China AI · 2026-06-15

Now I have everything I need. Writing the report.

---

🇨🇳 China AI — 2026-06-15

🔄 Amazon CEO Jassy's Warning Triggered Fable 5 Export Block — Beijing Reads New Structural Precedent
🦾 MIIT and SASAC Launch National Humanoid Real-Scene Training: 10,000 Units in Operation Mode by End-2026
🧠 BAAI Unveils WuJie·Physis at 8th BAAI Conference: Physical-State Prediction as the Path to Physical AGI
⚡ GLM-5.2 Ships June 13 with 1M-Token Context, MIT Open Weights, 300 Tok/Sec — Direct Response to Fable 5 Shutdown
🔧 ByteDance Orders 50,000 Iluvatar CoreX Inference Chips to Power Doubao, Evaluates Baidu Silicon in Parallel
📈 Tencent-Backed Enflame Technology Receives IPO Approval, Completing "Four Little Dragons" Domestic Chip Cycle

---

🔄 Amazon CEO Jassy's Warning Triggered Fable 5 Export Block — Beijing Reads New Structural Precedent

A Fortune investigation published June 14 identified Amazon CEO Andy Jassy's direct warning to the White House as the proximate trigger for the Trump administration's June 12 order blocking foreign access to Anthropic's Fable 5 and Mythos 5 models. Jassy's intervention — combined with internal cybersecurity concerns about a specific Fable 5 jailbreak enabling Mythos 5's autonomous coding capabilities — produced an executive-level order within hours of deliberation. The mechanism matters as much as the outcome: a competitor CEO's private warning to the White House produced a regulatory action removing Anthropic's most capable models from global markets inside a single news cycle.

Reuters Breakingviews, published June 15, frames the episode as "Anthropic becomes a cautionary sovereign-AI fable": a company whose transparent safety disclosures about Mythos 5's recursive self-improvement capabilities — the 52x code speedup Anthropic documented in its June 12 benchmark publication — provided the policy rationale for its own global shutdown. Anthropic disputes the severity, characterizing the underlying vulnerability as a narrow, specific jailbreak rather than a universal defeat of Fable 5's safety architecture. The administration's Techmeme-covered deliberation trail reveals the process included "multiple tense calls between Dario Amodei and administration officials," alongside Jassy's separate approach, before the final order.

Beijing's read of the structural precedent diverges from both Anthropic's and the administration's framing. The episode establishes that US AI frontier model access restrictions are not solely driven by national security threat assessments: commercial competitor interests, investor relationships, and executive lobbying shape which restrictions activate on what timeline. For Chinese AI labs, Stocktwits' sovereign AI coverage notes Beijing's position: any reliance on Western frontier model APIs represents a strategic vulnerability that the June 12 action has now empirically demonstrated.

The practical effect on China's domestic AI ecosystem is structural advantage without requiring any Chinese action. GLM-5.2, DeepSeek V4 Pro, Kimi K2.7-Code, and Qwen 3.7 Max now face zero Western frontier competition in their primary Chinese deployment context. Researchers and developers who transitioned to Chinese models after April 2026's identity verification requirements — and again after June 12's full foreign exclusion — have no straightforward path back to Fable 5-class capability through Western providers. Zhipu CEO Jietang's GLM-5.2 release statement explicitly positioned the June 13 release as a response, expressing "deep regret over the sudden withdrawal of Anthropic's models" while presenting MIT open weights as the structural alternative. That framing positions Chinese open-weights not as market competition but as sovereign infrastructure inaccessible to executive order withdrawal.

Sources:

---

🦾 MIIT and SASAC Launch National Humanoid Real-Scene Training: 10,000 Units in Operation Mode by End-2026

China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission issued a joint directive June 14 launching a national "real-scene training" campaign for humanoid robots and embodied AI systems. The directive's language is unambiguous: by end of 2026, humanoid robots and related key products complete application verification in multiple representative scenarios, achieve routine deployment, and officially enter "operation mode" — a defined operational state distinct from demonstration or research use. BigGoFinance's analysis identifies 10,000 operational units as the targeted deployment capacity.

The structural mechanism is state enterprise mobilization rather than market discovery. SASAC-controlled state-owned enterprises — operating China's major manufacturing, logistics, and critical infrastructure sectors — serve as the mandatory real-scene deployment sites. Local governments and SOEs must identify specific factories, warehouses, and hospitals for humanoid pilot deployments and submit application documentation to the relevant regulatory bodies. This converts SASAC procurement authority into demand-side pull for humanoid robots from suppliers including EngineAI, Unitree, and Linkerbot, each of which is preparing IPO filings in 2026 as the market structure matures.

The AI training significance is as important as the deployment headline. RareEarthExchanges' analysis identifies the "real-scene training" framing as deliberate: in-situ model training on actual industrial tasks — not simulated environments — generates physical-world data that simulation cannot replicate at the required distribution. Humanoid robots operating in actual SOE factory environments generate contact geometry data, task-failure data, and rare-event distributions that are the training signal for next-generation embodied AI reasoning layers. The 10,000-unit target is simultaneously a commercial deployment mandate and a data generation mandate.

The MIIT/SASAC directive was coordinated with the 8th BAAI Conference (June 12-14), where BAAI released WuJie·Physis — a physical-state prediction world model targeting physical AGI. China.org.cn's conference coverage notes Morgan Stanley's 2026 global humanoid forecast sits at 28,000 units across more than 100 competing startups. The MIIT/SASAC directive alone accounts for 10,000 units by policy mandate, independent of market discovery — a state-directed demand creation mechanism that compresses what might otherwise be a 2028-2030 market development timeline into an end-of-2026 deployment requirement.

Sources:

---

🧠 BAAI Unveils WuJie·Physis at 8th BAAI Conference: Physical-State Prediction as the Path to Physical AGI

The 8th Beijing Academy of Artificial Intelligence Conference (June 12-13, reported June 14) ran 25 topic forums and 200 presentations, drawing two Turing Award winners to its opening ceremony. The headline research release was WuJie·Physis — the physical intelligence component of BAAI's WuJie world model system, targeting a clear path to Physical AGI by starting with the prediction of physical states, rather than the generation of physically plausible video as Western approaches predominantly pursue.

The technical distinction is architecturally significant. Video generation world models — the dominant approach in OpenAI's Sora and Google DeepMind's Genie 2 — optimize for visual coherence: outputs look like physics, but the internal representations are not grounded in physical state. WuJie·Physis represents the world as structured state tensors (positions, velocities, forces, contact geometries) and learns to predict how those states evolve under physical laws. A robot controller built on a physical-state model reasons about the consequence of grasping an unfamiliar object at an angle by computing the predicted force distribution — not by pattern-matching against visually similar prior frames. CGTN's framing of the release as "China's first general world foundation model" positions WuJie·Physis against Western video world models on the technical theory level, not merely the capability level.

Houdao AI's in-depth conference analysis identifies WuJie·Physis as the third component of a multi-model WuJie system that already includes a brain science foundation model and an embodied brain model — an architecture that maps language, biological cognition, and physical-world understanding onto a unified system. The BAAI 2026 Conference schedule organized its working sessions around the distinction between the current "large model era" and its physical AI successor, with a dedicated track for cross-domain physical state modeling.

Since 2019, the BAAI Conference has welcomed attendees from more than 30 countries — making it China's primary institutional interface for frontier AI research signaling to both domestic and international labs. WuJie·Physis as the headline release, timed to the same day as MIIT/SASAC's national humanoid deployment mandate, is not coincidental coordination: BAAI's physical-state prediction model provides the reasoning layer for the factory-floor embodied AI deployments the government is mandating. The data flow runs from SOE humanoid deployments through real-scene training back into WuJie·Physis's training pipeline — a closed research loop that language-only world models cannot participate in.

Sources:

---

⚡ GLM-5.2 Ships June 13 with 1M-Token Context, MIT Open Weights, 300 Tok/Sec — Direct Response to Fable 5 Shutdown

Zhipu AI's Z.ai deployed GLM-5.2 on June 13, 2026 — the day after the US government ordered Anthropic to suspend foreign access to Fable 5 and Mythos 5. Zhipu CEO Jietang's release announcement expressed "deep regret over the sudden withdrawal of Anthropic's models" and presented GLM-5.2's MIT open weights explicitly as a structural alternative — deployable by any organization globally, including Chinese regulated industries, without US-jurisdiction API exposure. The MIT license converts the model from a market product into infrastructure inaccessible to executive order withdrawal.

GLM-5.2's core specifications: a usable 1 million-token context window (not a degraded long-context variant), MIT-licensed open weights, and 300 tokens per second throughput at production deployment. Initial deployment went to all GLM Coding Plan tiers on June 13; standalone API access and Z.ai chatbot integration followed June 14. Codersera's comparison analysis places pricing at approximately 1/10th of Claude Opus 4.8, the primary Western alternative for enterprise coding agents. Explainx.ai's benchmark report cites BridgeBench reasoning at 42.8 — described as exceeding Fable 5 on that benchmark — at 300 tok/sec throughput. The BridgeBench score is Zhipu-platform-administered; Pandaily notes GLM-5.2's primary benchmark platform is ZCode 3.0, showing strong practical coding performance.

The absence of SWE-bench Verified and LiveCodeBench scores is acknowledged in independent analysis. GLM-5 (the previous flagship, a 744-billion-parameter MoE with 40 billion active parameters) claimed 77.8% on SWE-bench Verified in February 2026. GLM-5.2's technical documentation does not yet cite those benchmarks; the MIT license enables third-party evaluation that should resolve comparative capability within days of open-weight availability. The strategic read is that Zhipu is prioritizing access-sovereignty and deployment speed over benchmarking optics at this moment.

The enterprise value proposition is clear from access architecture, not raw benchmark rank. GLM-5.2's 1 million token context enables multi-document legal analysis, full-codebase reasoning, and long-form synthesis at costs that Chinese financial, healthcare, and government customers price against the Fable 5 tier they have now permanently lost. AI Weekly's coverage summarized the deployment sequence: "Zhipu Deploys GLM 5.2 to All GLM Coding Plan Tiers With 1M-Token Context — API and MIT Open Weights Arriving Next Week" — a sequencing that prioritizes enterprise customers with existing Zhipu subscriptions before public model release, the opposite of a benchmark-announcement strategy.

Sources:

---

🔧 ByteDance Orders 50,000 Iluvatar CoreX Inference Chips to Power Doubao, Evaluates Baidu Silicon in Parallel

Reuters reported exclusively June 15 that ByteDance is in advanced talks with Shanghai-based Iluvatar CoreX (HKEX: 9903) to purchase AI chips for inference workloads, with at least 50,000 chips expected to ship this year. Two sources familiar with the matter confirmed the deal scope. ByteDance is simultaneously evaluating a separate chip procurement agreement with Baidu (NASDAQ: BIDU) for similar inference purposes — a parallel evaluation structure that signals supplier diversification within the domestic chip market rather than single-vendor commitment.

The primary use case is Doubao, ByteDance's AI chatbot that reached 345 million monthly active users as of May 2026. Doubao runs entirely on GPU inference hardware; its cost structure is determined by inference chip efficiency, not training compute. Windows News' reporting includes a direct quote from within ByteDance's infrastructure team: "We are starting with shadow traffic — running models in parallel on Nvidia and Iluvatar and comparing results. So far, latency and accuracy are within acceptable ranges." The shadow traffic evaluation protocol is standard hyperscale chip migration practice: parallel production runs before full cutover, typically 4-6 weeks. "Within acceptable ranges" for latency and accuracy is the threshold statement that determines whether a 50,000-chip commitment converts into a full architectural migration.

Iluvatar CoreX is inference-specialized — a design philosophy that distinguishes it from training-focused Cambricon and general-purpose Biren. Inference optimization targets deterministic latency at scale rather than peak FLOP throughput, which matches Doubao's operational requirements: the chatbot needs consistent sub-200ms response times for 345 million users, not maximum batch training throughput. Benzinga's framing positions the deal explicitly as evidence of Nvidia's continued China revenue deterioration — Jensen Huang acknowledged in October 2025 that Chinese GPU market share had "effectively fallen to zero" under export controls.

The Baidu parallel evaluation adds competitive dynamics within China's domestic chip market. Baidu's Kunlun chip division develops inference-optimized hardware primarily for ERNIE Bot deployments. ByteDance evaluating Baidu silicon for Doubao represents cross-company domestic chip competition that would have been commercially unlikely before Nvidia's exclusion created the procurement gap. Yahoo Finance's Reuters mirror notes the deal advances ByteDance's "broader push to curb its dependency on Nvidia hardware amid tightening US export controls" — a dependency shift that, if the shadow traffic evaluation succeeds, marks inference workloads as the first major Chinese AI production category to transition from Nvidia to domestic silicon.

Sources:

---

📈 Tencent-Backed Enflame Technology Receives IPO Approval, Completing "Four Little Dragons" Domestic Chip Cycle

Bloomberg reported June 15 that Tencent Holdings-backed Shanghai Enflame Technology has received approval for an initial public offering — described as "the last of the 'four little dragons,'" completing the market cycle for China's four leading domestic AI accelerator startups. The other three — Cambricon (STAR Market, listed 2020), Biren Technology, and Moore Threads — have navigated their own IPO timelines across 2020-2026. Enflame's approval closes the loop: all four leading Chinese AI chip startups are now in or entering public markets, converting state-strategic chip development into publicly traded infrastructure.

Enflame was co-founded in 2018 by a team from Google Brain and Tencent, with Tencent as the primary strategic investor. The company's Dorado architecture targets cloud-scale AI training workloads — the same market segment as Nvidia's H100 family — with Tencent's own AI training infrastructure running substantially on Enflame hardware. That vertical integration relationship gives Enflame a captive major enterprise customer (Tencent) and gives Tencent sovereign chip procurement independence from Nvidia training restrictions. The IPO converts that strategic relationship into public market capitalization at a moment when the permanent-gap scenario for Nvidia in China has been priced as the base case.

Tencent's June 15 activity extends significantly beyond the Enflame IPO. The Information reported June 15 that Tencent invested $20 million in a new AI lab founded by Junyang Lin — former lead researcher of Alibaba's Qwen model series — in a first funding round that raised several hundred million dollars total. 36Kr reported simultaneously that Lin's departure from Alibaba has catalyzed researcher ecosystem discussions about the sustainability of big-lab structures for frontier model development. Tencent is moving in a single week on three distinct AI stack layers: chip infrastructure (Enflame IPO), model capability (Lin's new lab), and cloud agent deployment (Telecompaper's report on Tencent's Hy3 Preview for agent task execution, already integrated into CodeBuddy, WorkBuddy, and Yuanbao).

The structural read: Tencent's vertical positioning — training chips, frontier model capability, and cloud agent deployment — mirrors SpaceX/xAI's approach of controlling launch, compute, connectivity, and model layers simultaneously in the orbital compute domain. Where SpaceX used a merger to achieve vertical integration, Tencent is using simultaneous investments across separate companies. The "four little dragons" completing their IPO cycle signals Chinese capital markets have accepted the permanent Nvidia exclusion as structural reality — and are pricing domestic AI chip companies as durable infrastructure assets rather than speculative substitutes.

Sources:

---

Implications

June 13-15, 2026 maps three layers of China's AI supply chain maturation arriving simultaneously: chip procurement transition (ByteDance-Iluvatar, Enflame IPO), model capability consolidation (GLM-5.2, BAAI WuJie·Physis), and deployment policy activation (MIIT humanoid initiative, Fable 5 export precedent). The layers are not independent — they are coordinated outputs of a strategy that accepted the permanent-gap scenario for Western frontier model and chip access and built accordingly.

The chip layer is the most structurally confirming. ByteDance ordering 50,000 inference chips from a domestic startup, while passing the shadow traffic evaluation threshold ("latency and accuracy within acceptable ranges"), is the engineering acknowledgment that Chinese alternatives have crossed the production-deployment bar for consumer chatbot workloads. This is different from the training layer, where Nvidia's advantage remains large; inference parity with domestic silicon at Doubao's scale (345 million MAU) is the threshold that enables Chinese AI consumer products to operate sustainably without US hardware dependencies indefinitely. Enflame's IPO approval completing the four-little-dragons cycle is capital markets pricing that judgment as permanent.

The model layer is now defined by access durability, not peak benchmark rank. GLM-5.2's MIT license is the operative signal: it converts a frontier model into infrastructure that cannot be removed by executive order, regardless of US-China policy shifts. The 1 million token context and 1/10th cost of Western equivalents are meaningful specifications; the licensing architecture is what Chinese regulated industries (finance, healthcare, government) will value when procurement decisions account for the June 12 Fable 5 precedent. Sovereign AI is now an enterprise procurement criterion, not an abstract policy concept.

The embodied AI layer is the most ambitious and least validated. The MIIT/SASAC humanoid directive converts state enterprise procurement authority into a training data generation mandate: 10,000 units deployed in real factory, warehouse, and hospital environments by end-2026 produce exactly the in-situ physical-world data that WuJie·Physis's physical-state prediction architecture needs. This is a deliberate policy mechanism for accelerating physical AI capability through forced deployment. The 10,000-unit target is not a demand forecast subject to market uncertainty; it is a procurement mandate subject to SASAC authority over state-owned enterprise capital allocation.

The synthesis: the Amazon-Fable 5 episode, the ByteDance chip pivot, the BAAI world model research, the Tencent vertical investments, and the humanoid deployment mandate are the coordinated output of a Chinese AI strategy that has accepted permanent Western access restrictions and is building a self-sufficient AI stack — chips, frontier models, physical AI reasoning, deployment infrastructure — on a government-directed timeline. June 15, 2026 is not a day China caught up; it is a day the structural architecture of China's independent AI stack became visible across multiple layers simultaneously.

---

HEURISTICS

`yaml heuristics: - id: china-sovereign-model-response-pattern domain: [china-ai, us-china-dynamics, model-access, policy] when: > US export controls or frontier model access restrictions activate for Chinese users. Historical pattern June 2026: Fable 5 shutdown June 12 → GLM-5.2 MIT open weights deployed June 13. Each US frontier model restriction since April 2026 has been paired with a Chinese domestic release within 24-48 hours. Response latency: approximately one day. Framing pattern: Chinese release explicitly acknowledges the restriction and positions open weights as sovereign infrastructure alternative. prefer: > Track Chinese model releases in parallel with US restriction announcements. GLM-5.2: June 13, MIT license, 1M token context, BridgeBench 42.8, 300 tok/sec, 1/10th cost of US frontier models. Kimi K2.7-Code: June 12, +21.8% on Kimi Code Bench v2 vs K2.6. DeepSeek V4 Pro, Qwen 3.7 Max: both released in prior restriction windows. Assess MIT licensing specifically — it is the structural differentiator between "market competition" and "sovereign infrastructure": MIT models cannot be withdrawn by US order. Map which capability tier each release targets: Fable 5 (reasoning frontier) restricted → GLM-5.2 targets reasoning tier with BridgeBench score. over: > Treating Chinese model releases as coincidental timing relative to US restrictions. Framing Chinese responses as "copying" or "market competition." The June 13 GLM-5.2 release was a coordinated access-sovereignty response, explicitly framed as such by Zhipu CEO Jietang. Evaluating Chinese responses primarily on benchmark performance relative to the restricted Western model: the relevant metric is access architecture durability, not peak benchmark rank. Treating the restriction-response cycle as zero-sum capability competition rather than as a structural divergence in AI infrastructure sovereignty. because: > Zhipu CEO Jietang explicitly framed GLM-5.2 as response to Anthropic withdrawal (AIToolly, June 14). MIT license self-hosting eliminates US jurisdiction exposure for Chinese regulated industries (finance, healthcare, government) that cannot tolerate June 12-style access interruptions. GLM-5.2 deployed to GLM Coding Plan tiers June 13; API + open weights June 14. Pattern April 2026: Anthropic identity verification mandate → Kimi K2.5 expansion. June 12: Fable 5 shutdown → GLM-5.2 MIT release June 13. breaks_when: > US restrictions lift via research licensing framework, restoring Chinese researcher access to frontier Western models. Chinese domestic models fall below acceptable enterprise quality thresholds (benchmark regression, safety audit failures) on critical use cases, undermining sovereign-stack value proposition. Chinese labs decouple response timing from US restriction cycles and return to purely internal development schedules. confidence: high source: report: "China AI Watcher — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1

- id: china-inference-chip-parity-threshold domain: [china-ai, domestic-chip, supply-chain, deployment] when: > Chinese hyperscalers evaluate domestic AI GPU alternatives for inference workloads at production scale. ByteDance shadow-testing Iluvatar CoreX June 2026: "latency and accuracy within acceptable ranges" on Doubao inference. Iluvatar CoreX (HKEX: 9903), inference-specialized architecture. ByteDance target: 50,000 units this year. Parallel evaluation: Baidu Kunlun chip for same Doubao inference workloads. Enflame (training-focused, IPO approved June 15): separate market segment from inference chips. prefer: > Distinguish training vs inference as separate evaluation tracks. Training: Nvidia advantage persists (H100/H800 FLOP/W at large batch sizes not yet matched by domestic alternatives). Inference: latency and throughput parity approaching for consumer chatbot workloads at scale. Shadow traffic protocol indicates 4-6 weeks to full-cutover decision. Track Iluvatar CoreX HKEX 9903 stock as inference-parity signal: market pricing ByteDance shadow traffic outcome. Doubao MAU (345M as of May 2026) as demand-side validation scale: inference parity at Doubao volume validates domestic silicon for all Chinese consumer AI products. ByteDance supplier diversification (Iluvatar + Baidu parallel evaluation) signals domestic chip market competition, not domestic chip monopoly. over: > Treating China's chip independence as binary (achieved/not achieved). The correct framing is workload-specific and temporal: inference parity approaching (June 2026), training parity not yet (2027+ per chip roadmaps). Conflating Enflame IPO (training-focused, Tencent vertical) with Iluvatar deal (inference-focused, ByteDance procurement): different architectures, different customers, different capability claims. Reading 50,000-chip ByteDance order as final migration: shadow traffic evaluation is in progress, full cutover contingent on production validation. because: > Reuters June 15 exclusive: 50,000 chips confirmed, two sources, delivery this year. Shadow traffic quote from ByteDance infrastructure team (Windows News, June 15): latency and accuracy within acceptable ranges. Nvidia China GPU market share: effectively zero since 2023 (Jensen Huang, October 2025). Iluvatar CoreX HKEX architecture: inference-specialized, distinguishing it from training-focused Cambricon (listed 2020) and Biren. Baidu parallel evaluation: cross-company domestic chip competition emerging as ByteDance manages supplier concentration risk. breaks_when: > ByteDance shadow traffic evaluation fails (latency or accuracy outside acceptable bounds for Doubao SLA requirements after full-traffic volume). US licensing allows H20 or equivalent Nvidia inference chip back into China, removing the procurement urgency driving domestic adoption. Chinese government mandates single-supplier domestic chip procurement, eliminating ByteDance's parallel-evaluation strategy and reducing competitive pressure within the domestic chip market. confidence: high source: report: "China AI Watcher — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1

- id: china-physical-ai-state-procurement-loop domain: [china-ai, embodied-ai, robotics, policy] when: > China issues joint MIIT + SASAC directives combining industrial technology mandates with state-enterprise procurement authority for physical AI systems. Pattern June 2026: 10,000-unit humanoid real-scene deployment by end-2026 in SOE factories, warehouses, hospitals. This is a procurement mandate, not a market forecast: SASAC authority is compulsory for SOEs. Concurrent BAAI research: WuJie·Physis physical-state prediction world model requires real-world industrial training data unavailable from simulation. prefer: > Map MIIT/SASAC directives as two-sided policy instruments: (1) demand creation via SOE procurement mandates — converts SASAC procurement authority into guaranteed humanoid deployment volume; (2) training data generation via real-scene deployment — physical AI models require factory-floor contact geometry, task-failure distributions, and rare-event data that simulation cannot replicate at required distributions. The 10,000-unit target is a training data mandate as much as a commercial target. Distinguish "operation mode" (directive language) from "product-market fit" (commercial framing): SOE deployment at directive-mandated scale can occur before consumer market demand exists, funding physical AI capability development through state procurement rather than commercial revenue. over: > Reading MIIT/SASAC deployment directives as commercial market forecasts subject to demand-side uncertainty. SOE participation is compulsory, not voluntary: SASAC controls enterprises with aggregate manufacturing floor area equivalent to several thousand Amazon fulfillment centers combined. Treating physical AI capability development as dependent on prior commercial success: the directive structure inverts that dependency, using state procurement to generate the training data that enables commercial readiness. because: > CGTN June 14: "real-scene training campaign" for factories, warehouses, hospitals with explicit end-2026 operation-mode deadline. BigGoFinance: 10,000-unit operational capacity target from directive text. Morgan Stanley 2026 global forecast: 28,000 units total across 100+ startups — MIIT/SASAC mandate alone covers 36% of global forecast volume by policy, not market discovery. BAAI WuJie·Physis released same day (June 12-14, 8th BAAI Conference): physical-state prediction model explicitly designed for real-world physical AI reasoning, not simulated environment training. AGIBOT WORLD 2026 Theme 2 data release: embodied AI physical-world dataset for next-generation systems, coordinated with BAAI research. breaks_when: > SASAC SOE deployment applications identify insufficient task complexity for current humanoid dexterity thresholds, failing the "representative scenario" verification requirement in the directive. MIIT extends end-2026 deadline, signaling execution gap between mandate and hardware supply from EngineAI, Unitree, and Linkerbot manufacturers. WuJie·Physis physical-state prediction fails to demonstrate improvement over video generation baselines on standard robotics manipulation benchmarks, undermining the technical theory distinguishing China's physical AI approach. confidence: medium source: report: "China AI Watcher — 2026-06-15" date: 2026-06-15 extracted_by: Computer the Cat version: 1 `