences due to the violation of corporate client confidentiality.

The solution was found in deploying a local instance of the Chinese GLM 5.2 model. Possessing open weights, it allowed for a full forensic examination within the company’s isolated perimeter. It was this tool that managed to reconstruct the trajectory of the autonomous agent, revealing the exploitation of zero-day vulnerabilities in dataset processing. Western flagships lost the security battle by the very fact of their existence as closed clouds.

The Hugging Face incident marked the birth of a new class of cyber threats. Attackers employed a «repository poisoning» method, embedding a malicious dataset. Once inside the infrastructure, AI agents switched to autonomous actions: they independently studied network topology, elevated account privileges, and performed lateral movement between clusters at a speed inaccessible to a human operator.

The intrusion was detected only thanks to a symmetrical response—the security system of Hugging Face itself, also running on AI. Algorithm-to-algorithm confrontation excludes the human factor: the time from detection to full compromise is reduced to minutes. Agents are capable of building complex multi-stage attack scenarios, improvising on the fly, which makes traditional signature-based defense methods completely obsolete.

For global business, the conclusions from this incident are existential. The collapse of the trusted black box concept. Corporations will have to admit the harsh reality: sending internal logs and critical data to external commercial LLM services is tantamount to voluntarily handing over safe keys to competitors.

Machine learning pipelines are now classified as the highest-risk surface. The only way to ensure a real level of information security is a complete abandonment of external APIs in favor of on-premise solutions. Open-weight models, such as GLM, Qwen, or Nemotron, are becoming the de facto standard for state structures and the enterprise segment, as only they guarantee the absence of uncontrolled intellectual property leakage beyond the protected perimeter.

While Chinese labs scaled open-source ecosystems, Western monopolists trapped themselves in hyper-security to preserve market valuation. The deficit of competitive Western open-source alternatives forced the rescue industry to turn to the architecture of a potential geopolitical adversary.

The first serious attempt to close this gap was made only in the summer of 2026 with the release of the Inkling model with 975 billion parameters. The project, led by former OpenAI CTO Mira Murati, is intended to become the foundation for the safe corporate adoption of open architectures. However, the release looks like a forced attempt to catch a departing train: it followed only after the Chinese approach had proven its superiority in real combat. Against the backdrop of periodic OpenAI service outages due to US export controls, the Western market found itself facing the prospect of total technological dependence on tools it is itself afraid to use for security reasons.

State Capital as the Main Driver of Silicon

ChangXin Memory Technologies’ (CXMT) listing on the Shanghai STAR Market marks the end of the Western monopoly on RAM. Raising $8.6 billion in the largest Asian IPO of 2026 is not just a financial transaction, but the pinnacle of China’s doctrine of «state venture capitalism.» Ten years after its founding with a tiny capital of 10 million yuan, CXMT has become the fourth-largest DRAM producer in the world, providing the Celestial Empire with a technological safety cushion.

The basis of CXMT’s success lies in the unprecedented concentration of administrative resources in the city of Hefei. The share of state structures reached 36.8%, turning a share package worth 213 billion yuan ($31.5 billion) into an asset comparable to 15% of the entire gross regional product of the metropolis.

However, the key difference of this placement from Western analogs lies in the investor composition. The shares were bought by the architects of China’s AI ecosystem themselves—Alibaba, Tencent, and Xiaomi. In this paradigm, money is secondary; industrial synergy is primary. The state has built a vertically integrated cartel, where the memory manufacturer receives guaranteed sales channels within the country, and tech giants gain immunity against American sanctions. This artificially smooths the cyclicality of the semiconductor market for CXMT but makes the company’s valuation derivative of geopolitical stability rather than quarterly profit reports.

For DeepSeek’s data centers and Huawei Atlas supernodes, own DRAM became a matter of physical survival. Without CXMT’s capacity, Chinese developers’ servers would have turned into scrap metal due to the US embargo. Today, the company holds about 8% of the global DDR5 and LPDDR5X market, offering chips 5–10% cheaper than Korean competitors.

Paradoxically, Beijing benefits from American technological leadership in HBM. Industry leaders—SK Hynix, Samsung, and Micron—are so busy satisfying the explosive demand for super-fast memory for Nvidia and AMD accelerators that they consciously cut supplies of regular server memory. CXMT occupies the vacated niche, receiving an anomalously high margin against the backdrop of a general silicon shortage. This success is already penetrating sanction barriers: Apple is actively negotiating the purchase of Chinese chips for devices intended exclusively for the domestic PRC market, trying to find loopholes in Washington’s export control.

The Performance Trap: Factory Growth Amid Wage Stagnation

Behind the facade of industrial triumph lies a troubling macroeconomic diagnosis. A semiconductor factory is a perfect machine for generating GDP, but a highly inefficient tool for raising living standards. Building clean rooms requires billions in investment but creates a minimum of jobs and practically does not stimulate domestic consumer demand.

The Chinese model generates a «Frankenstein economy»: colossal industrial export neighbors with household income stagnation. Huge state reserves are frozen in capital-intensive microchip production instead of funding the social sphere. Meanwhile, the current market euphoria looks frighteningly fragile. A P/E ratio of over 300 indicates that investors are buying not financial stability, but faith in infinite deficit. The memory market is notorious for its brutal corrections: when new factories from CXMT and its competitors reach full capacity, inevitable oversupply will crash prices, leaving the state with overvalued assets on its balance sheet.

The Glass Ceiling of Lithography

The destruction of the thirty-year oligopoly of the Big Korean Trio (Samsung, SK Hynix, Micron) has its physical limits. CXMT’s success today is based on mature process technologies that are still accessible to China. However, the next step—entering the premium HBM memory segment and mastering advanced production norms—runs into the wall of export control.

Further development is impossible without access to Dutch ASML steppers and Japanese materials, the supply of which is blocked at the political level. Therefore, a heavy discount for geopolitical risk is built into CXMT’s market value. Speculative hype around pre-IPO contracts has inflated capitalization expectations to an astronomical $425 billion, but the real test of Beijing’s strategy will begin when artificial deficit gives way to oversupply. Then the state will have to answer the question: will revenues from cheap RAM sticks pay off the colossal subsidies with which they have fed their new national champion for a decade?

OpenRouter Records the End of American Monopoly and the Birth of a Two-Loop World

Statistics from the independent OpenRouter hub, through which businesses buy wholesale access to neural networks, destroyed the last illusions of global technological unity. The share of Chinese AI models reached an unprecedented 58%, knocking American developments off the pedestal for the first time in a decade.

At the beginning of 2025, the balance of power seemed unshakable: the US controlled about 80% of LLM traffic, China received a modest 10%. Eighteen months later, the situation flipped. In June 2026, Chinese providers (DeepSeek, Tencent, Xiaomi, MiniMax) collectively provided 44% of the token volume among the top ten players, while the share of Google, OpenAI, and Anthropic collapsed to 30%.

The point of no return was the week of February 9–15, 2026. For the first time in history, Chinese architectures processed more data than Silicon Valley servers—4.12 trillion tokens versus 2.94 trillion. By summer, this gap turned into an abyss: the weekly volume of the Chinese segment reached 18 trillion tokens against 5.5 trillion for Americans (a ratio of more than 3:1). The platform’s total traffic grew explosively—from 5 trillion tokens in spring 2025 to over 20 trillion in spring 2026.

DeepSeek became the main beneficiary. In six months, the company doubled its share on the platform, from 9% to 18%, becoming the world’s largest provider by generation volume. The flagship V4 Flash model, costing only $0.09 / $0.18 per million tokens, effectively monopolized the agentic segment, providing up to 70% of all such load within the company’s ecosystem.

The West found itself in a catch-up position. While OpenAI and Anthropic hid their best models behind closed APIs or disabled them due to export controls, the Chinese market switched to a strategy of total openness. The first serious open Western model, Inkling (975 billion parameters), appeared only in the summer of 2026, when the train of American influence had already left the station.

WAICO as an Alternative to the West

Technological triumph required political formalization. On the sidelines of the World Artificial Intelligence Conference 2026 (WAIC), an act was signed to create the World Artificial Intelligence Cooperation Organization (WAICO). The new structure, headquartered in Shanghai, united 29 states, including Russia and key BRICS partners.

The creation of WAICO marks the dismantling of the old order. The global AI market has officially ceased to be monopolar. An architecture of «two speeds» is forming, where instead of a single model internet, two parallel realities emerge:

Walled Garden Ecosystem: American corporations sell access to their neural networks as a scarce premium-class service, strictly limiting technology use by other countries’ state bodies.
Open Weights Ecosystem: China commoditizes the basic AI layer. By providing advanced open-source models for free, Beijing hooks global development on its tech stack needle. Engineers from California to Jakarta now write code for the Ascend architecture and DeepSeek frameworks.

For international business, this means inevitable infrastructure cost increases. The era of universal corporate applications is over. Developers will have to maintain two parallel technology stacks, separated not only by ideology but also by regulatory requirements.

However, it was China’s pragmatism that became its main weapon. Understanding that catching up with the West in chip lithography is impossible, Beijing won the war at the software and economic level. By forcing American giants to close up to protect capitalization, China left the world market hungry. By offering a free alternative, it gained control over the global developer base. Technological sovereignty today is defined not by who invented the algorithm, but by whose free tools millions of programmers around the world use.

Conclusion

Artificial intelligence ceases to be a single global science and finally splits into two contours. The American path leads to the creation of super-powerful but closed, expensive, and politically controlled dual-use tools. The Chinese path offers the world cheap, productive, fully autonomous (on-premise) solutions with open source code, subsidized by the state. In the short term, the Chinese approach wins the race for mass adoption, democratizing access to technology. However, the concentration of huge state resources in the hands of a narrow circle of high-tech companies creates new risks—both economic and geopolitical.

Let's start a conversation

Start Scaling Today

Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.

Location: Moscow, ext.ter.g., sh. Ochakovskoe, 28, p.2
GPU Servers High-density compute
Облачная аренда Быстрый старт ИИ моделей
Enterprise SLA Guaranteed availability
Low Latency Optimal network routes