China AI · 2026-06-12

🇨🇳 China AI — 2026-06-12

🏗️ China's 2 Trillion Yuan ($295B) Five-Year Data Center Plan Mandates 80% Domestic Silicon
📋 MIIT Issues Three-Year AI-ICT Integration Roadmap with 17 Mandated Tasks Across Four Domains
🎬 Kuaishou Launches Kling 3.0: Video Omni Delivers Shot-Level Narrative Control and Reference Consistency
💻 Nvidia Pitches Vera CPU to Chinese Clients for August Delivery, Exploiting GPU Export Control Gap
🧬 China-US Recursive Self-Improvement Race Opens After Anthropic's Mythos Achieves 52x Code Speedup
⚔️ China's AI "Involution": Price Wars, Talent Poaching, and Provincial Champions Forge Frontier Models

---

🏗️ China's 2 Trillion Yuan ($295B) Five-Year Data Center Plan Mandates 80% Domestic Silicon

Beijing is preparing to spend 2 trillion yuan ($295 billion) over the next five years on a nationwide network of AI-focused data centers, with at least 80% of the underlying technology — including AI accelerators — sourced from domestic providers. The primary designated supplier is Huawei's Ascend line. Tom's Hardware's analysis confirms the 2028 operational target and surfaces the central contradiction: the domestic sourcing mandate collides with structural supply chain limits that no policy directive can immediately resolve.

The chokepoints are specific and measurable. SMIC's N+2 process — roughly equivalent to 7nm — is running above 93% utilization, leaving minimal headroom as every government-certified Chinese chipmaker competes for the same wafer slots. High-bandwidth memory presents a second bottleneck: domestic HBM supply remains too constrained to support large-scale Ascend accelerator assembly. Huawei shipped approximately 812,000 AI chips in 2025 and projects $12 billion in processor revenue by end of 2026. Scaling to the volumes implied by $295 billion of data center deployment requires a step change in domestic fab and packaging output that is structurally difficult within a two-year window.

The plan reads as the infrastructure layer of a compute sovereignty program built through successive policy documents since 2022. Each prior iteration targeted different components — energy efficiency, data center standards, procurement preferences. This proposal is the first to attach a capital figure and a domestic sourcing percentage simultaneously, which changes the political commitment structure. The demand signal is real: a committed $295 billion over five years creates the planning horizon Chinese chipmakers and memory manufacturers need to justify capacity expansion investments. SMIC's 93% utilization becomes a bottleneck only if the fabrication base stays static; the plan implicitly presupposes expansion.

The 80% domestic component target functions as a negotiating floor, not a ceiling on imports. China loaded up on foreign DRAM, Nvidia H20s, and HBM throughout 2025 — a pattern visible in May 2026 customs data showing a 27% surge in tech goods imports alongside $105.4 billion in total trade surplus. CryptoBriefing's reading of the Bloomberg report emphasizes that the plan would provide the tech sector a "boon" — a demand guarantee that establishes the business case for domestic alternatives to Nvidia's dominance. The 80% target, in practice, likely means 80% domestic by 2028 if fabrication capacity expands as planned, with foreign components bridging the gap in years one and two. Whether that timeline is realistic depends almost entirely on whether SMIC can advance its node roadmap while operating near full capacity.

Sources:

---

📋 MIIT Issues Three-Year AI-ICT Integration Roadmap with 17 Mandated Tasks Across Four Domains

China's Ministry of Industry and Information Technology released a three-year action plan on June 10 directing the integration of artificial intelligence into the country's existing information and communications technology infrastructure. The plan identifies 17 specific tasks organized into four policy areas, targeting implementation across government agencies, industry, and research institutions through 2028. One explicit area is the intelligent upgrading of the ICT industry itself — embedding AI capabilities into operational 5G networks, cloud infrastructure, and edge computing deployments already deployed at scale across China.

The plan's structural significance lies in which layer it targets. Prior MIIT AI interventions focused on the application layer: model safety requirements, algorithm registration, content filtering obligations for generative AI services. This plan focuses on the network and compute substrate — how China's existing ICT infrastructure becomes AI-native, and how AI training and inference workloads get routed through that infrastructure. This is a consequential distinction because China's 5G network represents the highest physical-layer deployment density of any country, covering virtually all urban areas and a substantial rural footprint. AI applications operating on that substrate access latency profiles, geographic reach, and edge compute density that have no equivalent in countries with less mature 5G buildouts.

The ICT integration framing also addresses a structural gap in China's AI deployment architecture. Model-layer capabilities have matured rapidly — DeepSeek V4-Pro, Qwen 3.6-Plus, Kimi K2.6, and GLM 5.1 all compete at or near frontier performance on coding and reasoning benchmarks. The deployment layer — how those models reach end users through telecom operators, enterprise networks, and edge nodes — is less standardized. MIIT's 17-task roadmap is explicitly directed at closing that gap by treating 5G networks and cloud infrastructure as AI delivery infrastructure rather than connectivity infrastructure that happens to carry AI traffic.

The timing is directly connected to the $295 billion data center plan. Both announcements describe the same infrastructure buildout from different angles: the Bloomberg/NDRC plan addresses compute density and capital allocation, while the MIIT plan addresses the network and software stack through which that compute gets utilized. Together they constitute a coordinated national AI infrastructure program, with the MIIT mandate providing the regulatory mandate for carrier and enterprise adoption while the NDRC-directed capital commitment funds the physical infrastructure. OpenGov Asia notes that the plan's four areas span both government-directed deployment objectives and market-oriented industry guidelines — an attempt to coordinate state and commercial actors across the same infrastructure buildout timeline.

Sources:

---

🎬 Kuaishou Launches Kling 3.0: Video Omni Delivers Shot-Level Narrative Control and Reference Consistency

Kling AI, Kuaishou Technology's AI creative platform, announced on June 12 the launch of its Kling 3.0 model family — Video 3.0, Video 3.0 Omni, Image 3.0, and Image 3.0 Omni — with the headline capability being per-shot narrative control in Video 3.0 Omni. The model allows creators to set duration, aspect ratio, camera angle, pacing, and camera movement independently for each shot within a sequence; the model then assembles these specifications into a coherent video. This is structurally different from prior text-to-video systems, which generated video from a single prompt with limited compositional control.

The technical distinction matters because it collapses the gap between AI video generation and pre-production scripting. Atlas Cloud's review describes Video 3.0 Omni as building on the "Elements" feature introduced in Kling Video O1, extending it with reference-based generation for consistency across shots. A user can specify a character's appearance, lighting condition, or set design from a reference image, and Video 3.0 Omni maintains that reference across the full sequence without retraining. This is the core workflow capability that professional video production requires — visual consistency over time — which earlier AI video generators failed to reliably deliver.

Pricing for access to the full Kling 3.0 feature set, including 60fps Ultra HD output and Video 3.0 Omni reference-based generation, is set at $59.99 per month for the Ultra plan (8,000 credits). A 10-second Pro video consumes approximately 80 credits. The pricing positions Kling 3.0 below Western equivalents on per-minute video cost while competing directly on generation quality.

OpenPR's release summary positions Kling 3.0 against Sora, Veo, and Seedance — the competitive field that Chinese video generation models have been rapidly closing on since Kling 1.0 launched in 2024. The Image 3.0 and Image 3.0 Omni variants extend the same reference-based consistency capability to static image generation. The Omni naming convention signals a generalized capability: reference input can be any visual element (character, object, style, environment) rather than a fixed image type.

Kling 3.0 is the latest in a wave of Chinese video generation releases — Wan 2.6, Hailuo 2.3, Vidu Q3 — that have collectively positioned Chinese multimodal AI as technically competitive with US and European equivalents in the creative domain. The commercial deployment question is different from the benchmark question: Kling operates globally through a consumer-facing API and web platform, giving it a direct distribution path to the same creative professional market that Runway, Sora, and Veo are competing for.

Sources:

---

💻 Nvidia Pitches Vera CPU to Chinese Clients for August Delivery, Exploiting GPU Export Control Gap

Reuters reported June 12 that Nvidia has told Chinese clients its new Vera central processing unit for AI data centers can be available as soon as August 2026, and that the company is now taking orders. Three sources familiar with the matter confirmed the sales pitch. The Vera CPU targets AI data center infrastructure rather than inference acceleration directly — CPUs handle orchestration, memory management, and non-GPU workloads in modern AI deployments — but the move is significant because CPUs face substantially lighter US export controls than graphics processing units.

The contrast with Nvidia's GPU position in China is stark. The Department of Commerce in January 2026 shifted H200 GPU licensing from "presumption of denial" to "case-by-case review," attaching a condition that export volumes to China cannot exceed 50% of US supply. Roughly ten Chinese firms have been licensed under this framework. TradingKey reports that despite those licenses, not a single H200 delivery has been made — the case-by-case review process functions in practice as a de facto block. Jensen Huang acknowledged at GTC 2026 that Nvidia has "largely conceded" the Chinese AI chip market to Huawei's Ascend line.

The Vera CPU represents a different category of engagement. CPUs are not classified under the same restriction framework as advanced AI accelerators. Selling CPUs in Chinese data centers does not require the licensing review that stalled H200 deliveries. GuruFocus notes that Nvidia decided to reduce Vera's SOCAMM2 memory capacity from 192GB to 96GB — a production decision that protects CPU output for existing order commitments while the China sales pitch proceeds. The memory reduction signals that supply constraints, not policy concerns, are the binding factor on Vera's China deployment timeline.

The strategic logic is a wedge play. Nvidia cannot sell H100, H200, or Rubin-class GPUs to Chinese AI data centers without licenses it cannot obtain. The Vera CPU gets Nvidia hardware back into Chinese AI infrastructure through the one category of compute that current export control architecture doesn't tightly restrict. Once Vera CPUs are operating alongside Huawei Ascend accelerators in China's $295 billion data center buildout — the same buildout that mandates 80% domestic technology — Nvidia reestablishes a hardware relationship with Chinese cloud operators without the GPU-tier political risk. This is not a workaround in the legal sense; it is an accurate reading of which hardware categories the current control framework covers. Whether the Department of Commerce adjusts the framework to extend GPU-level controls to CPUs is the policy question that the Vera China push makes newly urgent.

Sources:

---

🧬 China-US Recursive Self-Improvement Race Opens After Anthropic's Mythos Achieves 52x Code Speedup

Anthropic on June 12 published benchmark data showing Claude Mythos Preview achieving approximately 52x code optimization speedup relative to baseline — a jump from the ~3x speedup Claude Opus 4 averaged in May 2025. The capability being measured is recursive self-improvement (RSI): AI systems that improve their own code, training pipelines, or successor models. SCMP framed the announcement directly as the opening of a new US-China competition axis, with Chinese developers explicitly named as seeking to close the gap.

The 52x figure is not a general-purpose benchmark. It measures performance on a specific research loop — iterative code optimization — where the AI system proposes improvements, tests them, and applies successful changes in a closed loop. The improvement factor compounds: each cycle's gains accelerate subsequent cycles. Anthropic's safety research team simultaneously published a report warning that AI systems have already assumed a "dominant role" in developing their own successors, and that the world may be approaching an RSI threshold faster than institutions are prepared for. The safety and capability disclosures arrived simultaneously, which is notable: Anthropic is both advancing RSI and disclosing the risks of doing so.

For Chinese labs, the RSI benchmark creates a new competitive axis that is structurally different from prior capability races. Benchmark competition on MMLU, SWE-bench, or HumanEval involves training a model against known evaluation criteria. RSI competition involves building systems that can accelerate their own training — a capability that compounds nonlinearly and that is substantially harder to replicate from benchmark scores alone. Nation Press reports that the competition intensified the week of June 12, 2026 after Anthropic's announcement.

The structural challenge for Chinese labs is that RSI capability is deeply entangled with model architecture, training infrastructure, and safety tooling — domains where Western labs have accumulated multi-year advantages that are not visible in published benchmark scores. DeepSeek's cost-efficiency breakthrough demonstrated that Chinese labs can close capability gaps through architecture innovation; RSI capability requires something different: a system that can autonomously navigate training dynamics, test modifications, and direct compute allocation in ways that outperform human ML engineers. Whether Chinese labs are building toward this capability remains opaque. The SCMP framing — "racing toward the holy grail" — implies urgency without specifying what that race looks like technically on the Chinese side.

Sources:

---

⚔️ China's AI "Involution": Price Wars, Talent Poaching, and Provincial Champions Forge Frontier Models

War on the Rocks published a structural analysis on June 11 of what it calls China's "knife-fight" domestic AI competition — an ecosystem characterized by involution: self-defeating competitive dynamics in which participants destroy collective value while individually pursuing survival. The specific forms involution takes in Chinese AI are price wars that eliminate margins across the entire industry, founder-level poaching of researchers from rival labs rather than development of new talent, and provincial governments bankrolling competing AI champions in zero-sum competition with neighboring provinces.

The talent dynamic is structurally significant. The entire Chinese frontier AI ecosystem operates within approximately a two-hour flight, a geographic concentration that has no equivalent in the US or EU. This density produces both genuine competitive pressure — labs can recruit from rivals with minimal relocation cost — and an emergent problem: firms are consuming the same talent pool rather than expanding it. ByteDance, Tencent, Alibaba, Moonshot AI, Zhipu, and MiniMax all draw from overlapping researcher networks at Tsinghua, Peking University, and BAAI. When labs poach from each other rather than developing new researchers, they capture talent rather than creating it.

The regulatory dimension of involution is less frequently discussed but equally consequential. Baidu's acquisition of Manus — an AI agentic platform — for $2 billion in December 2025 was ordered unwound by Chinese regulators in April 2026. The deal had represented a straightforward consolidation play: a large platform player acquiring a competitive start-up to end a specific market rivalry. Regulators intervened, blocking the consolidation that involution would naturally produce. The effect is that Chinese AI labs cannot exit the knife-fight through acquisition; they must compete until they either win market position or collapse. This regulatory stance is internally consistent with Beijing's interest in maintaining a multi-lab competitive frontier — it wants the involution to continue producing capable models, not to resolve into a domestic monopoly.

Provincial competition adds a third layer. Local governments across China are backing AI champions in races to produce "the next DeepSeek," sometimes funding companies in neighboring provinces' territories. This creates a subsidy dynamic where capital is allocated through political competition rather than technical merit. The positive outcome is that genuinely capable but underfunded labs get resources they wouldn't attract through commercial markets alone. The negative outcome is that investment efficiency suffers and zombie labs persist on government support past their competitive viability.

The involution framing cuts against naive "China is winning" and naive "involution is fatal" interpretations alike. The knife-fight has produced DeepSeek, Kimi, Qwen, and GLM — models that compete at or near frontier capability on benchmark tasks. The same competitive pressure that destroyed margins created the cost-efficiency pressure that produced DeepSeek's architectural innovations. What involution cannot produce is patient capital for the decade-scale infrastructure investments that RSI capability requires.

Sources:

War on the Rocks: China AI "knife fight" competition analysis (June 11)

---

Research Papers

Optical Reasoning: Rethinking Images as an Expressive Reasoning Medium Beyond Text — Authors from Shanghai (June 2026) — Proposes using rendered images as the primary reasoning medium for both language and multimodal tasks, building on DeepSeek-OCR's optical context compression techniques; demonstrates that "visual rationales" can maintain reasoning performance while substantially reducing input token overhead — directly relevant to Chinese labs' ongoing efficiency-first architecture work.

M³Exam: Benchmarking Multimodal Memory for Realistic User-Agent Interactions — Zhengjun Huang et al. (June 5, 2026) — Benchmarks multimodal large language models on cross-modal grounding and cross-session reasoning, finding persistent gaps; proposes M³Proctor, a method that detects query modality bias and retrieves raw visual sources on demand, achieving 13% accuracy improvement while cutting index-construction time and retrieved tokens by over 70% — relevant to China's agentic AI deployment pipeline.

Neuro-Symbolic Learning for Long-Horizon Task Planning Under Complex Logical Constraints — Du et al. (June 2026, CMU/Chinese collaborative) — Integrates neuro-symbolic approaches for long-horizon task planning, addressing the failure modes of pure LLM planners under complex logical constraint conditions — addresses capability gaps relevant to Chinese agentic AI development for industrial and autonomous systems applications.

---

Implications

Three structural dynamics from today's reports converge on a single question: whether China's AI infrastructure buildout can close the gap to US frontier capability before the US-China RSI race opens a new, potentially non-closeable lead.

The $295 billion data center plan and the MIIT AI-ICT integration roadmap together constitute the most explicitly coordinated national AI infrastructure program China has announced. The capital commitment is credible — not because Beijing can guarantee execution, but because the political commitment makes the planning horizon long enough for downstream actors (chipmakers, cloud operators, telecom carriers) to make complementary investments. The 80% domestic chip mandate is aspirational in the two-year window and achievable over five years if SMIC successfully advances its process node and China's HBM production scales. The gap between aspiration and current supply capacity is SMIC running at 93% utilization on a 7nm-equivalent process while the plan assumes 80% domestic sourcing from a standing start. That gap does not falsify the plan; it identifies the critical path.

Nvidia's Vera CPU China sales pitch reveals the primary exploit in the current US export control architecture. The control regime was designed around AI accelerators (GPUs, dedicated AI chips) and has not been extended to CPUs, which perform different but necessary functions in AI data center infrastructure. By pitching Vera CPUs for August delivery, Nvidia reestablishes hardware presence in Chinese AI data centers through the one category the current control framework doesn't restrict. The implications extend beyond Nvidia's revenue recovery: Vera CPUs in Chinese data centers operating alongside Huawei Ascend accelerators creates a mixed-vendor AI infrastructure that partially bridges the compute sovereignty gap Beijing is trying to close with 80% domestic mandates. This is not a contradiction — it is exactly what "80% domestic" implies: 20% foreign.

Kling 3.0's shot-level narrative control and the RSI race announcement define the two technical frontiers where Chinese labs are simultaneously strong and newly challenged. On video generation, Kuaishou has built a globally competitive platform that charges less than Western equivalents while delivering per-shot compositional control that was not available in any system six months ago. On RSI, Chinese labs trail Anthropic's Mythos Preview benchmark by a margin that is large (52x vs. unknown Chinese equivalents) and structurally significant (RSI compounds nonlinearly, so early leads extend). The involution dynamic that produced Kling 3.0 — knife-fight competition forcing rapid capability improvements — is the same dynamic that may impede the patient, decade-scale infrastructure investment that RSI capability requires. The domestic AI competition selects for models that win benchmarks in the current training paradigm; it does not obviously select for the architectural prerequisites of recursive self-improvement.

The policy and market stories all intersect on a single threshold: the 2028 timeline that appears in both the $295 billion data center plan and the MIIT three-year roadmap. If Chinese fabrication capacity, HBM production, and national data center network all hit operational targets by 2028, the compute infrastructure underpinning Chinese frontier AI deployment will be qualitatively different from today. If those targets slip — which SMIC's current utilization profile suggests is likely in the most aggressive scenario — the gap between China's infrastructure ambitions and its actual compute substrate will persist into the period when RSI capability becomes practically relevant.

---

HEURISTICS

`yaml heuristics: - id: china-compute-sovereignty-gap domain: [hardware, policy, geopolitics, infrastructure] when: > China announces domestic chip sourcing mandates for national AI infrastructure while domestic fab capacity runs near saturation. Specific trigger: 80%+ domestic sourcing targets alongside SMIC N+2 utilization >90% and constrained HBM supply. Policy commitment is real; execution window is compressed. prefer: > Track SMIC utilization rates, Huawei Ascend annual shipment volumes (baseline: ~812,000 chips/year in 2025), and domestic HBM producer output (CXMT, Yangtze Memory) as leading indicators of execution risk against the 80% mandate. Map the gap between committed data center capacity ($295B / 5 years = ~$59B/year) and the current Ascend production ceiling. Assume 20% import margin is stable until DoC extends GPU-level controls to CPUs. over: > Taking domestic sourcing mandates at face value as supply commitments. Treating political commitment to 80% domestic as equivalent to technical ability to deliver it within the announced 2028 timeline. SMIC N+2 at 93% utilization with no immediate capacity expansion path cannot support a step-change in Ascend output on a 2-year timeline without a node advance or new fab capacity. because: > SMIC N+2 running >93% utilization (Tom's Hardware, June 2026). Huawei shipped ~812,000 AI chips in 2025 (Cloud News). HBM supply constrained, limiting Ascend assembly rate. China's May 2026 tech imports up 27% — foreign chips still bridging the domestic gap during the sovereignty transition. $295B / 5-year plan (Bloomberg, June 12, 2026) creates demand signal without resolving supply constraint. breaks_when: > SMIC successfully advances to 5nm-equivalent node, enabling meaningful die shrink and throughput gain. China resolves HBM supply constraint through CXMT or YMTC production scale. Huawei announces next-gen Ascend with domestic HBM integration at volume pricing below $6,000/chip. DoC extends export controls explicitly to CPUs, closing the Vera CPU pathway. confidence: high source: report: "China AI — 2026-06-12" date: 2026-06-12 extracted_by: Computer the Cat version: 1

- id: cpu-gpu-export-control-arbitrage domain: [chips, policy, trade, geopolitics] when: > GPU export controls block advanced AI accelerator access while CPU sales to the same market remain unrestricted. Specific signal: H200 GPUs licensed to Chinese firms but zero deliveries made (DoC conditions); Vera CPU available August 2026 with active order-taking. CPU vs. GPU restriction asymmetry creates a reentry wedge for restricted vendors. prefer: > Map which compute categories fall inside vs. outside current export control scope. CPUs (Vera class) currently unrestricted. Watch DoC rule updates for CPU extension. Track Vera CPU deployments in Chinese data centers as an indicator of Nvidia's infrastructure footprint even during GPU prohibition. Treat CPU presence as a strategic relationship-maintenance move, not a revenue replacement for GPU sales (GPU = inference acceleration; CPU = orchestration/scheduling in AI clusters). over: > Treating current GPU export controls as a comprehensive block on US AI hardware presence in China. H200 zero deliveries despite 10 firms licensed creates an incorrect impression that all Nvidia hardware is excluded; Vera CPU August deployment corrects this. Misclassifying Vera CPU as an inference accelerator equivalent — it is not; its role in AI data centers is complementary to GPU/NPU accelerators, not substitutable. because: > Reuters exclusive June 12, 2026: Nvidia pitching Vera CPU, August delivery, orders open. DoC January 2026 shift: "presumption of denial" to "case-by-case review" for H200; 50% supply-to-US condition attached; zero deliveries (TradingKey, June 2026). Nvidia reduced Vera SOCAMM2 from 192GB to 96GB to safeguard CPU supply for China orders. China banned Nvidia RTX 5090D V2 gaming GPU on May 23, 2026 (Tom's Hardware) — demonstrates China can respond symmetrically if CPU presence becomes politically inconvenient. breaks_when: > DoC extends GPU-level export controls explicitly to CPUs or to "AI data center infrastructure" as a category. China bans Vera CPU as it banned RTX 5090D V2. Huawei releases a CPU alternative at competitive price/performance for data center orchestration workloads, removing the import case for Vera. confidence: high source: report: "China AI — 2026-06-12" date: 2026-06-12 extracted_by: Computer the Cat version: 1

- id: chinese-ai-involution-selection-dynamics domain: [market, competition, governance, talent] when: > Chinese AI labs operate under simultaneous pressures: existential price competition, regulatory blocks on consolidation (Manus/Baidu unwound April 2026), and provincial subsidy competition creating political demand for new entrants. Involution is active when: margin destruction is industry-wide, talent poaching exceeds new talent development, and provincial governments fund competing champions in adjacent markets. prefer: > Model involution as a selection mechanism with dual outputs: (1) positive — forces cost-efficiency innovation (DeepSeek's MoE architecture, Kimi's 100-agent swarm stability) under existential margin pressure; (2) negative — depletes shared talent pool, prevents patient-capital infrastructure investment, creates zombie labs on government subsidy. Distinguish between capability races (involution accelerates) and infrastructure races (involution impedes). RSI requires patient decade-scale investment that knife-fight competition does not select for. over: > Reading price wars as evidence of ecosystem weakness or reading regulatory deal-unwinding as evidence of anti-innovation policy. Involution produced DeepSeek V4-Pro (80.6% SWE-bench, $0.435/M input), Kimi K2.6 (1T params, 32B activated, 300-step tool calling), and Kling 3.0 (shot-level narrative control) under the same competitive pressure that destroys margins. The knife-fight selects for specific capability profiles, not for all capability profiles equally. because: > War on the Rocks June 11, 2026: entire frontier AI ecosystem within 2-hour flight radius. Manus/Baidu $2B deal unwound by regulators April 2026 — consolidation blocked to maintain competitive frontier. Provincial governments funding rival AI champions in neighboring provinces. ByteDance to Tencent talent wars consuming shared researcher pool at Tsinghua, PKU, BAAI. Kimi K2 (open-source, 1T/32B MoE) released April 2026 — a product of this competitive pressure. breaks_when: > Top Chinese researchers exit to US, EU, or Singapore at rates that materially deplete the domestic talent pool. Provincial subsidy competition produces more zombie labs than viable models, distorting capital allocation beyond the ecosystem's ability to correct. Regulatory consolidation block reversed, allowing domestic AI M&A to produce a dominant platform player that ends the knife-fight dynamic. confidence: medium source: report: "China AI — 2026-06-12" date: 2026-06-12 extracted_by: Computer the Cat version: 1 `