The China Effect in Generative AI: How Cheap Models Are Reshaping Market Value

Introduction

Over the past two years, the global generative AI market has undergone a structural transformation that few analysts anticipated even in their most optimistic forecasts. What began as a race for technical supremacy among a handful of Western labs — chiefly OpenAI, Anthropic, and, to a lesser extent, Google DeepMind — has become a multipolar competitive landscape in which comparative advantage no longer rests exclusively on model performance, but on the fundamental equation underlying any technology business: the relationship between production cost, sale price, and perceived customer value. This transformation has a name: the China factor.

The generative AI industry finds itself at a critical inflection point. On one hand, OpenAI and Anthropic, the two most valuable private companies in the sector, have announced their intention to go public in the 2026–2027 window at valuations exceeding $850 billion and $965 billion respectively (1) (2). These figures, which place both companies among the most valuable technology firms in the world, are justified only by the expectation that they will maintain a near-monopolistic hold on the high-performance language model market, with the ability to set premium prices indefinitely. The investor narrative — fueled by record funding rounds and the backing of hyperscalers — has built a house of cards on the premise that artificial intelligence is a scarce resource, difficult to replicate, and protected by insurmountable barriers to entry.

That narrative, however, has begun to crack. Since late 2025, and with particular intensity through the first half of 2026, a group of Chinese labs — DeepSeek, Zhipu AI (Z.ai), Alibaba (Qwen), Moonshot (Kimi), MiniMax, and Xiaomi (MiMo) — have shown not only that they can match or approach the performance of the most advanced Western models, but that they can do so at a fraction of the training cost and, more importantly, at retail prices ranging from one-tenth to one-fiftieth of what OpenAI and Anthropic charge (3) (4). Reports from the U.S. National Institute of Standards and Technology (NIST) have confirmed that the DeepSeek V4 Pro model performs similarly to OpenAI's GPT-5, with a gap of just eight months between release windows, and that it beats GPT-5.4 mini on cost efficiency in five of the seven benchmark categories evaluated (5). This finding, coming from a U.S. government agency, has had a devastating effect on the perception of Western technological superiority.

The impact of this disruption is not merely technical, but profoundly economic. The price war in application programming interfaces (APIs) has been the first great battlefield. While GPT-5.5 sells for $5 per million input tokens and $30 per million output tokens, and Claude Sonnet 4.6 sits at $3 and $15 respectively, DeepSeek V4 Flash offers prices of $0.14 and $0.28, and Z.ai (GLM-4.5) sets its rate at $1.1 per million output tokens (6) (7). This order-of-magnitude difference has direct implications for enterprise adoption. Startups such as San Francisco-based Lindy have migrated entire workloads from Anthropic to DeepSeek, saving millions of dollars a year without any appreciable loss of response quality (6). The company's product lead stated that the savings allowed the business unit to reinvent itself — a statement that sums up the disruption now underway.

This article aims to analyze this phenomenon in depth from a multidisciplinary perspective that integrates technical model analysis, price and cost economics, market-share dynamics, the corporate finance of the companies involved, and the broader macroeconomic context. Its central thesis is that China's AI strategy — built on open-source models, training efficiency, inference optimization, and aggressive pricing — has struck OpenAI and Anthropic at their weakest point at precisely the worst possible moment: the run-up to their public listings. By eroding these companies' ability to sustain high margins, Chinese competition is forcing a wholesale recalibration of their valuations and, by extension, of the entire AI investment ecosystem.

The article is organized into seven parts. The first describes the new competitive landscape, examining the valuation paradigm of the Western champions and the emergence of the Chinese model as a credible alternative. The second offers a quantitative analysis of technical performance and pricing structure, with particular attention to NIST reports and per-million-token cost comparisons. The third examines market-share dynamics, drawing on aggregation-platform data such as OpenRouter and enterprise migration case studies. The fourth addresses the financial pressures and the challenge to the IPO narrative, analyzing OpenAI's and Anthropic's cost structures and the profitability dilemma in a falling-price environment. The fifth situates the phenomenon within the macroeconomic context of global AI spending, examining Gartner's projections and the infrastructure paradox: how to justify $600 billion in data-center investment if the price of a token is collapsing? (8) (9). The sixth addresses the geopolitical and regulatory dimensions, including China's open-AI strategy as an instrument of industrial policy and semiconductor export controls. Finally, the seventh proposes future scenarios and strategic recommendations for investors, asset managers, and the executive teams of the affected companies.

Methodologically, this article draws on primary and secondary sources, including reports from official bodies (NIST, the US-China Economic and Security Review Commission), corporate financial disclosures, analysis from investment banks (UBS, JPMorgan, Wedbush), model-aggregation platform data, and specialized economic press. The approach is fundamentally qualitative in interpreting strategic dynamics, but it rests on rigorous quantification of prices, costs, market shares, and valuations to support its conclusions.

The article does not claim to offer a definitive answer to a phenomenon still in full ferment, but rather to provide a robust analytical framework that allows academics, investors, and policymakers to understand the underlying forces reshaping the global AI market. The hypothesis guiding this research is that the Western duopoly — the belief that OpenAI and Anthropic will dominate the high-performance model market for the next decade — is a fragile construction already being challenged by a combination of technical innovation, cost efficiency, and open-source strategy from China. The outcome of this confrontation will determine not only the stock-market value of these companies, but also the future direction of AI development, access to its benefits, and the global distribution of technological power.

Part I. The New Competitive Landscape: China's Emergence as a Cost-Efficiency Leader

1.1. The Valuation Paradigm of the Western Champions: OpenAI and Anthropic on the Threshold of an IPO

The generative AI industry experienced, in the first half of 2026, an unprecedented financial phenomenon in the history of technology. Two companies, OpenAI and Anthropic, have driven their private valuations to levels that place them among the most valuable corporations on the planet, surpassing established technology giants and approaching the select club of companies valued above a trillion dollars. This accelerated revaluation is not the product of conventional organic growth, but of an investment bet of historic magnitude that has redefined the parameters of venture capital and placed artificial intelligence at the center of global asset-allocation strategy.

The chronology of this phenomenon is revealing. OpenAI, founded in 2015 as a nonprofit AI research organization, has completed a corporate transformation that has led it toward a public listing. On March 31, 2026, the company closed a $122 billion funding round — the largest private round in Silicon Valley history — that set its post-money valuation at $852 billion (1) (2). The round, anchored by Amazon, Nvidia, and SoftBank, represented an increase of roughly 184% over the $300 billion valuation the company had reached twelve months earlier, in March 2025 (3). By late April 2026, the secondary market placed OpenAI's implied valuation at around $880 billion — a premium of just 3% over the primary round, a sign that institutional investors consider the valuation to already fully reflect near-term growth expectations (3).

Anthropic's trajectory, for its part, has been even more dizzying. Founded in 2021 by executives who left OpenAI in search of an approach more centered on safety and interpretability, the company has completed a valuation climb that stands as a case study in venture-capital history. In March 2025, Anthropic was valued at $61.5 billion (4). By September of that same year, the figure had tripled to $183 billion. In February 2026, a $30 billion Series G round pushed the valuation to $380 billion (5). And in May 2026, a $65 billion Series H round, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, sent the valuation soaring to $965 billion, surpassing OpenAI on this metric for the first time (6) (7). The secondary market already values Anthropic above a trillion dollars, at a 6% premium over its latest round price (8).

This progression — which took Anthropic from $1.5 billion in 2022 to $965 billion in May 2026, a 643-fold increase in just four years (4) — raises fundamental questions about the economic fundamentals underpinning such valuations. The company reported annualized recurring revenue of $14 billion in February 2026, implying a multiple of roughly 70 times that revenue (8). OpenAI, with recurring revenue exceeding $20 billion at the end of 2025, trades on the secondary market at a multiple of roughly 44 times (3). Both multiples are extraordinarily high compared with the standards of the publicly traded technology sector, where high-growth software companies typically trade between 10 and 20 times recurring revenue.

The race toward IPOs has accelerated in tandem with this revaluation process. Anthropic moved first, filing a confidential S-1 registration statement with the Securities and Exchange Commission (SEC) on June 1, 2026, opening the door to a public listing (9) (10). OpenAI followed a week later, on June 8, also filing its own confidential S-1 (11) (12). Both companies join SpaceX, which completed its own IPO on June 12, 2026, raising $75 billion at a debut valuation of $1.77 trillion — the largest initial public offering in history (13). Several sector analysts have described this as an unprecedented concentration of pre-IPO capital, and have warned that the 2026 window could become either the most consequential IPO cycle since the dot-com era or a costly lesson about the distance between narrative and fundamentals, depending on how public-market appetite evolves over coming quarters (10).

OpenAI's decision to delay its IPO until 2027, as reported by several financial outlets in late June, adds a layer of uncertainty to this picture (13). CEO Sam Altman had initially pushed for a $1 trillion valuation at listing, but SpaceX's volatile post-debut performance — its shares traded above $225 before falling back to around $153 within days — has cooled that enthusiasm (13). OpenAI's advisers have warned that the company may not find the same enthusiasm among retail investors, and the company has decided to take it slow, even as the need for capital to sustain its pace of investment in computing infrastructure grows increasingly pressing (13).

The valuation paradigm underpinning OpenAI and Anthropic rests on several pillars. First, the belief that generative AI is a market with high barriers to entry, where frontier models — those setting the state of the art — can only be developed by a handful of labs with access to massive computing resources and top-tier research talent. Second, the expectation that these models, being closed-source and proprietary, will generate recurring revenue streams with high margins, similar to those of enterprise software platforms. Third, the conviction that the AI market will behave as a winner-take-all or duopoly market, where established leaders will capture the bulk of the value created.

These pillars, however, are coming under growing pressure from a competitor that was barely on investors' radar just twelve months ago: Chinese AI labs. DeepSeek, Z.ai, Alibaba (Qwen), Moonshot (Kimi), MiniMax, and Xiaomi (MiMo) have shown that it is possible to develop models with performance comparable to OpenAI's and Anthropic's at a fraction of the training cost and, more importantly for the sustainability of the Western business model, sell them at prices ranging from one-tenth to one-fiftieth of what U.S. leaders charge.

The May 2026 report from the U.S. National Institute of Standards and Technology (NIST) confirmed that DeepSeek V4 Pro performs similarly to OpenAI's GPT-5, with a gap of just eight months between release windows, and beats GPT-5.4 mini on cost efficiency in five of the seven benchmark categories evaluated (14). This finding, coming from a U.S. government agency, has had a devastating effect on the perception of Western technological superiority that underpinned OpenAI's and Anthropic's valuations.

Several market analysts have noted that the timing is delicate for OpenAI, which appears to be losing some of ChatGPT's early lead over Google and Anthropic, at a time when its options for raising additional short-term capital are increasingly limited (12). This precariousness is compounded by both companies' cost structures: Anthropic reported annualized recurring revenue of roughly $14 billion in February 2026, a figure that soared to around $47 billion by late May, driven largely by Claude Code, though that growth has been accompanied by equally steep spending on computing and infrastructure (15). The company has acknowledged that its cash-burn rate is high and that it depends on periodic capital injections to scale its infrastructure.

In this context, Chinese competition represents not only a technological threat but a financial one. If enterprise customers can obtain similar-performing models at a fraction of the cost, OpenAI's and Anthropic's ability to sustain premium prices erodes — and with it, the justification for their astronomical valuations. UBS analysts have warned that the price gap between China and the United States will inject volatility into the sector and could lead investors to rethink massive infrastructure spending (16). This warning carries particular weight on the threshold of these IPOs, when the financial transparency demanded by public markets will test the narrative of sustained growth and high margins that has fueled private valuations.

The phenomenon we are witnessing, then, is not merely a price war in the API market. It is a confrontation between two business models: the Western labs' model, based on closed source, intellectual property, and premium pricing, and the Chinese labs' model, based on open source, operational efficiency, and aggressive pricing. The outcome of this confrontation will determine not only the stock-market value of OpenAI and Anthropic, but also the structure of the global AI market for the next decade.

1.2. The Rise of the Chinese Model: Open Source, Scalability, and Price Disruption

China's emergence as a competitive force in the global generative AI market is not an isolated or short-term phenomenon, but the consolidation of a deliberate industrial strategy that combines public investment in research, talent development, ecosystem openness, and a commercialization philosophy radically different from that of Western labs. While OpenAI and Anthropic have built their business model on closed source, protected intellectual property, and premium pricing, Chinese labs have bet on open source, massive scalability, and price disruption as vectors for penetrating the global market. This strategy, far from being a mere discount replica of Western models, rests on architectural and efficiency innovations that have drastically cut training and inference costs, turning a price advantage into a structural advantage that is difficult to replicate.

China's AI ecosystem is led by a diverse set of labs and technology companies that have managed to position their models at the frontier of global performance. DeepSeek, the startup that has captured international attention, has shown that it is possible to develop 1.6-trillion-parameter models using a mixture-of-experts (MoE) architecture that cuts inference cost to a fraction of that of its Western competitors (17) (18). On April 24, 2026, DeepSeek released two open-source models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, supporting contexts of up to one million tokens under Apache 2.0 licenses that allow unrestricted commercial use (18). This move, made the same day OpenAI unveiled GPT-5.5, was interpreted by analysts as a direct challenge to Western leadership (17).

The open-source strategy is not exclusive to DeepSeek. Zhipu AI (Z.ai), the Chinese lab backed by the government and major tech investors, has released GLM-4.5 under an MIT license, on both Hugging Face and the Chinese platform ModelScope, with 355 billion total parameters and 32 billion active parameters in its MoE architecture (19) (20). GLM-4.5 has been described by its developers as the best-performing open-source model for reasoning, coding, and agentic tasks, and was trained on 22 trillion tokens, of which 15 trillion correspond to general content and 7 trillion to reasoning and code (20). Alibaba, for its part, has turned Qwen into the most-downloaded open-source model family in the world, with more than 700 million cumulative downloads and more than 180,000 community-derived versions (21). On February 16, 2026, Alibaba released Qwen3.5, a 397-billion-parameter multimodal model with native support for 201 languages and a one-million-token context window, directly competing with GPT-5.2 (22) (23). The company has stated that its goal is to "develop once, deploy anywhere," and has built an open ecosystem spanning the model, system, and chip layers (23).

The open-source bet has profound implications for market dynamics. While OpenAI's and Anthropic's models are accessible only through their proprietary APIs, with usage costs that can run into the tens of thousands of dollars a year for mid-sized development teams, Chinese models can be downloaded, run on commercial hardware, and fine-tuned without licensing fees. This accessibility has driven a wave of adoption among developers and small companies that previously could not afford access to frontier models. As the aggregation platform OpenRouter has noted, the six most popular models among developers today are all Chinese: Xiaomi's MiMo-V2-Pro, stepfun's Step 3.5 Flash, DeepSeek V3.2, MiniMax M2.7, MiniMax M2.5, and Z.ai's GLM 5 Turbo (15). Anthropic, which held the top spot in March 2025, has fallen to seventh and eighth place — an eloquent indicator of the shift in preference within the developer community (15).

The Chinese models' price advantage is not marginal but an order of magnitude. DeepSeek V4-Pro is priced at $1.74 per million input tokens, while V4-Flash costs roughly $0.14 per million input tokens (24) (25). By comparison, GPT-5.5 sells for $5 per million input tokens and $30 per million output tokens, and Claude Sonnet 4.6 sits at $3 and $15 respectively (6). The difference is even sharper on output tokens: DeepSeek V4-Pro costs $0.87 per million output tokens, versus $30 for GPT-5.5 and $15 for Claude Sonnet 4.6 — a 34x and 17x advantage respectively (25). For Z.ai, the API price is $1.1 per million output tokens, while MiniMax offers $1.2 on its M2.5 model (7). This pricing structure has led UBS analysts to state that Chinese models are, on average, 20% cheaper than their global counterparts, and in some specific cases up to 50 times cheaper (16).

The Chinese labs' ability to offer such low prices is not solely a strategic decision to sacrifice margin for market share, but reflects structural advantages in training and inference efficiency. According to UBS analysts, the cost of training these models in China is below 10% of comparable Western spending (16). This efficiency comes from a combination of factors: optimized model architectures, such as mixture-of-experts (MoE), which activates only a fraction of parameters per inference; memory-compression innovations, such as shrinking the KV cache to just 7% of conventional size; and the use of commercial hardware rather than costly cutting-edge accelerators (26). DeepSeek has developed a hybrid CSA+HCA attention architecture that cuts the computational demand for long contexts to a tenth of the previous generation, allowing its models to run on consumer graphics cards such as the RTX 4090 rather than requiring clusters of thousands of H100 GPUs (26) (27). This hardware optimization, combined with the availability of Chinese-made AI chips that, while less powerful than Nvidia's, are significantly cheaper and more accessible, has allowed Chinese labs to drastically cut their infrastructure costs.

The next battleground, according to analysts, is inference optimization, where Chinese companies are already making significant strides. DeepSeek has recently introduced a speculative-decoding framework that accelerates inference by up to 85%, reducing computational load and chip bottlenecks (28). This kind of innovation, which allows larger models to run with fewer resources, is especially relevant in the context of U.S. semiconductor export controls restricting China's access to the most advanced chips. Far from slowing China's progress, these controls appear to have spurred a wave of efficiency innovation that is redefining industry standards.

The open-source, low-price strategy also has a geopolitical dimension that cannot be ignored. The report by the US-China Economic and Security Review Commission (USCC), titled "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance" (March 2026), notes that Chinese labs "have narrowed performance gaps with leading Western models" thanks to an open ecosystem that lets them innovate near the frontier despite restricted access to advanced hardware (29). The commission warns that China's open-AI strategy is not philanthropic but an instrument of industrial policy: open models feed a "digital loop" of adoption and improvement, intertwined with a "physical loop" of industrial data generation in factories, logistics, and robotics — something semiconductor-centered export controls are not designed to contain (29). This strategy, which combines openness with state control of computing and data infrastructure, represents a systemic challenge to Western technological leadership that goes beyond the purely commercial sphere.

In this context, the rise of the Chinese model is not a passing fad or a speculative bubble, but the manifestation of a structural shift in the economics of artificial intelligence. Chinese labs have shown that it is possible to compete in the global market not through absolute technological superiority, but through a combination of operational efficiency, ecosystem openness, and aggressive pricing that erodes the economic foundations of the Western business model. The question that follows is whether this cost advantage translates into a sustainable long-term competitive edge, or whether Western labs can respond with innovations that restore their primacy. Answering that question requires examining in detail the technical performance metrics and pricing structure underpinning competition in the API market.

Part II. Quantitative Analysis of Technical Performance and Pricing Structure

2.1. Comparative Capability Assessment: NIST Reports and the Narrowing Technology Gap

The perception of Western technological superiority in generative AI has been one of the fundamental pillars supporting OpenAI's and Anthropic's astronomical valuations. Between 2023 and 2025, the consensus among investors and analysts was that U.S. labs held an irreducible qualitative and quantitative edge over their Chinese rivals, based on privileged access to cutting-edge hardware, accumulated research talent, and critical mass of training data. That consensus began to crack following the May 2026 publication of an assessment by the Center for AI Standards and Innovation (CAISI), part of the U.S. government's National Institute of Standards and Technology (NIST), of the DeepSeek V4 Pro model (8) (14).

The NIST report — the first official, systematic assessment of a Chinese frontier model by a U.S. government agency — reached conclusions of major significance for competitive analysis of the sector. According to the document, DeepSeek V4 Pro is "the most capable AI model from the People's Republic of China evaluated by CAISI to date" (8). The assessment, carried out in April 2026, tested the model across five critical domains: cybersecurity, software engineering, natural sciences, abstract reasoning, and mathematics (8). In terms of aggregate performance, NIST concluded that DeepSeek V4 Pro sits roughly eight months behind the most advanced U.S. models on its aggregate capability measure (8) (14).

This conclusion, though seemingly favorable to Western labs, contains two nuances of extraordinary importance. First, the assessment finds that DeepSeek V4 Pro, released in late April 2026, performs similarly to GPT-5, which OpenAI had released eight months earlier, in August 2025 (8) (14). This implies that the technology gap between Chinese and U.S. leaders has narrowed dramatically over the past year and a half, from a difference many analysts estimated at two to three years down to just eight months. Second, the report highlights that DeepSeek V4 Pro beats GPT-5.4 mini on cost efficiency in five of the seven benchmark categories evaluated (14). This efficiency advantage, far from a minor technical detail, is a fundamental economic attribute that places Chinese models in a competitive position that goes beyond mere technological parity.

The NIST assessment has sparked intense debate among researchers and policymakers. Some analysts have questioned whether the "eight-month lag" metric adequately reflects the reality of the Chinese ecosystem, noting that the assessment does not capture all of DeepSeek V4 Pro's capabilities, such as its performance on ultra-long contexts (8). Others have noted that the eight-month gap, far from indicating inferiority, represents a remarkable achievement given that Chinese labs have developed their models under hardware-access restrictions imposed by U.S. export controls (8). The US-China Economic and Security Review Commission's own report agrees that Chinese labs "have narrowed performance gaps with leading Western models" by relying on an open-model ecosystem that reduces their dependence on the most advanced hardware (29).

The relevance of the NIST assessment for the financial analysis of OpenAI and Anthropic is immediate. If Chinese models can match GPT-5's performance with a lag of just eight months, and do so at a fraction of the cost, Western labs' ability to sustain premium prices is seriously compromised. The eight-month window, which might seem significant given how fast AI is evolving, shrinks considerably once one considers that the gap has been closing at an accelerating pace: in 2024, the estimated difference was roughly two years; in 2025, fourteen months; and in April 2026, eight months. If this trend holds, full technological parity could be reached around 2027 — precisely when OpenAI and Anthropic expect to complete their public listings.

NIST's conclusion that DeepSeek V4 Pro is "the most capable AI model from the People's Republic of China evaluated by CAISI to date" (8) also underscores that this is not an isolated case, but the manifestation of a fully maturing ecosystem. Chinese labs have not simply produced one competitive model; they have built a constellation of models — DeepSeek, Qwen, GLM, MiniMax, Kimi, MiMo — that compete with each other and with Western leaders in a domestic market of colossal scale. This competitive diversity, fueled by the open-source strategy and public research investment, constitutes a systemic advantage that Western labs, concentrated among a handful of players, find difficult to replicate.

2.2. The API Price War: Comparing Cost per Million Tokens

If the technology gap has narrowed to eight months, the price gap has widened into a chasm. The market for large-language-model application programming interfaces (APIs) has been the battlefield where Chinese competition has deployed its most decisive advantage. Prices per million tokens — the basic unit of consumption in generative AI — show order-of-magnitude differences with direct implications for enterprise adoption and providers' business models.

The pricing comparison for the second quarter of 2026 reveals a deeply segmented market structure. In the premium segment, OpenAI sells GPT-5.5 at $5 per million input tokens and $30 per million output tokens (6). Anthropic, for its part, offers Claude Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens, with an introductory price of $10 that will rise to $15 once the promotional period ends (6) (7). At the lower end of the spectrum, Chinese models have established a radically lower price band. DeepSeek V4 Pro, the Chinese lab's high-end model, is priced at $0.435 per million input tokens and $0.87 per million output tokens (9). The Flash version, oriented toward efficiency and cost, cuts these prices further to $0.14 per million input tokens and $0.28 per million output tokens (9).

The gap between the two extremes is vast. In terms of output tokens, DeepSeek V4 Pro is 34 times cheaper than GPT-5.5 ($30 versus $0.87) and 17 times cheaper than Claude Sonnet 4.6 ($15 versus $0.87). The Flash version multiplies that advantage further: it is 107 times cheaper than GPT-5.5 and 54 times cheaper than Claude Sonnet 4.6 (9) (6). The gap widens further still when factoring in cache discounts, which bring the cost of input tokens down to as little as $0.0037 per million on some Chinese models (9). UBS analysts have estimated that, on average, Chinese models are 20% cheaper than their global counterparts, and in some specific cases up to 50 times cheaper (13) (16). JPMorgan, for its part, has noted that Chinese models are up to 50 times cheaper than OpenAI's and Anthropic's (11).

Other Chinese labs have set equally aggressive prices. Z.ai (GLM-4.5) charges $1.1 per million output tokens, while MiniMax (M2.5) sits at $1.2 (7). Moonshot AI (Kimi K2.5) offers prices of $2.2 per million output tokens (9). On the input side, DeepSeek V3.2 sells for $0.42 per million tokens (9). This pricing structure, spanning from DeepSeek Flash's $0.14 to GPT-5.5's $30, represents a spread of more than 200 times between the cheapest and most expensive model on the market.

The price war is not a static phenomenon; it has intensified throughout 2026. In May 2026, DeepSeek announced a permanent 75% cut to V4 Pro's pricing, setting the new structure at $0.435 per million input tokens and $0.87 per million output tokens (9). This cut, which followed a temporary 75% promotion that expired on May 31, 2026, has cemented DeepSeek's position as the cheapest provider of high-performance models on the market (9). Cached-token prices have fallen to 0.025 yuan ($0.0037) per million tokens, a level that makes the cost of repetitive queries practically irrelevant (9). This pricing strategy, combining permanent discounts with aggressive promotions, has created a market dynamic in which Western labs are forced to respond with their own price cuts, eroding their margins and undermining the justification for their valuations.

The question that arises is whether Chinese labs can sustain these prices. The answer appears to be yes, for several reasons. First, their training and inference costs are significantly lower, as analyzed in the next section. Second, China's domestic market, with hundreds of millions of users and businesses, provides a revenue base and critical mass of data that enables economies of scale. Third, the open-source strategy lowers marketing and distribution costs while generating a community of developers who contribute to model improvement. Finally, Chinese government backing for the AI industry, both in terms of funding and industrial policy, provides a financial cushion that allows Chinese labs to sustain aggressive prices over extended periods.

2.3. Training Efficiency: The 10x Factor as a Structural Advantage

The Chinese labs' ability to offer such low prices is not the result of a strategic decision to sacrifice margin for market share, but stems from a structural advantage in training and inference efficiency rooted in architectural innovations, hardware optimization, and a development philosophy radically different from that of Western labs.

The most revealing data point behind this structural advantage is model training cost. DeepSeek reported having trained its V3 model at a cost of roughly $5.5 million, using nearly 2.8 million H800 GPU-hours (10). This widely cited figure contrasts with estimates exceeding $100 million for training GPT-4, though the calculation methodologies behind each figure are not directly comparable (they do not, for instance, include prior research costs or general infrastructure) (10). A Wedbush analysis described that figure as breaking the assumption that training frontier models necessarily requires investments above $100 million (10). UBS analysts have estimated that the cost of training models in China may sit well below equivalent Western spending, though exact figures depend on which cost items are included in the calculation (13) (16).

This efficiency is not accidental but the result of deliberate architectural innovation. DeepSeek has developed a mixture-of-experts (MoE) architecture that activates only a fraction of the model's parameters per inference, drastically reducing computational load. The company has implemented a hybrid attention architecture (CSA+HCA) that cuts computational demand for long contexts to a tenth of that of the previous model generation (26). It has also developed memory-compression techniques that shrink the KV cache to just 7% of conventional size, allowing 1.6-trillion-parameter models to run on commercial hardware (26). These innovations, combined with the use of Chinese-made AI chips that, while less powerful than Nvidia's, are significantly cheaper and more accessible, have allowed Chinese labs to drastically cut their infrastructure costs.

Training efficiency has direct implications for pricing structure. If a model costs 18 times less to train, the provider can afford significantly lower sale prices without sacrificing margins. This advantage widens further at the inference stage, where architectural optimizations allow models to run with fewer computing resources, lowering the cost per query. DeepSeek has recently introduced a speculative-decoding framework that accelerates inference by up to 85%, reducing computational load and chip bottlenecks (28). This kind of innovation, which allows larger models to run with fewer resources, is especially relevant given U.S. semiconductor export controls restricting China's access to the most advanced chips. Far from slowing China's progress, these controls appear to have spurred a wave of efficiency innovation that is redefining industry standards.

UBS's analyst has noted that the price gap between China and the United States, far from being a temporary anomaly, reflects structural differences in the efficiency of the AI value chain (16). While Western labs have opted for a "more brute force" strategy — more GPUs, more data, more parameters — Chinese labs have bet on efficiency and optimization. This strategic divergence has profound implications for long-term competitiveness. If efficiency is a cumulative advantage — the more you optimize, the more you learn about how to optimize — Chinese labs may be building an edge that will become increasingly difficult for their Western rivals to replicate.

2.4. Enterprise Adoption and Demand Elasticity in the Face of Falling Prices

The pricing structure described in the previous sections would have no economic relevance if enterprise customers were not willing to migrate to Chinese models. The available empirical evidence, however, indicates that adoption of Chinese models by U.S. and European companies is accelerating at a pace that outstrips even the most optimistic analyst forecasts.

The most emblematic case of this migration is Lindy, a San Francisco-based AI startup that, in June 2026, decided to move 100% of its AI traffic from Anthropic's Claude models to DeepSeek V4 (11). Lindy's CEO, Flo Crivello, said the decision would let the company save millions of dollars a year in inference costs (11). Crivello noted that the company's monthly AI bill had surpassed its payroll costs, and that the migration to DeepSeek not only drastically cut costs but in some scenarios improved model performance (11). Although the migration required significant engineering effort — the CEO said the migration work was 100 times greater than expected — the savings obtained more than justified the investment (11).

Lindy's case is not isolated. According to a UBS report, 60% of companies are cutting their AI spending and shifting toward open-source Chinese alternatives (12). Data from the aggregation platform OpenRouter, which routes developer queries to different models, shows a seismic shift in developer preferences. In 2024, open-source Chinese models accounted for less than 1.2% of weekly token consumption on OpenRouter (12). By late 2025, that figure had risen to roughly 30% at peak moments, with an annual average of 13% (12). In April 2026, traffic from Chinese providers reached 51% of all tokens processed on the platform (12). The tipping point came during the week of February 9–15, 2026, when Chinese models processed 4.12 trillion tokens on OpenRouter, surpassing the 2.94 trillion tokens processed by U.S. models (12). Since then, Chinese models have outpaced their U.S. counterparts in API call volume for six consecutive weeks (12). By July 2026, Chinese models accounted for as much as 60% of token traffic on OpenRouter, with DeepSeek topping the list of most popular models (12).

The composition of the most popular models on OpenRouter is particularly telling. The six most-used models on the platform are all Chinese: MiMo-V2-Pro (Xiaomi), Step 3.5 Flash (stepfun), DeepSeek V3.2, MiniMax M2.7, MiniMax M2.5, and GLM 5 Turbo (Z.ai) (12) (15). Anthropic, which held the top spot in March 2025, has fallen to seventh and eighth place (15). This decline in popularity among developers — the most valuable customer segment for any AI platform — is a leading indicator of the erosion of Western labs' competitive position. Developers, who are the earliest adopters of new technologies and the most price-sensitive, are voting with their feet, and their vote is unambiguously in favor of Chinese models.

Demand elasticity in response to falling prices is extraordinarily high in the AI API market. A company consuming millions of tokens a day can save hundreds of thousands of dollars a year by migrating from a Western model to a Chinese one. In an environment of tight technology budgets, where chief technology officers face increasing pressure to demonstrate AI's return on investment, such savings are hard to ignore. As a JP Morgan report has noted, the first signs of migration toward cheaper alternatives are already visible, and this trend is expected to accelerate as more companies become aware of the price gap (11).

Enterprise adoption of Chinese models is not limited to startups. Large technology corporations and Fortune 500 companies are actively evaluating the possibility of incorporating Chinese models into their technology stacks. According to UBS, DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, Zhipu AI's GLM, and MiniMax have been gaining ground in corporate evaluations from mid- to late June 2026 (13). These evaluations, many of which have already moved past the proof-of-concept stage into pilot implementation, pose an existential threat to OpenAI's and Anthropic's business model. If the largest and most profitable enterprise customers begin migrating their workloads to Chinese models, Western labs' recurring revenue will be eroded — and with it, the justification for their astronomical valuations.

Part III. Market-Share Dynamics and the Displacement of Established Leaders

3.1. Popularity Analysis on Aggregation Platforms: OpenRouter as the Market's Thermometer

Market-share dynamics in the generative AI sector have undergone a radical transformation over the past year and a half, and no indicator reflects this shift more clearly than data from OpenRouter, the aggregation platform that routes developer queries to different language models. OpenRouter, which functions as a neutral marketplace where developers can compare and select models from multiple providers, has become the most reliable thermometer of preferences within the developer community — the most valuable and strategic customer segment for any AI provider.

OpenRouter's data reveals a story of the unstoppable rise of Chinese models and an equally pronounced decline of their U.S. rivals. At the end of 2024, open-source Chinese models accounted for less than 1.2% of weekly token consumption on the platform (0). Throughout 2025, that figure grew explosively, reaching nearly 30% at peak moments, with an annual average of 13% (0) (3). The tipping point, however, came in the first quarter of 2026. During the week of February 9–15, 2026, Chinese models processed 4.12 trillion tokens on OpenRouter, surpassing the 2.94 trillion processed by U.S. models (0). Since then, Chinese models have consistently maintained their lead, outpacing their U.S. rivals in API call volume for consecutive weeks (0). By April 2026, Chinese providers' share reached 51% of all tokens processed on the platform (0) (3).

The phenomenon's acceleration has been such that, in the week of May 25–31, 2026, weekly call volume on OpenRouter reached 25 trillion tokens, with Chinese models already processing 18 trillion tokens weekly, versus 5.5 trillion for U.S. models (0). This ratio — more than three to one — is unprecedented in the platform's history. The most recent data, from June 2026, confirms the consolidation of this trend: Chinese models' share reached 61% of total token consumption on OpenRouter, while the U.S. share fell to 33% (1) (3). This figure represents a complete reversal of the situation a year earlier, when Google, OpenAI, and Anthropic models together accounted for 72% of the platform's traffic (1) (3).

The breakdown by individual model is even more revealing. In the week of May 25–31, 2026, DeepSeek-V4-Flash topped the ranking with 4.41 trillion tokens processed, followed by Tencent's Hunyuan model, Xiaomi's MiMo-V2-Pro, and DeepSeek V3.2 (0). The six most popular models on OpenRouter are all Chinese: MiMo-V2-Pro (Xiaomi), Step 3.5 Flash (stepfun), DeepSeek V3.2, MiniMax M2.7, MiniMax M2.5, and GLM 5 Turbo (Z.ai) (1). Anthropic, which held the top spot in March 2025, has fallen to seventh and eighth place — a decline reflecting the erosion of its competitive position in the market's most dynamic segment (1) (15). The ranking of the ten most-used models shows Chinese models concentrating 5.3 trillion tokens out of a total of 8.7 trillion — a 61% share (3). U.S. models, which once dominated this ranking, have been relegated to secondary positions, with the exception of a few Google and OpenAI models that retain a token presence.

The explanation for this phenomenon lies in several converging factors. First, the value-for-money ratio of Chinese models is incomparably superior. Developers, the earliest adopters of new technology and the most cost-sensitive, have made a rational calculation: if they can get performance similar to Western models at a fraction of the cost, the decision to migrate is economically irresistible. Second, Chinese labs' open-source strategy has created a developer ecosystem that contributes to model improvement, generating a positive feedback loop that accelerates innovation and reduces costs. Third, Chinese models have demonstrated particularly strong performance on coding tasks and agent-based workflows — precisely the use cases generating the largest volume of token consumption on the platform (0).

The preference developers have shown for Chinese models carries first-order strategic implications. Developers are not only the most intensive users of AI APIs, but also the influencers who shape the purchasing decisions of the companies they work for. As a growing cohort of developers becomes familiar with and trusts Chinese models, the likelihood that these models get adopted in enterprise settings rises significantly. OpenRouter, then, is not just a market thermometer, but also an engine of change that accelerates the diffusion of Chinese models throughout the global software-development ecosystem.

3.2. The Enterprise Migration Case Study: Testimonials and Purchasing Decisions at Tech Startups

The migration of individual developers to Chinese models has been followed, with a short lag, by the migration of entire companies. The most emblematic and well-documented case of this phenomenon is Lindy, a San Francisco-based AI startup specializing in AI agents for business-process automation. In June 2026, Lindy's founder and CEO, Flo Crivello, announced a decision that would have seemed unthinkable just a few months earlier: the company had moved 100% of its AI traffic from Anthropic's Claude models to DeepSeek V4 (2). The decision, according to Crivello, was not taken lightly, but followed a thorough evaluation process spanning several months that included performance testing, cost analysis, and security considerations (2).

The results of the migration were dramatic. Lindy managed to cut its inference costs by millions of dollars a year — savings that, according to Crivello, transformed the business unit's economics and allowed the company to reinvest resources into R&D (2) (11). What is most notable about the case is that the migration did not come at the cost of quality. On the contrary, in some specific scenarios, DeepSeek V4's performance surpassed Claude's, particularly on reasoning and code-generation tasks (2). Crivello said the company's monthly AI bill had surpassed its payroll costs, and that the migration to DeepSeek not only drastically cut costs but in some cases improved model performance (11). The CEO also warned that the migration work was 100 times greater than expected — a testament to the technical challenges of switching model providers in a production environment (2). Even so, the savings obtained more than justified the investment, and Crivello strongly recommended that other startups seriously consider migrating to Chinese models (2).

Lindy's case is not isolated but the spearhead of a broader trend affecting the entire enterprise spectrum. According to a UBS report, 60% of companies are cutting their AI spending and shifting toward open-source Chinese alternatives (12). This figure, which includes both startups and mid-sized companies, reflects a profound mindset shift among technology leaders: generative AI, until recently seen as a scarce, premium resource, is being reconceived as a commodity subject to the laws of price competition. In this new paradigm, loyalty to a specific provider is a luxury few companies can afford, and purchasing decisions increasingly rest on objective cost-and-performance criteria.

The most significant case, however, is still taking shape. In June 2026, Axios revealed that Microsoft is considering integrating DeepSeek V4 into its enterprise tool Copilot Cowork as a low-cost alternative to OpenAI's and Anthropic's models (2). The decision, expected to be announced in the coming weeks, would send shockwaves through the sector. Microsoft is OpenAI's largest customer and has invested tens of billions of dollars in the company. If Microsoft — OpenAI's most strategic ally — begins replacing its models with Chinese alternatives, the message to the market would be unambiguous: not even the hyperscaler ecosystem is immune to the pricing pressure of Chinese models (2). Microsoft is evaluating DeepSeek V4 for its low operating cost and its good reputation among developers (2). The company plans to offer a lower-cost version of Copilot Cowork in the coming weeks, and integrating DeepSeek would be a core part of that strategy (2). The move, if confirmed, would carry geopolitical implications, as it could put Microsoft at odds with the Trump administration, which has imposed export controls on AI technology (2).

Enterprise migration toward Chinese models is not limited to startups and big tech. According to UBS, DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, Zhipu AI's GLM, and MiniMax have been gaining ground in corporate evaluations from mid- to late June 2026 (13). These evaluations, many of which have already moved past the proof-of-concept stage into pilot implementation, pose an existential threat to OpenAI's and Anthropic's business model. If the largest and most profitable enterprise customers begin migrating their workloads to Chinese models, Western labs' recurring revenue will be eroded — and with it, the justification for their astronomical valuations.

3.3. The Erosion of Anthropic's Market Share and the Shift in Developer Preference

The market-share dynamic described in the previous sections has a name and a surname: Anthropic. The company founded by former OpenAI executives has been, paradoxically, the most affected by the rise of Chinese models, despite its focus on AI safety and interpretability positioning it as the ethical alternative to OpenAI. OpenRouter's data shows a decline in Anthropic's market share that can fairly be called catastrophic.

In March 2025, Anthropic held a 29.1% share on OpenRouter, making it the most popular model provider among developers (1). Twelve months later, in March 2026, that figure had collapsed to 13.3% — a drop of more than half in a single year (1). The trend has not stabilized but accelerated in subsequent months. In June 2026, Anthropic's OpenRouter share fell to 17.6% in the final week of the month, after having reached a floor of 17.7% in the third week (1). Even the launch of Fable 5, a new Anthropic model, only managed a temporary bump to 20.7% in the second week of June, before the share fell again (1). This pattern suggests that new model launches, which once drove waves of adoption, now have an increasingly limited and short-lived impact on the company's market share.

Anthropic's decline is particularly significant because the company had managed to position itself as the preferred alternative for programming and software-development tasks, where Claude had captured more than 50% in some sub-segments (4). Yet this niche, precisely because it is the most token-intensive, is also the most price-sensitive. Chinese models, especially DeepSeek and Qwen, have demonstrated competitive coding performance at a fraction of the cost, driving mass developer migration. As an Exponential View analysis has noted, developers are abandoning U.S. models in favor of Chinese alternatives because they prioritize cost-effectiveness and response speed for simple, high-frequency, standardized tasks (1) (3). In this segment, the Chinese models' advantage is overwhelming.

The financial impact of this market-share erosion is hard to overstate. Anthropic has reported $5 billion in revenue against $10 billion in spending on inference and training alone (15). With market share in free fall and a cost structure showing no signs of shrinking at the same pace, the company faces a scenario of increasingly compressed margins. If the current trend holds, Anthropic could be forced to cut prices to retain customers, which would further erode its margins and undermine the justification for its $965 billion valuation (6).

The shift in developer preference is not a phenomenon likely to reverse in the short term. Developers who have migrated to Chinese models and built their workflows around them have no incentive to return to more expensive Western models unless those offer a performance edge that justifies the extra cost. The available evidence suggests that such an edge does not exist for most use cases. As the NIST report has noted, DeepSeek V4 Pro sits roughly eight months behind the most advanced U.S. models on its aggregate capability measure (5). Yet on specific tasks like coding and reasoning, the gap is even smaller, and on cost efficiency, Chinese models beat U.S. models in five of seven categories (5). In this context, the decision to migrate is rational and hard to reverse.

The erosion of Anthropic's market share also has a snowball effect. As more developers migrate to Chinese models, Anthropic's user base shrinks, which in turn reduces revenue and the company's ability to invest in research and development. This dynamic, if not interrupted, could lead to a vicious cycle of decline that would be difficult to stop. Exponential View analysts have noted that the combined share of the three big U.S. models (Google, OpenAI, and Anthropic) on OpenRouter has fallen from 72% to 33% in a year, with no sign of the trend slowing (1) (3). If this trend holds, U.S. models could become a minority in their own market — a scenario that would have been considered unimaginable just two years ago.

Part IV. Financial Pressures and the Challenge to the IPO Narrative

4.1. OpenAI's and Anthropic's Cost Structure: Infrastructure Spending Versus Recurring Revenue

The financial sustainability of OpenAI and Anthropic, in the context of Chinese competition, can only be understood through a detailed analysis of their cost structure. Both companies operate a capital-intensive business model in which computing-infrastructure expenses — training and inference — constitute the largest line item and, at the same time, the hardest to reduce in the short term. The nature of this cost structure determines their vulnerability to a prolonged price war.

The available data on Anthropic's cost structure is particularly revealing. The company reported annualized recurring revenue of roughly $14 billion in February 2026, which soared to around $47 billion by late May of that same year, driven largely by adoption of Claude Code (2). However, that revenue growth has been accompanied by equally intense spending on compute, training, and infrastructure, placing the company in a position of heavy cash burn despite the rapid expansion of its billings (2). Anthropic's cash-burn rate is therefore high, and the company depends on periodic capital injections from its investors to maintain its pace of investment.

Anthropic's spending structure reflects the dynamic of a company that has prioritized growth and market share over near-term profitability. The company has stated that its goal is to reach breakeven around 2027, but that goal rests on revenue-growth assumptions that may not hold if Chinese competition keeps eroding its market share and forcing price cuts (3). According to projections reported by CNBC, Anthropic expected revenue of $10.9 billion in the second quarter of 2026 — up from $4.8 billion in the first quarter — with a positive operating result of around $559 million (3). That would still be a thin margin relative to the company's valuation, and it depends on revenue continuing to grow at the current pace, an assumption Chinese competition calls into question.

OpenAI, for its part, has kept a tighter lid on its cost structure, but analysts estimate its situation is not substantially different. The company has invested billions of dollars in computing infrastructure, including agreements with Microsoft for use of its data centers, and faces training costs exceeding $100 million per model (7). OpenAI's recurring revenue, which exceeded $20 billion at the end of 2025, has grown impressively, but operating costs have grown as well, especially on the inference side, where user query volume has multiplied computational demand (4). The company has acknowledged that its cash-burn rate is high and that it needs additional capital to scale its infrastructure (7). The decision to delay its IPO until 2027, despite investor pressure, suggests the company is not confident it can present financials that justify its valuation on public markets (6) (7).

The fundamental problem with OpenAI's and Anthropic's cost structure is its downward rigidity. Training and inference costs are largely determined by hardware and electricity — two factors that cannot easily be reduced. Chinese labs, by contrast, have developed model architectures and optimization techniques that drastically cut these costs. While a Western model can cost more than $100 million to train, an equivalent Chinese model may cost less than $6 million (7). This difference, which is structural rather than cyclical, places Western labs at a competitive disadvantage that will be difficult to escape.

4.2. The Profitability Dilemma: Margins in a Falling-Price Environment

The price war launched by Chinese labs has put OpenAI and Anthropic in a first-order strategic dilemma: cut prices to hold market share, eroding margins in the process, or hold prices and risk losing customers to cheaper alternatives. Both options are damaging to profitability and, by extension, to the justification for their valuations.

The current pricing landscape shows a spread of more than 200 times between the cheapest and most expensive model on the market (9). In this context, maintaining premium prices has become a high-risk strategy. Enterprise customers, especially those with high consumption volumes, are running rational cost calculations. If a company consuming 100 million output tokens a month (roughly 150,000 pages of text) can save $2.9 million a year by migrating from GPT-5.5 to DeepSeek V4 Pro, the decision to migrate is economically irresistible (9). Even if the migration requires significant development investment, the return on investment is reached within months.

The profitability dilemma is compounded by the cost structure described in the previous section. If OpenAI and Anthropic cut prices to match their Chinese competitors, their margins would compress to levels that make profitability unviable. A simple analysis illustrates the problem: if GPT-5.5 cut its price from $30 per million output tokens to $0.87 (DeepSeek V4 Pro's price), its revenue per million tokens would fall by 97.1%. To maintain the same level of revenue, the company would need to multiply its token volume by 34 times — growth that, while not impossible, is highly improbable in the short term. The alternative — holding prices and losing customers — is equally damaging, since the loss of volume reduces total revenue while fixed infrastructure costs do not fall proportionally.

UBS analysts have warned that the price gap between China and the United States will inject volatility into the sector and could lead investors to rethink massive infrastructure spending (16). The Swiss bank has noted that widespread adoption of cheaper models could "curb some of that spending" — an observation implying that infrastructure providers' revenue, and that of the companies that depend on them, could be affected (16). This warning is particularly relevant for OpenAI and Anthropic, which depend on hyperscalers for both infrastructure and financing.

The profitability dilemma also has a time dimension. OpenAI's and Anthropic's public listings are slated for 2026–2027, a window in which the price war is likely to intensify. Public investors, unlike private ones, are less tolerant of losses and more demanding on profitability. If OpenAI and Anthropic cannot show a clear path to profitability by the time of their IPOs, their valuations are likely to face a significant adjustment. SpaceX's volatile post-debut performance — shares that rose above $225 before falling back to around $153 within days — has cooled enthusiasm for tech IPOs and made it harder for OpenAI to justify a $1 trillion valuation at listing (10).

4.3. Consequences for Pre-IPO Valuation: The Scarcity Premium in Doubt

OpenAI's and Anthropic's private-market valuations have been driven, to a large extent, by a scarcity premium: the belief that frontier AI models are a scarce resource, difficult to replicate and protected by insurmountable barriers to entry. The rise of Chinese models has cast doubt on that scarcity premium, showing that credible alternatives exist at a fraction of the cost. This has profound implications for both companies' valuations on the eve of their IPOs.

The scarcity premium shows up in OpenAI's and Anthropic's valuation multiples. OpenAI, with recurring revenue of roughly $20 billion, is valued at $852 billion — a multiple of 42.6 times (2). Anthropic, with recurring revenue of $14 billion, is valued at $965 billion — a multiple of 68.9 times (3). These multiples are extraordinarily high compared with the standards of the publicly traded tech sector, where high-growth software companies typically trade between 10 and 20 times recurring revenue. The scarcity premium justifies this differential: private investors are paying a premium for exclusive access to companies that, per the established narrative, are the only ones capable of developing frontier models.

That narrative has been seriously undermined by NIST reports showing that DeepSeek V4 Pro sits roughly eight months behind the most advanced U.S. models on its aggregate capability measure (5). The eight-month gap, which might seem significant given AI's rapid pace of change, shrinks considerably once one considers it has been closing at an accelerating rate: in 2024, the estimated difference was roughly two years; in 2025, fourteen months; and in April 2026, eight months (5) (8). If this trend continues, full technological parity could be reached around 2027 — precisely when OpenAI and Anthropic expect to complete their IPOs. In that scenario, the scarcity premium would evaporate entirely, and OpenAI's and Anthropic's valuations would converge toward the multiples of conventional technology companies.

The NIST assessment has sparked intense debate among researchers and policymakers. Some analysts have questioned whether the "eight-month lag" metric adequately captures the reality of the Chinese ecosystem, noting the assessment does not incorporate all of DeepSeek V4 Pro's capabilities, such as ultra-long-context performance (8). Others have noted that the eight-month gap, far from indicating inferiority, represents a remarkable achievement given that Chinese labs developed their models under hardware-access restrictions imposed by U.S. export controls (8). The US-China Economic and Security Review Commission's own report agrees that Chinese labs "have narrowed performance gaps with leading Western models" by relying on an open ecosystem that reduces their dependence on the most advanced hardware (29).

4.4. Defensive Strategies: Aggressive Discounts, Lite Versions, and the Cost of Retaining Customers

Facing competitive pressure from Chinese models, OpenAI and Anthropic have begun deploying defensive strategies aimed at retaining customers and protecting margins. These strategies include aggressive discounts, the launch of lite versions of their models, and the strengthening of developer ecosystems. Yet each of these strategies carries a cost that erodes profitability and weakens the case for their valuations.

The most immediate strategy has been price cuts. Anthropic has launched "lite" versions of its Claude Sonnet 4.6 model at significantly reduced prices (9). OpenAI, for its part, has introduced GPT-5.5 Mini, a lighter and cheaper version of its flagship model (9). These lite versions, however, are not fully comparable to Chinese models in terms of performance, and their price remains higher than DeepSeek's or Qwen's. Moreover, market segmentation — premium versions for customers willing to pay more, lite versions for price-sensitive customers — erodes average revenue per customer, shrinking overall margin.

A second strategy has been strengthening developer ecosystems. OpenAI has invested in its application platform and partner network, seeking to build a network effect that makes migration to other models harder. Anthropic has done the same, emphasizing its focus on safety and interpretability as differentiators that justify a premium price. Yet these strategies are costly to implement and take time to bear fruit. In the short term, developer decisions are dominated by cost considerations, not ecosystem loyalty.

The most ambitious — and also the riskiest — strategy is investment in innovation to re-establish the technology gap with Chinese models. OpenAI and Anthropic have said they are working on new models that, they claim, will restore their performance leadership. However, this strategy faces two obstacles. First, innovation investment is costly and requires capital both companies are raising through funding rounds that dilute existing shareholders. Second, recent history suggests Chinese labs are closing the gap at an accelerating pace, so any technological edge gained may be short-lived. The NIST report, which places the gap at eight months, suggests that even if OpenAI and Anthropic manage to launch new models with significant advantages, Chinese labs could match them within months (5).

The cost of retaining customers in a price-war environment is high. Aggressive discounts cut revenue without cutting costs, compressing margins. Lite versions cannibalize premium versions, reducing average revenue per customer. Investments in innovation and ecosystems require capital that, at best, delays the breakeven point and, at worst, dilutes shareholders. Add to this the loss of market share to Chinese models, and the financial outlook for OpenAI and Anthropic is far from rosy.

Lindy's case — the startup that migrated completely from Claude to DeepSeek — illustrates the challenge Western labs face (2). Lindy not only saved millions of dollars, but in some scenarios improved its application's performance. The migration, though costly in engineering effort, delivered a positive return on investment within months. This case, replicable across hundreds or thousands of companies, suggests that retaining customers in a price-war environment is an uphill battle for Western labs. The only way to retain customers is to offer value that justifies the extra cost, and the available evidence suggests Chinese models offer comparable value at a fraction of the cost.

In this context, OpenAI's decision to delay its IPO until 2027 can be read as a defensive strategy aimed at buying time to restore its competitive edge before facing public-market scrutiny (6) (7). But time is working against OpenAI and Anthropic: with every passing month, Chinese labs close the technology gap and consolidate their market position. If this trend holds, delaying the IPO will not be enough to avoid a valuation adjustment for both companies.

Part V. Macroeconomic Context: Global AI Spending and the Infrastructure Paradox

5.1. Gartner's Projections and the Trillion-Dollar Bet on AI Infrastructure

The phenomenon of Chinese AI competition cannot be fully understood without situating it in the macroeconomic context of global AI spending, which has grown explosively over the past three years. Projections from major analysis and consulting firms paint a picture of unprecedented investment, but also raise questions about the sustainability of that spending in an environment of free-falling token prices.

Gartner, the technology analysis and consulting firm, published its forecasts for global AI spending in May 2026. According to its report, global AI spending will reach $2.59 trillion in 2026, representing year-over-year growth of 47% (1). This figure, which exceeds the firm's own prior forecasts, reflects the accelerating adoption of AI across every sector of the economy. AI infrastructure — servers, data centers, semiconductors, and storage systems — accounts for more than 45% of the market, an estimated $1.43 trillion in 2026 (1). This segment is by far the most dynamic and has experienced the fastest growth, driven by demand from AI labs and hyperscalers.

The geographic distribution of AI infrastructure spending shows an uneven picture. The United States remains the largest investor, with roughly 40% of global spending, followed by China with roughly 25% and Europe with 18% (1). However, China's spending growth rate is significantly higher than that of the United States, reflecting the strategic priority the Chinese government has assigned to artificial intelligence. Gartner's report notes that, if the current trend holds, China could overtake the United States in AI infrastructure investment around 2028 (1).

Gartner's projections rest on the assumption that demand for tokens and computing capacity will keep growing at an accelerating pace, driven by AI adoption across every sector. That assumption, however, faces an unexpected challenge: the price war launched by Chinese labs. If token prices fall dramatically, infrastructure providers' revenue could be affected, and infrastructure spending might not grow at the projected pace. The paradox is that falling prices, which should stimulate demand, could reduce infrastructure providers' revenue and, therefore, their capacity to invest in new capacity.

5.2. Hyperscaler Capital Spending: The $600 Billion Bet

AI infrastructure spending is dominated by the hyperscalers: Amazon, Alphabet (Google), Microsoft, Meta, and Oracle. These five companies, which operate the world's largest data centers, have made an unprecedented bet on artificial intelligence, investing amounts that exceed the gross domestic product of many countries. According to a report by Wedbush Securities and S&P Global published in January 2026, the combined capital expenditure (CAPEX) of the five hyperscalers will exceed $600 billion in 2026 — a 36% increase over the already-record 2025 figure (2) (3). Roughly 75% of this spending, or $450 billion, will go toward AI infrastructure, including GPU purchases, data-center construction, and the development of interconnection networks (2).

The distribution of this spending among the hyperscalers reflects their differing AI strategies. Microsoft, which has been OpenAI's strategic partner and has invested tens of billions of dollars in the company, is channeling a significant share of its CAPEX toward the infrastructure supporting OpenAI's models. Amazon, which has invested in Anthropic and has developed its own AI chips (Trainium and Inferentia), is building specialized data centers for language-model inference. Google, which has bet on its own model ecosystem (Gemini) and its own chip development (TPU), is investing in infrastructure to support both its internal models and third-party ones. Meta, which has bet on open source with its Llama family, is building the world's largest AI infrastructure to support its models and social platforms (2). Oracle, finally, is investing in a global network of AI-specialized data centers, competing directly with the established hyperscalers.

The hyperscalers' AI bet is not philanthropic; it reflects the expectation that artificial intelligence will generate substantial returns in the future. Language models, per this logic, will become the platform on which future applications are built, and the hyperscalers that control the infrastructure will capture a significant share of the value created. This logic, however, runs into the paradox described in the previous section: if token prices fall dramatically, infrastructure providers' revenue may not be enough to justify the investment made.

5.3. The Infrastructure Paradox: Investment Without Return If Token Prices Collapse?

The fundamental paradox of the AI market is that falling token prices, which should stimulate demand and accelerate adoption, could have the opposite effect on infrastructure investment. Hyperscalers are investing hundreds of billions of dollars in building data centers and acquiring GPUs, based on the assumption that demand for computing capacity will grow exponentially. However, if token prices fall dramatically due to Chinese competition, infrastructure providers' revenue may not grow at the expected pace, and the return on investment could fall short of projections.

Analysis of token-demand elasticity is crucial to understanding this paradox. Token demand is price-elastic: when prices fall, consumption rises. DeepSeek has shown that its price cuts have stimulated significant growth in token consumption (0) (12). However, elasticity is not infinite. There is a limit to the volume of tokens users can consume, determined by available use cases and applications' capacity to integrate AI. If token prices fall to very low levels, the increase in consumption may not be enough to offset the drop in per-unit revenue.

UBS analysts have warned that the price gap between China and the United States will inject volatility into the sector and could lead investors to rethink massive infrastructure spending (16). The bank has noted that widespread adoption of cheaper models could curb some of that spending — an observation implying that the revenue of infrastructure providers, and the companies that depend on them, could be affected (16). This warning is particularly relevant for the hyperscalers, which have committed enormous amounts of capital to AI infrastructure.

The infrastructure paradox also has a time dimension. Hyperscalers are investing now, but the return on investment will materialize over a horizon of several years. If the price war intensifies in coming years, revenue generated by AI infrastructure could fall short of expectations, and hyperscalers could face underutilized capacity and reduced profitability. This situation, which would be especially damaging for the hyperscalers that have bet hardest on AI, could have systemic repercussions across the technology sector.

5.4. Lessons from the 1990s Telecom Bubble: Similarities and Differences

The 20minutos.es article that serves as the starting point for this analysis draws a parallel between the current state of the AI market and the telecommunications bubble of the 1990s (10). This parallel, though imperfect, offers valuable lessons for understanding the risks facing the sector.

The telecom bubble of the 1990s was characterized by massive investment in network infrastructure (fiber optics, undersea cable, switching equipment) based on the expectation that bandwidth demand would grow exponentially. Companies such as WorldCom, Global Crossing, and Level 3 invested hundreds of billions of dollars building fiber-optic networks, convinced that data-communication demand would justify the investment. When the bubble burst in 2000–2001, many of these companies went bankrupt or were absorbed, and the infrastructure built sat underused for years. The irony is that bandwidth demand eventually grew enough to justify the investment, but did so with a lag of several years, and the earliest investors lost all their capital.

The current state of the AI market bears unsettling similarities to the telecom bubble. As in the 1990s, we are seeing massive infrastructure investment — data centers, GPUs, interconnection networks — based on the expectation that demand for computing capacity will grow exponentially. As in the 1990s, investors are paying steep premiums for companies that promise to capture future growth but are not yet profitable. And as in the 1990s, there is a risk that demand will not grow at the expected pace, or that competition will erode providers' margins.

There are, however, significant differences that qualify the parallel. First, generative AI has already demonstrated real use cases and tangible value for businesses, unlike many telecom applications of the 1990s, which were largely speculative. Second, the hyperscalers investing in AI infrastructure are enormously profitable companies with solid balance sheets, unlike the highly leveraged telecom companies of the 1990s. Third, the Chinese government is actively backing AI development, adding a geopolitical dimension that did not exist in the 1990s.

The most important lesson from the telecom bubble is that even technologies that eventually transform the economy can generate massive losses for early investors. Fiber optics and broadband internet eventually revolutionized the global economy, but investors who bought WorldCom and Global Crossing shares at their peak lost everything. Similarly, AI may transform the global economy, but investors who buy OpenAI and Anthropic shares at their current valuations could suffer significant losses if Chinese competition erodes their margins and market share. Caution, therefore, is advisable for institutional investors considering exposure to this sector.

Part VI. Geopolitical and Regulatory Dimensions

6.1. China's Open-AI Strategy as an Instrument of Industrial Policy

China's emergence as a competitive force in the global AI market is not the result of historical chance or the mere accumulation of scattered technical capabilities. It reflects, instead, a deliberate industrial-policy strategy designed and implemented by the Chinese government over the past decade, combining public research investment, talent development, ecosystem openness, and, crucially, a philosophy of technology diffusion radically different from that of Western labs. This strategy, which the US-China Economic and Security Review Commission (USCC) has called an "open-AI strategy," was analyzed in depth in a March 2026 report titled "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance" (29).

The USCC report, which reports to the U.S. Congress, notes that Chinese labs "have narrowed performance gaps with leading Western models" by relying on an open ecosystem that reduces their dependence on the most advanced hardware (29). The commission warns that China's open-AI strategy is not philanthropic but an instrument of industrial policy designed to reinforce China's technological dominance, capture the global developer base, and set technical standards that favor its companies. This strategy, which combines openness with state control of computing and data infrastructure, represents a systemic challenge to Western technological leadership that goes beyond the purely commercial sphere.

The mechanisms of China's strategy operate on multiple levels. At the technological level, the Chinese government has actively promoted the development of open-source models as an alternative to OpenAI's and Anthropic's closed models. The "develop once, deploy anywhere" philosophy adopted by Alibaba with its Qwen family (21) (22) reflects a conception of AI as a global public good, but also as a vehicle for spreading Chinese technical standards. At the economic level, the Chinese government has provided direct and indirect funding to AI labs through subsidies, soft loans, and public contracts. At the regulatory level, China has established an AI regulatory framework that, while restrictive in some respects, provides companies with certainty and lets them operate with a degree of predictability not always found in other markets.

The impact of this strategy has shown up in the rapid adoption of Chinese models in the global market, as OpenRouter's data demonstrates (0) (12). The USCC has noted that China's open-AI strategy is working: developers worldwide are adopting Chinese models, and these models are setting de facto standards in key market segments. The commission has warned that, if this trend continues, the United States could lose its technological leadership in AI around 2030, with far-reaching strategic consequences for its national security and economic competitiveness (29).

The U.S. response to this strategy has so far been fragmented and reactive. Semiconductor export controls, examined in the next section, have been the main tool used by the U.S. administration to slow China's advance. However, these controls have had unintended effects and have failed to stop the progress of Chinese labs. OpenAI's proposal to give up a 5% stake to the U.S. government (0), examined in section 6.3, represents an attempt to involve the state in defending U.S. technological leadership, but raises questions about the wisdom of direct government intervention in the market.

6.2. Semiconductor Export Controls: Effectiveness and Unintended Effects

The main tool of U.S. technology policy for slowing China's advance in artificial intelligence has been the export control of advanced semiconductors. Since 2022, the U.S. administration has imposed progressive restrictions on exporting high-performance chips, semiconductor manufacturing equipment, and related technology to China. These restrictions, which affect companies such as Nvidia, AMD, and Dutch and Japanese lithography-equipment manufacturers, were designed to prevent Chinese labs from accessing the hardware needed to train and run frontier AI models.

The effectiveness of these controls has been debated among analysts. On one hand, it is undeniable that the restrictions have made it harder for China to access the most advanced chips, such as Nvidia's H100 GPUs and their successors. Chinese labs have had to rely on older-generation chips, chips manufactured in China with older process technology, or software optimization solutions that reduce hardware demand (26). On the other hand, the controls have failed to stop Chinese labs' progress. DeepSeek trained its V3 model at a cost of just $5.5 million using H800 GPUs, which are subject to export restrictions but remain accessible through unofficial channels or existing stockpiles (10). Chinese labs have developed architectural innovations, such as mixture-of-experts (MoE) and hybrid attention, that reduce computational demand and allow models to run on less advanced hardware (26).

The most significant, and unintended, effect of export controls has been the acceleration of Chinese innovation in computational efficiency. Far from slowing China's advance, the restrictions appear to have spurred a wave of efficiency innovation that is redefining industry standards. Chinese labs have developed memory-compression, speculative-decoding, and inference-optimization techniques that drastically cut the computational cost of their models (28). These innovations, adopted by the global developer community, are making Western labs' hardware advantage increasingly irrelevant. As one Wedbush analyst has noted, Chinese labs have shattered the assumption that frontier AI training costs necessarily exceed $100 million (10).

The limited effectiveness of export controls has led some U.S. policymakers to consider additional measures, such as restricting Chinese labs' access to open-source models hosted on U.S. platforms (Hugging Face, GitHub, etc.). Such measures, however, raise legal and technical challenges, and could have the unintended effect of fragmenting the global AI ecosystem into separate technology blocs. The USCC has warned that a strategy relying solely on export controls is insufficient, and has recommended a broader strategy including public research investment, talent development, and international cooperation (29).

6.3. OpenAI's Proposed Equity Stake to the U.S. Government: Techno-Nationalism in Action

Amid competitive pressure from China and the imminence of its own public listing, OpenAI has made a proposal that has surprised analysts and opened a debate about the state's role in the technology industry. According to reports published in July 2026, OpenAI has proposed giving up a 5% equity stake to the U.S. government, in a move interpreted as an attempt to secure political and regulatory support for its IPO, and to bring the state into the defense of its competitive position against China (0). The stake, valued at roughly $42.6 billion at the price of the latest funding round, would make the U.S. government one of the company's largest shareholders.

OpenAI's proposal has been met with skepticism by some analysts and enthusiasm by others. Its supporters argue that U.S. government participation in OpenAI would provide a strategic anchor guaranteeing national security and U.S. technological competitiveness in AI. The move, per this logic, would send a clear signal to the market and to foreign competitors that the United States is willing to mobilize all its resources to defend its AI leadership. Critics, on the other hand, argue that government participation in a private company poses risks of political interference, market distortion, and conflicts of interest. They also note that the move could deter other investors, especially international ones, who might fear excessive government influence over the company's management.

OpenAI's proposal is symptomatic of a broader trend toward techno-nationalism that has gained strength in recent years. The United States, China, and the European Union are competing for leadership in strategic technologies such as artificial intelligence, semiconductors, and quantum computing. In this context, technology companies are no longer mere economic actors but strategic assets that governments seek to protect and foster. U.S. government participation in OpenAI, if it materializes, would represent the most explicit manifestation of this trend, and could set a precedent for other technology companies facing Chinese competition.

OpenAI's proposal also has implications for its IPO. Government participation, if it materializes, would add a layer of regulatory complexity to the listing. The SEC, already reviewing OpenAI's S-1 filing, would need to assess the implications of government participation in the company's capital structure. Investors, for their part, would need to weigh the risks and benefits of having the U.S. government as a partner. It is possible that the proposal, far from facilitating the IPO, could complicate and further delay it, adding another layer of uncertainty to an already uncertain process.

Part VII. Future Scenarios and Strategic Recommendations

7.1. Equilibrium Scenario: A Two-Speed Market

The first possible scenario, which we will call the equilibrium scenario, envisions a market stabilizing around two distinct segments: a premium segment dominated by Western labs, and a commoditized segment dominated by Chinese labs. This scenario rests on the premise that AI models are not a homogeneous good, but present qualitative differences that justify market segmentation.

In the premium segment, OpenAI and Anthropic would retain their leadership in frontier models — those setting the state of the art in reasoning, creativity, multimodality, and safety. These models, which require massive R&D investment and offer qualitative advantages that are difficult to replicate, could continue to justify high prices for enterprise customers willing to pay for the best. Their advantage would lie not only in performance, but in the trust they inspire in regulated environments, integration with enterprise software ecosystems, and regulatory-compliance assurance. In this segment, margins would remain high, though lower than investors had anticipated before China's disruption.

In the commoditized segment, open-source Chinese models would dominate the mass-application market, where cost-effectiveness is the primary selection criterion. Classification tasks, information extraction, text summarization, standardized response generation, and automated workflows would be handled by Chinese models offering sufficient performance at a fraction of the cost. In this segment, margins would be thin, but transaction volume would be high enough to offset low per-unit profitability. Chinese labs, thanks to their structural efficiency, could sustain low prices and still earn reasonable profits.

This equilibrium scenario would require Western labs to accept a reduction in market share and margins, and investors to adjust their profitability expectations. Under this scenario, OpenAI's and Anthropic's valuations would converge toward the multiples of conventional technology companies, likely settling between 10 and 20 times recurring revenue — a significant reduction from their current valuations. This adjustment could be painful for investors who bought shares in the most recent funding rounds, but would provide a sounder foundation for long-term investment.

7.2. Disruption Scenario: AI Commoditization and the End of Premium Models

The second scenario, which we will call the disruption scenario, envisions a situation in which Western models' qualitative edge erodes to the point of disappearing, and artificial intelligence becomes an indistinguishable commodity where price is the only selection criterion. This scenario — the one investors fear and Chinese labs want — would result from full technological convergence between Chinese and Western models.

Indicators of this scenario are already visible in NIST reports, which place the technology gap at eight months and project convergence around 2027 (5). If the current trend holds, it is plausible that by 2027 Chinese models will match or even surpass Western models on most performance metrics. At that point, Western labs would lose their main differentiator, and their ability to charge premium prices would disappear.

In this scenario, OpenAI and Anthropic would face a dilemma: drastically cut prices to stay competitive, eroding their margins to unsustainable levels, or abandon the mass market and focus on high-value niches where customers are willing to pay for specific features (safety, regulatory compliance, ecosystem integration). The first option would lead to a price war that Western labs, with their higher cost structure, would likely lose. The second would mean a drastic reduction in their potential market and revenue.

The consequences of this scenario for OpenAI's and Anthropic's valuations would be catastrophic. In a commoditized AI market, valuation multiples would converge toward those of conventional technology-services companies — that is, between 1 and 5 times revenue. A valuation of this kind would cut OpenAI's value to under $100 billion, a fraction of its current $852 billion valuation. In this scenario, both companies' IPOs would face extreme difficulties, and investors who bet on them would suffer significant losses.

7.3. Fragmentation Scenario: Closed Regional Technology Blocs

The third scenario, which we will call the fragmentation scenario, envisions a situation in which the global AI market splits into regional technology blocs, each with its own standards, models, and ecosystems. This scenario, driven by geopolitical and regulatory considerations, would result from escalating tensions between the United States and China, and growing government intervention in the AI market.

In this scenario, the United States and its allies (Europe, Japan, South Korea, Australia) would restrict access to Chinese models, imposing export controls, sanctions, and regulatory barriers that would hinder their adoption in these markets. China, for its part, would do the same, creating an autonomous technology ecosystem based on its own models, chips, and platforms. The result would be a fragmentation of the global market into two separate ecosystems, with diverging technical standards, different prices, and unequal capabilities.

Fragmentation would have mixed consequences for OpenAI and Anthropic. On one hand, excluding Chinese models from the U.S. and European markets would eliminate price competition in their home market, allowing them to maintain high prices and healthy margins. On the other hand, fragmentation would shrink Western models' potential market, excluding them from China and from emerging markets aligned with China. Moreover, fragmentation would raise development costs, since Western labs would need to adapt their models to different regulatory standards and compliance requirements.

Fragmentation would also have implications for technological innovation. In a scenario of closed technology blocs, international collaboration in AI research would be severely limited, and technological advances would be concentrated independently within each bloc. History suggests that technological fragmentation tends to slow the pace of innovation and duplicate research costs, ultimately harming all players in the long run.

7.4. Recommendations for Institutional Investors and Asset Managers

Given the scenarios analyzed, institutional investors and asset managers considering exposure to the AI sector should adopt a cautious, diversified approach. Uncertainty around the outcome of the competition between Western and Chinese labs is high, and the risks of a valuation correction are significant.

First, investors should reassess OpenAI's and Anthropic's valuations in light of Chinese competition. The 40x and 70x recurring-revenue multiples applied to these companies are hard to justify in a scenario of price war and market-share erosion. Even under the equilibrium scenario, these companies' valuations should adjust downward toward the multiples of conventional technology companies. Investors considering participating in OpenAI's and Anthropic's IPOs should be aware there is significant risk that the listing price could fall below that of private rounds, and that early public investors could suffer losses.

Second, investors should consider exposure to Chinese labs, which are gaining market share at an accelerating pace. Companies such as DeepSeek, Alibaba (Qwen), Zhipu AI (Z.ai), and Moonshot AI (Kimi) are positioning themselves as leaders in the commoditized AI segment, and their revenue and market-share growth is likely to continue. Investors with access to Chinese markets may consider including these companies in their portfolios. However, they should be aware of the regulatory and geopolitical risks associated with investing in China, as well as these companies' lower financial transparency compared with Western firms.

Third, investors should consider exposure to AI infrastructure, which is experiencing explosive growth. Hyperscalers (Amazon, Google, Microsoft, Meta, Oracle) are investing hundreds of billions of dollars in data centers and GPUs, and semiconductor makers (Nvidia, AMD, Broadcom) are seeing demand surge. However, investors should be wary of the infrastructure paradox: if token prices fall dramatically, infrastructure providers' revenue may not grow at the expected pace, and returns could fall short of projections. Diversifying across different segments of the AI value chain (models, infrastructure, applications) is advisable to mitigate this risk.

Fourth, investors should pay attention to the sector's geopolitical and regulatory dimensions. Escalating tensions between the United States and China could lead to market fragmentation that would negatively affect the profitability of companies exposed on both sides. Investors should closely monitor regulatory developments, including export controls, sanctions, and foreign-investment policies.

Finally, investors should adopt a long-term perspective. Artificial intelligence is a transformative technology likely to generate significant economic value over coming decades. However, the history of disruptive technologies suggests that value creation does not always benefit early investors. The 1990s telecom bubble is a reminder that even technologies that eventually transform the economy can generate massive losses for early investors. Patience, discipline, and diversification are the best tools for navigating this uncertain environment.

7.5. Recommendations for OpenAI's and Anthropic's Executive Teams

OpenAI's and Anthropic's executive teams face a first-order strategic challenge: preserving their companies' value in an environment of intense competition and price war. The following recommendations, based on the analysis presented and lessons from business history, may help them navigate this uncertain environment.

First, OpenAI and Anthropic should accelerate efforts to cut operating costs. Training and inference efficiency is the most important competitive advantage held by Chinese labs, and Western labs must match or exceed that efficiency if they want to retain their ability to compete on price. This means investing in more efficient model architectures, compression and optimization techniques, and more cost-effective hardware. Adopting mixture-of-experts (MoE) models, hybrid attention, and speculative decoding, similar to what Chinese labs have developed, should be a priority (26) (28).

Second, OpenAI and Anthropic should differentiate their models through features Chinese labs cannot easily replicate. Safety, interpretability, regulatory compliance, integration with enterprise ecosystems, and privacy assurance are attributes enterprise customers value and are willing to pay a premium for. Investment in AI-safety research, a priority for Anthropic since its founding, could become a key differentiator in the premium market segment. OpenAI, for its part, should leverage its first-mover advantage in the consumer market and strengthen its application ecosystem and partner network.

Third, OpenAI and Anthropic should reconsider their pricing strategy in the commoditized segment. If they cannot compete on price with Chinese models, they should withdraw from that segment and focus on the premium segment, where their competitive edge is stronger. This would mean a smaller potential market, but also better margins and profitability. The segmentation into premium and lite versions, which both companies have already begun, should be deepened and refined.

Fourth, OpenAI and Anthropic should accelerate their IPO plans, but with a realistic pricing strategy. OpenAI's decision to delay its IPO until 2027 may be prudent if used to improve the company's efficiency and profitability. However, the delay should not become an excuse to procrastinate on difficult decisions. Public investors are less tolerant than private ones of losses and inflated valuations, and an IPO priced to reflect market reality would be preferable to a failed listing that would damage the company's credibility.

Finally, OpenAI and Anthropic must maintain constant vigilance over geopolitical and regulatory developments. Escalating tensions between the United States and China could significantly affect their business, both on the supply side (restricted access to hardware and software) and the demand side (restrictions on exporting their models to certain markets). The ability to anticipate and adapt to these regulatory changes will be a critical factor in their long-term success. OpenAI's proposal to give up a 5% stake to the U.S. government can be seen as an attempt to secure political support and mitigate regulatory risk, but it should be carefully weighed against its implications for the company's independence and investor confidence.

Conclusions

The analysis developed throughout this article has shown that the global generative AI market is undergoing a structural transformation driven by the rise of Chinese labs as credible competitors to Western leaders. This transformation, which has accelerated in the first half of 2026, carries profound implications for the valuation of the companies involved, competitive dynamics, market structure, and investor strategy.

The article's central thesis — that China's AI strategy, built on open source, training efficiency, and aggressive pricing, has struck OpenAI and Anthropic at their weakest point at precisely the worst possible moment, the run-up to their IPOs — has been confirmed by the available empirical evidence. Chinese models have shown performance comparable to Western models with a lag of just eight months, but at a fraction of the cost. This cost advantage, which translates into API prices 10 to 200 times lower, is eroding Western labs' market share and forcing a recalibration of their pricing strategies and cost structures.

OpenRouter data, showing a seismic shift in developer preference toward Chinese models, is the most eloquent indicator of this transformation. The six most popular models on the platform today are Chinese, and U.S. models' share has fallen from 72% to 33% in a year. Enterprise migration, exemplified by the Lindy case, is underway, and large technology corporations are actively evaluating whether to incorporate Chinese models into their technology stacks. If this trend holds, Western labs will face accelerating erosion of their revenue and their ability to justify their astronomical valuations.

OpenAI's ($852 billion) and Anthropic's ($965 billion) valuations rest on the premise that they will maintain a near-monopolistic hold on the high-performance model market, with the ability to set premium prices indefinitely. That premise has been seriously called into question by Chinese competition. Even under the most favorable equilibrium scenario, these companies' valuations should adjust downward toward the multiples of conventional technology companies. Under the disruption scenario — the one Chinese labs want — the value of these companies could shrink to a fraction of their current valuation.

The macroeconomic context of global AI spending, with projections of $2.59 trillion in 2026, adds a further layer of complexity to the analysis. The infrastructure paradox — massive investment in AI infrastructure that may not generate the expected return if token prices collapse — is a systemic risk affecting not only OpenAI and Anthropic, but also the hyperscalers and semiconductor makers. The lesson of the 1990s telecom bubble is that even technologies that eventually transform the economy can generate massive losses for early investors.

The geopolitical and regulatory dimensions add further uncertainty to the picture. China's open-AI strategy, analyzed by the US-China Economic and Security Review Commission, is an instrument of industrial policy designed to reinforce China's technological dominance and capture the global developer base. Semiconductor export controls, the main tool the United States has used to slow China's advance, have failed to stop the progress of Chinese labs and have spurred a wave of efficiency innovation that is redefining industry standards. OpenAI's proposal to give up a 5% stake to the U.S. government is a manifestation of the techno-nationalism gaining strength on both sides of the Pacific.

The future scenarios analyzed — equilibrium, disruption, and fragmentation — offer a framework for strategic decision-making. In the equilibrium scenario, the market stabilizes around two distinct segments: premium for Western labs and commoditized for Chinese ones. In the disruption scenario, technological convergence erases Western models' qualitative advantage and AI becomes a commodity. In the fragmentation scenario, the global market splits into regional technology blocs, with diverging standards and separate ecosystems.

Recommendations for institutional investors and asset managers include reassessing OpenAI's and Anthropic's valuations, considering exposure to Chinese labs, exercising caution in AI infrastructure investment, and monitoring geopolitical and regulatory developments. Diversification and a long-term perspective are the best tools for navigating this uncertain environment. Recommendations for OpenAI's and Anthropic's executive teams include accelerating cost-cutting efforts, differentiating their models through non-replicable features, segmenting the market, and adopting a realistic pricing strategy ahead of their IPOs.

In short, the generative AI market finds itself at a critical inflection point. Chinese competition has shown that artificial intelligence is not a scarce resource protected by insurmountable barriers to entry, but a playing field where efficiency, openness, and pricing strategy matter as much as technical capability. OpenAI and Anthropic, the Western champions that have dominated the AI narrative in recent years, face an existential challenge that will test not only their technical capability, but also their business model and financial strategy. The outcome of this confrontation will determine not only the stock-market value of these companies, but also the future direction of AI development, access to its benefits, and the global distribution of technological power.

References

1. 20minutos.es (2026). "China's AI Has Found OpenAI's and Anthropic's Weak Spot." July 5, 2026.

2. Exponential View (2026). "Chinese AI models overtake US in OpenRouter token consumption." June 2026.

3. PitchBook (2026). "The $600 Billion Bet: Hyperscalers Turn to Massive Debt to Fuel the AI Revolution." January 28, 2026.

4. Economic Times (2026). "Why IPO valuations, SpaceX's included, may not conclusively capture the current state of AI play." June 9, 2026.

5. The Register (2026). "Cheap Chinese models are overtaking Anthropic." March 28, 2026.

6. National Institute of Standards and Technology (NIST) - CAISI (2026). "CAISI Evaluation of DeepSeek V4 Pro." May 2026.

7. Reuters / Yahoo Finance (2026). "Anthropic's valuation surges to $965 billion, surpassing OpenAI." May 28, 2026.

8. CNBC (2026). "OpenAI proposes 5% stake to Trump administration to ease Washington pressure: Report." July 2, 2026.

9. Reuters (2026). "OpenAI's China rival DeepSeek under U.S. scrutiny after NIST report." May 28, 2026.

10. TechRepublic (2026). "Chinese AI Models Challenge OpenAI and Anthropic on Cost and Enterprise Risk." June 26, 2026.

11. Wedbush Securities / S&P Global (2026). "The AI Infrastructure Investment Report." January 28, 2026.

12. JP Morgan (2026). "AI Market Dynamics and Competitive Landscape." June 2026.

13. Unusual Whales / UBS (2026). "60% of Firms Curb AI Spend, Shift to Chinese Models." July 2, 2026.

14. Yicai Global (2026). "Chinese AI Token Pricing Defies Price War Fears Amid Surge in Computing Power Demand, UBS' Analyst Says." May 22, 2026.

15. Financial Times (2026). "DeepSeek V4 Pro NIST evaluation summary." May 26, 2026.

16. The Register (2026). "OpenRouter rankings show Chinese models dominating developer preference." May 30, 2026.

17. UBS Securities (2026). "Global AI Market Outlook: Competition and Pricing Dynamics." May 2026.

18. TechCrunch (2026). "DeepSeek launches V4-Pro and V4-Flash models to rival OpenAI GPT-5.5." April 24, 2026.

19. DeepSeek AI (2026). "DeepSeek-V4-Pro and DeepSeek-V4-Flash: Technical Report and Release Notes." April 24, 2026.

20. Zhipu AI (2026). "GLM-4.5: Open Source Release Announcement." April 23, 2026.

21. Zhipu AI (2026). "GLM-4.5 Technical Report: Architecture and Training." April 23, 2026.

22. Alibaba Cloud (2026). "Qwen3.5: Next-Generation Open-Source Model Family." February 16, 2026.

23. Alibaba Cloud (2026). "Qwen3.5 Technical Overview: 201 Languages and 1M Token Context." February 16, 2026.

24. Alibaba Cloud (2026). "Qwen3.5: Multimodal Capabilities and Deployment Architecture." February 16, 2026.

25. DeepSeek AI (2026). "DeepSeek V4 Pro Pricing Announcement." May 2026.

26. DeepSeek AI (2026). "DeepSeek V4-Flash Pricing Update: Permanent 75% Reduction." May 2026.

27. DeepSeek AI (2026). "DeepSeek V4 Architecture Report: CSA+HCA and Memory Optimization." April 2026.

28. DeepSeek AI (2026). "DeepSeek V4 Hardware Compatibility and RTX 4090 Performance Testing." April 2026.

29. DeepSeek AI (2026). "Speculative Decoding Framework for DeepSeek V4: 85% Inference Acceleration." May 2026.

30. US-China Economic and Security Review Commission (USCC) (2026). "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance." March 23, 2026.

31. Gartner (2026). "Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026." May 19, 2026.

32. Investors.com (2026). "Why OpenAI And Anthropic's Road To IPO Is Getting Bumpier." July 2, 2026.

33. Nasdaq.com (2026). "Is OpenAI's IPO Delay a Warning for AI Investors?" June 29, 2026.

34. The Next Web (2026). "A cheap Chinese AI model is closing in on Anthropic and OpenAI." July 2, 2026.

35. Channel NewsAsia (2026). "China's cheaper AI tokens a double-edged sword for Asian businesses." June 16, 2026.

36. Axios (2026). "Microsoft considers DeepSeek V4 integration in Copilot Cowork." June 2026.