Is China's AI API Pricing Sustainable? Verifying the Cost Structure Through Disclosures

Sat, 18 Jul 2026 11:00:00 +0900

Context This post is a follow-up to What China’s Open-Model Convergence Actually Changes. Where that piece covered the value-chain implications of performance convergence, this one verifies through Hong Kong disclosures whether that low-cost API pricing is sustainable at the cost-structure level. Best read alongside Kimi K3 Resets the AI Price Curve and The Real Debate in Semiconductors. Related hubs are the AI HBM Hub and the Exclusive Analysis Hub.

TL;DR

Current Chinese AI API pricing is a blend of structurally lower cost and strategic loss-leading competition. It is unlikely that every ultra-low price holds, and equally unlikely that prices fully revert to US frontier levels.
Hong Kong disclosures split the answer. Zhipu’s API gross margin improved to 18.9% (from 3.3% a year earlier), meaning it has started clearing marginal inference cost. But its total gross profit covers only 9.3% of R&D, so fully loaded cost remains unrecovered.
China’s cost edge is real, but it is not Huawei silicon or cheap electricity. The order is model architecture and reduced active parameters > batching, caching and quantization > cloud GPU utilization > low target margins > power and domestic NPUs.
Even generously sized, the power effect amounts to only about 2% of compute price. The fact that Qwen 3.5 Flash is priced identically in Beijing, Frankfurt and Virginia backs this up.
Unlike EVs, AI models have low switching costs. If winner-take-all emerges, it will be in the cloud and distribution layer, not the model layer. The most likely path is a low-price oligopoly (60%). [Analysis scope]

Key Framing

Chinese API prices are low not because they are already self-sustaining at a normal price level, but because improving inference efficiency and massive capital raising are propping them up together. That is a burden on the API margins of Western model vendors, but because the capital raised flows back into training and computing infrastructure, it does not mean total semiconductor demand is falling.

1. How Cheap Is It

Prices per 1M tokens as of July 2026. [Fact: official pricing pages]

Model	Input	Output
DeepSeek V4 Flash	$0.14	$0.28
DeepSeek V4 Pro	$0.435	$0.87
Gemini 3.5 Flash	$1.50	$9
GPT-5.6 Sol	$5	$30

DeepSeek V4-Pro is about 11 times cheaper than GPT-5.6 Sol on input and about 34 times cheaper on output. That does not, of course, mean performance, latency, SLA, security and tooling are equivalent.

The Price Gap Is Large Even Within China

Qwen 3.7 Max: RMB 12 input, RMB 36 output, currently 50% off
Doubao Seed 2.1 Pro: RMB 6 input, RMB 30 output
DeepSeek V4-Pro: far more aggressive than even its Chinese rivals

So DeepSeek’s pricing should not be read as the normal price for China as a whole. [Inference: price dispersion reading]

2. The Sustainable Part: Marginal Cost

DeepSeek V4-Pro is an MoE architecture that activates only 49 billion of its 1.6 trillion total parameters. Combining high utilization with batch processing and caching can sharply lower the actual compute cost per token. The cache-hit price for repeated input is just $0.003625 per 1M tokens. [Fact: DeepSeek official materials]

So for the following traffic, even current pricing can be sustainable on a marginal-cost basis.

Batch processing
Cache hits
Traffic without a latency guarantee
High-volume use that can sustain high utilization

3. The Hard-to-Sustain Part: Fully Loaded Cost

Published pricing alone struggles to recover model training cost, R&D, failed experiments, idle capacity, and security, sales and enterprise SLA costs.

Price normalization has already begun. [Fact: company actions and reporting]

DeepSeek cut V4 pricing 75%, then doubled peak-hour pricing
Alibaba also raised prices on some AI services by up to 34%
Dedicated enterprise throughput and low-latency service are priced above the public API

This is evidence that ultra-low prices do not hold across every time slot and customer segment.

Service	Sustainability
Cache, batch and non-priority traffic	High
Peak hours, low latency, high concurrency	Low
Enterprise security and SLA	Settling into separate, higher pricing
Full cost recovery for a standalone API business	Difficult at current prices

4. Verifying Through Disclosures: Zhipu and MiniMax

This is where the real analysis begins. Rather than speculate, we can check this against the disclosures of Hong Kong-listed companies.

2025	Zhipu (02513)	MiniMax (00100)
Revenue	RMB 724m	$79.04m
API/platform revenue share	26.3%	32.8%
API or company-wide gross margin	API 18.9%	Company 25.4%
Prior-year gross margin	API 3.3%	Company 12.2%
R&D / revenue	439%	320%
Adjusted net loss / revenue	439%	317%
Gross profit / R&D	9.3%	7.9%

[Fact: Zhipu 2025 results, MiniMax 2025 annual report]

Zhipu: Marginal-Cost Recovery Is Confirmed

API and open-platform revenue rose 293%, from RMB 48.48m to RMB 190.4m, and gross margin climbed from 3.3% to 18.9%. The company attributed this to inference efficiency, economies of scale and price increases.

There is a more important fact. As of March 2026, the company had raised API prices 83% versus the end of the prior year, and demand still exceeded supply. GLM-5’s current official price is RMB 4 input and RMB 18 output per 1M tokens. The top-tier API appears to have moved past a structure that dumps below direct inference cost. [Inference: implication of the price hike with excess demand]

MiniMax: Plausible but Unconfirmed

The company has not disclosed an API-only gross margin. Company-wide gross margin improved from 12.2% to 25.4%, which it attributed to model and system efficiency and optimized infrastructure allocation.

M2.5 is currently priced at $0.30 input and $1.20 output per 1M tokens, with a fast tier at $0.60/$2.40. Because M2.5 launched in 2026 while the disclosed financials run only through 2025, however, the standalone API cost ratio at current pricing has not yet been verified. [Blocked]

Fully Loaded Cost Remains Unrecovered at Both Companies

Zhipu’s total gross profit of RMB 297m covers only 9.3% of its RMB 3.18bn in R&D. MiniMax fared similarly: gross profit of $20.08m covered only 7.9% of its $253m in R&D, and operating cash outflow was $280m.

Services MiniMax purchased from the Alibaba Group in 2025 totaled $75.88m, or 96% of revenue. That figure blends training and serving cloud spend, so it cannot be read directly as API cost, but it is evidence of heavy reliance on external computing infrastructure. Alibaba holds a 17.06% stake in MiniMax.

Importantly, the disclosure states that related-party transactions were conducted on the same public pricing and terms offered to general customers. Capacity assurance and integration convenience from the strategic relationship are plausible, but there is no basis to conclude that a hidden price subsidy existed. [Inference: reading of the disclosure language]

Even after listing, Zhipu raised a net HK$31.375bn, and MiniMax raised a combined HK$15.957bn through new shares and convertible bonds. Most of the proceeds go into R&D and computing infrastructure. The fundraising itself is not evidence of an API loss, but frontier-model competition is not, at this stage, sustained by internal cash alone. [Fact: disclosure]

Verdict

Marginal-cost sustainability of paid, top-tier APIs: confirmed for Zhipu, plausible but unconfirmed for MiniMax
Sustainability of fully loaded cost including training: unmet at both companies
Sustainability of the price war: high. Capital-market fundraising can keep loss-making prices in place for a long time
Market structure: rather than winner-take-all, more likely segment-by-segment oligopoly, for example Zhipu in Chinese enterprise and on-premises deployment, MiniMax in overseas, consumer and multimodal

5. Decomposing the Cost Advantage: What Is the Real Reason

China’s cost advantage in API serving is real. But the core driver is not cheap electricity or Huawei silicon alone, it is the following combination.

Small active-parameter count and quantization
+ High batch-processing rate and cache utilization
+ GPU utilization at large-scale clouds like Alibaba
+ Low target margins
+ Cheap power in western China

Factor	Cost advantage	Judgment
Model architecture, quantization, caching	Very large	Core of China’s API price competitiveness
Cloud scale and utilization	Large	Alibaba’s most defensible moat
Low target margins	Large	Why prices are cheap, but a weakness for long-run profitability
Huawei Ascend	Moderate, conditional	Advantageous on optimized workloads, weak on general-purpose use
Western China power	Moderate	Effective for batch inference, limited for global real-time APIs
Government subsidies and policy finance	Present	Favorable for industrial buildout, but per-service amounts are opaque

6. Is Power Really the Deciding Factor

It is true that Chinese power is cheap. The landed power price at the Ningxia Zhongwei data center is RMB 0.36/kWh as of 2026, about 45% of the level in eastern China. The 2025 US industrial average is 8.62 cents/kWh, which works out to about RMB 0.62/kWh at an assumed exchange rate of 7.2 RMB/USD. Zhongwei is about 42% cheaper. [Fact: Chinese local government data, US EIA]

But power alone does not make API prices 5 or 10 times cheaper. Generously sizing the power effect, assuming a 1kW IT load per card and a PUE of 1.1, gives this.

Zhongwei: 1kW × 1.1 × RMB 0.36 = RMB 0.40/hour
US average: 1kW × 1.1 × RMB 0.62 = RMB 0.68/hour
Difference: about RMB 0.29/hour

Alibaba’s L20 on-demand price is RMB 14.4/hour, so even under this upper-bound assumption, the power difference is only about 2% of compute price. [Inference: own calculation]

Power matters, but it is not the main driver of the API price gap.

The Decisive Counter-Evidence

The western-power advantage suits batch inference and training well, but low-latency APIs must be deployed in eastern or overseas regions close to users. And Alibaba’s Qwen 3.5 Flash is priced identically at $0.029 input and $0.287 output per 1M tokens in Beijing as well as in Frankfurt and Virginia. [Fact: Alibaba Model Studio pricing]

That is the most direct evidence that low-cost APIs are not simply a result of Chinese power. [Inference: implication of identical regional pricing]

7. Alibaba’s Real Moat Is Utilization

Alibaba Cloud’s real strength lies less in power than in its ability to keep GPUs from sitting idle. [Fact: Alibaba official documentation]

Batch inference: 50% off list price
L20 spot instances: typically around RMB 2.88, versus RMB 14.4 on demand, a saving of roughly 80%
Idle GPUs: $0.000007 per CU idle rate versus $0.000018 per CU active rate
Mixing on-demand, spot and autoscaling to handle peak traffic

Spot capacity carries reclaim risk, however, so a real-time API with a strict SLA cannot run entirely on spot. The structure keeps base demand on on-demand or dedicated resources and routes only elastic demand to spot. [Inference: operational constraint]

8. Is Huawei Ascend Unconditionally Cheap

Strictly speaking, Ascend is not a GPU but an NPU.

Huawei makes the following claims for CloudMatrix 384. [Fact: Huawei official materials]

Single-card-equivalent inference throughput of 2,300 tokens/s
About 4x versus a non-supernode configuration
MFU improved by more than 50%
3-4x average per-card inference performance versus H20

It is a system-design approach that pools resources and uses MoE expert parallelism to compensate for the weaknesses of lower-spec individual chips.

But It Should Not Be Read as Low Cost Right Away

Huawei has not published an externally verified TCO on a like-for-like basis of model, latency and power. [Blocked]
Because CloudMatrix raises throughput by binding together many NPUs and network fabric, per-card throughput and total cluster cost are different things.
A 2026 field study of Ascend 910 needed 12 source patches, disabling of some high-throughput features, and repeated fault-handling measures to achieve stable inference. Even if the hardware is cheap, engineering and operating costs can rise. [Fact: Ascend field study]

Conditional Verdict

Condition	Verdict
Models deeply optimized for Ascend, such as Qwen, DeepSeek and GLM	Cost competitiveness possible
CUDA-based models ported as-is	Cost advantage uncertain
Large-scale batch inference within China	Favorable
Global enterprise real-time service	Unconfirmed once software and SLA costs are counted

9. How Is This Different From EVs

The similarities are a strategic industry, large-scale investment, large-enterprise subsidies, and capturing share through price cuts.

The difference is that switching costs for AI models are far lower. API specifications are similar, and customers can route across multiple models simultaneously or self-host open-source models. If one vendor raises prices, competing models and self-hosting create a price ceiling.

Platforms that combine cloud, data, security, payments, advertising and productivity tools, by contrast, have high switching costs. If winner-take-all appears, it will be in the distribution and cloud layer rather than the model layer. [Inference: switching-cost structure]

The most likely structure looks like this.

Alibaba: cloud, e-commerce, enterprise customers
ByteDance: consumer traffic, advertising, content
Tencent: WeChat, enterprise services
Huawei and Baidu: domestic infrastructure plus public-sector and enterprise customers
One or two independent model companies such as DeepSeek: setting the technical reference price

10. The Expected Path

Scenario	Assessed probability	Outcome
Low-price oligopoly	60%	Consolidates to 3-5 players, base API stays cheap, peak and SLA pricing rises
Full commoditization	25%	Open source, MoE and in-house chips drive further price declines
Large-scale normalization	15%	Chip, power and funding burdens push prices up 2-5x

[Inference: scenario estimate]

Even a 5x price increase would put DeepSeek V4-Pro’s output price at about $4.35, still below US frontier models. Current pricing may not be the floor, but the direction toward lower prices is hard to reverse.

11. Semiconductor Implications

Cheap APIs raise inference usage, which favors demand for server DRAM, NAND, networking and power. At the same time, they lower compute per token and HBM intensity, which pressures the per-unit economics of premium GPUs and HBM.

Target	Judgment
Samsung Electronics	Relatively cushioned by its large general-purpose DRAM and NAND exposure
SK Hynix	A long-run warning sign for the HBM scarcity premium
NVIDIA	Pressure on premium GPU unit pricing, offset by rising total inference volume
Alibaba, Tencent	Benefit from cloud and distribution ecosystem expansion rather than API margin

The key variable is not the rate of API price decline but the growth rate of total token usage. If usage grows faster than prices fall, the Jevons effect wins; if not, the multiple compresses first for premium GPUs and HBM. [Inference: directional judgment]

12. The Decisive Test Is the First-Half 2026 Disclosures

Does Zhipu’s API gross margin climb above 25-30% following the 83% price increase
Does MiniMax disclose an API cost ratio
Does revenue growth keep outpacing computing-cost growth
Does the gap between peak/SLA pricing and public API pricing become entrenched
Does Huawei publish a like-for-like, externally verified TCO

These five questions separate “sustainable cost” from “prices propped up by capital.”

Closing

Chinese API prices are low not because they are already a self-sustaining, normal price level. It is because improving inference efficiency and massive capital raising are propping them up together.

China has a structural cost advantage in domestic, batch-oriented, optimized model serving. But the extremely low API prices observed today cannot be explained by Huawei silicon and cheap power alone. The largest cost advantages, in order, are model size and reduced active parameters, batching/caching/quantization, cloud GPU utilization, and low target margins, with power and domestic NPUs coming after those.

From an investment standpoint, China’s cheap APIs do not immediately mean that “Chinese AI has secured an overwhelming cost moat.” The more accurate reading is that as inference service rapidly commoditizes, model companies’ margins fall, and value is increasingly likely to shift toward cloud, power and semiconductor infrastructure.

That is a burden on the API margins of Western model vendors, but because the capital raised flows back into training and computing infrastructure, it does not mean total semiconductor demand is declining.

This post synthesizes Hong Kong listing disclosures (Zhipu 02513’s 2025 results and share placement, MiniMax 00100’s 2025 annual report and fundraising), official company pricing pages (DeepSeek, OpenAI, Google, Alibaba Model Studio, Volcano Engine, Zhipu BigModel, MiniMax Platform), Huawei official materials and CloudMatrix-related research, Chinese local government power data, and US EIA statistics. MiniMax’s standalone API cost ratio and Huawei’s like-for-like, externally verified TCO remain undisclosed, and the scenario probabilities and power calculations are author estimates based on assumptions at the time of writing. The stocks mentioned are examples used to illustrate cost structure and are not a recommendation to buy or sell any specific security. Prices and disclosed figures are as of the announcement date and may change afterward. Investment decisions and responsibility for them rest with the individual investor.

Cost Structure on Korea Invest Insights