Why Is DeepSeek Raising Another $7 Billion
Just Months After Its First Round?

DeepSeek restarted Round 2 fundraising on August 4–5, 2026, targeting ~50 billion yuan (~$7 billion) at a reported ~500 billion yuan (~$70 billion) pre-money valuation — roughly +43% above Round 1's post-money, with signing planned for late August. If both rounds close, total raises in under five months would exceed 100 billion yuan. For engineering and product teams, the headline number matters less than what it funds: ~70% of each 10 billion yuan tranche is earmarked for compute. This article maps the full timeline, Round 1 vs Round 2 tables, cap-table mechanics, P/S multiples vs OpenAI and Anthropic, a China LLM competitor matrix, and a six-step framework for reading AI funding news without mistaking rumor for closed deals.

01 Timeline: from "no fundraising" to a 7x valuation jump in four months

The real risk is not "another raise" — it is reading unconfirmed terms as settled fact. Round 2 numbers come from anonymous dealmakers, not DeepSeek press releases. Governance is unusual: most outside money has no vote and a five-year lock. Valuation multiples (~140–150x P/S) far exceed OpenAI (~65x) and Anthropic (~21x) on trader estimates. Meanwhile the actual bill is compute. Cross-read with our coverage of DeepSeek's custom chip push and V4 Flash benchmarks — fundraising pace tracks infrastructure expansion, not API hype alone.

• High signal noise: Round 2 size, valuation, and signing dates are media-sourced from unnamed dealmakers. DeepSeek has not officially confirmed terms.
• Opaque control: Most external capital sits in a Liang-controlled LP with no voting rights and a five-year lock. The National AI Industry Investment Fund is the lone direct, voting, unlocked exception.
• Valuation vs revenue gap: ~140–150x P/S on ~$400–500M ARR (first public estimate, unverified) dwarfs Western peers — bubble or optionality is still untested.
• Compute is the line item: ~70% of each 10 billion yuan raise goes to chips, data centers, bandwidth, and liquid cooling — the pressure engineers feel is supply and cost, not paper valuation.

Chronology:

• April 2026: Registered capital rises from 10 million to 15 million yuan. Liang Wenfeng's direct stake jumps from 1% to 34%; combined with Ningbo Cheng'en Enterprise Management Consulting LP, he controls ~84.29%. Round 1 talks begin; V4 preview ships the same month.
• June 2026: Round 1 closes — ~50 billion yuan (~$7.4 billion) raised, post-money valuation above 350 billion yuan ($52–59 billion range across sources). China's largest-ever first LLM round.
• July 14–17, 2026: Foreign press reports STAR Market IPO prep alongside Round 2 talks at ~480 billion yuan pre-money (~$71 billion), ~+37% vs Round 1 post-money. ARR of $400–500 million surfaces publicly for the first time — not formally audited by DeepSeek.
• July 25–26, 2026: Round 2 talks pause. Bloomberg and others cite Liang's anger over leaked comments from a closed-door Round 1 investor meeting.
• August 4–5, 2026: Round 2 restarts — ~50 billion yuan target, ~500 billion yuan (~$70 billion) pre-money, ~+43% vs Round 1 post-money, late-August signing targeted. Both sides reportedly want a low profile.

As of publication, Round 2 amounts, valuation, and timing come from anonymous dealmakers and financial press — not official DeepSeek confirmation. Terms may change before signing.

02 Round 1 vs Round 2: the core numbers

Side-by-side beats a standalone "$70B" headline:

03 What a ~$70B pre-money valuation is actually pricing

1. Why the money: compute is the real invoice

Weeks after Round 1, DeepSeek reportedly planned to double headcount in data-center and AI Agent teams. Reuters has covered custom inference chip hiring. Analysts cited in Chinese press estimate that of every 10 billion yuan raised, ~7 billion yuan flows to compute — chip purchases, owned and leased data centers, bandwidth, liquid cooling. Fundraising is racing infrastructure build-out, not chasing vanity metrics.

2. Unusual cap table: most LPs cannot vote

Most Round 1 outside capital entered through a limited partnership controlled by Liang Wenfeng. Those investors have no voting rights and accept a five-year lock-up. The National AI Industry Investment Fund is the exception — direct equity, voting rights, no lock-up. That preserves Liang's ~84% effective control while keeping governance under scrutiny, especially for institutions sensitive to founder dominance and state-fund exceptions.

3. ~140–150x P/S: cash flow or call option?

At ~500 billion yuan pre-money and ~$400–500M ARR, implied P/S sits near 140–150x — well above OpenAI's ~65x and Anthropic's ~21x on trader math. One participating investor quoted in press put it plainly: LLM pricing is an option bet, not a DCF on today's revenue. Buyers are underwriting DeepSeek as potential infrastructure in China's enterprise Agent stack — not a SaaS multiple exercise.

# media-reported back-of-envelope (NOT official)
pre_money_cny   ≈ 500_000_000_000
arr_usd         ≈ 400_000_000 … 500_000_000
implied_ps      ≈ 140x … 150x
status          = UNCONFIRMED  # Round 2 not signed
# OpenAI ≈65x · Anthropic ≈21x (trader estimates)

04 Competitor matrix and a six-step funding-news framework

China's LLM valuation race in one view:

You do not need a primary allocation to use this deal as a signal. Six steps for engineering and product leads:

1. Tag deal status first: Separate closed / in negotiation / rumor. Round 2 is still in talks — do not treat target raise as cash on hand.
2. Align units: Pre-money vs post-money, yuan vs USD, ARR disclosed or press-estimated. This piece treats ARR and gross margin as media-cited unless DeepSeek publishes filings.
3. Map the compute bill: Apply the ~70% rule to each 10 billion yuan tranche — chips, facilities, bandwidth, cooling — then compare to your own Agent or CI bottlenecks.
4. Read governance terms: Voting rights, lock-ups, state-fund exceptions determine who actually controls roadmap and capex priorities.
5. Benchmark horizontally: Place DeepSeek next to Moonshot, Zhipu, MiniMax, OpenAI, and Anthropic multiples — isolated "$70B" headlines mislead.
6. Translate to engineering choices: If you run local or private Agents, iOS CI, or Apple Silicon inference, model vendor capex waves affect API price and availability — but your TCO still hinges on whether you rent native Mac bare metal or absorb hypervisor tax on generic cloud VMs.

05 Controversies, context, FAQ, and engineering takeaway

Controversies: leaked meetings, voting rights, and bubble talk

• Leaked closed-door comments: Round 2's July pause reportedly traces to Liang's reaction after Round 1 investor-meeting remarks spread online — a governance and communications stress test as raise size and participant count grow.
• Voting-rights asymmetry: Most external Round 1 money has no vote and a five-year lock; the National AI Industry Investment Fund does not. Forbes and others have questioned governance optics. No official DeepSeek statement as of publication.
• Extreme multiples: 140–150x P/S exceeds even high-growth SaaS norms (often 30–50x at the top). Whether it holds depends on converting model leadership into scaled B2B revenue post-IPO — still unproven.

All Round 2 figures here trace to anonymous dealmakers and outlets including Caijing, Reuters, Bloomberg, and Forbes. DeepSeek has not publicly confirmed Round 2 terms. Treat numbers as negotiation references, not settled facts.

Context: STAR Market reform, the abandoned "three nos," and global AI capital

• STAR Market fifth standard expanded to AI (June 17, 2026, Lujiazui Forum): Shanghai Stock Exchange widened the fifth listing pathway to artificial intelligence — profit not required, large revenue not required if tech credentials pass review. That underpins DeepSeek's reported plan to file by end-2026 and list in 2027.
• Five-year "no fundraising, no IPO, no commercialization" — abandoned: DeepSeek ran on High-Flyer capital alone until Round 1 in June 2026. The shift mirrors peers: Zhipu and MiniMax are HK-listed; Moonshot's round cadence is similarly aggressive. The sector is moving from pure tech narrative to capital markets narrative.
• Global AI funding context: Premium multiples are not China-only. Market chatter put OpenAI near $300B in 2025; Anthropic reportedly passed OpenAI's June 2026 mark. Paying up for DeepSeek is partly a bet that Chinese labs keep global competitiveness against a massive domestic market.
• Compute sovereignty subtext: Custom inference chips and owned data-center expansion — covered in our chip strategy piece — explain why revenue-light labs still need back-to-back mega-rounds.

Citable hard data (as of publication)

• Round 2 targets: ~50 billion yuan raise; ~500 billion yuan (~$70B) pre-money; ~+43% vs Round 1 post-money — not signed.
• Round 1 closed: ~50 billion yuan (~$7.4B); post-money >350 billion yuan ($52–59B range); Tencent ~10B yuan, CATL ~5B yuan among reported tickets.
• ARR and multiples: ~$400–500M ARR (press); implied P/S ~140–150x; gross margin reportedly >50% (unaudited).

FAQ

Sources and further reading (verify against latest official filings after publication):

Sina Finance (Caijing): DeepSeek resumes fundraising at ~500B yuan pre-money valuation

The Standard HK: DeepSeek resumes second funding round at 500B yuan pre-money

Bloomberg: Round 2 pause coverage (leaked investor-meeting comments)

Reuters: DeepSeek custom chips and funding pace reporting

When top labs pour raises into owned data centers and custom silicon, API supply and pricing wobble — but pinning every workload on a single hyperscaler VM stack still means hypervisor overhead, weak Apple Silicon and iOS toolchain fit, and long-cycle cost drift. For teams that need zero-loss native compute, stable iOS CI/CD, and 24/7 AI Agent automation, ZUKCLOUD bare-metal Mac mini cloud nodes are usually the better production fit: dedicated Apple Silicon hardware, no hypervisor tax, always-on, elastic daily/weekly/monthly billing. See pricing or order; architecture rationale in our bare-metal architecture manifesto.