How Chinese AI labs pay for it: Alibaba capex, DeepSeek and Kimi prices

Chinese AI labs fund themselves in different ways. Alibaba, which builds the Qwen models, spent RMB 67.678 billion (about US$9.975 billion) on capital expenditure in the quarter ended 30 June 2026. DeepSeek and Moonshot AI, the maker of Kimi, are private and publish no audited spending figures, only price lists and model pages.

This post focuses on money: what is filed, what is published, and what is missing. For the model by model story, including sizes, licenses and release dates, see the companion post on releases and stated direction. Nothing here is investment advice, and this post does not predict what any company will do.

Key numbers at a glance

Alibaba figures come from its results release for the quarter ended 30 June 2026, furnished to the SEC on 20 August 2026. The company states US dollar amounts as convenience translations. Prices are the labs' own published prices, checked 19 October 2026, and can change.

• Alibaba revenue, quarter ended 30 June 2026: RMB 268.953 billion (US$39.639 billion), up 9 percent year over year.
• Alibaba AI Cloud and Compute Services revenue: RMB 48.437 billion (US$7.139 billion), with growth in total and external customer revenue both accelerating to 45 percent, per the company.
• Alibaba AI related product revenue: RMB 12.376 billion (US$1.824 billion), the twelfth consecutive quarter of triple digit growth, per the company.
• Alibaba capital expenditure, same quarter: RMB 67.678 billion (US$9.975 billion), up 75 percent.
• Alibaba free cash flow, same quarter: an outflow of RMB 44.670 billion (US$6.584 billion), against an outflow of RMB 18.815 billion a year earlier.
• Alibaba net income, same quarter: RMB 10.444 billion (US$1.539 billion), down 75 percent.
• Alibaba cash and other liquid investments, 30 June 2026: RMB 474.505 billion (US$69.933 billion), as the company defines the measure.
• DeepSeek V4 Pro API price: US$0.66 per million input tokens on a cache miss and US$1.98 per million output tokens, off peak. Peak prices are double.
• DeepSeek V4.1 Flash API price: US$0.15 per million input tokens on a cache miss and US$0.60 per million output tokens, off peak. Peak prices are double.
• DeepSeek V3.2-Exp price cut, 29 September 2025: "DeepSeek API prices drop 50%+, effective immediately," per DeepSeek.
• DeepSeek Search and Fetch APIs: US$0.75 per 1,000 calls for each endpoint, billing from 31 October 2026 at 16:00 UTC.
• Kimi K3 API price: US$3.00 per million cache miss input tokens, US$0.30 for cache hits and US$15.00 per million output tokens.

Three funding models in one country

It is tempting to treat "Chinese AI labs" as one group. The public pages show three different setups.

A listed conglomerate: Alibaba. Qwen is built inside a company that sells cloud computing, runs online marketplaces and files public accounts. That matters because the accounts show spending. A conglomerate can fund AI from the cash its other businesses make, and from borrowing, and it has to report the result every quarter.

Private labs with little disclosed finance: DeepSeek and Moonshot. Neither lab's own pages state funding rounds, valuations, investors or total spending. DeepSeek's documentation covers models, prices and developer tools. Moonshot's Kimi K3 post lists a price and model details but no funding amounts or investors. No primary source on either lab's funding was available for this post, so it reports none. Treat figures about their valuations or backers that appear elsewhere as unverified until a primary source confirms them.

Open weights as a business choice. Both private labs publish model weights, the downloadable trained files, and make money from hosted access and developer services. That is a different bet from selling only closed API access: the lab gives away the ability to self-host and competes on the price and convenience of its own hosted service.

Alibaba: the filed numbers

Revenue and the AI cloud line

Alibaba's quarter ended 30 June 2026 shows total revenue of RMB 268.953 billion, up 9 percent. That is modest growth for the whole group. Inside it, the company says its Cloud Intelligence Group no longer reports separately and was combined with T-Head, its chip design unit, into a segment called AI Cloud and Compute Services. Revenue for that segment was RMB 48.437 billion, and Alibaba says both total and external customer revenue growth accelerated to 45 percent.

A separate line, AI related product revenue, was RMB 12.376 billion. Alibaba calls this its twelfth consecutive quarter of triple digit growth. The release does not define exactly which products count, so treat it as a company defined category.

The other segments show where the group's cash comes from and where it is growing. Per the release, China Quick Commerce revenue was RMB 53.295 billion, up 45 percent; China E-commerce was RMB 110.900 billion, down 8 percent; International E-commerce was RMB 27.761 billion, down 1 percent; and a segment called AI Labs and Applications was RMB 3.338 billion, up 16 percent. The largest revenue line is still domestic e-commerce, and it shrank, while the AI cloud segment grew at five times the group rate.

Chips inside the cloud segment

Putting cloud and chip design in one segment is itself informative. The release says "T-Head has established a portfolio of proprietary silicon, spanning GPU, CPU, storage and networking chips," and that its Zhenwu chips have reached more than 650 external customers across over 20 industries. A reasonable reading is that Alibaba wants to sell chips and cloud capacity together, though the release does not give that reason in so many words. For the wider hardware context, see the post on domestic Chinese AI chips.

Capital expenditure and cash

Capital expenditure, or capex, is spending on long lived assets such as data centres, servers and chips. Alibaba's capex in the quarter was RMB 67.678 billion (US$9.975 billion), up 75 percent from a year earlier. Free cash flow, a non-GAAP measure the company defines as operating cash flow less capex with its own adjustments, was an outflow of RMB 44.670 billion (US$6.584 billion). A year earlier the outflow was RMB 18.815 billion.

An outflow means the company spent more on investment than its operations produced in the quarter. It is not a loss in the accounting sense, and one quarter does not make a trend. It does show that AI infrastructure is being paid for partly out of balance sheet strength.

That balance sheet is large. Alibaba reports "cash and other liquid investments" of RMB 474.505 billion (US$69.933 billion) at 30 June 2026. That company defined measure includes treasury investments beyond cash, so it is not the same as net cash. Within it, cash and cash equivalents were RMB 142.914 billion and short-term investments RMB 154.653 billion. The company also kept buying back stock, though at a small scale: it repurchased 13.4 million ordinary shares for US$162 million in the quarter.

Profit under pressure

Spending this heavily shows up in profit. Net income was RMB 10.444 billion (US$1.539 billion), down 75 percent. Non-GAAP net income, the company's adjusted measure, was RMB 20.715 billion (US$3.053 billion), down 38 percent from RMB 33.510 billion. The release does not attribute the whole decline to AI, and a single quarter mixes many effects, so the honest reading is narrower: capex rose 75 percent while profit measures fell.

The release does not state a multi-year infrastructure investment total. Figures for Alibaba's multi-year plan that circulate elsewhere should be checked against the company's own statements.

Qwen in the same filing

The release says Alibaba launched its flagship foundation model Qwen3.8-Max "within three months of its prior version" and opened model weights of 2.4 trillion parameters. The open model and its license are described in the releases post. The release also says "250 million users have had their first AI-driven shopping experience through Qwen app's agentic features," which is a company metric, not a count of active users. It gives no monthly active user figure for the Qwen app.

The business logic is visible here even without a stated strategy: open weights spread Qwen's architecture and attract developers, while Alibaba earns on cloud capacity, hosted Qwen APIs and commerce traffic routed through the Qwen app.

DeepSeek: pricing as the message

DeepSeek publishes the clearest picture of how it prices. Its API documentation lists two current models.

• deepseek-flash (DeepSeek-V4.1-Flash-1017): 1 million token context. Off peak prices are US$0.003 per million cache hit input tokens, US$0.15 per million cache miss input tokens and US$0.60 per million output tokens. Peak prices are exactly double.
• deepseek-v4-pro (DeepSeek-V4-Pro-0813): same context. Off peak prices are US$0.022, US$0.66 and US$1.98. Peak prices are double.

A cache hit is when the service recognizes input it has already processed and reuses the work, which is why it costs far less. Peak hours are listed as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours, including weekends, are off peak.

How the price list got here

The dated record on DeepSeek's own pages shows a steady pattern of cutting price as the architecture gets cheaper to run:

• 29 September 2025. The V3.2-Exp release introduced DeepSeek Sparse Attention, aimed at cheaper long inputs, and announced "DeepSeek API prices drop 50%+, effective immediately." The new price table was published only as an image.
• 13 August 2026. With V4-Pro's general release, DeepSeek said it would "adopt peak/off-peak pricing, with off-peak prices set at half of the peak-hour prices," effective 16:00 UTC on 16 August 2026. Time of day pricing pushes demand into hours when machines are less busy, which is a common way to raise utilization; DeepSeek does not state its reasons.
• 10 September 2026. The V4.1-Flash release said "API prices have been reduced accordingly."
• 17 October 2026. The V4.1-Flash-1017 upgrade left the calling method and pricing unchanged, and DeepSeek opened its Search and Fetch APIs at US$0.75 per 1,000 calls for each endpoint. Calls made through server side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts at 16:00 UTC on 31 October 2026, which is midnight on 1 November in Beijing.

The pricing page adds a caution: "Product prices may vary and DeepSeek reserves the right to adjust them."

What the prices do and do not tell you

US frontier model prices are covered in the post on Google's Gemini 4 Argon. The gap between those lists and DeepSeek's is large, but the models differ in size, test results and purpose, so price per token alone does not tell you which costs less per finished task. The comparison that matters is pass rate against cost per task on your own work.

DeepSeek's Hugging Face page for the V4-Pro-0813 checkpoint lists the MIT license for the repository and weights, one of the most permissive common licenses. V4-Pro's stated size, 1.6 trillion total parameters with 49 billion active, is set out in the releases post. On hardware, the 0813 page's serving example runs the model "on a single 4×GB300 node," an NVIDIA system; it does not mention Huawei hardware or say what DeepSeek trains on.

Moonshot and Kimi: a higher price, and candid about limits

Moonshot's Kimi K3 post, with images dated 16 and 17 July 2026, lists K3's API price at US$0.30 per million cache hit input tokens, US$3.00 per million cache miss input tokens and US$15.00 per million output tokens. That is several times DeepSeek's V4-Pro list price for output. K3 is also a much larger model to serve: Moonshot recommends supernodes of 64 or more accelerators, a reminder that open weights at this size are not something a small team runs casually. Its size and design are in the releases post.

The post calls K3 "the world's first open 3T-class model," a company claim that has not been independently checked. The Kimi-K3 model card on Hugging Face names a "Kimi K3 License" for both code and weights. The card does not summarize the terms, so read the license file before commercial use. That differs from DeepSeek's MIT license and from Qwen's mix of Apache 2.0 and custom licenses, which is why license reading belongs in any model selection.

The post is also unusually frank, conceding that the model's "overall performance still trails the most powerful proprietary models." The model card notes that some evaluations ran on NVIDIA H20 GPUs rather than the H100 in the official setting, and some scores use custom harnesses.

What this means for a buyer

• Price is the headline difference, and it is large across these labs' lists. Run your own tasks before concluding anything.
• Licenses differ: MIT, Apache 2.0, community licenses and custom licenses are different sets of rights. Read each.
• Hosting location matters. Using a lab's own API means your data goes to that provider. Running open weights yourself keeps data in your control but adds your own security duty, such as checking the model files and the serving software.
• Prices are not promises. DeepSeek states that it reserves the right to adjust prices, and it changed its prices three times between September 2025 and September 2026.
• Disclosure varies. An SEC filer like Alibaba publishes regular, standardized figures. Private labs do not. If your procurement needs financial stability evidence, you will need to ask the vendor.

On export controls, which shape which chips these labs can buy, see the timeline from official sources.

What is firm and what is uncertain

Firm, because it is in a filing or a lab's own page:

• Alibaba's quarterly revenue, segment revenue, capex, free cash flow, net income and liquidity figures.
• The DeepSeek and Kimi published prices, the dates of DeepSeek's pricing changes, and license names as of 19 October 2026.

Uncertain or not disclosed:

• Funding, valuations and total spending of DeepSeek and Moonshot. No primary source is public.
• Alibaba's multi-year AI investment plan. Not stated in the June quarter release.
• The size of DeepSeek's September 2026 price cut, which the updates page describes without numbers.
• Training hardware for all three labs.
• Benchmark scores. Self-reported, with custom harnesses in places.

What to watch next

• Alibaba's next quarterly results, for capex, free cash flow and the AI Cloud and Compute Services growth rate.
• DeepSeek's Search and Fetch billing start on 31 October 2026, and any further price changes.
• The Qwen4 release, which the Qwen3.8-Flash-Next card names as the target of its architecture preview.
• Independent tests of Kimi K3 and DeepSeek V4-Pro against closed models.
• Any primary disclosure on funding from DeepSeek or Moonshot.

If you run downloaded models or open source packages in your own stack, Safeguard offers SBOM and dependency scanning, and a free scan is available at https://gold.safeguard.sh/scan.

Quick check

Sources

• Alibaba Group, results for the quarter ended 30 June 2026 (Form 6-K, Exhibit 99.1), 20 August 2026: https://www.sec.gov/Archives/edgar/data/1577552/000110465926099220/tm2623667d1_ex99-1.htm
• SEC EDGAR list of Alibaba 6-K filings, checked 19 October 2026: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001577552&type=6-K&dateb=&owner=include&count=10
• DeepSeek API documentation, pricing page: https://api-docs.deepseek.com/quick_start/pricing
• DeepSeek API documentation, updates page: https://api-docs.deepseek.com/updates
• DeepSeek, V4.1-Flash-1017 and Search API release note, 17 October 2026: https://api-docs.deepseek.com/news/news261017
• DeepSeek, V3.2-Exp release note, 29 September 2025: https://api-docs.deepseek.com/news/news250929
• Hugging Face, DeepSeek-V4-Pro-0813 model page: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
• Hugging Face, Kimi-K3 model page: https://huggingface.co/moonshotai/Kimi-K3
• Moonshot AI, Kimi K3 post: https://www.kimi.ai/blog/kimi-k3
• Moonshot AI, blog listing: https://www.kimi.ai/blog/

Last reviewed: 19 October 2026.