DeepSeek closed in on a ~$7.4 billion funding round at a ~$74 billion valuation, hired investment banks, and set its sights on a 2027 STAR Market IPO — while reporting API gross margins of 82.9% that would make any Western AI lab envious. Alibaba completed the largest follow-on offering in Hong Kong history — HK$80 billion (~$10.2 billion) — with all proceeds earmarked for full-stack AI and an explicit promise to investors of a 3-year payback. Tencent released its Hy4 Preview open-source model and posted the week’s standout number: Q2 capex up 176% year-over-year, with the company’s first-ever negative free cash flow. And Beijing formalized roughly 200 AI standards, including a mandatory pre-development ethics review for every AI R&D organization in the country. Meanwhile, the office-agent war escalated on four fronts, Taiwan indicted nine people over a banned Nvidia-server smuggling ring, and Nvidia quietly resumed H200 sales to China. Here is this week’s China AI Weekly.

DeepSeek: $74B Valuation, 82.9% API Margins, IPO Prep

The Second Round

DeepSeek is finalizing its second funding round of roughly RMB 50 billion (~$7.4B) at a valuation near RMB 500 billion (~$74B), expected to close by the end of August. The first round — RMB 50B at a $50B valuation, completed in June — brought in Monolith Management, Tencent, JD.com, NetEase, and CATL, plus a state AI fund. Round two adds CPE, Legend Capital, Shixiang Capital, and Stony Creek Capital, along with local-government-backed funds.

The numbers put DeepSeek’s trajectory in context. High-Flyer, Liang Wenfeng’s quant fund, bankrolled the lab from the start — and the first round alone represented more than 60% of High-Flyer’s RMB 80B in assets under management. DeepSeek has decisively outgrown the “side project” model.

The raise briefly paused in July after leaked investor-meeting remarks — in which Liang said China’s AI gap versus the U.S. is mainly a compute constraint — triggered a compliance review. It is now back on track, and so is the IPO plan: DeepSeek has hired investment banks and is preparing for a possible STAR Market (Shanghai) listing in 2027.

Financials: Revenue 10x, Losses Narrowing

The Information, via syndication, reported DeepSeek’s first public financial picture:

  • Revenue of RMB 475M ($70M) in the first seven months of 2026 — roughly 10x full-year 2025
  • Net loss narrowed to RMB 715M (from RMB 935M in 2025)
  • Gross margin of 44.6% overall, with API gross margin at 82.9% — versus roughly 39% at OpenAI and 63% at Anthropic

That API margin number is the story. DeepSeek raised API prices in August — V4-Pro peak output now runs $3.96/M tokens, off-peak $1.98/M, with input at $1.32/$0.66 peak/off-peak (effective August 16) — and the market absorbed it. Chinese LLM API prices have risen across the board this year per Morgan Stanley; the price-war era is over, and monetization has begun.

Model Updates: V4 Pro GA and V4 Flash

DeepSeek’s V4 Pro went to general availability on August 13 — 1M-token context, MIT license — and sits in the top tier of domestic models on OpenCompass’s mid-July rankings. V4 Flash is now in public beta: the lightweight tier priced at $0.22/$0.44 input and $0.66/$1.32 output per million tokens.

The Security Story That Won’t Go Away

The week’s notable negative: research from TeamT5, reported by Bloomberg on August 24, found that Chinese state-affiliated hacking groups more than doubled their attack volume after integrating DeepSeek and other open-source models into their operations — spanning reconnaissance, exploit creation, and malware development. Named groups include Grimfengxi (exploit code), Huapi (Taiwan email attacks), and Teleboyi (IP/domain mapping). Analysts cite DeepSeek’s low cost and comparatively weak guardrails versus Western models; Moonshot’s Kimi K3 is considered more powerful but costlier to run.

This is the double-edged nature of open weights: the same accessibility that built DeepSeek’s developer ecosystem is now a weaponization vector. Expect this thread to keep surfacing in Western policy debates about open-source AI.

Alibaba: The HK$80B Raise and the 3-Year Payback Pledge

The Largest Follow-On in Hong Kong History

Alibaba completed a HK$80B (~$10.2B) placement — 710 million shares at HK$112.70 — the largest follow-on offering ever on the Hong Kong exchange and the third-largest globally this year. Sovereign wealth funds from the Middle East, Europe, and Asia subscribed to more than 40% of the deal.

All net proceeds are earmarked for full-stack AI: semiconductors, data centers, and proprietary models. The raise is part of Alibaba’s 3-year, RMB 380B (~$56.5B) AI investment pledge.

The boldest detail is the investor timeline Alibaba committed to: the company expects a 3-year payback on AI capex at current gross margins. With equipment useful life of five years, that implies mid-teens ROIC — Morgan Stanley had penciled a 13% base case.

Q2: Profits Down, AI Cloud Up

The earnings narrative from the prior week is still driving coverage. Q2 revenue came in at RMB 269.0B (+9% YoY), but adjusted net profit fell 38% YoY to RMB 20.7B — eroded by AI capital expenditure of ~$10B in the quarter, up 75% YoY.

The payoff is visible on the other side of the ledger: AI cloud + compute revenue hit RMB 48.4B, up 45% YoY, with EBITA +133%. Analysts now call Alibaba the “water seller” of the AI boom — the company converting AI spending into revenue fastest. Qwen remains the world’s most-downloaded open model family, and Qwen3.8-Max (2.4T-parameter MoE, 1M context, $2/$6 per million tokens) anchors the premium tier.

Qianwen Work: The Enterprise Office-Agent Play

Alibaba launched Qianwen Work (Tongyi Office) in public beta on August 30 — desktop client first, then web and DingTalk. It is the first product supporting desktop agents, cloud agents, and enterprise collaboration agents simultaneously, positioned squarely at B2B/enterprise — a deliberate contrast to Tencent’s WorkBuddy, which leads on the consumer-desktop side.

Apple’s China-Exclusive Model

The biggest distribution story in Chinese AI lands next month: Apple and Alibaba co-trained a China-exclusive model (confirmed by Reuters August 14). Apple’s on-device model handles real-time and privacy-sensitive tasks; Qwen handles long-text and multimodal workloads on the cloud. Rollout is expected with iOS 27 and iPhone 18 in September 2026.

Context matters here: Apple vetted eight Chinese labs over three years — Baidu, ByteDance, Tencent, DeepSeek, Baichuan, Moonshot, and Zhipu all fell out at various stages before Qwen won the slot. That makes this the biggest real-world distribution event for a Chinese LLM yet: tens of millions of iPhones, shipping with Qwen inside.

Tencent: Hy4 Preview and the CapEx Blowout

Hy4 Preview (August 28)

Tencent released Hy4 Preview, an open-source Mixture-of-Experts model with 770B total parameters and ~49B active, plus a 1M-token context window. Tencent claims it outperforms Z.AI’s GLM-5.3 and Moonshot’s Kimi K3 in internal blind engineering tests, with mixed results against Alibaba’s Qwen3.8-Max on third-party benchmarks.

The model targets software engineering, research, and financial analysis, and will integrate with Tencent’s CodeBuddy (developer tools) and WorkBuddy (office agents). It is the first major release under ex-OpenAI researcher Yao Shunyu, who now leads Hunyuan — following Hunyuan 3.0 in April.

The Standout Number: Capex +176% YoY

Tencent’s Q2 capital expenditure hit RMB 52.7–59.3B (+176% YoY) depending on cash-basis versus commitment-basis accounting — roughly $9B USD, about 28x the year-ago quarter per Nikkei. That drove Tencent’s first-ever negative free cash flow: -RMB 13.8B, despite RMB 100B+ annual profits. Roughly RMB 51.4B of Q2 payments were prepayments for compute procurement.

Priority list, per the company: (1) bigger Hunyuan models, (2) WorkBuddy inference capacity.

Tencent is also leading the domestic-chip shift: industry surveys expect Tencent’s share of domestic-chip procurement to rise above 65% — Enflame for training, Cambricon for inference. Q2 revenue passed RMB 200B for the first time (RMB 204.8B, +11%).

WorkBuddy: The Office-Agent Leader

Tencent’s WorkBuddy leads China’s office-AI desktop market at 11.15M MAU in July (AICPB), ahead of Baidu’s Dazi at 6.74M. The overall office-agent market is projected to grow from RMB 21.2B in 2025 to RMB 44.9B in 2026 and RMB 332B by 2029 (IDC) — the prize everyone is now spending for.

Baidu: Dazi Explodes, No Raise Planned

Baidu’s Dazi office agent — launched March 2026 — hit 6.74M desktop MAU in July, up 1,063.79% month-over-month, the fastest growth of any office AI agent. The numbers are dramatic:

  • ~150 iterations since launch (near-daily releases)
  • Daily questions up 60x since launch
  • Site visits +845% MoM
  • Enterprise edition bundles 15 suites / 96 skills (finance, legal, product/R&D, ops, HR)

Baidu’s August 27 product event confirmed roughly 9x user growth in a month, framing Dazi as the centerpiece of CEO Robin Li’s “application-driven” ERNIE strategy, alongside AI search, digital humans, Miaoda, and Famou.

Financially, Baidu is the disciplined outlier: Q2 capex of RMB 11.3–11.4B (~3x YoY) is the smallest of the BAT trio, and Baidu is the most reliant on selling compute and tokens. A spokesperson confirmed no equity offering is planned — cash reserves and operating cash flow are deemed adequate. A deliberate contrast to Alibaba’s record raise, and a quiet vote of confidence in Baidu’s own numbers.

ByteDance: Doubao Work and the 10T-Parameter Ambition

Doubao Work Launches

ByteDance launched Doubao Work on August 24–25 — a standalone AI office product under the Doubao brand, with the TRAE and Coze teams folded into Doubao (TRAE IDE/CLI remain a developer product line). It integrates with Feishu/Lark, can take over virtual desktops and browsers, and ships with a 30-day free trial. This is a direct assault on Tencent’s WorkBuddy from the company with the most aggressive spend.

The Most Aggressive Capex in China

ByteDance raised its full-year 2026 capex plan to RMB 200B (from RMB 160B) — the most aggressive of any domestic peer, and the clearest signal that the office-agent war and the model race are now inseparable.

A 10-Trillion-Parameter Model?

The Financial Times reported ByteDance is pretraining a model of up to 10 trillion parameters — more than 3x Moonshot’s Kimi K3 at 2.8T — a scale that would rival Anthropic’s most advanced systems. Leadership told staff the company will not distill, and intends to push parameter competition into the 5–10T range.

Also this month: SeedRealtime (August 5), a native full-duplex audio-visual LLM with a single unified architecture handling perception, reasoning, and response in parallel — already deployed in Douyin and Doubao. Seedance 2.0 video models are widely used by studios and generating high-margin cloud revenue.

One compliance note: ByteDance removed AI-companion functionality from Doubao entirely, complying with the July 15 rules banning AI companions for minors.

The Rest of the Field: Moonshot, Zhipu, MiniMax

  • Moonshot AI: Kimi K3 (2.8T params) remains the largest open-weights model; Kimi K3 Max ranks #2 on LMArena’s coding preference ranking (Elo 1674) — the top Chinese lab, ahead of Qwen3.8-Max at #3. Kimi K2.6 is reported to match GPT-5.5 on SWE-bench Pro. Moonshot is seeking cloud-revenue deals with U.S. giants.
  • Zhipu AI (Z.AI): GLM-5.3 shipped August 14 at the same price as GLM-5.2 — a 744B-parameter MoE under an MIT license. Zhipu remains the first foundation-model company to IPO (Hong Kong, January 2026). Community teardowns traced an anonymous “牛来” model to Zhipu lineage (unconfirmed).
  • MiniMax: The M3 open model launched with a permanent 50% price cut ($0.30/$1.20 per million tokens) — the price curve keeps resetting even as DeepSeek raises rates at the top.

The ecosystem read from 未盡研究 (via 虎嗅): DeepSeek, Zhipu, and Moonshot have crossed from follow-the-leader into frontier competition — and the real differentiator now is manufacturing frontier capability under tighter resource limits.

Regulation: Ethics Review Before Development

The Mandatory Pre-Development Ethics Review

The biggest regulatory story of the week: the Administrative Measures for the Ethical Review and Services of AI Science and Technology (Trial) — issued by MIIT and nine other departments in March, effective April 3 — were formalized August 26–27 as roughly 200 standards take effect.

The core requirement: any AI R&D organization in China — universities, institutes, companies — must pass an ethics review BEFORE development begins, pre-training included.

The structure is three-tiered:

  1. Internal ethics committees at each organization
  2. Accredited external ethics-review centers — MIIT-affiliated institutes are already positioning themselves as compliance providers, drawing comparisons to the post-GDPR consultancy boom in Europe
  3. A mandatory second review by a government-assigned expert panel for high-risk projects — defined as those influencing behavior, emotion, or health; algorithmic opinion-mobilization; or highly autonomous safety/health decision-making

Decisions come within a 30-day window (or a formal extension). Follow-up reviews run at least every 12 months — every 6 months for high-risk systems.

Agent Interconnectivity Standards

Alongside the ethics regime, seven national standards on “AI — Interconnectivity of Intelligent Agents” were published, and public feedback opened on 121 more industry standards, including an AI model context protocol. China is building the plumbing for interoperable agents — while regulating the ethics of building them in the first place.

MIIT Vice Minister Xin Guobin said on August 26 that the ethics-review pilot has scaled from provinces to direct city-level operation — a step toward national rollout.

The 2026–2030 Cyberspace Action Plan

The Central Cyberspace Affairs Commission’s plan for cyberspace/informatization enterprises (issued August 21, reported this week) pairs support with control:

  • Support side: high-end AI chips, high-performance training clusters, LLMs, multimodal and world models, AI agents and embodied intelligence, open-source collaboration, financing channels, government funds, and IPO encouragement
  • Control side: algorithm controls, personal-info enforcement, AI-misuse governance, data-export compliance, antitrust, anti-excessive-subsidies, a crackdown on “involutionary” cutthroat competition, exclusivity bans, and platform fee regulation

AI Companions, Payments, and Industrial Policy

  • AI companions: The Guardian covered the July 15 rules banning AI companions for minors and restricting adult-facing chatbots over “emotional dependence” and birth-rate concerns; providers are warned against services that “replace social interaction.” Doubao removed its companion feature entirely. Loopholes remain for educational services and non-continuous emotional interaction.
  • Payments: PBOC Deputy Governor Lu Lei called for unified standards for AI agents in payments, warning of opaque decision-making and fragmented protocols — proposing a “know your agent” self-discipline convention. Hong Kong regulators expanded their GenAI Sandbox++ to 36 projects across banking, insurance, and pensions.
  • Industrial policy: NDRC is accelerating embodied-intelligence deployment across manufacturing, healthcare, consumer services, and public safety; Guangzhou’s Huangpu bonded zone approved a RMB 950M computing complex for video-generation and coding AI firms, including token exports; NSFC is funding ~70,000 projects in 2026 (+31% YoY); semiconductor industry profits are up 18.5x YoY in the first seven months; and AI daily token usage reportedly surpassed 500 trillion by June.

Chips and Export Controls: The Squeeze Tightens

Taiwan Indicts Nine Over Banned Nvidia Servers

On August 24, Keelung prosecutors charged nine people — including one Nvidia Taiwan employee and two Super Micro employees — over the illegal export of 130 B300 servers (banned for China). Of those, 74 were delivered to China (50 via Indonesia, 16 direct, 8 via Japan to Hong Kong); 56 were seized at customs. Charges include breach of trust, document forgery, and embezzlement, with sentences of up to five years sought.

This is the first known criminal crackdown on the AI-server black market — and it signals enforcement is moving from policy to prosecution.

Nvidia Resumes H200 Sales — Sort Of

Nvidia quietly resumed H200 sales to China on August 26 — the first AI-chip sales since early 2025 — shipping a “trivial” number of units under a January U.S. license, less than 1% of data-center revenue. But Beijing’s objections prevented full utilization, forcing a $400M excess-inventory charge. Nvidia still projects zero China data-center revenue going forward.

The Apex Logistics Probe

Bloomberg reported August 27 that the U.S. is probing Apex Logistics (a Singapore-based Kuehne+Nagel unit) over suspected Nvidia-chip smuggling to China — potentially the first enforcement action against a transporter, widening enforcement from chipmakers to the logistics chain.

Domestic Substitution Accelerates

Every enforcement headline is mirrored by a domestic-substitution headline: Tencent’s shift to >65% domestic-chip procurement (Enflame for training, Cambricon for inference); CXMT posting blockbuster results on the AI memory crunch; China’s chip-sector profits up 18.5x year-over-year. The compute bottleneck is binding — and it is reshaping the entire supply chain.

The Week in Perspective

Four storylines define this week:

1. Capital is validating Chinese AI at platform scale. In seven days, Alibaba raised $10.2B and DeepSeek closed in on $7.4B at a $74B valuation — roughly $17.6B of fresh equity, with IPO pipelines forming behind both (DeepSeek targeting the STAR Market in 2027). Combined with Tencent’s +176% capex and ByteDance’s RMB 200B plan, the “side project” era of Chinese AI is definitively over. This is now an industry being built with institutional capital, sovereign wealth funds, and explicit payback timelines.

2. The office-agent war is the new competitive front. Four major product moves in seven days — Tencent’s WorkBuddy (11.15M MAU), Baidu’s Dazi (+1,064% MoM), Alibaba’s Qianwen Work public beta, ByteDance’s Doubao Work — all chasing a market IDC projects at RMB 332B by 2029. The battleground has shifted from model benchmarks to desktop distribution, and the winner will be the platform that owns the enterprise workflow.

3. Regulation is moving upstream — before development, not after. China’s pre-development ethics review, ~200 standards, agent-interconnectivity standards, and the 2026–2030 cyberspace action plan add up to something new: an AI governance regime that gates research before it starts, not just products after they ship. The compliance industry forming around it is already being compared to post-GDPR Europe.

4. Compute is the bottleneck that ties everything together. Nvidia’s H200 resumption stalled on Beijing’s objections. Taiwan indicted nine people for smuggling B300 servers. The U.S. is probing logistics firms. Tencent is shifting procurement above 65% domestic. Guangzhou is building RMB 950M compute parks. And token usage has reportedly passed 500 trillion a day. Every strategic thread this week — funding, models, regulation — ultimately traces back to one question: who controls the compute?

The throughline: Chinese AI has entered its capital-intensive phase. The labs that shocked the world with low-cost models are now raising billions, promising investors payback on datacenter-scale infrastructure, and building agents that will sit inside the country’s offices. The question for the West is no longer whether China can build frontier AI — it is whether the capital, compute, and regulatory machine behind it can be matched.

Follow @kkaminsk on X for daily China AI updates. This post is part of the China AI Weekly series.