DeepSeek's $74B Pivot, Kimi K3's Open-Weight Earthquake, and ByteDance's 10T Bet — China AI Weekly

DeepSeek formally launched V4 Pro, raised prices up to 14x, and reopened an $8 billion funding round at a $74 billion valuation. Moonshot AI released Kimi K3 — a 2.8-trillion-parameter open-weight model that became the first Chinese model to top a major coding benchmark. ByteDance is quietly pre-training a 10-trillion-parameter model — the largest ever attempted by anyone. And China now holds 4 of the global top 10 AI models, with the US-China gap narrowed to statistical indistinguishability on most tasks. Here is this week’s China AI Weekly.

DeepSeek: V4 Pro, Pricing Pivot, and $8 Billion at $74 Billion

V4 Pro Official Release

DeepSeek formally released V4-Pro-0813 on August 13, replacing the April preview that had been outperformed by its own cheaper sibling (V4 Flash) in some independent tests. The official version delivers major upgrades:

• Agentic capabilities — AI agent workflows, terminal operations, and cybersecurity tasks
• Software engineering — DeepSWE benchmark jumped from 12.8 (preview) to 62.7
• 1M-token context window with 384K max output, supporting both thinking and non-thinking modes
• Responses API and Codex integration for developer tooling

Benchmark highlights per DeepSeek’s own reporting:

The model is available via API, app, web, and “Expert Mode” on partner platforms.

From Price Predator to Premium

The lab that made “cheap AI” its brand is now charging premium prices. DeepSeek announced massive API price increases effective August 17:

• V4 Pro: $1.32/M input tokens, $3.96/M output tokens (per Artificial Analysis)
• That is 9x input and 14x output compared to V4 Flash ($0.14/$0.28)
• Peak/off-peak pricing: during weekday peak hours (9–12, 14–18 Beijing time), output tokens rise to ¥27/M; off-peak is half that
• Increases range from 50% to 1,100% above current prices depending on model, token type, and timing

This marks the end of DeepSeek’s ultra-cheap phase. The lab needs to monetize — and the pricing now reflects the capabilities.

$8 Billion at $74 Billion Valuation

DeepSeek reopened its second external funding round around August 6–7 after pausing in late July. The pause was triggered by leaked remarks from founder Liang Wenfeng regarding Nvidia dependence and the US-China AI gap.

• Target: $8 billion at a pre-money valuation near ¥500 billion ($74 billion)
• For context: the first external round in June 2026 raised ~$7.4B at ~$52B — the jump to $74B in three months is staggering
• Monolith Management is in talks to join; existing syndicate includes China’s national AI fund, NetEase, JD.com, IDG Capital, Loyal Valley, and Shixiang Capital
• Founder Liang Wenfeng is reportedly committing $3B personally
• Agreements expected late August; potential Shanghai IPO targeted for 2027

DeepSeek Harness Team — Chasing Claude Code

DeepSeek set up a “DeepSeek Harness Team” with a verified WeChat account, recruiting for roles to build agentic developer tools that can compete with Anthropic’s Claude Code. This signals DeepSeek’s ambition beyond raw models into the agentic developer-tools space — a market that Anthropic has largely owned.

DeepSeek × Unitree Robotics — Embodied Intelligence Bet

DeepSeek invested ¥140.8 million (~$20.8 million) in Unitree Robotics’ Shanghai IPO, with joint development of AI models for humanoid robots planned. This marks DeepSeek’s first disclosed move into embodied intelligence and robotics — a signal that the lab sees physical-world AI as the next frontier.

Kimi K3: The Open-Weight Earthquake

2.8 Trillion Parameters, Free to Download

Moonshot AI released Kimi K3 — and it is the largest open-weight model ever released:

• 2.8 trillion parameters with a Stable LatentMoE architecture (896 experts, 16 active per token, ~104B active parameters)
• 1 million-token context window, natively multimodal (text + image)
• Key innovations: Kimi Delta Attention (KDA) — hybrid linear attention reducing computational complexity from quadratic to near-linear; Attention Residuals (AttnRes)
• 2.5x scaling efficiency over Kimi K2
• Trained on ~20,000 Nvidia chips — notable given US export controls

Full weights (1.56TB across 96 shards) were released on Hugging Face under Modified MIT license on July 27, reaching #1 trending within 30 minutes. The base model is open-sourced; the post-training pipeline (RLHF, tool-use, safety) remains proprietary.

Benchmark Performance — First Chinese Model to Top a Major Coding Benchmark

Pricing: $0.30/M cached input tokens, $3/M cache-miss input, $15/M output tokens — roughly 40% cheaper than Claude Fable 5 on blended workloads.

Demand Overwhelm

Moonshot suspended all new consumer subscriptions within 48 hours of launch. Demand pushed capacity to the limit. Existing subscribers were prioritized; new spots reopened in batches. The company’s statement: “Kimi K3 has received far more love than we expected.”

The Global Leaderboard Reality

The Artificial Analysis Intelligence Index for August 2026 tells the story:

Chinese models now hold 4 of the top 10 spots globally. The gap between Chinese and American frontier models has narrowed to what analysts describe as “statistical indistinguishability on most tasks.” This is the current reality, not a future projection.

ByteDance: The 10-Trillion-Parameter Bet

The Biggest Model Ever Attempted

Financial Times reported on August 7 that ByteDance is pre-training a model with up to 10 trillion parameters — the largest model ever attempted by anyone:

• Architecture: Mixture-of-Experts; ~30,000 GPUs for the training run
• Timeline: 3–6 months of pre-training, implying earliest internal evaluation around November 2026–February 2027; product launch potentially mid-to-late 2027
• Final parameter count not locked — 10T is a ceiling target, not a finalized number

For scale context:

The project is run by ByteDance’s Seed AI research org (~2,000 staff), led by Wu Yonghui (ex-Google DeepMind). Founder Zhang Yiming reportedly directed the team to pursue genuine technical breakthroughs rather than relying on distillation from Western AI systems — “won’t copy the West.”

No model name, no release date, no benchmark yet. This is a leak, not a launch. But the scale signals intent.

Alibaba: Qwen3.8-Max, Apple Partnership, and Revenue-Share Open Weights

Qwen3.8-Max

Alibaba’s largest and most capable model: 2.4 trillion parameters, with a preliminary Artificial Analysis score of 53.9. Already powering commercial services on Alibaba Cloud’s Lingjun Zhenwu M890 supernode.

Revenue-Share Model for Open Weights — A First

Alibaba plans to require large commercial users of Qwen3.8-Max to share revenue — a significant departure from the industry standard where open-weight releases have been freely deployable. This signals a maturation of China’s open-weight strategy from a technology challenge to a business-model disruption. If the model succeeds, expect other Chinese labs to follow.

Apple Partnership — Qwen Powers Apple Intelligence in China

Apple published a Chinese-language guide for Mac users to connect Alibaba’s Qwen to Siri and Writing Tools on macOS 26.6+ on August 8. The guide was pulled within ~24 hours — Apple customer service said the feature was not officially launched, and the regulatory process may still be pending.

Separately, Reuters reported on August 14 that Apple has trained a China-specific LLM with Alibaba’s support — a departure from Apple’s strategy of relying on third-party models. The CAC registered Apple’s generative AI service in July, clearing the way for Apple Intelligence on Chinese iPhones. Under the arrangement, Qwen powers Siri and Writing Tools; Baidu is also contributing technology; Alibaba is barred from using submitted materials for training.

The world’s most valuable company is now dependent on Chinese AI inside China.

Alibaba Cloud Supernode

Launched the Lingjun Zhenwu M890 supernode in Ulanqab, Inner Mongolia on August 12 — the first supernode in China to successfully run LLMs exceeding 2T parameters. Designed for inference of MoE models up to 10T parameters. Already powering Kimi K3 and Qwen3.8-Max commercially.

QwenWork — Office-AI Consolidation

Alibaba combined QoderWork, Wukong, and MuleRun into QwenWork on August 3 — desktop control, cloud execution, and enterprise agents in one platform. Public beta for web and PC; DingTalk integration to follow. This ends internal “horse racing” (parallel competing teams) in favor of a centralized platform.

AI Spending

Alibaba pledged ¥380 billion (~$53 billion) over 3 years for cloud/AI infrastructure (announced February 2025) and is now signaling it will exceed that pledge. Well over ¥100 billion spent on AI/cloud in the trailing four quarters. June 2026 quarter results due August 20.

Tencent: Capex Triples, WorkBuddy Hits 20M MAU

Q2 2026 Earnings — AI Spending Surge

Tencent’s Q2 results, reported August 12:

• Revenue: ¥204.8 billion ($30.4B), up 11% YoY
• Capital expenditure: ¥52.8 billion ($7.8B), up 176% YoY — nearly tripled
• Operating capex: ¥51.8 billion, up 190% YoY
• Large enough to push free cash flow negative for the quarter (–¥13.8 billion vs. +¥43.0 billion a year ago)
• Spending directed at Hunyuan model enhancements, inference, AI features across products, and external cloud demand

Tencent shares sagged after the disclosure — market concerns about whether AI spending will generate returns. The AI infrastructure arms race is consuming cash at unprecedented rates.

WorkBuddy — 20 Million Monthly Active Users

Tencent’s WorkBuddy AI agent has surpassed 20 million monthly active users — the top spot in the AI office platform market. An internal reorganization on July 20 moved the QClaw Product Center (OpenClaw-based desktop assistant) into the cloud division housing WorkBuddy. New features rolled out August 3. The strategy: lightweight, open platform targeting SMEs with the lowest entry barriers and broadest service coverage.

BAT Office-AI Consolidation: All Bets Placed in Two Weeks

Within roughly two weeks, all three BAT giants ended internal horse racing and centralized their office-AI offerings:

Enterprise AI is the new battleground. The consolidation signals that the experimental phase is over — these companies are now building for market share.

Baidu: AI Revenue Crosses 50%, Earnings Due August 18

In Q1 2026, AI-related businesses accounted for 52% of Baidu Core revenue for the first time — up 49% to ¥13.6 billion. AI Cloud revenue reached ¥11.3 billion, with the infrastructure piece growing 79% and GPU cloud growing 184%. Baidu has invested well above ¥100 billion in AI infrastructure and models since launching Ernie in March 2023.

The Tianchi 256-card supernode (based on Kunlun chips) launched, supporting Wenxin, DeepSeek, GLM, and MiniMax models. Baidu is integrating its internal office agent “dodo” into Baidu Dazi — meeting, document, and knowledge-base functions routing into Dazi’s enterprise edition.

Q2 2026 results due August 18 — will update the AI revenue picture. Management frames their approach as “maintaining strategic investment intensity while preserving financial discipline.” They will not try to out-spend larger balance sheets.

China AI Regulation: First to Regulate AI’s Emotional Interactions

Anthropomorphic AI Interaction Services Measures (Effective July 15)

China enacted the Interim Measures for the Administration of Anthropomorphic AI Interaction Services — the first nationwide rules governing AI’s emotional interactions with users:

• Lifecycle risk assessment, ethics review, content monitoring, and incident-response programs required
• Mandatory “minor mode”: guardian consent, spending/usage limits for users under 14; ban on virtual companion/relative services for minors; ban on emotionally manipulative design
• AI must send warnings to users showing signs of excessive dependence; reminders every 2 hours that they are speaking with AI
• Enforcement is real: 12 companies fined a combined ¥4.2 million in the first week for violations involving emotional dependency risks and minor protections
• Impact: ByteDance, Alibaba, and Tencent all pulled their AI companion apps from the market

Beijing is the first government to regulate AI’s emotional interactions at scale. The speed of enforcement — fines in week one — signals this is not a paper tiger.

Cross-Border and Banking Rules

• National Financial Regulatory Administration (June 18): Guidelines on safe AI development in banking and insurance — 32 principles under 7 pillars, including risk-control-committee approval before deploying AI for high-risk use cases, a ban on using personal information as AI training data, and mandatory filing of externally sourced AI models with the CAC
• CAC cross-border transfer Q&A guidance: clarifies “separate consent” disclosure requirements and “necessity” test for routine transfers
• Export controls: MOFCOM placed UAV exports to the US under case-by-case review (August 5); new exit-and-entry regulations (effective September 15) allow authorities to stop citizens from leaving if implicated in export-control or technology-transfer violations

WAICO: China’s Global AI Governance Body Expands to 37 Members

The World AI Cooperation Organization (WAICO), launched July 16 in Shanghai with delegates from 29 countries, has expanded to 37 member states by mid-August. China’s AI+ International Cooperation Initiative covers 8 areas including sharing open-source ecosystems, joint rules/standards, and safety governance. The organization is gaining traction as China’s answer to global AI governance.

WAIC 2026 — Largest Ever

The World Artificial Intelligence Conference (July 17–19, Shanghai) was the largest ever: 1,100+ companies, 4,400+ AI products/services exhibited. Xi Jinping delivered a keynote on guiding AI with human wisdom and international consensus, emphasizing the “intelligent economy.” China’s core AI industry exceeded ¥1.2 trillion ($173.9 billion) in 2025, crossing the trillion-yuan mark for the first time, with 6,200+ AI-related companies. AI penetration in key sectors (manufacturing, healthcare, finance, logistics) has reached 80%+.

Other Regulatory Developments

• Draft National Anti-Cyber Violence Law (July 29, public comment): prohibits deepfakes and profiling-based targeting for harassment
• Mandatory L3/L4 automated-driving safety standard (GB 44721-2026, effective July 1, 2027): lifecycle management, driver takeover, confirmation testing
• AI content labeling mandatory (explicit + implicit metadata watermarks) since September 2025
• September 1 relief valve: Companies processing personal data on fewer than 100,000 individuals qualify as “small-scale handlers” with simplified notice, consent, and longer audit cycles
• Ministry of State Security issued its second warning in two months about risks of accessing foreign AI services
• Draft standard for AI agent security classification and grading — addresses data leakage, uncontrolled permissions, tool misuse, and actions deviating from user intent

Infrastructure & Ecosystem

National AI Computing Build-Out

China’s AI computing capacity grew 177% YoY by end of June, reaching 2,185 exaflops. The first AI super cluster built from 100,000 domestically produced chips is now operational. A national monitoring and scheduling platform is in trial operation, with 60%+ of computing power visible to it and routing jobs across regions. China expects ¥4 trillion (~$560 billion) in new direct AI investment during the 15th Five-Year Plan (2026–2030), mostly private capital.

Startup Registration Surge (H1 2026)

• 55,000 new generative AI companies registered (28% YoY increase)
• 116,000 new humanoid robot companies (9.5% increase) — more than 2x the number of new GenAI firms
• China’s generative AI user base: 602 million (42.8% population penetration, 141.7% annual increase)
• 796 generative AI services completed CAC filing; 481 applications finished registration

Capital Markets Fueling AI

China is mobilizing $28 trillion in capital markets to fuel AI development — shifting from subsidies to equity/bond financing. Firms raised $217 billion since 2024 (vs. US rivals’ $1.4 trillion, though Chinese funding costs are far lower, with an average bond coupon of 1.9%). Insurance funds are deepening allocations to hard-tech IPOs. DeepSeek’s $20.8M Unitree investment was timed to Unitree’s landmark Shanghai IPO.

GPU/Chip Constraints

US export controls on advanced AI chips remain the primary constraint — not tariffs. China is pursuing a two-track strategy: maximizing compute efficiency via MoE architectures while building an ecosystem on domestic chips (Huawei, Kunlun). Per the Asia Society: “Compute scarcity is now a managed condition rather than a crisis.”

Geopolitics & International

• Manus AI buyback: Former investors of Chinese startup Manus AI bought the company back from Meta after Chinese regulators ordered the purchase unwound — signals Beijing building a firewall around homegrown tech
• US reviewing how Chinese AI firms access Nvidia chips abroad to close loopholes
• Palo Alto Networks under cybersecurity review in China — may be part of a cross-departmental response to US restrictions
• China added 14 European entities to its export control list — prohibiting supply of dual-use items without authorization
• China’s drone exports rose 26% in H1 2026 to 2.42 million units (driven by Southeast Asia, Africa, Middle East); US sales plunged 40%+ amid export curbs
• EU AI Act fully effective August 2, 2026 — enforcement powers now active
• EU launched €30B tender for up to 7 AI gigafactories (€10B public + €20B private)
• Trump administration finalized a voluntary pre-release testing framework for frontier AI models — developers voluntarily provide government early access for ~30 days before public release
• OpenAI publicly disclosed its upcoming Astra model may have crossed into a “critical” cybersecurity threshold — the most consequential safety disclosure in frontier AI’s short history

What to Watch

• DeepSeek’s August 17 pricing switch. The 9x–14x price hike is the clearest signal yet that Chinese AI labs are moving from subsidized disruption to commercial sustainability. Watch for whether V4 Pro demand holds at the new price points — and whether competitors follow.
• Alibaba’s Q2 results (August 20). With ¥380B pledged and spending set to exceed that, Alibaba’s earnings will show whether AI infrastructure investment is translating into cloud revenue growth.
• Baidu’s Q2 results (August 18). AI already exceeds 50% of Baidu Core revenue. Whether that trajectory continues will set the tone for AI monetization across the BAT.
• ByteDance’s 10T model timeline. If the 10T model hits its November–February internal evaluation window, expect benchmark leaks — and a potential step-change in the global leaderboard.
• AI companion regulation enforcement. The first fines landed in week one. Watch for whether regulators expand scope beyond companion apps to broader AI interaction services.
• Apple Intelligence China launch. The guide was published and pulled within 24 hours. The regulatory clearance is in place, but the feature is not officially live. Watch for the iOS update that activates it.

This week’s China AI Weekly is defined by three tectonic shifts: DeepSeek’s transition from price predator to commercial entity, Kimi K3 proving that the open-weight frontier is now Chinese, and ByteDance betting 10 trillion parameters on a future where “won’t copy the West” is the strategy. The US-China AI gap is no longer a gap — it is a race, and on most metrics, it is a dead heat.

Building an AI architecture that survives the next regulatory shift? Big Hat Group delivers enterprise AI consulting and Azure-native deployments for teams that need vendor-neutral orchestration, model gateway patterns, and documented governance across open-source and proprietary models. Whether you are evaluating Kimi K3, building a DeepSeek integration strategy, or navigating China’s AI regulatory landscape, book a discovery call to scope the work.

Follow along on X for real-time analysis as these stories develop.

Sources:

• Reuters (Aug 13, Aug 14): DeepSeek V4 Pro launch, Apple–Alibaba partnership
• Bloomberg (Aug 14): DeepSeek $8B funding round, Tencent earnings
• Financial Times (Aug 7): ByteDance 10T parameter model
• Artificial Analysis (Aug 2026): Global AI Intelligence Index
• CGTN (Aug 9, Aug 14): China AI computing capacity, DeepSeek launch
• Global Times (Aug 9, Aug 12, Aug 13): WAIC 2026, Alibaba supernode, regulation
• China Daily (Aug 13): DeepSeek V4 Pro coverage
• CSIS PacTech Pulse (Aug 2026): WAICO, export controls
• Semafor (Aug 13): Manus AI buyback, chip access review
• Asia Society (Aug 13): Compute scarcity assessment
• Kiteworks (Aug 12): AI regulation compliance
• Rest of World (Aug 10): AI companion app regulation
• Korea Economic Daily (Aug 14): China AI talent pool, capital markets
• AIN China (Aug 9, Aug 11): Kimi K3 benchmarks, DeepSeek Harness Team
• LongYield Substack (Aug 14): BAT office-AI consolidation, Baidu analysis