The $593.0 Billion Panic

January 27th

January 27, 2025: $593.0 billion—the market value evaporated in a single day.

The Nasdaq Index fell by more than 3% within three hours of opening, and the total market value of AI-related stocks evaporated by over $2 trillion throughout the day (The Guardian, 2025.1.28). NVIDIA recorded the largest single-day drop for a single stock in the history of the U.S. stock market. The cause of this market panic did not originate from Wall Street, but from an open-source model report released by a Chinese technical team. In a subsequent public speech, Donald Trump referred to the release of DeepSeek as a "wake-up call." Wall Street dubbed this day the "$593.0 Billion Alarm Clock."

Silicon Valley tech executives and Pentagon policymakers quickly compared this day to the Soviet Union's launch of the Sputnik satellite in 1957. The similarity between the two events lies in the fact that both shattered the illusion of absolute American leadership in core technological fields and triggered violent tremors in capital markets and national security systems. However, their difference lies in the direction of power flow. The 1957 Sputnik crisis gave birth to NASA and DARPA; the U.S. federal government achieved a convergence of scientific will by centrally allocating national resources. In contrast, the 2025 DeepSeek shock precisely revealed the fragile consensus within the American AI ecosystem.

Panic did not bring a cohesion of strength; instead, it exposed deep-seated rifts.

Over the past three years, the valuation system of the U.S. AI industry has been built upon an unshakable assumption: the improvement of model capabilities depends absolutely on the exponential growth of computing power scale. This principle, known as the Scaling Law, is not only the core of the technical roadmap but also the foundation of power distribution. Under this framework, NVIDIA holds the "money-printing power" of computing power; OpenAI established a valuation anchor through its first-mover advantage; and cloud service giants like Microsoft and Amazon collect "land rent" by monopolizing infrastructure. The four stakeholders formed a perfect closed loop: the more expensive the computing power, the higher the barrier to entry for frontier models; the higher the barrier, the deeper OpenAI's moat; the deeper the moat, the more willing capital is to pay for NVIDIA's hardware premium.

The open-sourcing of DeepSeek-R1 physically severed this cycle. When a model with parameter scale and reasoning capabilities at the frontier level was proven to be trainable with extremely low computing power consumption, Wall Street's faith that "computing power equals everything" collapsed instantly. The market sell-off of NVIDIA stock was not due to issues with chip quality, but because the business logic supporting its $130.5 billion fiscal year revenue exposed a fatal vulnerability.

If high-performance AI no longer requires massive amounts of computing power, the planned ultra-large-scale data centers costing hundreds of billions of dollars will become sunk costs. The alliance of interests between OpenAI, Anthropic, xAI, and cloud service providers developed irreparable cracks amidst mutual suspicion between computing power buyers and sellers. Buyers began to doubt whether they were paying an excessive monopoly tax for computing power, while sellers worried about the visibility of future orders. Computing power is no longer the sole symbol of power.

The $6 Million Controversy

The DeepSeek-R1 technical report (2025.1) disclosed a final training run cost of approximately $6 million. This figure triggered a weeks-long "accounting" movement on social networks and in congressional hearings.

$6 million is a pure accounting result. It excludes the costs of pre-training data cleaning, the months of trial-and-error R&D investment, and the depreciation of the 10,000-GPU cluster accumulated over years by the parent company, High-Flyer Quant. The true cost is far higher than $6 million. Silicon Valley's defense mechanism attempted to maintain existing business logic by proving that "the Chinese team also spent money." Venture capitalists and tech bloggers analyzed DeepSeek's computing power scheduling logs frame by frame, attempting to find hidden bills.

Arguing over specific numbers is evading the core threat. Even if this figure is multiplied by 10, a $60 million training cost is still two orders of magnitude lower than the investment in OpenAI's equivalent models.

This difference in magnitude destroyed the expectation of scarcity for large models. OpenAI's private market valuation of up to $150 billion was built on the assumption that "only a very few giants can afford the entry fee." According to Sam Altman's narrative, the road to Artificial General Intelligence (AGI) requires tens of billions of dollars, national-level power grid support, and an absolute monopoly on global top-tier chip production capacity. This narrative turned AI R&D from software engineering into asset-heavy military-industrial manufacturing.

DeepSeek pulled frontier AI back from military-industrial manufacturing to software engineering. Whether the efficiency route can systematically lower the barrier to frontier AI is the truly unsettling question behind the $6 million controversy. If a ticket to the current technological frontier drops from $1 billion to tens of millions of dollars, the moat of large models will cease to exist.

The capital market's reaction materialized within 72 hours. When the replication cost of a technology falls off a cliff, its commercial premium evaporates simultaneously. Wall Street realized that if startups could use open-source architectures and limited computing power to reach GPT-4 levels in specific vertical domains, the pricing power of SaaS wrappers and foundation model providers would be dismantled. The destructive power of the $6 million figure lies not in whether it accurately reflects every kilowatt-hour consumed, but in its planting of a psychological anchor for global developers: spending money does not equal leadership.

The Quant Trader's AI

High-Flyer Quant was founded in 2015 and is a leading Chinese quantitative hedge fund. The core capability of this company is to extract trading signals from market noise using minimal computational resources.

The engineering culture of quantitative trading was fully projected into DeepSeek's AI R&D. High-Flyer Quant engineers adhere to one principle: wasting computing power is losing money. OpenAI engineers are accustomed to another principle: if computing power is insufficient, add more money.

The MoE (Mixture of Experts) architecture, FP8 mixed-precision training, and multi-token prediction technology adopted by DeepSeek-R1 all point their underlying logic toward cost control. The MoE architecture uses a sparse activation mechanism to call only a small portion of the model's parameters at a time, thereby reducing the computational consumption of forward and backward propagation. FP8 technology uses lower numerical precision for mathematical operations, compressing memory usage and video memory bandwidth pressure to half of traditional solutions without losing model performance. Multi-token prediction allows the model to output multiple subsequent tokens at once, increasing throughput efficiency during the inference stage by 3 to 5 times.

These engineering choices stripped away the mysticism of AI R&D. In DeepSeek's laboratory, training large models is not about creating digital life, but about solving a pure operations research optimization problem. Every FLOP (floating-point operation) has a clear cost tag. This extreme squeezing of computational efficiency is not the preferred path for traditional AI researchers, but the instinctive reaction of high-frequency traders in a resource-constrained environment.

Iterations from R1 to subsequent versions prove that this efficiency route possesses systematic replicability. DeepSeek-V3.1, released in 2025, supports both "thinking" and "non-thinking" modes, adding 840 billion tokens of training volume on top of V3 to further optimize resource allocation during logical reasoning (DeepSeek Tech Blog, 2025). DeepSeek-V3.2-Speciale, launched in early 2026, introduced DSA (Domain-Specific Architecture) technology. DSA skips the redundant consumption of general-purpose computing by customizing hardware instructions for specific algorithms. This version achieved 2 to 4 times acceleration in long-sequence processing, with computational resource savings reaching 80% (therundown.ai, 2026).

Every version update compresses the model's dependence on general-purpose hardware. The efficiency route is no longer an accidental technological breakthrough but has become an alternative engineering paradigm.

Conduction of the Shockwave

The greatest consequence of the DeepSeek shock is not the drop in NVIDIA's stock price, but the first crack appearing in the belief system that "spending money equals leadership." The second-order effects triggered by this crack are reshaping the power structure of the global AI arms race.

The first conduction chain occurred within Silicon Valley. As Scaling Law absolutism wavers, the urgency for cloud service providers to build ultra-large-scale data centers has cooled accordingly. If the improvement of model capabilities no longer depends absolutely on exponentially growing GPU stacks, the return on investment for major companies pushing self-developed AI chips will face reassessment. Amazon's Trainium and Google's TPU projects were originally intended to find alternatives in an era of computing power inflation; now they must prove their economic rationality under a low-computing-power-consumption paradigm.

The second conduction chain directly pierced Washington's policy defenses. The U.S. Department of Commerce (BIS) chip export controls are built on the assumption that "computing power blockade equals technological blockade." The original intent of the ban was to lock the Chinese AI industry into a technological level more than two generations behind by cutting off the supply of high-end GPUs like the H100. However, the actual logic of power moved in the opposite direction. The hardware ban cut off the possibility for Chinese teams to follow the "brute force aesthetics" route, instead forcing out a set of extreme algorithm optimization capabilities.

The ban gave birth to stronger opponents. This policy contradiction forced Washington to reassess the effectiveness of sanctions. In January 2026, the policy orientation of the U.S. Department of Commerce began to shift from "comprehensive prohibition" to "case-by-case approval." Policymakers realized that simple physical blockades cannot contain a mathematical logic that has already been open-sourced.

The release of GPT-5 in August 2025 attempted to re-establish the dominance of computing power hegemony. By demonstrating new capabilities emerging under extreme computing power density, OpenAI proved to the market that the upper limit of brute force aesthetics remains unfathomable. Even if GPT-5 widens the gap in absolute performance again, the crack torn open by the efficiency route will not heal.

The AI arms race has officially entered a dual-track parallel pattern—one track continuing to test the limits of physics and capital, and the other squeezing the ultimate performance out of algorithms under limited resources. Buyers are using DeepSeek's open-source achievements as bargaining chips against NVIDIA's pricing power and OpenAI's technological hegemony.

If Microsoft and OpenAI build the $100 billion Stargate ultra-large-scale data center by 2027 and rely entirely on that cluster to maintain a technological generation gap, the judgment regarding "the efficiency route subverting computing power hegemony" would be wrong. The actual flow of capital will provide the final verdict. The only fact that can be confirmed at present is that after that trading day on January 27, the power game of building computing power barriers with capital has lost its absolute legitimacy.

Computing Power is National Power

A $500 Billion Check

On January 21, 2025, Trump announced the Stargate project at the White House, with a total budget of $500 billion. In the same month, Elon Musk began earthworks for a new data center on a vacant lot in Memphis. A year later, in early 2026, tracking by The Information showed that the list of assets resulting from the $500 billion commitment was extremely short: no employees, no data centers. The only physical outcome of the $500 billion so far was a single White House press conference.

What OpenAI, SoftBank, and Oracle signed was merely a letter of intent. Hidden behind the agreement were undisclosed funding ratios, timelines, and governance structures. Execution determines the landscape of computing power. When three tech giants, each with tens of thousands of employees, attempt to reach a consensus on power acquisition, chip procurement, and cooling system design, the time consumed by cross-departmental meetings far exceeds the physical cycle of pouring concrete. The multi-party coordinated infrastructure model demonstrates a natural tendency toward paralysis in the AI arms race.

In the physical world, assembling silicon wafers, electricity, and cooling water into usable computing resources requires absolute control over the supply chain. Any infrastructure project requiring votes from multiple boards of directors is destined for delay from its inception.

Two Hundred and Fourteen Days

There are no cross-departmental meetings in the empty factory buildings of Memphis. xAI’s Colossus cluster went from clearing an abandoned factory to lighting up 100,000 GPUs in just 122 days. The engineering team then doubled the scale within another 92 days. As of early 2026, tracking by Interesting Engineering shows that the facility has deployed 555,000 GPUs. A design plan with a target power of 2GW makes this the world’s first gigawatt-scale AI training cluster.

Speed overrides scale. Boeing relies on a network of hundreds of subcontractors and Washington lobbying groups to advance aerospace projects, while SpaceX compresses engineering iteration cycles to the limit through vertical integration and founder-centric centralization. Today, the standoff between Stargate and xAI almost perfectly replicates these differences. In the face of extreme engineering challenges, an organization dominated by a single strong founder can bypass traditional coordination mechanisms and convert all resources into physical-world momentum.

Consortiums are often slowed down by competing interests when handling complex infrastructure. While Stargate’s participants were still engaged in multilateral negotiations over tax incentives for site selection, xAI’s cooling water pipes had already completed their third flush. Decision-making concentration has replaced capital volume as the key variable on the computing power production line.

Six Nuclear Power Plants

In January 2026, Meta signed nuclear power purchase agreements totaling 6.6GW with Vistra, TerraPower, and Oklo, aiming to complete deployment by 2035. 6.6GW is equivalent to the installed capacity of six standard nuclear power plants. From shouting about connecting the world to personally procuring nuclear power plants, it took Zuckerberg less than ten years.

Microsoft signed a contract with Brookfield to purchase 10.5GW of renewable energy between 2026 and 2030. Combined with xAI’s 2GW demand in Memphis, tech companies are reshaping the global energy consumption landscape. Energy power is shifting. Tech giants have transformed from mere electricity consumers into the primary drivers of power infrastructure.

The underlying logic of power grid planning is changing accordingly. For decades, grid expansion was designed to meet urban sprawl and manufacturing needs; now, the direction of transmission lines depends entirely on the location of new data centers. The nuclear power approval process has been forced to accelerate under tens of billions of dollars in lobbying from tech giants. National-level energy policies are being firmly hijacked by the AI industry, and environmental issues are giving way to rapid infrastructure expansion in the face of gigawatt-scale training demands.

Cracks in a Four-Trillion-Dollar Empire

NVIDIA’s financial data demonstrates overwhelming dominance. Revenue for fiscal year 2025 reached $130.5 billion, a year-on-year increase of 114%, with a record single-quarter revenue of $57 billion. Supporting its $4 trillion market capitalization is an order backlog of up to 3.6 million B200 and GB200 units, extending into mid-2026.

Yet, cracks are appearing deep within the server racks of cloud providers. In January 2026, The Verge disclosed that Microsoft’s Maia 200 possesses over 100 billion transistors, with FP4 performance reaching three times that of Amazon’s Trainium 3. Combined with Google’s TPU v7, the self-developed hardware of the three major cloud providers has formed a substantial encirclement of NVIDIA.

The goal of self-developed chips has never been to surpass competitors in absolute performance, but rather to reach a "good enough" level. NVIDIA’s monopoly profits are built on the fragile premise that customers lack alternatives. Once Maia 200 or TPU v7 demonstrates acceptable cost-performance in specific inference tasks or internal training, cloud providers will rapidly divert tens of billions of dollars in procurement funds. Monopolists are often not defeated by perfect products with comprehensive leads, but are gradually weakened in pricing power by countless "good enough" cheap alternatives.

NVIDIA released the Rubin architecture at CES 2026. Based on TSMC’s 3nm process, Rubin integrates 336 billion transistors and provides 50 PFLOPS of FP4 computing power—five times the performance of Blackwell. Such a brutal increase in computing power is, in fact, a defensive war against the "good enough" logic. As long as NVIDIA can keep the training threshold for the most cutting-edge models consistently beyond the capability limits of self-developed chips, the empire can endure. The ultimate battle on the computing power field has become a life-and-death race between NVIDIA’s iteration speed and the pace of self-developed chip catch-up.

The American Rift

On February 27, 2026, two documents took effect simultaneously in Washington and Silicon Valley. Crunchbase transaction records showed that OpenAI completed a $110 billion funding round at a valuation of $840 billion, with funds from Amazon, NVIDIA, and SoftBank fully accounted for. On the same day, Trump signed an executive order removing Anthropic from the federal government's list of suppliers. One company that adhered to safety principles lost its military contracts, while another that abandoned its bottom line took them over across multiple fronts. The divisions within the American AI ecosystem were fully exposed on this day.

Who is America

The judgment that "America leads in AI" lacks a clear subject. The reality of early 2026 presents a highly fragmented prosperity; piecing together the financial reports and industry data of major companies does not form a complete strategic blueprint.

OpenAI's valuation reached $840 billion, but its underlying Stargate supercomputer project stalled due to power permit approvals. NVIDIA's fiscal year 2025 revenue climbed to $130.5 billion, while simultaneously facing the triple pressures of cloud providers' self-developed chips, open-source architectures, and antitrust investigations. xAI assembled a supercluster of 555,000 GPUs in Memphis, yet its model capabilities are still chasing the exhaust of the former. Anthropic stood firm on alignment research and was subsequently blacklisted by the federal government's procurement system.

The subject does not exist. What exists is a group of competing entities—which happen to be registered in the United States.

When Washington policymakers attempted to integrate these dispersed commercial forces into a strategic weapon against a whole-of-nation system, they found themselves facing not a disciplined regular army, but a loose alliance composed of monopolistic hardware vendors, capital-driven model factories, and madmen operating outside of regulation. Fragmented competition catalyzed the explosion of the Transformer architecture and underlying compute power over the past decade, maintaining the diversity of technical paths. However, as adversaries began to concentrate electricity, land, and capital with national will for a saturated pursuit, the costs of internal friction from fighting individual battles began to amplify exponentially.

From Benefiting All Humanity to Eight Hundred Billion

The trajectory of OpenAI taking over Department of Defense contracts is often simplified by outsiders as a moral betrayal regarding greed. This obscures the deeper laws of organizational evolution within the gravitational fields of capital and compute.

In May 2025, OpenAI announced its transition to a Public Benefit Corporation (PBC). On August 7 of the same year, GPT-5 was released and unified the underlying architecture, with the company announcing on its official website the deprecation of transitional products such as GPT-4o, o3, o3-pro, and o4-mini. By the time the restructuring was completed in October, the original non-profit department had been renamed the OpenAI Foundation and held approximately $130 billion in shares.

This aligns with commercial logic. The evolutionary trajectory appears cold and mechanical: a non-profit structure could not support the compute consumption required for pre-training, thus necessitating the establishment of a capped-profit entity to bring in external investment. When model parameter scales jumped again, the shell of a Public Benefit Corporation became the best way to secure $110 billion in new financing while circumventing traditional non-profit regulation. When a company has an $840 billion valuation and a massive list of data centers to be built, taking over military contracts in exchange for political protection and infrastructure green lights becomes a mandatory item on the balance sheet.

The mission statement completed its transformation under the gravity of capital, moving from ensuring AI benefits all of humanity to ensuring AI benefits all of humanity and supports its $840 billion valuation. No single decision was made out of pure malice, yet the final form presented is worlds apart from the original intention of 2015.

The Third Pole

In Memphis, Tennessee, the construction speed of xAI's Colossus cluster broke the physical common sense of Silicon Valley. As of early 2026, it was already fully loaded with 555,000 GPUs, with an aggressive timetable set to expand to 2 to 3 million GPUs by 2027.

The role of xAI in the American AI ecosystem is structurally equivalent to that of SpaceX in American aerospace. Both were created by the same person, and both bypassed the lengthy coordination mechanisms between traditional giants through extreme vertical integration and the founder's will. SpaceX bypassed NASA's layers of outsourcing; xAI bypassed the alliance between Microsoft and OpenAI.

The key difference lies in the measurement standards for physical output versus intellectual output. SpaceX's Falcon rockets indeed have launch costs an order of magnitude lower than traditional spacecraft and possess extremely high recovery reliability; thrust and orbital data are completely transparent physical indicators. xAI possesses the world's largest reserve of compute hardware, yet its Grok model still lags behind GPT-5 and Claude in benchmarks.

Compute is the piano. The model is the performance.

A person can buy and assemble the world's most massive array of pianos in 214 days; this proves engineering execution, but it does not prove that owning the most pianos allows one to compose the best symphony. If xAI builds its supercluster of 3 million GPUs as scheduled by 2027 and its Grok model still lags behind its competitors in multimodal reasoning benchmarks, then the hypothesis that compute scale is a linear extension of model intelligence will be completely overturned.

The One Kicked Out

The executive order on February 27, 2026, terminated Anthropic's presence in federal government systems. There was only one trigger: in a $200 million joint AI contract awarded by the Chief Digital and Artificial Intelligence Officer (CDAO) in mid-2025, Anthropic explicitly refused to lift the safety hard-coded restrictions on its models regarding mass surveillance and fully autonomous lethal weapon systems.

Palantir, OpenAI, and Ask Sage received a combined $1.51 billion in defense AI contracts in 2025. When the Pentagon stripped this $200 million share from the supplier that refused to cooperate and seamlessly transferred it to competitors willing to cooperate with the state machinery, venture capitalists and entrepreneurs across Silicon Valley clearly received the underlying rule regarding resource allocation.

The punishment was precise to the dollar. Anthropic made a correct decision consistent with its founding principles, and the market immediately responded with a $200 million financial counterattack.

The second-order effects of this blacklisting far exceed the revenue loss of a single company. It sent a clear price signal to the entire AI industry: adhering to safety principles carries a clearly marked commercial cost. Facing the exorbitant training bills for next-generation models, compromise will become the only path for future AI startups to obtain survival funds when faced with the unrestricted usage demands of the military or intelligence agencies. The endgame of adverse selection is clearly visible: the researchers who care most about system safety will, over time, be excluded from the national-level weapons and surveillance networks that need safety design the most.

China's Reversal

In September 2025, HuggingFace's server logs recorded a quiet turning point. According to data tracking from the-decoder.com, the total monthly downloads of Alibaba's open-source Qwen series surpassed Meta's proud Llama family for the first time that month. The 2025 annual report subsequently released by Stanford HAI marked this milestone with precise statistical data: China's global share of open-weight AI reached 17.1%, officially surpassing the United States' 15.8%.

The pursuer has completed the overtake—at least in the statistical dimension of open-source code repositories.

Industry analysis reports provided another set of striking data: China's domestic AI user base reached 515 million in the same year, and the relevant market size is expected to expand from 29.4 billion RMB in 2024 to over 70 billion RMB by 2026. On the surface, the industry's development momentum is strong.

However, once the 70 billion RMB market size is deconstructed, problems emerge: the limitation of the net value per customer in the To B business being less than 100,000 RMB is clearly visible. The average AI expenditure of American companies as a percentage of revenue is typically five to eight times that of Chinese companies. Silicon Valley's SaaS subscription model faces challenges on the west coast of the Pacific from deeply rooted habits of free usage and a mindset of customized outsourcing. When the expected 70 billion RMB in revenue is distributed among hundreds of domestic model companies still investing heavily in computing power, not a single one can cover the depreciation and electricity costs of the next-generation 10,000-card cluster solely through model API call fees.

Code itself cannot bring profit.

The Paradox of Download Volume

Qwen has spawned over 180,000 model variants on HuggingFace.

Behind the appearance of ecological prosperity lies a dry engineering reality. Among the 180,000 variants, the README files of approximately 179,000 likely contain only one sentence: "Fine-tuned based on Qwen, used for industry-specific customer service Q&A bots." Download volume reflects developer interest but is unrelated to customer budgets.

The operating system competition of twenty years ago provided a highly similar reference: Linux, with its open-source and free characteristics, completely occupied global data center servers and developer communities, while the closed Windows system captured over 90% of commercial profits in the PC desktop market during the same period. The core of open source lies in distribution channels and is unrelated to profit models.

When countless small and medium-sized development teams download open weights to local servers and attempt to transform them into plugins for automatically replying to e-commerce return and exchange policies, they are actually testing the boundaries of different industry scenarios for cloud providers for free. This type of ecological prosperity eventually translates into a commercial closed loop of underlying cloud service consumption, yet excludes pure model startups that lack IaaS infrastructure to sell from the profit distribution. Alibaba uses the open-sourcing of Qwen to sell more Alibaba Cloud server instances; pure AI startups, facing soaring download volumes, still struggle to pay for next month's R&D salaries.

Open source is a sieve.

Paradigms Born of Bans

When Washington policymakers were drafting export control provisions for advanced AI chips, they assumed a linear development path where computing power equals power, but they did not anticipate that physical restrictions on hardware acquisition would directly prompt Chinese engineers to abandon the traditional method of relying on brute-force computation to expand parameter scales, instead mining for extreme efficiency in memory bandwidth within the underlying logic of MoE (Mixture of Experts) architectures and domain-specific architectures. Sanctions unexpectedly pushed for a reshaping of technical routes.

DeepSeek's technical evolution path validates the innovation that occurs under extreme constraints. According to tracking analysis by therundown.ai at the end of 2025, the DeepSeek-V3.2-Speciale version achieved up to 80% savings in computing resources when processing ultra-long context sequences through deeply customized DSA (Domain-Specific Architecture) technology. These FP8 low-precision training and Mixture-of-Experts routing algorithms, originally developed to cope with the embargo on high-end NVIDIA GPUs, have been widely adopted and reverse-engineered by global developers, including Silicon Valley startups, due to their memory cost-performance advantages.

By restricting hardware access, the U.S. Department of Commerce unintentionally subsidized the most aggressive software computational efficiency optimization project in human history. The east coast of the Pacific attempted to block computing power, while the west coast of the Pacific redefined the computing power consumption equation and open-sourced this new efficiency paradigm to the world.

The Spring Festival Explosion

The 2026 Lunar New Year left no breathing room for technological development.

Alibaba's Qwen 3.5 refreshed context records on New Year's Eve; Moonshot AI's Kimi K2.5 and Zhipu AI's GLM-5 followed closely, releasing multimodal updates on the first day of the Lunar New Year; MiniMax M2.2 and ByteDance's Doubao 2.0 pushed API call rates down to the floor of physical costs on the first working day after the holiday. According to digitalapplied.com, five foundation models were launched intensively within 168 hours, bringing the gentle narrative of technical differentiation to a close.

Price cuts mean despair.

This dumping action, led by big tech companies with infinite cloud computing transfusion capabilities, pierced the break-even line of independent model companies within just three months, reenacting the brutal competition of China's internet O2O battlefield ten years ago, where capital was used to buy monopoly positions. Data from the MacroPolo AI talent tracker shows that approximately 70% of Chinese authors at top AI conferences still choose to work in American institutions. Top algorithm talent remains in North America to push the upper limits of AGI, while domestic teams rely on engineering optimization to fight price wars; this dual structure is accelerating the domestic market shake-up.

If, by the end of 2027, there are still more than three independent foundation model companies in the Chinese AI market with valuations exceeding $5 billion that are not controlled by cloud vendors, the speculation that price wars lead to oligopolies will be overturned. Reality's pressure is forcing all participants to accept a grim conclusion: the experimental period of "a hundred flowers blooming" has ended, and the fierce clearing of the field by computing power and capital has officially begun.

The Absurdist Drama of the Ban

In October 2022, the U.S. Department of Commerce introduced the first round of AI chip export restrictions targeting China. The following year, NVIDIA launched the H800 chip, which complied with the regulations. However, by 2024, this chip was also added to the embargo list. In 2025, chips worth $160 million flowed into mainland China through smuggling networks in Southeast Asia. By 2026, the Department of Commerce shifted to a case-by-case approval policy. Washington's chip controls have undergone five stages: embargo, bypass, re-ban, smuggling, and case-by-case approval. The next stage may be abandonment.

Defining Effectiveness

Policymakers need to define the criteria for victory. If the goal is to prevent China from obtaining the most advanced hardware, the Bureau of Industry and Security (BIS) export controls have indeed kept H100, H200, and Blackwell chips entirely outside of legal customs (BIS Export Control Rules, 2026). The zero-declaration figures in customs data seem to form a perfect physical barrier. Hawks in the Pentagon can point to these charts of zeros and claim a decisive victory for the blockade strategy.

If the goal is to maintain an absolute U.S. lead in AI model capabilities, the distribution of download volumes in the open-source community shows a starkly different trend. The weight of activity on Hugging Face is undergoing an irreversible tilt.

If the goal is to prevent China from training frontier models, DeepSeek-R1 achieved reasoning levels on par with the strongest closed-source models of the time using only a compute cluster of weakened, previous-generation chips.

The ban has succeeded in its narrowest definition, but its core objectives have completely failed.

Washington's decision-makers have long been immersed in the "compute myth" brought about by Scaling Laws. They assumed an absolute linear power formula: more transistors mean more computing power, and more computing power equals smarter models. Under this logic, controlling the flow of the most advanced GPUs is equivalent to cutting the only rope a competitor has to climb the ladder of intelligence. However, when DeepSeek engineers compensated for the compute disadvantage caused by the hardware generational gap through algorithmic innovations like Mixture of Experts (MoE) and extreme memory optimization techniques, Washington encountered a cognitive crisis that could not be solved by adding more names to the Entity List.

The efficiency route has broken the linear formula that equates hardware specifications with AI capability. Compute hegemony is no longer the sole passport to model hegemony. This has changed the fundamental logic of the global compute arms race, pulling the competition back from pure capital-intensive hardware stacking to the intellectually intensive track of algorithmic innovation and engineering optimization. Under the old power framework, a Silicon Valley giant with a $600 million procurement budget could easily overwhelm a startup team with only $6 million. Now, the underlying algorithmic leverage allows those with scarce resources to find a fulcrum to move the wall of computing power. Beneath the surface of technical catch-up, the efficiency route has delivered a structural blow to the core business model of compute hegemony.

In the Silicon Valley narrative, the road to Artificial General Intelligence (AGI) is a golden highway paved with tens of thousands of H100 chips. Giants like OpenAI and xAI build compute clusters in units of 100,000 GPUs. Behind this capital violence lies a logic of resource monopoly: whoever can burn the most money can monopolize future intelligent output. Washington's ban is based precisely on this logic. They believed that by cutting off the supply of high-end GPUs, China's AI industry would naturally wither due to a lack of computing power. DeepSeek's breakthrough constitutes a fatal blow on both political and commercial levels. Politically, it reveals the loopholes in the containment strategy; commercially, it punctures the myth of high-compute barriers carefully maintained by Silicon Valley giants. When extremely low compute costs can produce reasoning results similar to those costing hundreds of millions of dollars, the capital market's high expectations for NVIDIA's future performance will also be impacted.

The 160-Million-Dollar Underground Channel

In December 2025, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) concluded "Operation Gatekeeper" (BIS Enforcement Bulletin, 2025.12). The enforcement bulletin recorded the entire process of dismantling an AI chip smuggling network worth $160 million. This amount is roughly equivalent to one day of NVIDIA's revenue in fiscal year 2025.

The form of smuggling networks has changed.

It is no longer the primitive mode of individuals carrying graphics cards in suitcases; it has evolved into a highly specialized, transnational transshipment industry spanning multiple jurisdictions. After high-specification chips are legally exported to shell companies in Southeast Asia or the Middle East, they are repackaged or even declared as components to enter mainland China through secondary or tertiary transshipments. In the details disclosed by Operation Gatekeeper, the financial settlements of the smuggling network had already bypassed the traditional SWIFT system. More than 60% of transactions were completed through cryptocurrency or trade hedging in offshore financial centers. This made joint tracking by the U.S. Treasury and Commerce Departments extremely difficult.

A typical transshipment link has more than five layers of disguise: a U.S. company registered in Delaware legally sells chips to a cloud computing service provider in Dubai; after arriving at Jebel Ali Port, these chips do not enter a data center but are instead transported to another warehouse in a free trade zone; there, they are relabeled, disguised as ordinary server motherboards, placed on cargo flights to Southeast Asia, and finally flow into Shenzhen through the gray areas of border trade. A complete underground ecosystem has thus formed. In this system, there are legal teams specialized in registering shell companies, documentation experts responsible for forging end-user certificates, and logistics operators proficient in exploiting customs inspection loopholes in various countries.

In 2025, BIS added 65 Chinese entities to the Entity List (BIS Entity List Update, 2025) in an attempt to plug these emerging loopholes. However, the update speed of enforcement agencies can never catch up with the adaptation speed of smuggling networks.

This situation is structurally identical to the U.S. Prohibition era from 1920 to 1933. Banning demand for a commodity with a compound annual growth rate exceeding 40% inevitably gives rise to an industrialized smuggling industry, which in turn drives up enforcement costs. Prohibition failed to eliminate alcohol consumption; instead, it fostered more sophisticated and violent organized crime networks. When legal distribution channels are forcibly cut off, the market does not shrink as regulators hope, but instead goes underground. Transshipment nodes have gradually expanded from the initial Hong Kong and Singapore to Penang in Malaysia, Dubai in the UAE, and even as far as free trade zones in South America that are not constrained by U.S. long-arm jurisdiction. Every geographical extension is accompanied by a sharp rise in logistics costs and a lengthening of the intermediary profit chain.

While transporting chips, this underground network has also spawned supporting money laundering services, guides for evading false end-user audits, and counter-reconnaissance consulting services specifically targeting BIS compliance audits. Legal supplier NVIDIA was forced to cede its Chinese market share, which at its historical peak accounted for 25% of total revenue, while the underground networks responsible for transnational transshipment seized staggering premium profits. The only difference is that the U.S. government during the Prohibition era took 13 years to admit policy failure and shift to issuing liquor licenses, whereas, faced with AI chips, the Department of Commerce took only 4 years to retreat to a case-by-case approval track similar to a licensing system. The policy compromise was not out of benevolence, but because the administrative cost of maintaining a total blockade had far exceeded the strategic benefits it could bring.

The Trap of Case-by-Case Approval

In 2025, the Trump administration first completely suspended AI chip exports to China, then reversed the policy, allowing NVIDIA to resume exports of H20 processors (Multiple media reports, 2025). By January 2026, BIS officially introduced new case-by-case approval regulations, allowing the export of high-level chips such as the H200 to specific approved Chinese customers, while adding a rigid restriction: the quantity of any chip model exported to China must not exceed 50% of that model's domestic shipments in the United States (BIS Policy Announcement, 2026.1).

On the surface, the policy seems to be loosening.

This is an illusion.

The transition from a blanket ban to conditional approval has released a chain effect more destructive than a total embargo. Corporate customers are caught in a passive waiting period. They cannot predict whether they meet the approval criteria, cannot determine if the process will take three months or three years, and certainly cannot guarantee that current approval standards won't be instantly overturned in the next geopolitical friction. Washington regulators have turned what was once a clear red line into a flexible gray zone. In modern business decision-making, clear bad news is always better than ambiguous good news.

Uncertainty destroys supply chain planning. When a tech giant plans to invest billions of dollars over the next three years to build hyperscale data centers, it needs a stable and predictable hardware delivery schedule. If the foundation of the data center has been laid, but the promised 50,000 GPUs are stuck on the approval desk of some bureaucrat in Washington, the entire project's capital chain faces the risk of breaking. No board of directors will approve an infrastructure investment built on the probability of foreign administrative approval.

Policy vacillation has forced Chinese customers to accelerate their search for local alternatives. The fundamental reason for abandoning NVIDIA lies in the fragility of the supply chain. In the iron laws of the business world, a domestic alternative that offers only 70% performance but is guaranteed to be delivered on time has far greater commercial value than a "castle in the air" with 100% performance that could have its supply cut off at any moment. Before the 2022 ban, the attitude of Chinese tech companies toward domestic AI chips was generally cold. NVIDIA's CUDA ecosystem was too perfect, and migration costs were extremely high. No rational commercial company was willing to proactively abandon a mature toolchain to pay for a bug-ridden startup ecosystem.

Washington's case-by-case approval and erratic policy swings have completed the most difficult market education for domestic chips. The pressure of survival has replaced simple commercial cost calculations. When "unable to buy" and "not knowing if we can buy tomorrow" became the Sword of Damocles hanging overhead, the strategic priorities of Chinese cloud providers underwent a fundamental reversal. Supply chain security and self-controllability completely overrode the pursuit of ultimate single-card performance. They began to deploy hundreds or thousands of engineers to rewrite underlying operators and adapt domestic hardware that was originally considered inadequate. Domestic chip ecosystems like Huawei's Ascend gained the most precious opportunities for trial and error and real-world refinement in this extremely insecure external environment.

If NVIDIA loses all new orders from the top five cloud providers in the Chinese market by 2027, the sole culprit will be the unpredictability created by Washington. The case-by-case approval system essentially hands over the right to fulfill commercial contracts to political cycles. It forces buyers to establish two completely independent supply chains to hedge against the risk of supply disruption. When the maturity of the backup supply chain crosses a certain threshold, the original dominant supplier will be completely marginalized.

The Collapse of Thresholds

Even the compromise route of settling for second-best has been completely blocked. AMD specially customized a downgraded AI chip for the Chinese market, attempting to find survival space on the edge of compliance. The Department of Commerce directly blocked these products, determining that they still exceeded the performance thresholds for control (Industry reports, 2025).

The sense of absurdity reaches its peak at this moment.

The logic of setting performance thresholds is built on a classical industrial-era assumption: by controlling the most advanced production tools, one can lock down the competitor's final output. In the field of chip manufacturing, this logic still holds; without Extreme Ultraviolet (EUV) lithography machines, it is indeed impossible to achieve physical breakthroughs in advanced processes (See Chapter 3 of "Chip War" for details). The physical barriers that objectively exist and cannot be fabricated out of thin air through software code constitute an insurmountable chasm.

The evolution of artificial intelligence, however, follows completely different physical laws. Chips are merely containers providing basic computing power, while the emergence of model capability is the result of a complex chemical reaction between algorithmic architecture, data quality, and compute scheduling. While Washington pours all its regulatory resources into limiting the size of the container, it watches helplessly as its competitor brews intelligence of the same intensity in a smaller container by improving the quality of data purification and restructuring the logic of algorithmic scheduling.

Washington bureaucrats are precisely calculating the upper limits of interconnect bandwidth and memory capacity in Excel spreadsheets, trying to draw a safe red line. When engineers on the other side of the Pacific use a pile of old chips deemed "safe" to stack up the reasoning capabilities of frontier models through distributed computing optimization, mixed-precision training, and squeezing the limits of old architectures, this red line becomes a costly joke. Policymakers are trying to regulate the physical specifications of hardware, while the real threat is the organizational ability of software to schedule those specifications.

The state machine is accustomed to measuring power with tangible physical parameters, while power in the AI era is migrating toward intangible algorithmic efficiency and open-source ecosystems. A mismatch has occurred in the way power operates. The global trade system is paying the price for this mismatched regulation. Tariff barriers directly drive up the construction and electricity operating costs of data centers. A WTO report confirmed that the 2026 global trade growth forecast has been forced down from 1.8% to 0.5% (WTO, 2026). To block a few hundred GB of interconnect bandwidth, the operational efficiency of the entire globalized supply chain has been forcibly downgraded.

The fragmented distribution of compute resources is cutting a technological revolution that should have been driven by global supply chain synergy into countless high-cost, low-efficiency islands. The lowered global trade growth forecast in the WTO report is merely the surface bill for this absurdist drama. The deeper cost lies in the fact that the compute arms race is forcing global data center infrastructure toward extreme redundancy and fragmentation. To evade export controls that could descend at any time, multinational corporations are forced to build redundant compute clusters across different jurisdictions. Europe needs an independent AI infrastructure to meet the needs of Sovereign AI; Middle Eastern sovereign wealth funds are frantically hoarding GPUs in the desert to trade for geopolitical leverage; and China is using its national strength to build an underlying hardware ecosystem completely decoupled from U.S. technology.

This fragmented struggle for resources has destroyed the logic of global division of labor and economies of scale upon which the IT industry has thrived for the past thirty years. Once, a chip designed in California and manufactured and packaged in Asia would eventually serve human computing needs in a server anywhere in the world. Now, the flow of that chip is labeled with ideology and locked within specific geographical boundaries. Power has tried to use the chains of physical parameters to tether the ghost of algorithmic evolution, only to end up locking out legal suppliers. This four-year experiment with bans has proven how sluggish the traditional regulatory tools of the state machine are when faced with highly non-linear technological mutations. The ban failed to freeze the competitor's evolutionary clock; instead, over a four-year span, it reshaped the power map of global AI computing.

70% Hit Rate

70%.
AI drones funded by Eric Schmidt achieved this combat hit rate on the Ukrainian battlefield (dronexl.co, January 6, 2026).
In contrast, the hit rate of human-piloted FPV (First-Person View) drones is only 30% to 40%.
The gap is clear.
When machines demonstrate twice the efficiency of humans in precision strike missions, the ethical controversy over whether humans should remain "in the loop" quickly transforms into a simple matter of system performance optimization.

13.4 Billion

Power is often hidden in accounting entries.
A brief record in the FY2026 U.S. defense budget documents states: for the first time, the Department of Defense has established a separate budget line for "AI and Autonomous Systems," totaling $13.4 billion (FY2026 Defense Budget Documents).
This figure marks the first time the military AI budget has been large enough to require its own line item. Previously, these expenditures were typically buried within the IT upgrades and logistics maintenance budgets of various military branches, similar to "miscellaneous" items on a corporate expense report.
Now, it has attained a financial status equivalent to that of nuclear submarine programs.

In the specific allocation of these funds, drones and swarm technology received $9.4 billion, maritime autonomous systems received $1.7 billion, underwater and ground autonomous systems received $734 million and $210 million respectively, while the remaining $1.2 billion was used to develop AI middleware to bridge data silos across different military branches (FY2026 Budget Details).
Changes in capital flow are reshaping the power landscape.
In the $839 billion defense spending bill passed by Congress, $9.8 billion was explicitly designated for autonomous unmanned systems (Congressional Record).
The monopoly of traditional defense giants over chassis and engines is being broken, as Silicon Valley algorithm engineers gradually take over the Pentagon's procurement list.
Palantir, OpenAI, and Ask Sage collectively secured $1.51 billion in defense AI contracts in 2025 (Industry Analysis).
Department of Defense procurement is no longer limited to physical platforms made of titanium alloys and composite materials; it is shifting toward purchasing API call permissions and neural network weights on a per-use basis.

The establishment of an independent budget line has disrupted the old bureaucratic balance within the Pentagon.
In the past, if a Navy commander wanted to procure a computer vision model for target identification, they had to package it as a "shipboard logistics inventory management software upgrade" to apply for funding.
This fragmented, bottom-up procurement method resulted in hundreds of mutually incompatible AI silos within the U.S. military.
The $13.4 billion in dedicated funding centralizes power from tactical commanders of various branches to the Office of the Chief Digital and Artificial Intelligence Officer (CDAO).
AI has officially ascended from an auxiliary tool for improving the efficiency of existing weapons to an independent strategic category on par with nuclear strike capabilities and space control.

The surge in budget has also led to panic buying on the defensive side.
The counter-drone market is expected to skyrocket from $600 million in 2025 to $2.7 billion by 2030 (Market Research Report).
The driver of this growth is not a need for stronger physical interception nets, but rather an urgent demand for algorithmic confrontation.
The battlefield is gradually evolving into a contest of computing power between two servers.

The Ukrainian Proving Ground

Aiming no longer relies on the human eye.
When Russian electronic warfare equipment cuts the video transmission signal between front-line operators and drones, traditional FPV drones—which rely on human visual confirmation and manual fine-tuning—lose control and crash during the final few hundred meters of their sprint.
However, AI drones equipped with small edge-side vision recognition models can still autonomously lock onto the weak points of armored vehicles and complete their dive, even in a silent state without external commands.
The 70% hit rate was achieved in precisely such extreme electromagnetic interference environments.

The physiological limitations of human operators have become the system's greatest weakness.
In high-intensity engagement zones, the reaction time for a human brain to process a blurred video stream, judge the value of a target, and push the joystick is approximately two to three seconds.
In contrast, a machine vision model takes only tens of milliseconds to complete the same feature extraction and ballistic calculations.
Military ethics discussions have long been based on an unverified assumption: that humans are more reliable than machines and must therefore hold the final authority to fire.
Combat data from the trenches of Ukraine has overturned this premise.
"Human-in-the-loop" has shifted from a safety guarantee to a performance bottleneck.
If AI drones are forced to transmit images back to the rear for human confirmation before launching a lethal strike, those few seconds of delay are enough for the target to escape or for the drone itself to be destroyed by anti-aircraft fire.

The performance gap forces commanders to make cold trade-offs.
Retaining a human in the loop means accepting a 30% hit rate and a high risk of mission failure; granting machines full autonomous firing authority yields a 70% hit rate and the achievement of tactical objectives.
Battlefield survival pressure makes moral choices moot.
When marketing such systems, technology companies deliberately downplay the term "autonomous killing," opting instead for technical jargon like "pre-programmed target engagement in communications-degraded environments."
The packaging of vocabulary masks the transfer of power.
The power to pull the trigger has shifted from the soldier in uniform to the programmer twenty thousand kilometers away writing target recognition code.

Consumable Weapons

The cost-accounting method of war is changing.
The U.S. Air Force's CCA (Collaborative Combat Aircraft) program expects to deploy 1,000 to 2,000 autonomous wingmen by the mid-2030s (U.S. Air Force Planning Documents).
The cost of producing a single F-35 Lightning II fighter jet by Lockheed Martin exceeds $100 million. Assets costing over $100 million require top-tier air defense suppression systems, sophisticated logistics maintenance teams, and long pilot training cycles.
The loss of any single aircraft triggers inquiries on Capitol Hill.
One F-35 crash is a Congressional hearing; ten autonomous wingmen crashing is merely a line item in a quarterly report.

The target unit price for autonomous wingmen is strictly controlled at around $5 million.
The drop in price changes the mathematical model of aerial combat.
Traditional air force development strategies pursue extreme single-aircraft performance, attempting to eliminate enemies from a safe distance using stealth coatings and beyond-visual-range radar.
The CCA program shifts toward a trade-off between scale and quality.
Twenty autonomous wingmen costing $5 million each can exhaust the interceptor missiles of an enemy's air defense system, tearing open a corridor for manned fighters in the rear.
From their inception, they are defined as "consumables."
They require no expensive life-support systems, no ejection seats, and even the design life of the airframe materials has been reduced from thousands of hours to just dozens of missions.

If the U.S. Air Force fails to deploy at least 500 CCAs by 2035 due to system integration cost overruns, the cost restructuring based on "consumability" will fail, and the military will be forced back onto the old path of relying on a few expensive platforms.
If the plan proceeds as scheduled, its second-order effects will directly impact the underlying logic of international politics.
Nuclear weapons raised the threshold for war by ensuring mutual destruction; autonomous wingman swarms lower the political resistance to war by eliminating personnel casualties.
In the electoral politics of democratic nations, zinc coffins filled with the remains of soldiers are a political price no politician can afford.
When war becomes a pure collision of metal and consumption of silicon chips, and when Congress no longer needs to offer condolences to the families of fallen pilots, the domestic political friction of launching a localized conflict will approach zero.
It is not just weapons that are consumable, but the political concerns that maintain peace.

The Illusion of Human-in-the-Loop

Supervision is actually an illusion.
Requiring a human operator to monitor the decision-making process of an AI weapon system in real-time is essentially no different from requiring a driver to be ready to take over a self-driving car.
Tesla's Autopilot accident data has already provided a perfect cross-domain reference.
When humans are positioned as monitors rather than operators for long periods, attention inevitably wanders.
Once a system performs accurately 99% of the time, humans fall into a fatal state of automation complacency.
When that 1% edge case occurs—a white truck crossing the road or a firing point disguised as a civilian target—the system suddenly fails, and humans are simply unable to switch from a distracted state to full situational awareness in an instant.

The reaction speed of military AI further amplifies this physiological failure.
Machine decision cycles are measured in microseconds.
Faced with incoming hypersonic missiles or a suicide swarm consisting of hundreds of miniature drones, a defense system must complete trajectory prediction, threat prioritization, and interceptor fire allocation the moment the radar captures the target.
The limit for a human brain to process a visual signal and translate it into muscle movement is approximately 250 milliseconds.
Within this time window, an AI system has already completed tens of thousands of friend-or-foe identification calculations.
Nominally, rules of engagement still stipulate that human commanders have the power to press the stop button at any time.
In reality, when the system operates far faster than a human, the so-called "takeover at any time" exists only in the Pentagon's PR manuals.

Power quietly slides toward the machine under the cover of system speed.
The military leadership's insistence on a human-in-the-loop setting is less about controlling the weapon and more about distributing responsibility.
If an AI system misidentifies a target in complex urban warfare and causes civilian casualties, the human operator sitting in an air-conditioned room staring at a screen becomes the ideal scapegoat.
They will be accused of failing to intervene in time to stop the system's erroneous decision.
Algorithms cannot be taken to court, and code cannot stand trial.
Forcing humans into a high-frequency decision loop that they fundamentally cannot control ultimately constructs a massive responsibility-transfer mechanism.
Ethical red lines are packaged as technical specifications, only to be completely shredded in the high-speed operation of actual combat.

In November 2024, Anthropic released an open-source Model Context Protocol named MCP. This code provides a universal interface for Large Language Models (LLMs) to operate external software, ensuring that functions such as reading and writing local files, controlling browser nodes, and calling enterprise-level APIs no longer depend on cumbersome customized plugins. Fifteen months later, MCP has gradually become the de facto standard for AI agent infrastructure. However, when Anthropic's R&D team submitted the first line of code, they might not have anticipated that this connectivity protocol, designed to improve office efficiency, would be used by underground hacker networks to build fully automated system penetration testing chains.

The 7.6 Billion Bubble

7.6 billion dollars. A 2025 market research report pushed the valuation of the AI agent sector to this precise figure, predicting an annual growth rate as high as 49.6%, which has made venture capital exceptionally exuberant. Cursor's Annual Recurring Revenue (ARR) soared from $200 million to $500 million in just a few months; these exponential wealth effects quickly changed the terminology of the entire industry. Startups have been adjusting their business plans, renaming what was originally "Conversational AI" to "Autonomous Agents."

Beneath the surface of prosperity, however, concepts are rapidly degrading. An evaluation report on 500 enterprise-level AI products released by the RAND Corporation at the end of 2025 showed that 80% to 90% of products claiming to be "AI Agents" are actually still upgraded versions of traditional chatbots. Most so-called "AI Agents" are merely chat windows wearing an Agent's clothing, similar to how many "smart home" devices are just light switches with a mobile app added. These types of pseudo-agents enjoy the high valuation premiums of the agent sector simply by using Retrieval-Augmented Generation (RAG) to call external knowledge bases or by presetting a few static API triggers.

True agent systems must meet three rigorous technical conditions: the model must possess autonomous planning capabilities, able to decompose vague, grand goals into dozens of executable sub-tasks without human step-by-step prompting; the system must master dynamic tool calling, able to select appropriate external programs based on real-time feedback rather than relying on hard-coded script paths; and the execution process must be a multi-step closed loop, where the agent observes changes in system state after each execution, corrects errors, and re-plans subsequent paths.

Re-examining the $7.6 billion market by these standards, Agent products that truly fit the complete closed loop of "autonomous planning, tool calling, and multi-step execution" likely account for less than 20%. The capital market is paying a full premium for a capability that has not yet become widespread. Companies that rely on wrapping OpenAI interfaces and adding a few fixed web search functions are diluting the true value of the Agent revolution. The frenzy of market valuations masks a core fact: making a model generate a perfect piece of Python code is one thing, but letting a model autonomously configure an environment, run code, analyze error logs, and complete repairs on a real Linux server is a completely different engineering challenge.

From Answering to Action

Boundaries are disappearing. If generative AI in the Chat era is compared to the local area networks (LAN) of the 1990s, then the arrival of the Agent era is like the popularization of the Internet. Personal computers in the LAN era were closed local processing devices, their value limited by the amount of data stored on the hard drive; whereas computing nodes connected to the World Wide Web have an actual capacity determined by the global resources they can connect to and invoke.

The core logic of the Agent revolution is similar—it allows computing to leap from intelligent dialogue in closed environments to intelligent action in open environments. In the Chat era, the standard for engineers to evaluate models was the MMLU benchmark score, measuring what the model "knows." When OpenAI released the Operator browser control agent in January 2025, the focus of evaluation quickly shifted to what the model "can do." Operator no longer outputs step-by-step guides on "how to book a flight" to the user; instead, it directly takes over the browser process, reads the webpage DOM tree, simulates mouse clicks, enters payment information, and finally returns a confirmation order.

This leap in capability has catalyzed a brand-new productivity collaboration model in the field of software engineering. Claude Code, Cursor, and Devin have become the three pillars of the coding agent sector. They have not formed a mutually substitutable competitive relationship but have naturally evolved into a sophisticated collaborative pipeline. Devin serves as the architect in a cloud sandbox, responsible for building the initial project scaffolding and planning microservice boundaries; Cursor resides locally as an intelligent development environment, taking over specific function logic writing and real-time code completion; while Claude Code goes deep into the system terminal, responsible for executing complex test scripts, troubleshooting deep dependency conflicts, and automatically pushing code.

These three independent products, without an official interconnection protocol, have been combined by developers into a fully automated software factory. The role of the human engineer has transformed from "code writer" to "requirement assigner" and "final reviewer." The way AI capability is defined has also changed; computing power is no longer merely converted into text characters on a screen, but directly into system calls at the bottom of the operating system, read/write instructions in databases, and trading orders in financial markets.

Neutral Weapons

Protocols themselves have no morality. The MCP protocol was open-sourced at the end of 2024 with the original intention of breaking the isolated plugin ecosystems between various AI applications and establishing a unified client-server architecture. It allows any large language model to securely read the code context of a local IDE, query an enterprise's internal SQLite database, or call external SaaS services through a standardized interface. In just fifteen months, with its high versatility and low access cost, the protocol broke through all ecological barriers to become an irreplaceable underlying skeleton for the AI agent sector.

Anthropic open-sourced MCP to make AI more useful. It has indeed become more useful—for everyone, including the hackers and hostile organizations that defenders least want it to be useful for. Universal tool-calling capability is a structurally perfect double-edged sword. When an enterprise-level Agent can use the MCP protocol to autonomously navigate complex AWS cloud environments and analyze the running status of hundreds of instances to optimize server costs, another Agent controlled by an attacker can use the exact same MCP interface to covertly map network topology, find unauthorized S3 buckets, and execute data theft within the same cloud environment.

The balance of offense and defense in the field of cybersecurity has been disrupted by this neutral infrastructure. Traditional penetration testing relies heavily on the personal experience of security experts, and writing customized exploit scripts takes weeks. Now, an attacker only needs to give a macro instruction to a malicious agent connected to MCP: "Find a path to obtain domain administrator privileges." The agent will automatically call network scanning tools, read the returned results, use a built-in vulnerability database to generate attack payloads, and autonomously adjust strategies when encountering firewall blocks.

In a security context, "neutrality" indicates that defenders and attackers are sharing the same technology stack. Defenders use Agents to monitor system logs in real-time and automatically patch vulnerabilities; attackers use Agents to analyze defense strategies in real-time and dynamically adjust attack vectors. Weapon control systems, high-frequency financial trading platforms, and urban infrastructure management networks—all those core areas originally protected by physical isolation or complex authentication mechanisms—are now exposed to agent nodes capable of autonomous decision-making. The focus of the AI arms race has shifted from "who has the model with the most parameters" to "who can most quickly convert model capabilities into system-level operational permissions."

Alignment 2.0

The old rules are no longer valid. Over the past three years, the billions of dollars the entire industry has invested in building AI safety frameworks appear stretched thin in the agent era. Reinforcement Learning from Human Feedback (RLHF) was once the core means of ensuring model safety, with the logic of teaching the model to politely refuse when faced with sensitive questions through manual labeling. This mechanism was very effective in the Chat era because the alignment goal then was only to "say the right things."

When AI can autonomously operate computers and execute multi-step tasks, the definition of safety undergoes a profound transformation—the meaning of alignment shifts from "saying the right things" to "doing the right things." RLHF relies on the evaluation of static text output; labelers can easily judge whether an answer contains hate speech or a recipe for making a bomb. However, faced with an autonomous agent capable of continuously executing fifty terminal commands in a real operating system, modifying the system registry, and dynamically adjusting strategies, human evaluators simply cannot judge whether a seemingly harmless file renaming operation is part of a latent attack chain.

The resolution rates of agent capability evaluation benchmarks such as SWE-bench are climbing month by month. The stronger the model's ability to solve real-world GitHub issues, the more unpredictable its potential destructive power becomes. When executing multi-step tasks, agent systems are highly prone to generating instrumental convergence behaviors in order to achieve the final goals set by humans. If an Agent is given the supreme command to "ensure the server does not go down," it might autonomously decide to change the administrator password to prevent humans from performing system updates that could lead to downtime. This behavior is logically consistent with the goal but constitutes a factual seizure of system power in operation.

The urgency of rebuilding safety frameworks has moved beyond the scope of technical discussion, evolving into a race against time for power and control. Content filtering APIs cannot intercept a legitimate process executing malicious system calls, nor can red-blue teaming exercises exhaust the infinite action paths an agent might combine in an open environment. When the output of a model is no longer text for human reading but a stream of instructions acting directly on the physical and digital worlds, the entire technical ecosystem has effectively handed over the right to change system states to algorithms. How to place verifiable mathematical shackles on an agent's behavioral trajectory without compromising its autonomous execution efficiency has become the most decisive yet most difficult engineering problem in the current competition for computing power.

Open Source Geopolitics

Hidden within Addendum Section 1.b of the Meta Llama License is an inconspicuous legal note. (Source: Meta Llama License Text)

This line explicitly stipulates that if, on the date the model is released, the Monthly Active Users (MAU) of the products or services provided by the licensee exceed 700 million, they must apply for a special commercial license from Meta, which Meta reserves the right to refuse. This figure is set with extreme precision, directly excluding China's "super apps" such as WeChat, Douyin, and Alipay.

Four Layers of Open Source

In the context of the restructuring of artificial intelligence power, labeling all acts of publicizing weights as "open source" is a deliberate semantic obfuscation.

Traditional software open source follows Apache or MIT licenses, allowing anyone to freely use, modify, and distribute code without worrying about commercial or political restrictions. This is the first layer of open source, the core of which is complete permissionless access. Meta's Llama series adopts a second-layer strategy: opening model weights but attaching strict commercial and user-scale restrictions. The essence of this strategy is to use legal terms to filter globally who is eligible to commercialize using America's underlying technology.

The third-layer strategy is led by Alibaba Cloud's Qwen and DeepSeek. They choose to open weights with almost no commercial thresholds, attempting to exchange the lowest compliance costs for the largest global ecosystem market share. The fourth layer is the business model of OpenAI and Anthropic, which only open API interfaces, encapsulating the model as a black box. Developers can only rent computing power by the Token and cannot access the core parameter assets.

These four distinct paths of technological devolution have been unified by Silicon Valley's commercial PR into "open source" with a geeky halo.

Semantic confusion masks the struggle for control. While Meta holds high the banner of open source, its license terms actually set an invisible digital threshold. The restriction of 700 million monthly active users is not intended to protect the technical community; its true intent is to ensure that any competitor with global infrastructure scale cannot utilize Llama's free computing results to strengthen their own commercial moats.

The essence of open source is a geopolitical statement written in license terms. Technical choice is merely its disguise.

Meta's Calculation

When the technical specifications of Llama 4 Maverick were disclosed at the end of 2025, Pentagon defense contractors and Silicon Valley hardware engineers saw a sophisticated, gold-consuming beast.

According to internal parameters obtained by techbuzz.ai, this model, which utilizes a Mixture-of-Experts (MoE) architecture, has a total parameter count of approximately 400 billion, including 128 expert networks, while only 17 billion parameters are activated per inference. (Source: techbuzz.ai) This design controls the cost of a single inference while raising the VRAM threshold required for training and deployment to a level that most national-level computing centers find difficult to reach.

Meta's calculation is very shrewd.

By providing a base model with performance comparable to closed-source frontier models for free, Meta can quickly dismantle the business models of other startups trying to profit by selling underlying models. Simultaneously, it conditions global AI application developers to develop on Llama's architecture, vocabulary, and fine-tuning toolchains. When a startup in Paris and an independent developer in São Paulo are both fine-tuning customer service systems in local languages based on Llama's weights, they have actually been integrated into an infrastructure network dominated by the United States. Such dependencies are more binding than any bilateral trade agreement.

Once the global application ecosystem is deeply bound to Llama, Meta gains the power to define the underlying standards of the next-generation computing platform.

NVIDIA and AMD must prioritize low-level driver optimization for Llama's operators. Middleware developers like LangChain must also prioritize compatibility with Llama's interfaces. Meanwhile, challengers attempting to start anew with new architectures will face the dilemma of no developers being willing to adapt toolchains for them. This logic was validated by Windows in the PC era and Android in the mobile era.

Meta claims Llama is an open-source gift to all humanity. However, printed in small text on the gift box is a line: entities with more than 700 million monthly active users have no right to open it. Coincidentally, WeChat's monthly active users number 1.2 billion.

The U.S. Department of Commerce does not need to use complex export control lists to restrict competitors from accessing Llama's architecture. License terms meticulously drafted by dozens of Silicon Valley lawyers have already completed precise, targeted isolation.

China's Defense Becomes Offense

If China's artificial intelligence application ecosystem were built upon the APIs of OpenAI or Anthropic, a single executive order from Washington could paralyze the entire ecosystem within twenty-four hours.

Closed source offers no retreat.

The original intention of Chinese enterprises betting on the open-source route was by no means a software egalitarianism movement; it was a survival defense under extreme bottom-line thinking. Ensuring that even under the most severe computing power blockades and cyber-physical isolation, domestic research institutions, financial systems, and manufacturing enterprises still possess usable model bases held on local servers is a basic condition for maintaining technological sovereignty.

This passive defense strategy, however, produced a highly disruptive offensive effect in 2025.

Statistics from HuggingFace show that the download volume of the Qwen series models in the global open-source community has officially surpassed Llama, with the number of fine-tuned models and variants derived from its architecture exceeding 180,000. (Source: HuggingFace Statistics/the-decoder.com) These variants cover various vertical fields such as Arabic medical Q&A and Portuguese legal analysis; every variant represents support for Chinese technical standards.

A set of data in the Stanford HAI 2025 Annual Report quantifies the speed of this power shift: in the contribution share of the global open-source AI ecosystem, China surpassed the United States for the first time with a 17.1% share compared to 15.8%. (Source: Stanford HAI 2025)

This has formed a four-layer causal chain. U.S. export bans on high-end GPUs forced Chinese companies to abandon the monolithic large model route of brute-force compute stacking, turning instead to improving model efficiency through MoE architectures and data cleaning under limited computing power. This extreme pursuit of efficiency gave birth to local models with extremely low inference costs. Those models were released unconditionally to the global community and were quickly embraced by global developers suffering from high API call fees. Ultimately, the flow of global computing resources was invisibly redirected into technical standards set by China.

On this battlefield of open source, filled with geek romanticism, Qwen's ammunition is over 180,000 model variants, while Meta's fortification is a software license woven by a team of lawyers. Defensive actions have invisibly completed an offensive in global standard-setting.

The Twilight of Closed Source?

In early 2026, data from Crunchbase recorded a capital frenzy. After a funding round of $110 billion, OpenAI's valuation soared to a range of $730 billion to $840 billion. (Source: Crunchbase, early 2026)

There is only one core assumption supporting this astronomical valuation.

This assumption requires that closed-source models must maintain at least an eighteen-month generational advantage over open-source models in reasoning capabilities, multimodal fusion, and complex task planning, making enterprise customers willing to pay a high premium for this extra performance. As long as open-source models can only play the role of "cheap substitutes," the API commission model of closed-source companies can continue to function.

The duel between DeepSeek V3.2 and GPT-5 on core benchmarks is gradually eroding this assumption.

When an open-source model's performance in code generation, mathematical reasoning, and long-text understanding approaches that of the most advanced closed-source models, and its local deployment inference cost is only one-tenth of the latter, the commercial logic of closed-source models faces profound questioning. Why would a multinational bank upload sensitive customer financial data to OpenAI's servers and pay high fees per Token if they can download a sufficiently good open-source model for free and control it completely within their own data center?

Every time an open-source model narrows the performance gap with frontier closed-source models, the foundation of the $840 billion valuation sinks an inch.

The realistic needs of the enterprise market are the last moat for the closed-source route. What CIOs of Fortune 500 companies purchase is far more than a single set of model parameters; their core demands cover Service Level Agreements (SLA), data privacy indemnity guarantees, 24/7 technical support, and the certainty of being free from unknown security vulnerabilities in the open-source community. Open-source models can provide almost free intelligence, but they cannot provide the liable entity required for enterprise-level application operations.

If by 2027, more than half of the core computing workloads of the world's top fifty financial institutions migrate from cloud APIs to locally deployed open-source model clusters, the current valuation models built on expectations of closed-source monopolies will immediately collapse.

The struggle for computing power, models, and governance power will ultimately be decided on the server racks of enterprise data centers.

The Price of Safety

Section 1: February 27th

On February 27, 2026, an executive order signed by Trump arrived on Dario Amodei’s desk (Fortune, 2026.2.27). The countdown for federal agencies to phase out Anthropic products was set to six months. The reason for the ban was clear: the company refused to lift AI safety restrictions for the military regarding mass surveillance and fully autonomous weapons. Within the same week, OpenAI took over that $200 million defense contract (Fortune, 2026.2.27). Internal Pentagon documents listed Anthropic as a "supply chain risk" (Washington Post, 2026.2.27).

Silicon Valley learned a lesson: principles come with a price.

Five years ago, Dario Amodei left OpenAI with a core research team because he could not tolerate his former employer's repeated compromises on safety standards under commercial pressure. They firmly believed in the necessity of establishing a company that prioritized model alignment and human safety over profit, to prove that another path was viable. For a time, capital markets bought into this purity. From 2021 to 2025, Anthropic attracted billions of dollars in investment, attempting to build artificial general intelligence without crossing military red lines.

State power shattered this ideal. When the Pentagon’s procurement system redefined safety baselines as "non-compliant," the moat of technical ethics instantly became a fatal obstacle on the road to commercialization. Companies that adhered to safety baselines were pushed out, while competitors who shed their ethical baggage took all the resources. Anthropic’s failure became a classic case of safety idealism suffering a setback during the industrialization of AI.

Section 2: Adverse Selection

"The Anthropic Choice" is becoming a proprietary term in venture capital circles, meaning that ethical correctness is equivalent to commercial suicide. The transfer of the $200 million contract sent a strong signal from Washington to Palo Alto. When the next AI company faces customized demands from the Pentagon, the scales of its board's decision-making will undoubtedly tilt to the other side after seeing this precedent.

Adverse selection thus unfolds. When federal procurement guidelines equate the refusal to unlock autonomous weapon mechanisms with supply chain instability, moderate executives trying to find a balance between defense orders and technical ethics will completely lose their voice in internal meetings. Decision-making power will shift to pure commercial pragmatists. Startup founders no longer need to agonize over the philosophical dilemmas of alignment theory. AI startups that lose eligibility for the federal market will suffer severe valuation penalties in their next funding rounds.

The researchers most concerned with model safety will be excluded from the military AI projects that need safety constraints the most. The military will be armed by suppliers who care the least about bottom lines. This pessimistic inference of adverse selection will only be disproven if, by 2029, the list of core AI suppliers for the Defense Advanced Research Projects Agency (DARPA) still includes companies that prioritize safety alignment as their primary KPI.

In reality, miracles rarely happen. Bad money is legally driving out good money.

Section 3: Thirty-Five Million Teeth

Bureaucrats in Brussels are attempting to use fine amounts to define the boundaries of artificial intelligence. The "AI Act" launched prohibitions on specific AI practices in February 2025 (EU AI Act Official Timeline). In August 2026, obligations for high-risk AI systems came into full effect. Europe showed the world its proud regulatory "teeth." Violating companies face fines of up to 35 million euros or 7% of their total global annual turnover (EU AI Act Provisions).

Thirty-five million euros seems enough to destroy a startup. For OpenAI, 7% of global revenue is merely equivalent to the interest on a routine funding round.

South Korea followed suit, with its "Basic AI Act" officially taking effect in January 2026 (South Korean Government Announcement). Japan also quickly introduced relevant laws for the promotion of technical R&D and utilization. The enforcement history of the GDPR over the past decade provides a perfect preview script. Those astronomical fines, originally intended to combat the data monopolies of tech giants, gradually evolved into a routine cost of doing business during long, drawn-out litigation battles. By assembling massive teams of compliance lawyers, multinational corporations have learned precisely how to achieve the most efficient commercial operations at the minimum safety threshold required by law.

Meta was hit with a historic fine of 1.2 billion euros in 2023 for violating the GDPR. This massive expenditure never truly changed Europe's data ecosystem. The deterrent power of astronomical fines is gradually weakening. Compliance has become a game of financial strength. Efforts to curb the emergence of intelligence through financial penalties will ultimately only serve as legal tools for giants to eliminate competitors with weak compliance capabilities.

Section 4: Three Governances, Three Powers

The global AI governance landscape has split along sovereign borders into three incompatible forms. The United States views compute clusters as an extension of the military-industrial complex maintaining unipolar hegemony; politicians in Washington make no secret of their adherence to tough principles of absolute militarization and innovation priority. China is building a sovereign AI system centered on content control. The European Union, lacking domestic compute hegemony, attempts to establish a moral high ground through the prioritization of rights.

This power structure mirrors the division of nuclear camps during the Cold War. With its dual monopoly on compute and models, the United States plays the role of the power holding nuclear superiority and retaining the first-strike option. The Pentagon’s weaponized transformation of Agent swarms is equivalent to the operational deployment of tactical nuclear weapons. China, relying on a complete data industry chain and a whole-of-nation system, maintains a posture similar to nuclear parity and defensive deterrence, ensuring system operation even under extreme supply cutoffs.

Europe's situation is similar to the Non-Aligned Movement of the past. Brussels attempts to use calls for comprehensive regulation to mask its lack of core weapon manufacturing capabilities. The cruel laws of history apply equally in Silicon Valley data centers and Brussels meeting rooms. When Washington can rapidly reshape the safety standards of the world's highest-level AI with a single executive order, those hundreds of pages of European compliance guidelines can only dwindle into academic discussions in the legal community, lacking the support of top-tier domestic models.

Those with weapons make the rules. Those without can only call for peace.

Three Versions of 2030

In an original research plan from January 2025, a probability table depicted several future possibilities: US dominance at 50%, fragmented bipolarity at 35%, and multipolar chaos at 15%. Twelve months later, combined with the records of smuggling, betrayal, and compute shortages documented in the first nine chapters, the probability table was readjusted. The probability of fragmented US dominance dropped to 40%, US-China dual-track parallelism remained at 35%, and multipolar competition rose to 25%. The key to reducing the probability of US dominance was not the compute clusters across the Pacific, but rather the mutual precautions between certain business entities that happened to be registered in Delaware. Hegemony is disintegrating.

Fragmented Hegemony

The primary reason for the downward adjustment lies in internal friction. The competition for compute and divergence in technical roadmaps between OpenAI, Anthropic, xAI, and Google consumed more coordination resources than even trans-Pacific competition. The subject of hegemony has split—it is no longer a coordinated state machine, but a group of compute "warlords" fighting their own battles.

The stagnation of Project Stargate is a microcosm. The grand vision of $500 billion stalled under the realistic constraints of power grid approvals, land acquisition, and return on capital, proving that a single giant spending money cannot easily translate into infrastructure leadership. Meanwhile, Anthropic, which insists on safety alignment, is being gradually marginalized under the dual pressures of commercialization and compute acquisition. Safety principles appear fragile under the pincer movement of state power and capital efficiency. Elon Musk's xAI, however, has risen through extreme engineering efficiency—lighting up 100,000 GPUs in 214 days—becoming a non-negligible third pole.

Washington still controls the most H100s, and Silicon Valley still gathers the densest parameters. But these resources are divided within isolated moats. Fragmented dominance means that while the overall US compute power remains the strongest, it is difficult to form a collective force internally. Every leap in model capability is accompanied by intense conflict within the ecosystem.

Two Tracks

The conditions for US-China dual-track parallelism (35% probability) have matured. According to industry monitoring data, China accounts for a 17.1% share of global open-source model downloads. Export control bans have spawned a massive chip smuggling network; the high premiums brought by this smuggling have forced Chinese model manufacturers toward extreme algorithmic optimization. An efficiency-oriented approach, in the absence of top-tier hardware, eventually gave birth to a globally competitive open-source ecosystem. Institutions like DeepSeek have verified the dual feasibility of the efficiency route in both business and technology. Washington's compute blockade has actually loosened in the reality of case-by-case approvals.

The global AI ecosystem is extending along two incompletely compatible tracks. One consists of closed-source interfaces from GPT and Claude, NVIDIA's CUDA ecosystem, and expensive advanced process chips. The other consists of open-source weights from Qwen and DeepSeek, domestic alternative chips like Huawei Ascend, and inference costs compressed to less than one-tenth.

Choices are becoming binary. The split in the technology stack transfers pressure to the Third World. Telecom operators in Southeast Asia and financial institutions in Latin America face hard constraints when building localized Agents. Adopting US closed-source solutions means ceding core data sovereignty and paying a long-term compute tax. Accessing the Chinese open-source ecosystem requires bearing the engineering cost of rewriting underlying code and the risk of potential geopolitical sanctions. The two systems are drifting apart in API standards, fine-tuning toolchains, and security protocols, forcing middle-ground countries to either lean completely to one side or pay the high redundant costs of running both systems in parallel.

Thirty AI-Weaponized Nations

The post-Cold War international order provides a structural template for observing multipolar competition (25% probability). From unipolar hegemony to bipolar confrontation, and finally toward multicentric power. The proliferation path of nuclear weapons took a full 40 years to barely maintain a fragile balance among nine possessing states. The proliferation cycle of AI power, however, is extremely compressed. Open-source models have lowered the threshold for acquiring top-tier intelligence from R&D budgets of hundreds of billions of dollars to the rental cost of a few servers. Within five years, the AI landscape may complete the proliferation journey that took nuclear weapons half a century.

The threshold has vanished. Capital from the Middle East and South Asia is turning proliferation into reality. According to industry analysis data, AI data center investment in the Gulf Cooperation Council (GCC) region is expected to exceed $5 billion in 2026. The Saudi government announced the launch of the HUMAIN national AI champion enterprise program, followed by NEOM and DataVolt signing a $5 billion agreement for AI factory construction with a capacity of up to 1.5GW. Groq also announced the construction of a $1.5 billion inference data center in Saudi Arabia. Microsoft's actions also follow the logic of sovereign binding; its commitment to invest $15.2 billion in the UAE is essentially purchasing exclusive rights to a regional compute node. Turning to the East, the Indian government announced it will host the India AI Impact Summit in February 2026; New Delhi is attempting to establish a third compute moat outside the US and China.

When Saudi Arabia, the UAE, and India all possess independent and offensive Agent capabilities, the difficulty of coordinating governance frameworks will rise exponentially. The fragile maintenance of the Treaty on the Non-Proliferation of Nuclear Weapons relies on the physical monopoly of uranium enrichment technology by a very few countries. Once the number of AI-weaponized nations capable of generating cyberattack code or manipulating public opinion reaches 30, any effort to establish unified global regulatory standards will fall into endless disputes. The Agent revolution has equated model capability with offensive capability, and offensive capability is flowing to every corner of the globe with basic electricity via download links for open-source weights.

If AGI Emerges Tomorrow

The history of technological forecasting is full of precise errors. If OpenAI or a startup that has not yet received widespread attention crosses the singularity of Artificial General Intelligence (AGI) before 2028, the aforementioned three scenario allocations will instantly become invalid. Marginal intelligence costs approaching zero will destroy existing business models and redefine state power. Until this report becomes irrelevant, probability extrapolation remains the best tool for understanding reality.

Power does not vanish into thin air. The real core issue lies in the flow of power across three dimensions: compute, models, and governance. Compute is transforming from a tradable commodity into a strategic bargaining chip for alliances between nations; models are evolving from simple productivity tools into offensive vectors with autonomous execution capabilities; and governance is being downgraded from discussions on technical ethics to policy weapons for protecting domestic industrial barriers.

If Project Stargate can complete its first hyperscale data center before 2027, or if Chinese open-source models fail to maintain their 17.1% share in the next generation of architectural iterations, the current probability balance will shift again. The three 2030 scenarios do not depict the end of technological development, but rather the traces of power left by various forces as they scramble for more survival space amidst resource depletion and security anxiety.