From Xiaomi to Zhipu: The Q3 2026 Profitability Landscape for Chinese AI Model Vendors

Translated by TechFlow

TechFlow Brief: While everyone is comparing model scores, the real question is — how much are these models actually costing, and who is paying? This article breaks down the books of 19 Chinese model vendors and finds that none can profit solely from their models themselves. Whether you are an investor picking stocks or a technical leader selecting models, you should first understand this map of "who is losing money."

This article was co-authored by Robonaissance and Inside China's Machine.

The Landscape of Chinese AI Foundation Models | Q3 2026

Among all Chinese model vendors that have filed financial reports, none can prove that their models are self-sustaining. Parent companies plug the gaps with other businesses, while labs rely on investors to foot the bill—and investors' valuations keep shifting.

This is the first installment in our quarterly series on Chinese AI model vendors.

Xiaomi sells phones and cars. According to itself, nearly 30% of its 18.2 billion RMB half-year R&D budget went into AI. We estimate approximately 5.5 billion RMB ($800 million), which covers all its AI work including language models, so subsequent comparisons are not strictly apples-to-apples. In the same six months, Hong Kong-listed lab Zhipu's total R&D investment was 2.13 billion RMB ($316 million), making Xiaomi's spend more than double that.

On the independent benchmark Artificial Analysis Intelligence Index, the two language models differ by just one point. Xiaomi's MiMo-V2.6-Pro scored 46, and Zhipu's GLM-5.3 scored 45. To run the entire index, buyers spend $207 using MiMo and $2,503 using GLM. That is the buyer's cost. How much the model is worth to the vendor itself is another figure; this article asks: who is paying this bill? Xiaomi's answer comes from its phone and auto businesses. We view this profit alongside its AI expenditure here.

A model can break even only if revenue exceeds construction and operational costs, including R&D. Zhipu and MiniMax are two listed labs whose primary business is selling models; both report positive gross margins—meaning revenue minus direct customer service costs. This implies the gap does not lie within their reported revenue-cost figures.

Zhipu's books show a loss. Half-year revenue was 954 million RMB ($142 million), with a net loss of 2.07 billion RMB ($308 million). Based on our summary of company filings and reports, it raised or agreed to raise approximately 61 billion RMB ($9 billion) between July and mid-September, part of which has not yet been marked as completed.

Our answer: across all the financial reports we have read, not a single model supports itself. Gross margins from Zhipu and MiniMax customers cover only about one-eighth and one-fourteenth of R&D expenses, respectively. Among measurable parent companies, profits cover AI expenditures by roughly 2 to 7 times. The rest of the labs are funded by investors, whose stock prices are constantly fluctuating. Listed parent companies bury model costs within larger business segments, so standalone P&L statements for models are unavailable. Private labs and ByteDance do not file financial reports.

This is critical for investors assessing China AI exposure and operators selecting Chinese models to build products. Neither can judge a vendor's longevity solely from scores or prices.

Most Vendors Disclose Almost Nothing About Model Costs

Of the 19 vendors on the Figure 1 map, only five submitted financial reports containing model information: Zhipu, MiniMax, Sensetime, along with Alibaba and Baidu's AI business lines. Tencent, Xiaomi, and Kuaishou provided model data during earnings releases and conference calls; the remainder can only be tracked through company statements and media reports.

Who Covers the Losses: Chinese model vendors grouped by funding source, featuring flagship models, Artificial Analysis scores, disclosed funding data and reliability levels for each, plus stated methodology for burn rates.

Click image to enlarge. Reliability: Filing refers to reports submitted by the vendor themselves. Company refers to disclosures made at earnings calls or press releases. Reporting refers to named media, insiders, or unverified filings, which should be considered unverified. None means no information was disclosed. H1 refers to H1 2026; ARR refers to annual recurring revenue. Burn rates are not comparable across vendors. Zhipu's claimed $1.6 billion contrasts with its half-year revenue doubled to approximately $283 million. Ant Group and Huawei disclose only at the group level. Excluded from table: telecom operators (host and distribute models but do not disclose model revenue), DeepThaw, Shanghai AI Laboratory (non-profit institution).

Given such limited disclosure, buyers must start with what every model publicly offers: scores and prices. What do they tell us about the vendors?

Bills Within the Same Score Tier Differ by Up to 24 Times

The Artificial Analysis leaderboard rounds scores to integers. Five Chinese models score between 44 and 46: Xiaomi MiMo-V2.6-Pro, Alibaba Qwen3.8 Max, Zhipu GLM-5.3, Moonshot AI Kimi K3, and StepFun Step 5 Preview. These are the top five Chinese models on the Artificial Analysis leaderboard as of October 1. The highest-scoring US model on the same index scores 58, twelve points above the top Chinese model. Other Chinese vendors rank lower or are unranked.

Scores indicate model capability. Artificial Analysis also reports how much it costs to run the full index on each model—the price a buyer pays to run the same tests. Among the five, bills range from $207 for MiMo to $4,935 for Qwen, nearly a 24-fold difference within the same score tier. GLM's $2,503 is 12 times that of MiMo. Figure 2 shows the bills for all five.

Token counts may matter more than list prices. Tokens are text fragments, and models charge separately for input and output tokens. Qwen's output list price is lower than Kimi's, but Qwen cost $4,935 to run the test while Kimi cost $3,658, because Qwen consumed 9.9 billion input tokens during the test versus Kimi's 2.6 billion. Operators should evaluate model pricing based on their actual workloads.

Different Bills in the Same Score Tier: Costs for MiMo-V2.6-Pro, Step 5 Preview, GLM-5.3, Kimi K3, and Qwen3.8 Max to run the Artificial Analysis Intelligence Index, broken down by input and output, along with scores, token counts, output list prices, and three rows of US reference data.

Gray rows represent US reference data. Claude Opus 5.5 high setting scores 54 and costs $2,172, which is lower than GLM's score of 45 at $2,503.

MiMo's output list price is $0.87 per million tokens. Such a low price might reflect genuinely low costs, or it could simply be a customer-acquisition strategy, leaving buyers unable to tell. To understand how much a model costs a vendor and who pays, the evidence lies in the vendor's books.

Two Listed Labs Cannot Cover the Gap with Customer Revenue

Zhipu and MiniMax's financial reports consist almost entirely of their model businesses, with no parent company safety net. Zhipu's half-year revenue grew 399.7%, but it recorded a net loss of 2.07 billion RMB. Its gross margin was 26.4%, down from 50.0% year-over-year—revenue is growing rapidly, but margins are thinner. MiniMax revenue grew 283% to $116.6 million, with a net loss of $358 million, R&D spending of $296.9 million, and a gross margin of 17.9%.

Both labs earn positive gross margins on the products they sell. Whether a lab classifies computing power under cost of revenue or R&D expenses is its own choice, so the ratios below are for reference only. By our calculations, Zhipu's gross margin is approximately 252 million RMB against 2.13 billion RMB in R&D expenses; MiniMax's gross margin is approximately $21 million against $296.9 million in R&D: R&D expenses are roughly 8 and 14 times the amounts contributed by customers after service costs, respectively. R&D spending at both labs approaches the scale of net losses, making R&D the primary source of the gap.

Figure 3 places the gross margins of the two labs side by side with their R&D expenses. Sensetime is also Hong Kong-listed and develops the SenseNova model, accounting for 79.9% of total revenue for generative AI and 17.1% for visual AI. However, we cannot find model-level costs in its financials, so its accounts cannot demonstrate whether a model can sustain itself. Alibaba's AI segment will be discussed in the next section.

Comparison of gross margins vs. R&D expenses for Zhipu and MiniMax in H1 2026, accompanied by revenue and net losses: Zhipu's R&D is roughly 8 times its gross margin, MiniMax's roughly 14 times.

In all the accounts we have reviewed, none show that payments from customers cover the costs of building and running models, especially once R&D expenses are included. Someone else makes up the difference.

Parent Companies Only Pay When Core Businesses Are Solid

Three parent companies disclosed enough information to calculate coverage multiples, defined as profit after deducting AI expenses divided by those expenses. Others cannot be measured. Figure 4 places the labs' ratios side by side with the parent companies'.

Coverage Multiples: Ratio of customer gross margins to R&D expenses for Zhipu (0.12x) and MiniMax (0.07x), and ratio of profit after AI expenses to those expenses for Alibaba (Group 2.0x, Commerce 2.9x), Xiaomi (at least 2.2x), and Tencent (7.2x), spanning varying periods and accounting methods.

In Q2, Tencent's operating profit before certain items—an adjusted metric defined by itself—grew 9% to 75.6 billion RMB, after already deducting approximately 10.5 billion RMB ($1.56 billion) in costs from new AI products, including Hunyuan models, Yuanbao assistant, and coding tools. This 10.5 billion RMB is a cost item that includes partial revenue, and Tencent does not disclose AI revenue separately. Cash flow is where pressure shows: its free cash flow—a company-defined cash metric—was negative 13.8 billion RMB ($2 billion) due to upfront computing payments; without these prepayments, it would have been positive 37.6 billion RMB. Its Hunyuan 3 model scores 25 on the index, while the top five score 44 to 46, and its updated Hunyuan 4 preview is unrated. Therefore, the parent company with the highest coverage multiple here does not hold the top-ranking models.

Alibaba's newly established AI Lab & Applications segment consolidates its various model labs with consumer apps like Tongyi and QwenWork products, meaning its revenue covers more than just model sales. For the quarter ending in June, the segment reported a loss of 13.86 billion RMB ($2.06 billion) in adjusted EBITA terms, with revenue of 3.34 billion RMB ($500 million). Its commerce group reported an adjusted EBITA of 39.7 billion RMB ($5.9 billion), roughly triple that loss, while Alibaba's consolidated adjusted EBITA including the AI loss fell 30% to 27.3 billion RMB ($4.1 billion), approximately double the loss.

Xiaomi—the test case introduced at the beginning of this article—reports profit after R&D expenses. Adjusted net profit for the first half was 12.3 billion RMB ($1.83 billion), down 42.8% year-over-year. Based on our estimate of a maximum AI R&D spend of 5.5 billion RMB, the coverage multiple is at least around 2.2x, close to Alibaba's group figure and far below Tencent's. Xiaomi told analysts that revenue from model access sales has begun to materialize and that monetization is not currently its primary goal. The $0.87 output price is something a profitable parent company can afford to maintain, regardless of service costs.

By our calculations, across different metrics and periods (quarterly for Tencent and Alibaba, semi-annual for Xiaomi), the coverage multiples are roughly 7x for Tencent, ~2x at the Alibaba group level, and at least 2.2x for Xiaomi. Under the same assumptions, if post-AI-expense profits were to drop further without changes in spending, Alibaba's group-level multiple would reach 1x only if profits fell 50%, Xiaomi's only if it fell at least 55%, and Tencent's only if it fell 86%. Xiaomi's adjusted net profit has already dropped 42.8%, and the 2.2x figure is calculated after that decline. Figure 4 aligns metrics from different frameworks onto the same scale: lab bars use gross profit before R&D, while parent company bars use profit after AI expenses, showing which side of the 1x line each company falls on rather than precise peer-to-peer gaps.

Baidu and ByteDance are the remaining two largest parent companies and cannot be measured. Baidu's financials report revenue lines for its AI business but provide no AI cost data. Its network marketing revenue fell 19% in Q2, while its AI-related business lines—primarily AI cloud infrastructure—grew 25%. ByteDance does not file financial reports. The company claims its Doubao model processes 180 trillion tokens daily, a usage metric, while The Information, citing insiders, reports its profits are declining.

A parent company's coverage multiple is a moving target. Xiaomi's H1 adjusted net profit fell 42.8%, Baidu's quarterly network marketing declined 19%, Alibaba's group profit including AI losses dropped 30%, Tencent grew 9%, and ByteDance's profits are reportedly slipping. Not a single parent company's filing indicates whether its models have recouped costs. With no parent profits to rely on, who foots the bill for the labs?

Labs Pay with Shares, Digital Readers Cannot Verify

With no other businesses to tap, labs pay with shares and bonds. Zhipu's trajectory illustrates this. Its stock closed at a peak on June 22. In July, media reported it completed a share sale raising approximately 27 billion RMB ($4 billion). By September 9, the stock had fallen 62% from the peak closing price. On September 12, it agreed to a placement of approximately 13.5 billion RMB ($2 billion)—issuing new shares to investors—and a convertible bond of 20.14 billion RMB ($3 billion)—a loan convertible into equity. This 33.6 billion RMB was secured after the stock price dropped and has not yet been marked as completed. The announcement states 60% of placement proceeds will fund R&D and infrastructure, 15% for expansion and strategic investments, and 25% for working capital. By September 30, the stock was down 73% from the peak close.

We calculate Zhipu's total financing at approximately 61 billion RMB ($9 billion), nearly 30 times its half-year loss. If the September deals close, it has sufficient cash relative to its H1 spending pace. At the H1 loss rate, this cash runway could last several years, meaning H1 spending wouldn't force another sale immediately. But losses have not stalled: Zhipu's adjusted loss—excluding share-based compensation and similar items—widened 12.1% compared to H1 2025. The unanswered question is the pricing of the next round. Using the September 30 close, a new round would price at a stock down 73% from the peak. MiniMax raised approximately $2 billion in July via placement and convertible bonds, roughly 5.6 times its half-year loss.

Beyond shares, the numbers accompanying financing are unverifyable. Zhipu's data is the most auditable. The company stated in August that annualized revenue on its model access platform reached $1.6 billion, nearly six times the $283 million derived from doubling its full-half-year product revenue of 954 million RMB ($142 million). Annualizing the most recent period inflates the rate compared to the half-year average in fast-growing businesses; Zhipu's full-year accounts will validate this figure.

Private labs publish nothing. Cited by other media, The Information reported on September 24 that DeepSeek's API—the paid channel for developers to call its models—achieved an 82.9% gross margin in the first seven months of the year. This is the margin on selling tokens, excluding research and training, differing from Zhipu's filed 26.4% and MiniMax's 17.9% metrics, and DeepSeek has not confirmed it. Such high margins still fail to answer the core question because DeepSeek does not disclose its research costs. According to Caixin, its seed round raised approximately 50 billion RMB ($7.4 billion), valuation set at 350 billion RMB ($52 billion). Bloomberg reports Moonshot AI reportedly closed a $3.5 billion round in July at a $35 billion valuation. Neither round came with filed accounts. These numbers—whether filed or reported—show no model covering its own costs. What accounts actually prove the point?

Only Model-Level Accounts Can Close the Case

One filing proves the point: a division where model revenue exceeds model construction and operational costs, inclusive of R&D. Alibaba's AI Cloud and Computing division sells compute and cloud services, so it does not qualify. Three closer readings can show whether the gap is narrowing. Full-year accounts for Zhipu and MiniMax will come out next year, revealing whether gross margins are rising relative to R&D spend. In the first half, Zhipu's gross margin covered about one-eighth of R&D expenditure, and MiniMax's about one-fourteenth. Reaching R&D spend with full-year gross margin—the point where gross margin alone covers research—or a model-level account showing revenue above costs would settle the debate. If the coverage ratio reaches at least double the H1 figure—one-quarter for Zhipu, one-seventh for MiniMax—but remains below 1x, it shows the gap is narrowing while the premise holds true. Coverage equal to or below H1 numbers indicates no progress. Listing documents for private labs, complete with audited gross margins and R&D spending, will test claims like DeepSeek's. If a parent company reports a line item pairing model revenue with model costs, it will prove the point from the opposite angle. Only the full-year accounts of labs have a deadline.

Until these accounts arrive, longevity can be gauged by two things: the extent to which a parent company's other businesses cover the model division—Tencent, Alibaba, and Xiaomi reports allow readers to verify—and the stock price when the lab next sells shares, which cannot be predicted in advance. Operators building on any particular Chinese model face the same dual questions; they should ask who funds their supplier and prepare a backup model. On July 19, Moonshot AI announced it paused new consumer subscriptions because request volume for Kimi K3 approached its compute cluster limits. Its statement did not mention APIs, so operators relying on Kimi should inquire whether API capacity is also constrained. Investors can gauge risk exposure with two questions: how much other businesses at the model side cover its AI spend, and when it will next need to sell shares.

Xiaomi—the lowest-accounting-risk vendor—covers its AI R&D with adjusted net profit at least 2.2x over, though that profit fell 42.8% in the half-year. Zhipu has raised or agreed to raise financing nearly 30 times its half-year loss, while customer gross margins cover only one-eighth of its research. Therefore, its endurance depends on whether these financings close and the pricing of the next round—its stock is already down 73% from the peak.

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps