Beijing Doesn’t Need Fewer AI Labs

2026 09 02

Summary

The Chinese AI lab landscape does not need a consolidated lab landscape to govern it. Its intervention repertoire under ordinary politics is campaign-style enforcement, where a crisis mobilizes the bureaucracy to regulate multiple firms at once and issue sector-wide rules. Compliance rises across the industry without reducing the number of firms. This enforcement window after a crackdown may support more safety practice adoption in labs should safety regulation be built into policy beforehand. Separately, governance proposals that presuppose a small number of state-legible projects mistake simplicity for a precondition.

Market attrition in the lab landscape has been offset by entry from platform incumbents. Along the current trajectory, standalone labs are most at risk of being priced out through frequent repricing, as opposed to inhouse labs from well-resourced big tech firms. Alongside existing enforcement mechanisms with platform incumbents, Beijing is increasing industry-wide legibility and enforcement capacity, and it shows no preference for a nationalized landscape.

Introduction

The safety community sometimes assumes that Chinese frontier AI will consolidate into more governable actors. We may distinguish two claims: Consolidation refers to a decrease in the number of frontier developers, via acquisition or merger. Rationalization refers to state-directed consolidation; the strongest form of this is nationalization. The assumption appears primarily in two forms: as a forecasting scenario of nationalized projects, or as a prescription in governance proposals that depend on a small number of state-legible AI development projects. A related assumption, that a sufficiently alarming capabilities incident would trigger a paradigm-shifting government response, is further explored in a companion piece.

Labs will be treated as projects in this memo. At present, the mapping of firms to projects is essentially one-to-one: each lab trains its own frontier models, and the labs are competing rather than pooling resources. The observable dynamics are talent movement between firms and redirection of investment toward stronger labs, contrary to an assumption of cross-firm collaboration.

This memo describes state behavior according to precedent under ordinary politics. If state leadership began to regard AI as a decisive strategic advantage, then ordinary processes are no longer constraints to addressing what is perceived as an urgent existential crisis. The nationalization scenario in scenario forecasts is a claim about exceptional politics where leadership believes AI is decisive. In critical junctures like these, decision-making breaks away from precedent-based reasoning, and this memo cannot address this scenario. However, the memo demonstrates that a small set of actors is likely not a precondition for regulatory enforcement.

Some proposals that assume a small number of projects focus on compute verification are better founded. Compute itself is plausibly more concentrated and state-legible than labs: through Huawei, the state-coordinated East Data West Computing hubs, and the broader national integrating computing power network. A small set of labs is not required for the legibility of the compute landscape either.

The potential pathways for consolidation are either market competition or state intervention. The memo will address both in the following order:

• The market has produced attrition, but entry by platform incumbents has offset reduction by exits. The trajectory suggests convergence on big tech inhouse labs rather than consolidation. (Section 1)
• The state’s intervention repertoire, compared against historical precedent, is firm enforcement and sector elimination, not industry rationalization. Firm-specific punishment, however, raises compliance across the industry. (Section 2)
• Beijing is building legibility and enforcement capacity without a crisis, and demonstrates no preference for nationalization. (Section 3)

Section 1: Attrition Was Offset by Platform Incumbents, but Standalone Labs Are Exposed

The number of labs training large models has increased. Attrition, where firms exited frontier training, occurred through market dynamics but was more than offset by entrants. As capability thresholds for “frontier” vary, my set of relevant labs consists of those training large models, observable from public releases and funding activity.

As of September 2026, two of the original six tigers have exited frontier training—Baichuan has pivoted to healthcare, and 01.AI (Lingyi Wanwu) has repositioned as a Palantir-style sovereign AI vendor for governments and large enterprises. ByteDance (Seed/Doubao) and Alibaba (Qwen) have entered.

Both exits were market-selected, having insufficient private venture capital, no patron, and no external revenue for costly training runs. Remaining labs are insulated from market dynamics in one of three ways: operating cash flow at platform incumbents (established technology firms), founder capital at DeepSeek, and public-market or state-linked capital at standalone labs.

Platform incumbents fund training from existing capital and revenue streams rather than external raises, and their continued participation in frontier training is largely contingent on internal capital-allocation decisions at least through the next eighteen months. Capex at current scale has already made free cash flow negative at Alibaba and Tencent, but a single training program remains a small portion of either firm’s spending.

DeepSeek is backed by founder capital and is insulated from decision-making of its investors, so it likely has a durable position at the frontier of Chinese AI development. Founder Liang Wenfeng contributed roughly $3B of $7B raised in its June funding round and retains control as general partner over the limited partnership of all commercial investors. Tencent and CATL accepted five-year lock-ups and no voting rights.¹ The sole voting direct-equity investor is the state-backed National AI Industry Investment Fund, a notable signal of government support.²

Standalone labs have public-market and state-linked capital, thus exposed to high-frequency repricing. Zhipu listed on the Hong Kong Stock Exchange in January 2026. Dependent on heavy demand from state-owned enterprises (SOE) and financial-institutions, it re-rated sharply upward through June, when its market capitalization briefly exceeded HK$1 trillion. However, it has since shed nearly half that value against releases from competitors. MiniMax, listed the following day, slid from approximately $28 billion in May to roughly $12.5 billion in August for the same reason. StepFun has filed for a Hong Kong listing and Moonshot is opening a final pre-IPO round.³

There are three significant observations.

First, exits so far are market-selected, with those remaining insulated through capital from the state and tech giants. State subsidization appears in SOE participation in funding rounds and state-linked demand on income statements. This should not be over-read as state control of privately founded labs through stakeholding, but standalone labs that rely on state-linked demand are plausibly more responsive to regulatory signals.

Second, the emergence of new players comes from a predictable class of platform companies. ByteDance (Seed/Doubao), Alibaba (Qwen), Meituan (LongCat), Ant Group (Ling-1T), Xiaomi (MiMo), and Baidu (ERNIE) are cash-generating incumbents for whom inhouse labs are a feasible strategic option.

Third, future exits will likely be the result of either internal resource allocation decisions or high-frequency valuation repricing. For the incumbents and DeepSeek, continued frontier participation is an internal capital-allocation decision rather than an exit forced by market dynamics. While the standalone labs remain sentiment-dependent to some degree, rapid repricing release-by-release against the outputs of strongest labs places firms in extremely volatile positions that ultimately threaten their survival as a frontier lab, just as MiniMax demonstrates. If a lab falls behind by a few months, it may be forced to exit.

As a caveat on observability, these resilience assessments are built from publicly visible releases and valuations, aided by measured input from industry insiders. Survivability, ultimately, is a product of capital, sustained internal commitment, and team quality, of which only the first is observable from outside. Assessments of the other two factors diverge. For instance, Alibaba’s cash flow dwarfs competitors, yet informed observers disagree about its team quality relative to ByteDance’s.

Section 2: Historical Interventions Are Enforcement Campaigns, Not Rationalizing

Historical cases in comparable Chinese sectors show two phenomena: market-driven mergers with no state role and crisis-triggered state interventions that restructured individual firms or eliminated sectors outright, not rationalization of a competitive private landscape into fewer, governable champions.

I selected consolidation and crisis-intervention events in e-commerce and fintech for case studies. As the Chinese AI industry consists of private firms, I select for privately founded Chinese sectors that experienced a competitive phase prior to state intervention: e-commerce and fintech. Pre-reform socialist industrial precedents are omitted due to the non-existence of private enterprise; the relevant institutional environment is post-1992 party-state capitalism, where private enterprises operate under the macro-economic and political control of the party-state.⁴ For the same reason, I chose to omit telecommunications as a case study despite the industry having experienced rationalizing state intervention, as it was explicitly placed under absolute state control while the AI sector is designated only to receive subsidies and guidance funds. While national security interest in AI is rising, the two categories are distinct in Chinese policy and the ownership instruments in telecommunications are not applied in AI. When focusing on market competition in ordinary politics, the other two cases are more suitable. I specifically analyze market structure and crisis-intervention junctures in e-commerce and fintech.

The only consolidation among competing firms was market- and capital-driven merger activity with no state influence nor crisis trigger. The Meituan–Dianping and Didi–Kuaidi mergers were coordinated under pressure from major investors to end unsustainable pricing wars. The state was absent during talks and asserted merger-review jurisdiction over these mergers only retroactively, in 2021.

Crisis-triggered interventions, in contrast, restructured individual firms or eliminated sectors.

In e-commerce and payments, crisis produced firm-specific restructuring. The regulatory foundations for intervention demonstrably predated the crisis trigger and subsequent enforcement. Ant Group, Alibaba’s fintech affiliate and operator of Alipay, was days away from a ~$34 billion Shanghai-Hong Kong dual listing when founder Jack Ma publicly derided Chinese financial regulators as running a ‘pawnshop mentality’ at the October 2020 Shanghai Bund Summit. The listing was suspended subsequently on November 3, and the following intervention became the defining case of platform-era enforcement.

The microlending and platform-antitrust draft documents released mere days after the speech suggested preparation for a crackdown long before the inciting event, which otherwise appeared to influence the timing and severity of intervention.⁵ Months prior to the intervention, the People’s Bank of China (PBOC) finalized the measures by which Ant would be restructured in September 2020, effective November 1. This was followed by the draft online-microlending rules released on November 2, 2020. The State Administration for Market Regulation also released draft platform antitrust guidelines on November 10. Ant Group was thereafter restructured into a supervised financial holding company and Alibaba faced a RMB 18.2 billion fine for exclusivity practices.

Eight months into the enforcement campaign sparked by the Ant crackdown, Didi’s defiance of state authority also triggered firm-specific punishment and introduction of formal enforcement mechanisms. Didi listed on the New York Stock Exchange after regulators suggested it delay. Following this, Beijing initiated a surprise inspection in a discretionary application of a cybersecurity review mechanism. After a year-long investigation, the CAC fined the firm $1.2 billion, and the episode motivated the state to build up formal enforcement capacities it previously lacked: a February 2022 revision of the Cybersecurity Review Measures requiring pre-listing review for platforms of over one million users. The November 2020 Ant incident created a focal point for regulators in banking, finance, antitrust, unfair competition, data security, and education to act against big tech companies simultaneously.⁶ The incident was notable for an abrupt correction after long regulatory tolerance, and compliance costs rose across the industry.

The most drastic restructuring was an elimination of the fintech sector. Ezubao was a Ponzi scheme launched in July 2014 and attracted roughly RMB 50 billion ($7.6 billion) from about 900,000 investors, before it collapsed December 2015 to February 2016 with 21 arrests made. The China Banking and Insurance Regulatory Commission (CBIRC) eventually declared in November 2020 that operating peer-to-peer (P2P) platforms had been zeroed out.⁷ The sector was thereupon eliminated rather than consolidated into supervised champions. This is the most severe intervention thus far. However, this intervention is largely inapplicable to AI given that the state concluded the P2P lending sector produced no legitimate value despite about $115 billion in outstanding debt. The state’s agenda on AI is the opposite.

A parallel response to a political crisis at a Chinese lab will look like the Cyberspace Administration of China (CAC) penalizing and restructuring the offending firm, not a nationalization of the lab landscape. We can reasonably assume a crisis will determine the timing and severity of intervention, which would be followed by enforcement on other labs, and the content of any escalation in Chinese AI will be visible in policy drafts released before intervention. Whether a penalty would render a firm unable to compete is uncertain: Didi remained an industry leader despite the $1.2B fine and eighteen-month registration ban, and the competitive landscape barely changed. Granted, should state intervention knock labs permanently offtrack in an intelligence explosion, for example, such a setback will render a firm uncompetitive and act functionally equivalent to attrition. In the industry as a whole, firm-specific penalization would be accompanied by sector-wide compliance investment at other labs. Whether this compliance would be safety-relevant depends entirely on whether safety requirements are built into the regulatory corpus beforehand.

Section 3: China is Building Regulatory Capacity, but Shows No Preference to Nationalize

Outside the possibility of crisis-triggered interventions, Beijing passes regulation that increases the legibility of the industry without a crisis prompting it. The 2023 interim generative-AI measures, the algorithm registry, 2025 content-labeling requirements, the standards body TC260’s iterative AI Safety Governance frameworks, the 15th Five Year Plan commitment to risk management systems, and the agent governance measures effective July 2026 demonstrate political willingness to proactively keep the AI sector monitorable. These compliance obligations are qualitatively different from interventions relevant to market rationalization.

The closest demonstration of structural intervention was the National Development and Reform Commission’s (NDRC) order blocking Meta’s acquisition of Manus in April 2026, which risked foreign ownership over a strategic asset but otherwise did not involve a crisis. However, the nature of this intervention was prevention of foreign ownership; it is not a signal of domestic restructuring.

Legibility measures indicate capacity for monitoring but not willingness for restrictive enforcement. The state’s heavy investment in AI as a policymaker, investor, and customer simultaneously means it also does not have a strong incentive to perform prohibitively strict regulation on the industry. The governance corpus signals pro-growth sentiment and stakeholder coordination, and regulatory content is often diluted in drafting. The Interim Measures removed a three-month fine-tuning deadline and softened obligations for training-data accuracy, for example, between draft and final document.

There is no indication the state prefers a nationalized landscape. Given the party-state’s unprecedented endorsement of AI development, a P2P-style sector annihilation or a rationalization of the lab landscape into champions is highly unlikely. Firm-specific intervention, escalating to sector-wide behavioral rules, is the modal path that the historical record of Chinese party-state capitalism supports.

The Chinese leadership’s revealed preference since late 2022 points the same way. Platform rehabilitation was framed explicitly in terms of international competition, and DeepSeek’s breakthrough was treated as a private, market-produced success. Broadly, official innovation policy has continuously emphasized the critical role of enterprises in innovation.

Conclusions

Industry-wide compliance in Chinese party-state capitalism is achieved by a crisis-intervention that mobilizes the bureaucracy against the sector, not by consolidation. The platform regulatory campaign raised compliance across e-commerce and fintech while the actor landscape stayed largely intact. Alibaba similarly remained an incumbent despite Ant’s restructuring. Currently, the lab count has not fallen. Market competition produced attrition but has been offset by entry from platform incumbents. The consolidation sometimes assumed has little support under the current trajectory or the precedent.

However, with consideration for the future of Chinese domestic AI policy, this is not a portrait of a state that is neglecting to regulate AI. China is building out regulatory capacity continuously and deliberately to monitor deployments and detect failure modes. Moreover, as the historical record shows, industry-wide compliance is likely achievable through the punishment of individual firms rather than through the reduction of their number. Proposals that frame a small number of actors as the precondition for effective domestic governance neglect the existing mechanisms of enforcement. Precedent shows the bottleneck on enforcement is bureaucratic inertia, slow information transfers to the top, and state investment in the industry that hinders commitment to policing it. A tractable followup question is what would make AI risk legible upward. The window for safety-relevant content to be included and defended (to survive dilution in drafting stages) is before a crisis-intervention enforcement campaign. Toward this end, drafts in circulation are worth tracking for safety-relevant content.

Appendix

Note: The purpose of this is to address claims that Chinese firms will be priced out of frontier training runs in the near future. The calculations are not meant to measure frontier costs precisely, just to determine whether costs could exhaust announced AI capex and price platform incumbents out through comparison of capex with upper-bound cost estimates. Exit through internal resource reallocation is the likelier outcome but not modeled.

Epoch AI’s 2024 cost model estimated amortized hardware and energy cost for a final frontier training run has grown at about 2.4x per year since 2016 (95% CI: 2.0–3.1x). As of February 2026, Epoch’s dashboard shows training costs climbing 3.5x annually.

I use C = $1.5B as an upper-bound run cost for the annual cost of a frontier training program. The annual cost of a frontier training program C consists of failed experiments, ablations, data pipelines, staff, and final runs. Epoch estimates the total development-cost to be 47–67% hardware, 29–49% R&D staff, and 2–6% energy; following this estimation, the total program costs roughly 2–4x the cost of a final run. Assume a 2026 US final training run estimation at $200–500M with the 2–4x multiplier: C ≈ $0.4–2B per year, where $1.5B is a rough upper estimate.

The actual costs for Chinese labs are likely below this bound on account of two structural conditions. First, research staff costs remain below those of the US even accounting for additional compensation. Second, the reported compute required is lower. Specifically, DeepSeek-V3’s disclosed 2.788M H800 GPU-hours for its training run is about one-eleventh of Llama-3-405B’s 30.8M.

Financial capacity.

An upper-bound cost of $1.5B is a small number for platform incumbents, and the cost would need to increase an additional 12–20x before a single program takes up announced capex entirely.

Platform incumbents are likely not going to be priced out of inhouse training programs within the 18-month window. The larger cost-growth rate of 3.5x/year shown by Epoch suggests that costs compound about 6.5x over 18 months and 12x over 24 months.

The comparison is a rough bound demonstrating ability to fund training programs; it should not be read as a precise threshold for three reasons. AI capex targets have been revised upward: ByteDance raised its 2026 capex target from ~$24B to over $29B. Alibaba expressed it would “overshoot” its commitment and designated an additional HK $80B ($10.2B) of new equity on AI. Tencent’s president similarly described overspending as a lump-sum for this year and the next, i.e. capex would not eat into earnings in the long term, though free cash flow turned negative in the June quarter. Secondly, announced capex is used on compute buildout that can be continually rented out and generates revenue. Tencent reports rental margins of roughly 30%. But more fundamentally, capex is a budget for hardware acquisition, not staffing. A third of C is R&D staff costs, which is a separate budget line. These mismatches mean the viable time window may shift by months, and it is difficult to measure a precise threshold.

Thus, while incumbents can choose to exit through internal resource reallocation, most likely from disappointment on returns from AI, this establishes a floor on forced exits. Based on numbers alone, the cost of training programs is unlikely to compel exit within eighteen months.