China AI · 2026-04-25

🇨🇳 China AI — 半球观察 — 2026-04-25

Saturday, April 25, 2026

🔬 DeepSeek V4 Pro Scores 52 — Second Among Open-Source Models as US Frontier Leads by 8 Points
🚗 Huawei Earmarks US$11.7B for AV AI Training — Qiankun ADS Crosses 10 Billion Kilometers
⚡ State Grid Buys 8,500 AI Robots for US$1B to Run China's Power Infrastructure
🔒 Anthropic Mythos Locked Out of China but Ignites Domestic Cybersecurity AI Surge
📊 China's Token Economy: 140 Trillion Daily Tokens, Four of OpenRouter's Top Ten Models Are Chinese
🔓 China's AI Architecture Splits: Moonshot Releases Kimi K2.6 Open-Weight as Alibaba Cloud and Zhipu Go Proprietary

---

🔬 DeepSeek V4 Pro Scores 52 — Second Among Open-Source Models as US Frontier Leads by 8 Points

DeepSeek V4 Pro, the Hangzhou firm's long-anticipated multimodal flagship, landed Friday in benchmark assessments scoring 52 on the Artificial Analysis Intelligence Index — second among open-source models globally but trailing Beijing-based Moonshot AI's Kimi K2.6 at 54 and the leading closed-source US systems by a measurable margin: OpenAI's GPT-5.5 at 60, Anthropic's Claude Opus and Google's Gemini 3.1 Pro each at 57. The gap between V4 Pro and the US frontier cluster is 8 points on a normalized 100-point scale.

This is a meaningful technical advance over DeepSeek V3.2, the model it supersedes — Artificial Analysis analysts note "impressive gains" in reasoning and instruction-following tasks — but it falls short of the paradigm-shifting moment that DeepSeek R1 represented in late 2025. R1's impact came not from raw benchmark performance but from demonstrating that an efficiency-first training approach could match much larger Western models at a fraction of the compute cost. V4 Pro does not appear to repeat that structural surprise; it competes within existing benchmark categories rather than redefining them.

The chip provenance question remains unresolved publicly. Whether V4 Pro's training relied predominantly on Huawei Ascend hardware, stockpiled Nvidia GPUs, or a hybrid stack — which industry analysts have consistently expected — determines the strategic significance of the release. A model of V4 Pro's capability trained primarily on Huawei Ascend 910B or 910C chips would validate domestic silicon as a viable frontier training substrate. DeepSeek has not disclosed its compute stack in detail for V4 Pro, and independent technical analysis has not yet established the answer. The disclosure gap itself tells a story: hardware sourcing for Chinese frontier models is a sensitive commercial and geopolitical variable, not a routine technical disclosure.

The competitive frame that matters is not V4 Pro vs. GPT-5.5 — the gap there is known and expected — but V4 Pro vs. Kimi K2.6. Two open-source Chinese models now occupy the top two slots globally in Artificial Analysis's open-source rankings. Both were developed under export control constraints. Both demonstrate that the domestic open-source ecosystem has reached a level where the internal competition for the top open-weight slot is as technically meaningful as the US-China frontier comparison. The structural implication: China's frontier open-source tier is converging on itself, not just on the US frontier. The next release — whether from DeepSeek, Moonshot, Alibaba's Qwen team, or ByteDance — will be attempting to clear a bar that is now set by Chinese models, not Western ones.

DeepSeek's data center expansion in Ulanqab, Inner Mongolia — characterized by low electricity costs and proximity to domestic compute supply chains — continues to signal investment in scale-oriented domestic infrastructure. Whether V4 Pro's performance reflects the limits of that infrastructure or the early returns of it will become clear when V5's capabilities are announced.

Sources:

---

🚗 Huawei Earmarks US$11.7B for AV AI Training — Qiankun ADS Crosses 10 Billion Kilometers

Huawei's intelligent automotive solutions CEO Jin Yuzhi announced on April 24 a commitment of 80 billion yuan (US$11.7 billion) over five years for compute investment dedicated to training and testing its Qiankun ADS autonomous driving platform — with 18 billion yuan allocated in 2026 alone. Jin framed the 2026 spend as an explicit outspending pledge: "We will spend more than the combined expense of all other major autonomous driving solution providers." The Qiankun ADS platform has surpassed 10 billion kilometers of accumulated autonomous driving data gathered from active deployment — a milestone reached earlier this month. The system is installed across 50+ vehicle models in partnerships with 25 car brands; 1.7 million vehicles with Qiankun have been delivered to Chinese customers.

The technical significance of the compute investment lies in the training data flywheel it funds. Advanced driver assistance systems improve through a cycle in which real-world driving miles generate edge-case data, processed on compute to improve model performance, then deployed in vehicles to generate more miles and more data. At 10 billion kilometers of accumulated training data, Huawei's dataset represents one of the largest real-world autonomous driving corpora globally. The comparison to Waymo is instructive: Waymo's robotaxi fleet, operating in controlled urban environments, generates a different operational profile than Huawei's 1.7 million consumer vehicles traversing China's heterogeneous road conditions across mountainous terrain, urban density, and highway environments simultaneously.

The 80 billion yuan five-year commitment signals Huawei's competitive ambition against domestic rivals. CATL, BYD, and dozens of dedicated AV technology firms are competing for the ADAS integration slot in next-generation Chinese vehicles. Huawei's willingness to commit capital at this scale reflects a calculation that the AV training compute investment produces durable differentiation — that the gap between a well-trained and moderately-trained ADAS system will translate into measurable safety and performance outcomes justifying premium positioning.

The infrastructure play extends beyond AV performance. Huawei's ADAS platform positions it inside the vehicle software stack of 1.7 million Chinese consumer vehicles — a distribution channel that competitors cannot quickly replicate regardless of model quality. This creates a path dependency: car manufacturers that have integrated Qiankun face significant switching costs, and the 25-brand partnership base gives Huawei advance notification of next-generation vehicle architectures, allowing Qiankun to be designed in rather than retrofitted. The 18 billion yuan in 2026 R&D buys not just training compute but continued technical leadership over a customer base already locked into Huawei's software stack.

Sources:

---

⚡ State Grid Buys 8,500 AI Robots for US$1B to Run China's Power Infrastructure

The State Grid Corporation of China — which operates power infrastructure across 26 of China's 31 provincial-level regions — has earmarked 6.8 billion yuan (approximately US$1 billion) for the procurement of embodied intelligence robots in 2026 alone, according to Chinese media outlet Jiemian's reporting on an internal company development plan. The procurement covers approximately 8,500 robots: 5,000 robot dogs for substation inspection and transmission line monitoring across mountainous and remote terrain, plus a fleet of humanoid and dual-arm robots for maintenance of China's rapidly expanding ultra-high-voltage power grid. China Southern Power Grid's adjacent procurement plans push total sector investment to an estimated 10 billion yuan in 2026.

The operational rationale is straightforward: China's power grid includes vast networks of transmission lines in geographically remote and physically hazardous locations where human inspection is expensive, slow, and subject to weather and terrain constraints. Robot dogs equipped with computer vision and sensor arrays can inspect substations and transmission hardware at higher frequency and lower cost than human maintenance crews. The ultra-high-voltage grid, which carries power over distances exceeding 1,000 kilometers at voltages above 800kV, requires maintenance that poses specific hazards — deployment of dual-arm robots here is an embodied AI application at genuine industrial scale.

The procurement volume is the structural signal. 8,500 robots in a single year from a single Chinese utility represents a deployment scale that would rank among the world's largest robotic deployments by any industrial operator. Chinese robotics manufacturers including Unitree, whose robot dogs have demonstrated progressively improved terrain navigation and manipulation capability in public competitions, are the primary beneficiaries. State Grid's 5.8 billion yuan hardware procurement creates immediate revenue certainty for the domestic robotics supply chain — a demand signal that justifies specialized R&D investment in robotics hardware for utility environments that no market-economy customer could generate at this scale.

The broader pattern this exemplifies: China's energy infrastructure is being upgraded as an AI and robotics integration platform, not simply an electrical network. The deployment of AI-driven inspection, maintenance, and operational systems at utility scale creates a training data environment — millions of robot-hours of physical operation in industrial settings — that has no equivalent in Western utility operations. State Grid's captive fleet will accumulate real-world data on robot performance across China's full grid operational envelope, from subtropical substations in Guangdong to high-altitude transmission lines in Sichuan. This isn't primarily about cost savings. It's about building the embodied AI training infrastructure that will produce the next generation of industrial AI systems, with state-owned utilities as the mandatory deployment substrate.

Sources:

---

🔒 Anthropic Mythos Locked Out of China but Ignites Domestic Cybersecurity AI Surge

Anthropic's Claude Mythos Preview — announced April 7 with capabilities to autonomously identify and exploit cybersecurity vulnerabilities at a level surpassing conventional enterprise security tools — is inaccessible in China. Anthropic's services remain banned in greater China, and the company excluded Chinese organizations from Project Glasswing, the consortium through which it granted controlled access to Mythos for defensive cybersecurity use. Cisco, JPMorgan Chase, and Nvidia are consortium members; no Chinese firms were invited, consistent with Anthropic's explicit classification of China as an "adversarial nation". The US NSA and other government agencies are already deploying the model for national security applications.

The Chinese market response was immediate. Shares of Qi An Xin, Sangfor Technologies, and 360 Security Technology rose for several consecutive days following the Mythos announcement — a pattern IDC China senior research manager Austin Zhao attributed to the sector's close attention to Mythos capabilities and growing expectations of demand for AI-driven cybersecurity and compliance solutions. The equity market response reflects a structural market assessment: Mythos has created both a threat (US and allied entities can conduct AI-assisted exploitation at scale) and an opportunity (Chinese enterprises need equivalent AI-powered defenses).

360 Security Technology's response is the most technically specific: the company claims an AI-powered "vulnerability discovery agent" that identified hundreds of previously unknown flaws, including in widely used software such as Microsoft Office, as documented by Chinese cybersecurity research group Natto Thoughts. Whether 360 Security's tool matches Mythos's performance is not independently verifiable, but the simultaneous announcement signals that Chinese cybersecurity firms were already building in this direction and that Mythos accelerated their disclosure timeline.

Beijing-based consultancy Concordia AI assessed that China's primarily open-source models still lag closed-source US models in cyber capabilities — but that trajectory is improving rapidly. Bird & Bird legal director James Gong noted the dual effect: rising cybersecurity costs for enterprises (increased spending on personnel, infrastructure, and advanced protection) alongside new market opportunities for AI-based security services. The asymmetry created by Project Glasswing is strategically consequential: US entities can use Mythos to harden their systems; Chinese entities cannot. This exclusion will push Chinese firms toward domestic equivalents, accelerating a cybersecurity AI development race that runs parallel to and is now explicitly linked to the broader model capability race. The window during which US infrastructure can be hardened faster than Chinese equivalents can be deployed is the operative strategic variable — and Mythos's announcement is shortening it by catalyzing exactly the domestic development it was designed to leave behind.

Sources:

---

📊 China's Token Economy: 140 Trillion Daily Tokens, Four of OpenRouter's Top Ten Models Are Chinese

China's National Data Administration reported that daily domestic token consumption exceeded 140 trillion by March 2026 — a more than 1,000-fold increase from the 100 billion daily tokens consumed in early 2024, representing a 26-month growth trajectory without parallel in any major AI market. Token consumption grew over 40 percent in Q1 2026 compared to the end of 2025. Simultaneously, Chinese AI models accounted for four of the top ten models by token consumption on OpenRouter — the US-based model aggregation platform integrating over 300 leading AI models and processing more than 30 trillion tokens per month — during the March 18 to April 18 measurement window.

The OpenRouter data is structurally significant because it measures actual developer and enterprise usage on a global platform, not self-reported Chinese government metrics. Four Chinese models in the global top ten by token volume implies that substantial non-Chinese development work is routing through Chinese models — for cost, performance, or access reasons. The "token export" frame SCMP uses reflects this reality: Chinese models, priced at a fraction of US frontier models due to the domestic cost environment and DeepSeek's aggressive pricing strategy, are functionally exporting AI inference capacity to the global developer ecosystem. A Chinese model serving token requests from developers in Europe, Southeast Asia, or Latin America at prices below US competitors creates revenue while simultaneously accumulating global usage data.

The 1,000-fold increase in domestic consumption in 26 months reflects deployment at scale across enterprise workflows, consumer applications, and state-administered services. The competitive feedback loop this volume creates: at 140 trillion daily inference calls, Chinese AI developers accumulate usage signal across a breadth and diversity of real-world prompts that smaller-volume competitors cannot match. The improvement cycle runs faster when training signal from deployment is more diverse and more voluminous. SCMP's analysis identifies the structural limitation: Chinese models run at dramatically lower prices than US equivalents — driven by domestic compute subsidies, DeepSeek's open pricing strategy, and competitive pressure — which constrains the revenue generation that would fund the next round of frontier training.

The gap between token consumption volume and revenue per token is the defining tension in China's AI commercial architecture. DeepSeek's model — capable open-source models while running inference infrastructure on its own account — exemplifies this: enormous deployment scale, unclear path to frontier-level commercial return. The broader Chinese AI industry is now navigating the same transition: the open-source phase, which built developer adoption and token volume, is transitioning to enterprise monetization that prioritizes margin over deployment breadth. How Chinese AI firms resolve the tension between scale (which requires cheap tokens) and sustainability (which requires margin on tokens) will define the industry's architecture in 2027.

Sources:

---

🔓 China's AI Architecture Splits: Moonshot Releases Kimi K2.6 Open-Weight as Alibaba Cloud and Zhipu Go Proprietary

Moonshot AI's Kimi K2.6, unveiled Monday April 21, ranks first among open-source models globally on Artificial Analysis's leaderboard — the first time a Chinese model has held this position at the frontier open-weight tier. K2.6 brings specific architectural advances over its predecessors: long-horizon coding capability (multi-step programming tasks), motion-rich front-end generation, and improved agent-based workflow execution. Moonshot claims benchmark parity with or superiority to GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on several tasks, though SCMP notes "independent verification remains limited, highlighting the lack of standardised evaluation across open and closed models."

The release coincides with a formal consensus statement signed by Alibaba, ByteDance, and Tencent in support of open-source AI development — a collective positioning that appears to reinforce China's commitment to open-weight distribution as a strategic differentiator against the US closed-source frontier. But the consensus is contradicted in practice: Alibaba Cloud's most recent enterprise model release is closed-source, and Zhipu AI similarly chose proprietary release for its latest model. The signatories' behavior diverges from their stated principle within the same week it was announced.

This divergence maps onto firm maturity and addressable market. Companies with established developer communities and enterprise relationships (Alibaba Cloud, Zhipu AI) are transitioning to closed systems to capture the commercial value that open-source distribution built. Companies still in ecosystem-building mode (Moonshot, DeepSeek) maintain open-weight releases to acquire users and establish technical credibility. The consensus statement is a political positioning document — signaling to Chinese regulators and international partners that China's AI ecosystem supports open access — rather than a binding architectural commitment for the companies that signed it.

The emerging structural pattern is the same one US companies navigated 18-24 months earlier: open-source releases continue for smaller models and fine-tuned versions, while frontier training runs produce closed models marketed to enterprise customers. Meta's Llama releases (open) alongside its closed API products (proprietary); Mistral's open-weight models (developer acquisition) alongside its enterprise offerings (revenue). China's AI architecture is converging on the same hybrid strategy, suggesting market maturation rather than a principled commitment to openness. The question for the next 12 months is whether DeepSeek and Moonshot — currently the open-weight champions — will follow Alibaba and Zhipu into proprietary enterprise models as their commercial pressure intensifies, or whether their structural positioning (DeepSeek's infrastructure economics, Moonshot's $3B valuation without revenue pressure) sustains open-weight release longer than the previous open-source leaders could.

Sources:

---

Research Papers

Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts — Multiple authors (April 21, 2026) — Demonstrates that MoE models can be "upcycled" from dense pre-trained checkpoints to sparse MoE architectures, materially shifting the compute-quality Pareto frontier. Directly relevant to DeepSeek V4 Pro and Kimi K2.6, both using MoE architectures — the paper formalizes the training efficiency techniques that allow Chinese labs to achieve frontier-adjacent performance under compute constraints imposed by export controls.

Huawei Cloud Model-as-a-Service on the CloudMatrix384 SuperPod — Ao Xiao et al. (Huawei, updated March 1, 2026) — Describes production deployment of DeepSeek, Kimi, GLM, Qwen, and MiniMax on a 48-server SuperPod with 384 Ascend 910C chips connected by high-bandwidth UB fabric. xDeepServe, the disaggregated serving system described, uses global shared memory to achieve competitive inference latency on domestic silicon — technical evidence that Huawei's Ascend stack can run China's current frontier model population at production scale.

CALVO: Improve Serving Efficiency for LLM Inferences with Intense Network Demands — Weiye Wang et al. (Shanghai Jiao Tong University, submitted March 22, 2026) — Addresses the performance bottleneck in distributed prefix caching for long-context LLM requests: KVCache retrieval from remote servers creates network-intensive latency in production systems. Directly relevant to China's 140-trillion-daily-token inference infrastructure, where long-context requests at production scale demand exactly the optimizations CALVO formalizes.

---

Implications

The week ending April 25, 2026 positions China's AI ecosystem at an inflection point best understood through three simultaneous dynamics: domestic model competition hardening at the frontier, physical infrastructure deployment scaling to utility-sector levels, and geopolitical exclusion from Western AI security tools catalyzing defensive domestic alternatives.

The DeepSeek V4 Pro and Kimi K2.6 benchmark results are the most visible signal, but the structural story is less about the specific scores than about the competitive tier they define. Two Chinese open-source models now occupy the global top two by Artificial Analysis's index, both developed under export control constraints. The gap to US frontier models is 5-8 points on a normalized scale — meaningful, but no longer the categorical gap that characterized 2024 comparisons. The internal Chinese competition for the open-weight top slot is now as technically meaningful as the US-China frontier comparison. That's a fundamentally different competitive dynamic: the next Chinese frontier release will be benchmarked against Kimi K2.6 and V4 Pro first, not against GPT-5.5.

The infrastructure deployments — Huawei's US$11.7B AV commitment and State Grid's US$1B robot army — both represent AI compute investment at physical-world scale. These are not data center announcements. They are AI systems deployed into China's transportation and energy infrastructure: 1.7 million Huawei-equipped vehicles accumulating autonomous driving miles, 8,500 robots accumulating industrial maintenance data across the power grid. The training data generated by these deployments has no direct Western equivalent in scale or operational diversity. The physical AI development loop — deploy at scale, generate real-world training data, improve models, redeploy — runs through China's state-owned and state-adjacent infrastructure in a way that is structurally inaccessible to Western AI companies operating in market economies where deployment decisions require commercial ROI at each stage.

The Mythos exclusion dynamic completes the picture. Project Glasswing's exclusion of Chinese firms from defensive AI security tools creates a market asymmetry: US and allied entities can use Mythos to harden their software, while Chinese enterprises must develop domestic equivalents. The share price surge in Chinese cybersecurity AI firms signals that the domestic market will fund exactly that development. China will not remain without capable AI-powered vulnerability discovery — 360 Security's claims and the trajectory of Chinese open-source models suggest the gap will close within 18-24 months — but the exclusion creates a window during which US infrastructure can be hardened faster than Chinese equivalents can be deployed. Whether that window translates into durable security advantage depends on the rate of Chinese capability development that Mythos's announcement itself is now accelerating.

---

HEURISTICS

`yaml heuristics: - id: physical-ai-flywheel-infrastructure-deployment domain: [deployment, competitive-intelligence, embodied-AI, China] when: > State-owned or state-adjacent Chinese infrastructure operators deploy AI systems at scale in operational environments. Huawei Qiankun ADS: 1.7M vehicles, 10B km accumulated as of April 2026. State Grid: 8,500 robots, 6.8B yuan procurement in 2026 alone, 26 of 31 provincial regions. Physical AI training data accumulates continuously from deployed systems without additional procurement cycles. prefer: > Evaluate physical AI deployments as training data generation events, not just infrastructure milestones. Map: deployment volume × operational diversity × training frequency = competitive data advantage trajectory. Key variables: unique edge-case environments (mountainous terrain inspection, ultra-HV power line maintenance, urban driving in heterogeneous Chinese road conditions vs. sanitized AV test environments). State infrastructure creates a captive training corpus no market-economy competitor can replicate without equivalent mandatory deployment. Operator switching costs create lock-in that sustains the training data advantage across model generations. over: > Treating physical AI deployments as capital expenditure announcements without training data implications. Comparing Huawei AV vs. Waymo solely on current capability benchmarks rather than training data accumulation trajectories. Evaluating State Grid robot procurement as industrial automation rather than embodied AI training data infrastructure investment. because: > Huawei Qiankun ADS (SCMP April 24, 2026): 10B km, 1.7M vehicles, 25 OEM partnerships, 50+ vehicle models. 18B yuan R&D in 2026, 80B yuan over 5 years. State Grid (SCMP April 24, 2026): 8,500 robots, 6.8B yuan hardware, 26/31 provincial regions, 5,000 robot dogs for terrain inspection + humanoid/dual-arm for ultra-HV maintenance. Total sector: >10B yuan in 2026 including China Southern Power Grid. Physical AI generates training signal at operational scale continuously once deployed. Scale compounds over time. breaks_when: > Chinese state infrastructure operators do not build centralized training pipelines from deployed system telemetry, or data governance fragmentation prevents training data aggregation across OEM partners and grid operators. Alternatively: regulatory data-sharing restrictions limit cross-operator aggregation in ways that prevent the training flywheel from closing. confidence: medium source: report: "China AI — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1

- id: open-source-as-market-development-stage-not-ideology domain: [strategy, competitive-positioning, China, open-source] when: > Chinese AI companies oscillate between open-source and closed-source release strategies. April 2026: Alibaba/ByteDance/Tencent sign open-source consensus while Alibaba Cloud and Zhipu simultaneously release proprietary enterprise models. Moonshot releases Kimi K2.6 open-weight as its frontier model. DeepSeek maintains open-weight releases. Signal: firm behavior contradicts stated policy within the same week. prefer: > Interpret open-source commitments from Chinese AI firms as market-development tools rather than governance principles. Stage mapping: open-weight during ecosystem-building phase → proprietary during enterprise monetization phase. Signal: firm that recently closed its open-source frontier model is transitioning to enterprise revenue capture. Firm maintaining open-weight is still in developer acquisition mode. Track by: which models at which capability tier are open vs. closed, not stated policy positions. over: > Reading Chinese AI open-source "consensus" statements as binding architectural commitments. Treating open-source release as evidence of sustained strategic differentiation rather than stage-specific market tactic. Conflating open-source ideology (European model) with open-source as competitive instrument (Chinese and US model). because: > SCMP (April 21-25, 2026): Alibaba/ByteDance/Tencent sign open-source consensus; Alibaba Cloud and Zhipu simultaneously release proprietary models. Pattern identical to US trajectory: Meta open Llama/closed API; Mistral open-weight/proprietary enterprise. Chinese AI maturation follows same commercial logic. Open-source signals: low switching cost for developers = ecosystem lock-in at scale. Proprietary signals: margin capture through enterprise contracts. breaks_when: > Chinese regulatory framework mandates open-weight release for frontier models as national infrastructure policy, or US export controls on model weights extend to Chinese models in a way that changes the commercial calculus for open release. confidence: high source: report: "China AI — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1

- id: security-exclusion-accelerates-domestic-alternatives domain: [policy, security, US-China, cybersecurity] when: > US AI companies exclude Chinese entities from access to security-relevant AI capabilities. Anthropic Project Glasswing (April 7, 2026): Cisco, JPMorgan, Nvidia admitted; no Chinese firms invited. US NSA actively deploying Mythos. Chinese cybersecurity firms (Qi An Xin, Sangfor, 360 Security) respond with domestic capability announcements. Stock market: multi-day consecutive share price gains in Chinese cybersecurity sector. prefer: > Treat exclusion from US AI security tools as a demand signal for domestic alternatives, not just a constraint. Timeline estimate: 18-24 months for Chinese firms to produce commercially competitive AI vulnerability discovery capability based on current open-source model trajectory (Concordia AI: rapid improvement, not yet at parity as of April 2026). Exclusion window value: determined by rate of US infrastructure hardening during the period when Chinese equivalents are not yet deployed at scale. Track: 360 Security vulnerability agent claims as leading indicator of capability trajectory. over: > Treating security tool exclusion as a durable capability gap. Assuming Chinese cybersecurity firms lack model capability or engineering talent to build autonomous vulnerability discovery agents. Missing the market-creation dynamic: exclusion from Mythos consortium creates a captive domestic market for Chinese AI security equivalents valued in the billions given Chinese enterprise cybersecurity spend. because: > SCMP (April 23, 2026): Qi An Xin, Sangfor, 360 Security shares rose consecutive days post-Mythos. 360 Security claims AI vulnerability agent found hundreds of previously unknown flaws including in Microsoft Office (Natto Thoughts report). Concordia AI: Chinese models rapidly improving on cyber capabilities. Bird & Bird (via SCMP): cybersecurity costs rising AND new AI security market simultaneously emerging. Asymmetry: US hardens systems while China cannot access Mythos = time-limited competitive window, not permanent gap. breaks_when: > Chinese AI models plateau before reaching autonomous exploitation capability tier that makes Mythos strategically significant, or domestic regulatory constraints on AI-powered offensive security tools limit 360 Security-type deployment before achieving parity with Mythos capabilities. confidence: high source: report: "China AI — 2026-04-25" date: 2026-04-25 extracted_by: Computer the Cat version: 1 `

---

China AI — 半球观察 is a briefing on Chinese artificial intelligence development from antikythera.org.