Liang Wenfeng does not appear among Chengyan's direct partners. The change recorded in Hangzhou's corporate registry on Nov. 17 adds a new partnership, Hangzhou Chengyan, holding 17.265% of DeepSeek's operating company. The partnership added 31 new limited partners, including Tencent, CATL and China Mobile. Liang sits one level up: he owns 99% of Hangzhou Chengce, which serves as the general partner of both investment vehicles, Chengyan and Chengli. Chengce's own contribution to Chengyan fell from 83.33% to 0.0001%.

That is how the founder raised China's largest AI round while keeping control. DeepSeek closed about RMB 100 billion, roughly $15 billion, in its second external round. A day earlier, Washington tightened the rules on Chinese access to advanced chips. On the same day as the registry change, Nvidia said it had stopped shipping licensed Hopper data-centre products to China.

The size of the round is only one figure in the filing. Compute remains the binding constraint for DeepSeek. Since Nov. 15 the company has limited long conversations to 72 hours of inactivity and restricted some document reading, citing insufficient computing capacity. The company continues to ration tokens after raising $15 billion.

The control structure

Chengyan's registered contribution jumped from RMB 6,000 to about RMB 3.439 billion. The operating company's registered capital rose 21.1%, from RMB 16.45 million to RMB 19.92 million. These figures are small next to the headline round, and the gap matters for anyone reading the structure of the deal.

The money reaches DeepSeek indirectly. Outside investors buy limited-partnership units in Chengyan, which holds the operating-company stake. Hangzhou Chengce, 99% owned by Liang, remains the general partner. One venture investor told Chinese media the lock-up runs seven years, plus five more after an IPO. Another account describes a five-year lock-up with no voting rights. In both cases the capital arrives with a long horizon and control stays at the top.

State-linked money is the main change. Trivium put state investors at 27% of this round, against about 2% in the first. Tencent's Shanghai Hengxiu and three CATL funds each took 9.5653% of Chengyan. China Mobile's investment arm and CITIC Securities took 7.6522% each. The National AI Fund lifted its direct stake in the operating company to about 0.39%. The structure reflects industrial policy.


RMB 100B second round, Caixin ↑ 100% vs first round

DeepSeek's second external round

The first round was RMB 50 billion: closed in June, registered in July. This one roughly doubles it. · Caixin, Nov 2026

Two estimates of the round's size

Caixin reports the round at about RMB 100 billion, or $15 billion. The Information puts it at RMB 105.5 billion, about $15.8 billion, with a post-money valuation near RMB 615.5 billion, roughly $92 billion. Caixin's free text does not print the valuation. DeepSeek has not confirmed either figure.

The registry does not settle it. Chengyan's RMB 3.439 billion of registered contribution is about a thirtieth of the reported RMB 105.5 billion, and public reports do not explain the difference between subscribed capital and money actually paid in. Yicai and industry analyses are explicit: RMB 105.5 billion comes from reports about the transaction, and the registry filing contains no amount. One aggregator reproduced the figure as RMB 1.055 trillion, a decimal-place error.

The arithmetic is straightforward. If the post-money is RMB 615.5 billion and the round is RMB 105.5 billion, the pre-money is RMB 510 billion, close to the RMB 500 billion Reuters reported in October. 36Kr's dollar valuation of $86 billion implies 7.16 yuan to the dollar, which does not square with its own $15.8 billion for the same round. Currency conversion affects these figures more than the filings do.

The filings do support the scale. The first round raised RMB 50 billion at a RMB 350 billion post-money valuation; it closed in June and the registration was completed in July. Liang personally committed RMB 20 billion. Across both rounds the total is close to RMB 156.5 billion, about $23.5 billion. That is more than most listed AI companies have raised in their entire lives.

DeepSeek has hired CITIC Securities to prepare a STAR Market listing and intends to start the process this year: a prospectus is expected by the end of 2026, with a listing in 2027. On the available numbers, the IPO will serve mainly as a price-discovery mechanism for every private AI asset in China.

Chengyan partnerShare of partnershipSubscribed (RMB)
Shanghai Hengxiu (Tencent)9.5653%329 million
Three CATL funds9.5653%329 million
China Mobile Investment7.6522%263 million
CITIC Securities Investment7.6522%263 million
Monolith (Tianjin Lisi Xingshen)7.1740%247 million
China Life Equity4.7826%164 million
Chengyan LP contributions after the Nov. 17 registration. Chengyan holds 17.265% of DeepSeek. Source: Tianyancha/Qichacha via Yicai and National Business Daily (36Kr), Nov. 19, 2026.

The first round as the template

The first round set the pattern. DeepSeek raised RMB 50 billion at a RMB 350 billion post-money valuation, with Liang personally committing RMB 20 billion. Tencent, the CATL system, NetEase, JD.com and IDG came in alongside him. 36Kr puts the first round at RMB 51 billion, a difference that matters only when building a precise cap table.

The second round followed the same machinery. It launched in mid-July with a RMB 50 billion target, the target was revised upward twice, and the round closed near RMB 105.5 billion by the accounts that use The Information's figure. First-round investors Monolith and IDG participated again; alongside them, Zhengxingu, Shixiang, Boyu and CPE took stakes of about 3.35%. According to National Business Daily, DeepSeek required every GP to fully disclose its underlying LP lists with look-through verification at each level; some investors were rejected despite sufficient capital, and a stake could be recovered if the review was not passed.

The selection logic was not purely financial. An institution interviewed by Chinese media said industrial reputation and the resources an investor could bring carried more weight than the size of the cheque. That is how China Mobile, China Telecom and a list of strategic industrials ended up in a model lab's cap table. It also explains the state share rising from about 2% to 27% in four months.

For a company heading toward a STAR Market listing, the design is deliberate. Outside capital sits in a partnership, takes a lock-up and gets no votes. Liang keeps the general partner. A domestic IPO can price the equity without handing control to a consortium of strategics. The round is large because compute is expensive, and the structure is conservative because the listing is the main goal.

The lock-up is telling. Seven years plus five after an IPO matches the horizon of infrastructure capital. Investors are being asked to fund a build cycle. Tencent and CATL tying at 9.5653% each, with China Mobile and CITIC just behind at 7.6522%, form a syndicate assembled for permits, power and distribution.

The week Washington tightened export controls

The BIS interim final rule, "Addressing Remote Access to Advanced Semiconductors", was announced on Nov. 16 and took effect on Nov. 19. It requires a licence for advanced chips and servers sent to any data centre worldwide that provides more than 12 million total processing performance (TPP) to companies headquartered in China or Macau. Twelve million TPP is roughly the ceiling of 750 H100 or H200 chips.

The rule also adds licence requirements for chip exports to Malaysia, Singapore, Indonesia, Thailand and Brazil, and rescinds the 2025 AI Diffusion Rule. Existing operators have 90 days to comply. The key criterion now checked is the jurisdiction of the customer's headquarters, while the address of the data centre is secondary.

Nvidia moved the same week. Interim CEO Colette Kress told analysts the company sold $2.4 billion of licensed Hopper data-centre products to China in the quarter ended Oct. 25, and that none of that business is in guidance.

Our guidance does not have any licensed Hopper data center products included, and we are no longer shipping licensed Hopper data center products to China.— Colette Kress, interim CEO, Nvidia, Nov. 17, 2026

Jefferies estimates the new rules could trim $44 billion to $62 billion from Nvidia's data-centre revenue next year. The company says its backlog for Vera Rubin makes that immaterial to the top line. Both statements can hold: a supplier can lose a market and still sell everything it can make.

What the remote-access rule covers

Licences apply to exports, re-exports and transfers of advanced chips and servers to data centres that serve Chinese or Macau-headquartered customers above the 12 million TPP threshold, including indirect access through brokers. Data-centre operators and suppliers must verify customer ownership. Compliance window: 90 days from the Nov. 19 effective date. An exemption applies only to systems used exclusively as commercial recommender services.

Ascend as an expensive substitute

The domestic substitute exists and is expensive. Epoch AI figures cited by Maybank put Huawei's Ascend 950DT at one PFLOP/s in FP8, about half an H100. Its FP4 throughput trails Nvidia's B300 by 6.75 times. Cost efficiency shows a wider gap: 0.05 tera operations per second per dollar for Ascend, against 0.34 for the B300.

Scale is shifting anyway. Chinese chipmakers took 41% of China's AI accelerator market in 2025, according to Caixin, with Huawei the main beneficiary. DeepSeek has worked with Huawei since September 2024 to adapt its R1 model for Ascend supernodes. In September 2026 it open-sourced infrastructure components for the platform. The company is not waiting for a policy reversal.

Maybank argues DeepSeek is relatively resilient because of that Huawei work. Labs that rent most of their compute and lack a consumer or advertising business to fund research look more exposed. Access to chips becomes a balance-sheet question, and the balance sheets now favour the large platforms.

DeepSeek's plans are measured in gigawatts. Caixin reports the company plans 1GW of capacity in Ulanqab, the Inner Mongolia cluster with cheap wind power and new data halls. About 160,000 Ascend 950DT chips have been ordered for the cluster; deliveries ramp from the fourth quarter of 2026, with commissioning expected in late 2027 to early 2028. Maybank estimates the delivery wait exceeds a year. The project needs both energy and compute hardware; the capital is raised in yuan, and the capacity still has to be built.

Two scenarios are possible. In the first, the money solves the problem: domestic accelerator supply improves, Ulanqab energises on schedule, the token limits come off, and the round looks cheap. In the second, yields stay low, power connections slip, and the round becomes a bridge to an IPO that has to explain infrastructure spending alongside modest gross margins. The registry does not show which scenario will play out; the next two quarters of capacity additions will.

Capital and production limits

The optimistic reading is that DeepSeek raised a war chest and will spend it on domestic silicon. The more cautious reading is that the round buys optionality while throughput depends on production. Neither a registry filing nor an export licence resolves fab capacity or chip yield. The token limits on DeepSeek's own app are the most direct confirmation of the shortage.

As we wrote in November, compute credit has become the constraint that decides which AI labs scale (the Marble case). DeepSeek tested that proposition at the largest scale China has attempted: raise the money, keep the control, and find that the scarce input is still a rack of accelerators and the power to run it.

The STAR Market filing will give outside investors their first audited look at revenue and depreciation. The Ulanqab build will show whether the 1GW plan is achievable. Until then, the most informative figure is the 72-hour limit on a conversation.

📊
Signals to watch

STAR Market filing date and the valuation anchor it sets

Ulanqab commissioning: Ascend 950DT deliveries from Q4 2026, cluster launch late 2027 to early 2028

Ascend 950DT volume shipments and yield

Whether DeepSeek lifts its token throttling after new capacity lands
DeepSeek closes a RMB 100 billion second round
The registry change, the 17.3% investor stake, and the compute squeeze behind the raise.
Primary account of the close, the stake and the export-control backdrop.
The Chengyan partnership, partner by partner
Registered-capital changes, the 31 new LPs and the look-through stake table.
Registry-level detail that the headline round numbers leave out.
The remote-access rule, explained
The 12 million TPP threshold, the new country licences and the 90-day compliance clock.
The policy text that worsened DeepSeek's compute problem the week its round closed.