DeepSeek's $7.3B Mega-Round: China's AI Funding Frenzy Hits Historic Heights

In the span of seven days, China's AI industry raised more money than most countries spend on AI in a decade.

DeepSeek — the Hangzhou-based research lab that upended global AI markets in January 2025 with its shockingly efficient R1 reasoning model — is now seeking 500 billion yuan ($7.35 billion) in its first-ever external funding round. If completed, this will be the largest single fundraising event in Chinese AI history, nearly triple the previous record.

But DeepSeek is only the headline. In the same week:

- Moonshot AI (Kimi) closed a $2 billion round at a $20 billion valuation, led by Meituan's Longzhu Capital

- StepFun raised approximately $2.5 billion and filed for Hong Kong IPO

- InfiniGence AI (无问芯穹) secured 700+ million yuan for AI infrastructure, bringing total funding to 2.2+ billion yuan

- Salvatore Sanfilippo — the creator of Redis — built a dedicated inference engine for DeepSeek V4, achieving 468 tokens/second on a Mac Studio

Add to this Sichuan province's plan to grow its AI industry to 400 billion yuan by 2030, Guangdong's province-wide AI+ transformation initiative, and China Mobile's trillion-token service package, and you get a clear signal: China's AI industry is transitioning from research phase to capital-intensive scale phase, and the world is watching.

Executive Summary: The Week China AI Went All-In

Source: The Information, Bloomberg, 36Kr, QbitAI, Sina Finance, ESM China, 21CBR, official company announcements

What makes this week different from previous funding cycles: In 2024 and early 2025, Chinese AI funding was scattered, speculative, and focused on "me too" model companies. This wave is concentrated, strategic, and infrastructure-heavy. The investors are state funds, industry giants, and founders themselves — not speculative VC money. The capital is going to compute, talent retention, and productization, not just research.

1. DeepSeek's $7.3B Round: Anatomy of a Historic Deal

1.1 The Numbers That Matter

Let's put DeepSeek's fundraising in context:

Source: The Information, 36Kr, 21CBR, Bloomberg

The scale is unprecedented. At $51.5 billion, DeepSeek would be valued higher than Twitter ($41B), Snap ($18B), and Uber at its IPO ($82B) — and this is a company that, until now, has had zero external funding and minimal commercial revenue.

1.2 Why DeepSeek Is Raising Now — The Real Story

DeepSeek's founder Liang Wenfeng has been famously hostile to external capital. In 2024, he told interviewers: "The problem we face has never been money." His hedge fund parent, High-Flyer Quant, managed ~700 billion yuan with a 56.6% average return in 2025, generating roughly 40 billion yuan in profits. DeepSeek had a money pipeline.

So what changed? Three structural forces:

Force 1: Talent Hemorrhaging

The V4 technical report — all 58 pages of it — contains a haunting footnote. Ten names carry asterisks marked "已离职" (departed). These aren't junior engineers. The talent drain includes researchers who went to Xiaomi (Luo Fuli), Tencent (Wang Bingxuan), and ByteDance (Guo Daya).

Without external equity, DeepSeek couldn't offer competitive stock packages. When competitors could offer 2-3x cash plus equity, staying at a bootstrapped research lab became financially irrational for top talent.

Force 2: Compute Constraints

DeepSeek V4's full training run required ~10 million GPU hours. The next generation — V4.1 with multimodal capabilities, enterprise tools, and MCP protocol support — will need significantly more. Export controls on advanced AI chips mean every training run must be carefully planned, and domestic alternatives (Huawei Ascend) require substantial optimization investment.

Force 3: Productization Pressure

DeepSeek's technical reputation is unmatched. V4 scores #1 among open-source models on coding benchmarks. But commercially, the company is an API provider charging 12 yuan per million tokens — a race-to-the-bottom pricing strategy that requires massive volume to become profitable.

At $51.5B valuation, using the AI industry standard 100x P/S ratio (Kimi's current multiple), DeepSeek needs $515M in annual revenue to justify its price tag. At current API pricing, that requires ~43 trillion tokens per year — roughly 20% of China's entire 2025 token consumption.

The math doesn't work without a product breakthrough.

1.3 The Investor Lineup: National Team + Tech Giants

Source: The Information, 21CBR, Sina Finance

Notably: Alibaba's negotiations reportedly hit obstacles, with some sources suggesting DeepSeek is wary of becoming too dependent on a single cloud provider. Tencent's proposal for an acquisition stake was also reportedly declined. DeepSeek wants capital partners, not strategic controllers.

1.4 What the Money Buys: 12 Months of Competitive Time

DeepSeek has reportedly told investors the funds will be deployed across four priorities:

Source: Company investor communications reported by QbitAI and 36Kr

The V4.1 release — scheduled for June 2026 — is the first test. If it delivers multimodal capabilities (image + audio + text), enterprise tooling, and MCP (Model Context Protocol) integration, DeepSeek can begin transitioning from "best open-source model" to "best AI platform."

2. The Ecosystem Funding Wave: Kimi, StepFun, and the Infrastructure Layer

2.1 Moonshot AI (Kimi): From $4.3B to $20B in Six Months

While DeepSeek dominates headlines, Moonshot AI's funding trajectory is equally impressive:

Source: Bloomberg, Caixin, 36Kr

The revenue story is what justifies the valuation. Kimi K2.5, launched in late April 2026, generated more revenue in its first 20 days than the entire year of 2025. Monthly ARR broke $100M within weeks and hit $200M by May. On Stripe's global payment rankings, Kimi shot from outside the top 100 to #9 — the first Chinese AI product to crack the top 10.

What's driving this revenue surge? Kimi has aggressively productized around three use cases: long-document analysis (its signature strength), AI search, and coding assistance. Unlike DeepSeek's API-first approach, Kimi built a consumer product with clear upgrade paths and enterprise tiers.

2.2 StepFun: The Quiet Giant Preparing for IPO

StepFun (阶跃星辰), founded by former Microsoft Research Asia VP Yuan Jing, raised approximately $2.5 billion in a pre-IPO round and is preparing to list on the Hong Kong Stock Exchange.

Source: 36Kr, Sina Finance

StepFun has been less flashy than DeepSeek or Kimi in consumer markets but has built strong enterprise relationships. Its IPO will be a litmus test for whether public markets will accept Chinese AI companies at tech-giant valuations.

2.3 InfiniGence AI: The Infrastructure Layer Nobody Talks About (But Everyone Uses)

While model companies grab headlines, InfiniGence AI (无问芯穹) is building the plumbing that makes them work. Its latest 700+ million yuan round brings total funding to 2.2+ billion yuan — making it the best-funded AI infrastructure company in China.

Source: Official announcements, QbitAI, 163.com

The token economy thesis: InfiniGence CEO Xia Lixue describes his company's role as the "refinery in the token economy" — converting electricity into tokens, the fundamental resource of the AI age. Their platform's token volume has grown 20x since late 2025, suggesting the entire Chinese AI market is scaling at similar rates.

The company has also formalized an AI Productivity Formula:

AI Productivity = Intelligence Scale × Token Production Efficiency × Token Value Conversion

This framework attempts to quantify how infrastructure investments translate to economic output — a useful lens for understanding why investors are pouring billions into compute layers.

3. The Global Open Source Moment: Redis Creator Builds DeepSeek's "Private Highway"

Three weeks after DeepSeek V4's release, something unusual happened in the open-source world.

Salvatore Sanfilippo — known to programmers worldwide as antirez, the creator of Redis (74,000 GitHub stars) — released ds4.c: a dedicated inference engine built from scratch for a single model, DeepSeek V4 Flash, on a single platform, Apple Silicon.

3.1 Why This Matters

Source: GitHub repository analysis, QbitAI

Performance numbers on Mac Studio M3 Ultra (512GB RAM):

Source: antirez GitHub repository benchmarks

For a 284B parameter MoE model running locally, 468 tokens/second prefill is a breakthrough. It makes coding agents, long-document analysis, and multi-turn conversations genuinely usable on a desktop machine.

3.2 Three Technical Innovations

1. Asymmetric Quantization

Rather than blanket 2-bit quantization across all layers, ds4.c applies aggressive compression only to routing expert layers (the bulk of parameters), while keeping shared experts, projection layers, and routing layers at Q8 precision. This preserves model quality where it matters.

2. Disk-Based KV Cache

Most LLM clients are stateless — they resend the entire conversation history with every request. ds4.c persists KV states to disk, keyed by SHA1 hash of token sequences. For agents like Claude Code that send 25K-token initialization prompts, subsequent sessions skip prefill entirely.

3. Dual API Compatibility

The engine supports both OpenAI (`/v1/chat/completions`) and Anthropic (`/v1/messages`) API standards out of the box, with optimized tool-calling support. Configuration examples for Claude Code, opencode, and Pi are included in the README.

3.3 The Symbolic Significance

When a legend like antirez — who built one of the most important pieces of infrastructure in modern computing — chooses to dedicate months of work to optimizing for your model, it signals something: DeepSeek V4 has crossed into the territory where the global developer community invests its scarcest resource (attention and engineering time) voluntarily.

This is the truest measure of open-source success. Not GitHub stars. Not download counts. When brilliant engineers build *for* your model, not *around* it.

4. Policy Tailwinds: Sichuan, Guangdong, and the National AI+ Push

Capital flows don't happen in a vacuum. The policy environment in early May 2026 is aggressively pro-AI:

4.1 Provincial AI+ Plans

Source: Sichuan Government, Guangdong Government, State Council documents

4.2 China Mobile's Trillion-Token Gambit

At the 2026 Mobile Cloud Conference (May 8), China Mobile announced it would open a trillion-token service experience package to enterprises. The state carrier is building gigawatt-level AI data centers and positioning itself as the default compute provider for businesses that can't afford their own infrastructure.

Source: Xinhua, China Mobile announcement

The strategic implication: State-owned enterprises are building the rails. Private companies (DeepSeek, Moonshot, etc.) are building the trains. The government is ensuring the tracks exist and are affordable to ride.

5. Global Context: How China's AI Funding Wave Compares

Source: Bloomberg, The Information, Financial Times, company disclosures

Three observations:

1. Chinese AI companies are closing the valuation gap. DeepSeek at $51.5B is roughly 1/18th of Anthropic's target valuation — but Anthropic has 100x the revenue. The market is pricing DeepSeek's *potential*, not its current business.

2. The "product gap" is the real risk. Anthropic's revenue explosion came from Claude Code — a product that transformed it from an API company to a solutions company. DeepSeek has no equivalent product yet. V4.1 in June is the first serious attempt.

3. State capital is replacing VC capital. Unlike Western AI funding (dominated by Microsoft, Amazon, VC firms), China's wave is led by National IC Fund, local government funds, and founder capital. This changes the incentive structure — less pressure for immediate returns, more alignment with national strategic goals.

6. Social Pulse: What Chinese Tech Communities Are Saying

**"DeepSeek 500亿融资，梁文锋自己出200亿。这哪是融资，这是给全员期权定价。"

— Zhihu top comment, 12K upvotes**

*"DeepSeek's 50B fundraising, Liang Wenfeng putting in 20B himself. This isn't fundraising — it's setting a market price for employee stock options."*

**"估值515亿美元，要撑起这个数，年收入得5亿美元。现在API定价12元/百万token，得卖多少token？"

— V2EX discussion, 847 replies**

*"$51.5B valuation means $515M annual revenue at 100x P/S. At current API pricing of 12 yuan/million tokens — that's a LOT of tokens."*

**"antirez给DeepSeek写专用引擎，这比任何评测都有说服力。"

— Hacker News Chinese mirror, top comment**

*"antirez building a dedicated engine for DeepSeek is more convincing than any benchmark score."*

**"无问芯穹的Token每两周翻一番，上次见到这曲线还是3G时代手机流量爆发。"

— WeChat tech blogger, 100K+ reads**

*"InfiniGence's token volume doubling every two weeks — last time I saw this curve was the 3G mobile data explosion."*

**"大厂吃不透也放不下DeepSeek。腾讯想收购被拒，阿里谈判僵局。好的AI公司从不缺买家。"

— 36Kr comment section, editor's pick**

*"Big tech can't swallow or let go of DeepSeek. Tencent's acquisition proposal rejected, Alibaba talks stalled. Good AI companies never lack buyers."*

**"Kimi两个月ARR从1亿涨到2亿美元，Stripe全球第九。这成绩比估值更值得关注。"

— Twitter/X Chinese tech circle, viral thread**

*"Kimi's ARR going from $100M to $200M in two months, Stripe global #9. This matters more than the valuation number."*

7. The Critical Questions Ahead

7.1 Can DeepSeek Productize?

The existential question for DeepSeek is whether it can build products, not just models. Its technical excellence is proven. Its commercial execution is unproven. The $7.35B round buys 12 months of competitive time — but Google I/O (May 19), where Gemini 4 is expected, and Anthropic's relentless Claude Code expansion, won't wait.

7.2 Will the Talent Stay?

Stock options priced at a $51.5B valuation are a powerful retention tool. But if V4.1 underperforms, or if the product roadmap slips, those options could lose their golden shine. The 10 departures marked in the V4 paper were a warning shot. This round is the response.

7.3 Is China's AI Infrastructure Ready?

InfiniGence's token growth is impressive, but 20x growth in four months also means the infrastructure is being stress-tested at unprecedented scale. With NVIDIA effectively banned from China (0% market share per Jensen Huang's admission), the domestic chip ecosystem (Huawei Ascend, Moore Threads, Biren) must scale rapidly. Can it?

7.4 What Happens to Smaller Players?

If DeepSeek, Kimi, and StepFun are sucking up all available capital and talent, what happens to the long tail of Chinese AI startups? The market is consolidating around 3-5 national champions, mirroring the US dynamic (OpenAI, Anthropic, Google, Meta). Is this healthy for innovation?

Conclusion: The Month Everything Accelerated

May 2026 will be remembered as the month China's AI industry transitioned from a research curiosity to a capital-intensive, globally competitive force. The numbers are staggering: $14+ billion in funding announced in a single week across just four companies. A founder putting $2.9 billion of his own money on the line. A Redis legend building dedicated infrastructure for a Chinese model. Provincial governments committing hundreds of billions to AI transformation.

But beneath the headlines, the real story is structural: China's AI ecosystem has achieved something that looked impossible two years ago — a self-sustaining flywheel of talent, capital, policy, and infrastructure that doesn't depend on Western technology. The chips are domestic. The models are world-class. The capital is flowing. The policy is supportive. The developers are building.

DeepSeek's $7.35B round isn't just a fundraising event. It's a declaration that China intends to compete at the highest level of AI — not as a fast follower, but as a co-leader. The next 12 months will determine whether that declaration becomes reality.

The funding is done. The building begins now.

*Published May 11, 2026. Data current as of publication date. Funding figures based on reported negotiations; final terms may differ.*

Related Reading:

- China's Agent Era Begins: National Policy Framework and the Trillion-Dollar Race (May 10, 2026)

- China's AI Infrastructure Awakens: DeepSeek, Kimi, and the Token Economy (May 9, 2026)

- China's AI 'War of the Gods': DeepSeek, Kimi, and Open-Source Dominance (May 8, 2026)