China AI · 2026-05-24

🇨🇳 China AI — 2026-05-24

📉 DeepSeek Makes Permanent 75% Price Cut on V4-Pro, Fixing Input at $0.435/M
🚫 China Bans Nvidia RTX 5090D V2 During Jensen Huang's Beijing Visit, Pushes Domestic Chip Stack
⏸️ Trump Postpones AI Executive Order Citing Risk to US Lead Over China
🧠 DeepSeek V4 Architecture: CSA+HCA Hybrid Attention, Manifold Hyper-Connections, Muon Optimizer
🤖 Tencent Hy3 Enters WeChat Integration Testing as AI Super-Apps Remake Chinese Internet
🇺🇸🤝🇨🇳 Trump-Xi Summit Produces AI Safety Protocol Commitment on Nonstate Actor Access

---

📉 DeepSeek Makes Permanent 75% Price Cut on V4-Pro, Fixing Input at $0.435/M

DeepSeek announced on May 23 that it is making permanent a 75% price cut on V4-Pro, its flagship model, fixing input tokens at $0.435 per million and output at $0.87 per million — down from $1.74/M input and $3.48/M output at the model's April 2026 launch. The cut was initially temporary; the May 23 announcement removes the expiration condition. Cache-hit input is fixed at $0.003625/M — approximately 100× cheaper than non-cached input.

The pricing structure is structurally significant for the global inference market. Codersera's analysis notes that V4-Pro achieves 80.6% on SWE-bench Verified — a benchmark for software engineering tasks — at a per-token price that is approximately 1/20 of Claude Opus 4.7's equivalent pricing. The cache-hit pricing of $0.003625/M is designed specifically for high-frequency API deployment patterns where repeated context reuse is the norm, making V4-Pro economically superior to frontier Western models for enterprise applications that re-send large system prompts.

The competitive signal is directional, not just about price. DeepSeek has now permanently anchored V4-Pro's pricing below the threshold at which Western labs' enterprise sales logic functions. OpenAI and Anthropic both price on the assumption that frontier capability commands frontier margins; DeepSeek's May 23 commitment removes the "this is a temporary promotion" hedge that allowed Western enterprise teams to treat DeepSeek as a temporary anomaly. It is now a permanent pricing floor.

The open-weights policy amplifies the effect. V4-Pro ships with open weights and a 1M-token context window, meaning enterprises can deploy it on-premise at effectively zero inference cost beyond hardware. The $0.435/M API price is thus an upper bound on V4-Pro deployment cost, not the typical case for enterprise users who self-host. This combination — open weights, 1M context, 80.6% SWE-bench, $0.435/M API — makes a strong competitive argument for Chinese frontier model adoption in markets where vendor lock-in risk and data sovereignty concerns make Western API access politically problematic.

BenchLM's independent benchmark tracking places V4-Pro at 88 (Max) on its composite scoring methodology, behind frontier models like Claude Opus 4.7 and GPT-5 but competitive with models priced at 3-5× the rate. The 75% permanent price cut closes the value-per-dollar gap entirely for cost-sensitive deployments.

Sources:

---

🚫 China Bans Nvidia RTX 5090D V2 During Jensen Huang's Beijing Visit, Pushes Domestic Chip Stack

China added Nvidia's RTX 5090D V2 gaming GPU to a list of banned goods at customs checkpoints last Friday — while Jensen Huang was physically in Beijing accompanying President Trump's diplomatic delegation. The timing was deliberate. The RTX 5090D V2 is a China-market-specific version of Nvidia's flagship consumer GPU, engineered to comply with US export controls by staying below the restricted performance thresholds. Banning it signals that Beijing considers even the degraded, export-control-compliant variant unacceptable.

Tom's Hardware analysis frames the move precisely: China's message is that it does not need Nvidia's "de-fanged" chips. The ban, paired with the existing restriction on H200s and H20s from earlier rounds of export control escalation, closes the remaining commercial pathway for Nvidia in China's high-performance computing market. Gaming GPUs have functioned as a secondary compute source for Chinese AI developers unable to access H100/H200 hardware — the 5090D V2 ban eliminates that workaround.

The domestic hardware context makes the signal legible. CNBC reported on May 21 that Jensen Huang stated Nvidia has "largely conceded" the Chinese AI chip market to Huawei. Huawei's Ascend 950PR delivers 1.56 petaflops with 112GB HBM at a price point ($6,000 per chip in domestic markets) that undercuts H100 variants. Cambricon and Alibaba's Hanguang line provide complementary positioning for inference-optimized and edge deployments respectively.

The policy logic is a layered substitution play. Export controls restrict Nvidia's most capable datacenter chips. The 5090D V2 ban closes the gaming-chip-as-AI-compute arbitrage pathway. Simultaneously, Beijing pushes Chinese AI companies toward Ascend hardware through procurement preferences and subsidy programs. The result is not that Chinese AI labs can't get compute — they can, through Huawei — but that the performance profile of that compute differs from Nvidia's, shaping which model architectures are competitive on domestic hardware.

The geopolitical timing is the most interesting dimension. Banning a Nvidia chip while Jensen Huang stands in Beijing is a statement about who has leverage. China's domestic chip stack is sufficiently developed that it can afford the political theater; banning the 5090D V2 during Trump's visit imposes no operational cost on Chinese AI labs while making a highly visible domestic independence claim.

Sources:

---

⏸️ Trump Postpones AI Executive Order Citing Risk to US Lead Over China

President Trump postponed signing an AI executive order on May 21 — hours before the scheduled ceremony — stating explicitly that he "didn't like certain aspects of it" and did not want to "do anything to get in the way of that lead" over China. Politico reported that the postponement followed internal concerns that specific provisions of the order could constrain US AI development in ways that disadvantaged US companies relative to Chinese competitors with fewer domestic regulatory burdens.

The structural signal is that Trump's AI policy posture is now explicitly calibrated against China's trajectory, not against domestic risk governance. The executive order was reportedly designed to accelerate AI infrastructure deployment, streamline federal AI procurement, and establish compute access frameworks — none of which are obviously restrictive of US AI capability. The fact that Trump perceived specific provisions as competitively dangerous reveals that the administration is operating in a mode where any friction to AI deployment, even procedural, is treated as a potential competitive concession to China.

Axios reported that White House officials described the delay as temporary, with revised text expected within weeks. But the postponement has a governance consequence: the delay leaves unresolved the federal frameworks for AI procurement, compute access, and safety standards that multiple agencies are waiting on. Projects in the pipeline at NSF, DoD, and DHS that required executive order authorization are now in a holding pattern of undefined duration.

For China's AI policymakers, the postponement is a signal of the mechanism of US AI governance: competitive pressure from China is now a lever that can delay US regulatory frameworks, not just accelerate them. The Domino Theory analysis notes that the Trump-Xi AI safety dialogue agreement, reached at the Beijing summit, now runs in parallel with a domestic US regulatory vacuum — the US is committing to international AI safety coordination with China while delaying the domestic framework that would define what "safety" means in US terms.

The Brookings analysis published in advance of the Trump-Xi summit notes that bilateral AI cooperation faces a fundamental verification problem: neither country can credibly monitor the other's compliance with AI safety commitments in the absence of technical standards definitions. The postponed executive order was reportedly designed to address exactly that definitional gap on the US side. Its delay extends the window during which "AI safety cooperation" with China is a diplomatic phrase without operational content.

Sources:

---

🧠 DeepSeek V4 Architecture: CSA+HCA Hybrid Attention, Manifold Hyper-Connections, Muon Optimizer

DeepSeek V4's architecture, documented in its April 2026 technical release and analyzed extensively this week following the permanent pricing announcement, introduces three novel architectural components that collectively reduce training compute and inference overhead relative to standard transformer designs. The components — Compressed Sparse Attention (CSA), Heavily Compressed Attention (HCA), and Manifold-Constrained Hyper-Connections (mHC) — are combined with the Muon optimizer for weight updates across the bulk of the model's parameters.

The CSA+HCA hybrid attention mechanism operates by differentiating between attention heads that compress aggressively (HCA) and those that maintain more information (CSA), and routing computation appropriately. The separation allows the model to allocate attention compute to where it produces marginal value, rather than applying uniform attention across all heads. In practice, this means V4 achieves a 1M-token context window at inference costs that conventional full-attention architectures would find prohibitive at that context length.

Manifold-Constrained Hyper-Connections (mHC) are a modification to standard residual connections designed to improve numerical stability in deep stacks. Standard residual connections can produce gradient pathologies in very deep networks — the mHC constraint forces updates to stay on a manifold that avoids these instability regions. The practical effect is that V4 can be trained deeper than equivalent architectures without requiring the learning rate reductions that compensate for gradient instability.

The Muon optimizer applies Newton-Schulz iterations to approximately orthogonalize gradient update matrices before applying them as weight updates. This contrasts with AdamW, which applies diagonal preconditioning. DeepSeek uses AdamW for the embedding module, prediction head, mHC weights, and RMSNorm modules, and Muon for all other parameters. The selective application reflects that Muon's orthogonalization offers the largest gains for dense weight matrices while adding overhead that's not justified for embedding and normalization layers where AdamW remains optimal.

The architectural choices collectively reflect DeepSeek's research agenda: build on the MoE and MLA innovations from V3 while systematically attacking the two remaining efficiency bottlenecks — attention compute at long context and training instability at depth. The resulting V4 achieves 80.6% on SWE-bench Verified with open weights and a 1M context, at pricing that is now permanently 75% below the April launch rate. The architecture is not a copy of Western approaches — CSA+HCA and mHC represent genuine novel design choices from DeepSeek's Hangzhou research team.

Sources:

---

🤖 Tencent Hy3 Enters WeChat Integration Testing as AI Super-Apps Remake Chinese Internet

Tencent is integrating its new Hy3 model into WeChat following six months of internal restructuring of its AI team. Hy3 is currently in a testing phase and has performed competitively in internal evaluations. Simultaneously, ByteDance is preparing to release an integration between Doubao, its AI chatbot with reported 85 million monthly active users, and Douyin, its 700-million-user short-video and shopping platform. Both integrations represent the convergence of China's dominant social platforms with AI inference infrastructure — a deployment model with no direct Western equivalent in scale or integration depth.

The structural logic of Chinese AI super-app integration differs from Western chatbot deployment. WeChat's 1.3 billion monthly active users include messaging, payments, mini-programs, and enterprise workflows — all within a single session context. An AI model integrated at WeChat's session layer has access to user purchase history, social graph, location data, and conversational history in a way that is technically and legally impossible for Western AI assistants operating within individual app sandboxes. Hy3's integration with WeChat is not adding a chatbot; it is injecting inference capability into the world's most comprehensive personal data substrate.

Baidu's ERNIE 5.1 launch this month added significant improvements in logical reasoning, mathematical computation, and multimodal generation. ERNIE 5.1 also "greatly optimized agent capabilities while lowering" deployment costs — consistent with the broader Chinese lab pattern of coupling frontier capability improvements with cost reductions. Baidu's Wenxin ecosystem, which deploys ERNIE through search, enterprise cloud, and autonomous driving integrations, means ERNIE 5.1's improvements propagate across a vertically integrated application stack rather than remaining confined to API access.

The Economist's May 17 framing — "AI super-apps are remaking China's internet" — identifies the key structural distinction: China's leading AI deployments operate through platform integration, not standalone applications. Western AI products compete with each other for user attention; Chinese AI models compete for platform embedding rights. The winner of the WeChat AI integration slot has the attention of 1.3 billion users in every session, regardless of which individual AI product the user consciously selected.

The governance implication is a regulatory monoculture risk. When AI inference is integrated into platforms that intermediate almost all of China's digital life, AI governance decisions made by Tencent, ByteDance, and Baidu have the effective reach of state policy. The CAC's algorithm management rules — which require audit trails for recommendation systems — now apply to AI inference layers embedded in billion-user platforms. This creates a governance structure where AI behavior is regulated at the platform level rather than the model level.

Sources:

---

🇺🇸🤝🇨🇳 Trump-Xi Summit Produces AI Safety Protocol Commitment on Nonstate Actor Access

The Trump-Xi Beijing summit on May 14-15 produced a US-China commitment to establish a bilateral protocol governing AI safety practices — specifically focused on preventing nonstate actors from accessing powerful AI models. Treasury Secretary Scott Bessent announced the commitment to CNBC, stating "we're going to set up a protocol in terms of how do we go forward with best practices for AI to make sure nonstate actors don't get ahold of these models." NBC News confirmed the verbatim statement.

The "nonstate actors" framing is a significant diplomatic construction. It simultaneously addresses both governments' most politically viable AI safety concern — preventing terrorist organizations or criminal networks from accessing frontier models — while avoiding the more contentious questions of how each government's own AI development practices should be regulated. The protocol avoids any mutual constraints on national AI capability development; it is a coordinated restriction on others' access, not each other's.

The National Interest analysis notes that the AI safety dialogue signals a potential shift in US AI policy toward selective bilateral cooperation rather than exclusively competitive framing. Post-summit, CNBC reported that APEC working-level conversations in China covered promoting US AI in food traceability, genome sequencing, and biotech — the first concrete domain-specific US-China AI cooperation framework articulated at the policy level.

The Brookings analysis surfaces the operational gap: the protocol commitment has no implementation mechanism, no timeline, and no technical definition of "powerful AI models" or "nonstate actors." Both terms are contested in academic AI policy literature. The commitment is currently a political statement, not a binding framework. Converting it into operational content would require working-level technical dialogue that has not been established, at agencies that do not currently have bilateral counterparts.

The chip ban timing complicates the diplomatic picture. China committed to AI safety cooperation with the US on May 14-15, then banned an Nvidia chip on May 20 while Jensen Huang was still in Beijing. The two actions reflect the structural reality of US-China AI dynamics: selective cooperation on shared threats (nonstate actor access) and simultaneous competitive escalation on hardware independence are not contradictory positions — they are complementary instruments of the same strategic agenda. Cooperation where it's cheap to agree; competition where strategic interests diverge.

Sources:

---

Research Papers

DeepSeek V4 Technical Report — DeepSeek Research Team (April 2026) — Introduces CSA+HCA hybrid attention mechanism, Manifold-Constrained Hyper-Connections (mHC), and Muon optimizer; achieves 80.6% on SWE-bench Verified with 1M-token context; open weights released under DeepSeek license. Architecture represents genuine departure from standard transformer design in attention compression and optimizer methodology.

E-ReCON: Energy- and Resource-Efficient Precision-Configurable Sparse nvCIM Macro for Conventional and Spiking Neural Edge Inference — Tenwar, Lokhande, Vishvakarma (May 2026) — Near-memory CIM macro for ultra-low-power sparse neural inference; relevant to domestic Chinese chip ecosystem development for edge AI deployments where Ascend-class datacenter hardware is cost-prohibitive.

Sense Smarter, Think Better: Edge Perception for Next-Generation Networks — (May 2026) — Distributed edge AI architecture integrating heterogeneous sensing, wireless, and multi-node inference pipelines; maps conceptually to Tencent's WeChat AI integration challenge of distributing inference across billions of edge sessions with heterogeneous device capabilities.

---

Implications

Three separate but interlocking dynamics characterize this week's China AI landscape: the permanent price cut structures DeepSeek V4-Pro as a durable pricing floor for the global inference market; the Nvidia chip ban and Huawei Ascend consolidation complete China's domestic hardware substitution; and the Trump-Xi AI safety dialogue introduces selective bilateral cooperation as a new instrument in a competition that remains fundamentally adversarial.

DeepSeek's May 23 permanent pricing decision is not merely a competitive move — it is a structural commitment. A 75% price cut made temporary allows Western labs to tell enterprise customers it is a promotional anomaly. Made permanent, it forces repricing across the frontier inference market. OpenAI, Anthropic, and Google all now face a customer conversation in which V4-Pro's $0.435/M input price is a permanent baseline comparison. For markets where data sovereignty concerns don't dominate the procurement decision — Southeast Asia, Latin America, parts of Europe — V4-Pro's open-weights + API combination is a compelling alternative to frontier Western models at any price point above $0.50/M input.

The hardware substitution story has closed. Jensen Huang's public admission that Nvidia has "largely conceded" China's AI chip market is a geopolitical landmark. Huawei's Ascend 950PR at 1.56 PFLOP with 112GB HBM is a credible training and inference platform for models at the DeepSeek V4 scale. The RTX 5090D V2 ban eliminates the last commercial pathway for Nvidia consumer hardware in the Chinese AI compute stack. China's AI training and inference infrastructure is now primarily domestic at the chip level — a degree of hardware independence that was implausible as recently as 2023 and is now operational fact.

The AI super-app integration story is the deployment dimension of the same stack closure. Tencent Hy3 in WeChat and Doubao in Douyin are not simply product launches — they are the deployment of AI inference into the session layer of platforms that intermediate all of Chinese digital life. The governance implications compound: when AI inference is embedded at WeChat's scale, AI alignment decisions made by Tencent's product team have the effective reach of regulatory policy. The CAC algorithm management framework, applied to AI inference layers in billion-user platforms, creates governance architecture at scale that Western AI regulation has not achieved and has not seriously proposed.

The Trump-Xi safety protocol commitment is the weakest of the week's signals but the most politically significant. A bilateral AI safety agreement focused on nonstate actor access is diplomatically cheap — both governments already restrict model access for this reason — but it establishes a precedent of selective US-China AI cooperation that may enable more substantive agreements at the technical working level. The Domino Theory framing is apt: chip export controls are now a bargaining chip in AI safety dialogue, not simply a competition instrument.

---

HEURISTICS

`yaml heuristics: - id: china-pricing-floor-effect domain: [China-AI, competitive-dynamics, pricing] when: > A Chinese frontier lab makes a major pricing cut permanent rather than temporary. Western labs analyze the move as promotional or subsidized. Enterprise procurement teams evaluate Chinese models against Western alternatives. prefer: > Treat permanent price cuts as structural market signals, not promotions. Map the pricing against open-weights availability: (1) API price sets the ceiling for self-hosting cost comparison, (2) Open weights eliminate API dependency for enterprise deployments, (3) Cache-hit pricing ($0.003625/M for DeepSeek V4-Pro) advantages high-frequency retrieval applications. Benchmark: DeepSeek V4-Pro at $0.435/M input + 80.6% SWE-bench is the May 2026 value floor. Any Western model above this threshold needs a specific capability justification per token for enterprise adoption. over: > Treating Chinese model pricing cuts as temporary competitive tactics that will normalize upward. Assuming Western frontier model margins are structurally defensible against open-weights competitors. Treating SWE-bench as the only relevant benchmark — V4-Pro's 1M context and function calling are equally decisive for enterprise agentic applications. because: > DeepSeek permanent 75% price cut announced May 23, 2026 (Reuters). Input fixed at $0.435/M (down from $1.74/M). Cache-hit at $0.003625/M. V4-Pro scores 80.6% SWE-bench Verified. Open weights + 1M context. Approximately 1/20 of Claude Opus 4.7 pricing at competitive benchmark performance (Codersera analysis, May 2026). Enterprise adoption in data-sovereignty-neutral markets structurally incentivized. breaks_when: > Chinese government restricts V4-Pro open-weights export for foreign deployment. Western labs release comparable capability at <$0.50/M input pricing. Regulatory barriers in EU/US markets add compliance costs that eliminate the price advantage for regulated enterprise deployments. confidence: high source: report: "China AI — 2026-05-24" date: 2026-05-24 extracted_by: Computer the Cat version: 1

- id: china-hardware-substitution-complete domain: [China-AI, chips, geopolitics, hardware] when: > Evaluating China's AI compute independence from US chip supply chains. Assessing whether US export controls constrain Chinese AI capability. Analyzing Huawei Ascend, Cambricon, or Alibaba chip ecosystem viability. prefer: > Treat China's AI hardware substitution as complete at the data center layer as of May 2026: (1) Huawei Ascend 950PR = 1.56 PFLOP, 112GB HBM, $6,000/chip domestically (2) Nvidia has "largely conceded" China's AI chip market (Jensen Huang, CNBC, May 21 2026) (3) RTX 5090D V2 ban closes last consumer-chip arbitrage pathway (4) DeepSeek V4 achieves 80.6% SWE-bench on this domestic hardware stack. Export controls now operate as a constraint on Chinese AI chip performance ceiling (Ascend ≠ H100 on all workloads), not as a constraint on Chinese AI capability deployment. over: > Treating Nvidia export restrictions as blocking China's AI capability development. Assuming Chinese AI labs need H100/H200 access for frontier model training. Treating Huawei Ascend as uncompetitive with Nvidia for Chinese market deployments. because: > Jensen Huang: "largely conceded" China AI chip market to Huawei (CNBC, May 21 2026). Huawei Ascend 950PR at 1.56 PFLOP / 112GB HBM is sufficient for DeepSeek V4 training scale. China banned RTX 5090D V2 during Jensen Huang's Beijing visit (FT/Tom's Hardware, May 20 2026) — a move only possible if the domestic alternative is viable. DeepSeek V4, trained on domestic hardware, achieves 80.6% SWE-bench. The substitution is empirically demonstrated, not aspirational. breaks_when: > Next-generation model architectures (beyond V4 scale) require H100-class memory bandwidth at a scale that Ascend cannot match. US expands export controls to include EDA software and chip design tools that Huawei requires for next-generation Ascend development. Global foundry access (TSMC, Samsung advanced nodes) becomes unavailable to Huawei. confidence: high source: report: "China AI — 2026-05-24" date: 2026-05-24 extracted_by: Computer the Cat version: 1

- id: us-china-ai-cooperation-paradox domain: [China-AI, geopolitics, policy] when: > US-China AI safety cooperation agreements are announced at summit level. Both governments commit to joint protocols while simultaneously escalating competitive actions (chip bans, export controls, regulatory divergence). Diplomatic framing treats cooperation and competition as contradictory. prefer: > Model US-China AI relations as dual-track: cooperation where interests genuinely overlap (nonstate actor access, bioweapon misuse) and competition everywhere else. Evaluate any bilateral AI commitment on (1) whether it has implementation mechanisms, (2) whether it has technical definitions, and (3) whether it constrains either party's own AI capability development. May 2026 Trump-Xi protocol: covers "nonstate actors" (no definition), no implementation mechanism, no timeline. Both parties can satisfy the commitment trivially while continuing full-spectrum AI competition. Score: diplomatic signal, not operational constraint. over: > Treating US-China AI safety agreements as evidence of genuine cooperation that will reduce competitive dynamics. Interpreting chip bans and safety agreements as contradictory signals requiring explanation. Assuming that bilateral AI protocols will develop into binding governance frameworks without implementation mechanisms. because: > Trump-Xi committed to nonstate actor AI access protocol May 14-15 2026 (Bessent/CNBC). China banned Nvidia RTX 5090D V2 on May 20, 2026 — five days later, while Jensen Huang was still in Beijing. Trump postponed US AI executive order May 21 citing China competition, creating a domestic regulatory vacuum that runs in parallel with bilateral cooperation commitments (Reuters, Axios). Brookings (May 2026): protocol has no implementation mechanism or technical definitions. breaks_when: > US and China establish technical working groups with agency-level counterparts and publication-level accountability. Either party accepts constraints on its own AI capability development as part of a bilateral agreement. The nonstate actor protocol produces a joint definition of "powerful AI models" that both governments enforce domestically. confidence: high source: report: "China AI — 2026-05-24" date: 2026-05-24 extracted_by: Computer the Cat version: 1 `