The Dark Horses of Tokenmaxxing Era Threaten an Inference Price War

The tokenmaxxing phenomenon has rattled both AI lab CEOs and customers. The former have acknowledged that token pricing is a major issue, while the latter increasingly face pressure to control inference costs.

But Chinese AI models seem to be at an advantage. Their APIs are often priced 5–10x lower than comparable US offerings while delivering similar performance.

Several founders and developers have reported significant savings after switching from Western models to DeepSeek.

Recently, Florent Crivello, CEO and Co-founder of Lindy, an AI platform building customisable, no-code agents, wrote on X that the company shifted 100% of its traffic to DeepSeek V4, moving away from Anthropic’s models. According to Crivello, the transition is saving the company millions of dollars while improving performance across several core use cases. He added that Chinese models are likely to exert margin pressure on AI APIs for the foreseeable future.

The trend is beginning to show up in spending data as well. Ramp, the American fintech giant, reported that DeepSeek was among the fastest-growing software vendors on its spend management platform in June 2026. Its workloads are using DeepSeek’s first-party API rather than self-hosted deployments.

The cost differential is stark. While DeepSeek has priced its V4 Pro at $0.435 per 1 million input tokens and $0.87 per 1 million output tokens, Claude models are priced much higher. Sonnet 4.6, the most widely used Claude model, is priced at $3 per million input tokens and $15 per million output tokens. Meanwhile, the now suspended Fable 5 is priced at $10 and $50, respectively.

The growing adoption of lower-cost Chinese models is also increasing competitive pressure on US providers. According to a WSJ report, OpenAI is considering significant token-price reductions to compete more aggressively, while Google has consistently expanded its portfolio of lower-cost Gemini offerings.

I Went From $3,000/Month on Claude to $5/Week on DeepSeek

And honestly? 80% of my work is identical.

For the past two months, I was burning $3-5K monthly on Claude Code. Every idea from design to development to testing - full end-to-end automation, even simulating users to test…

— Xiaoyin Qu (@quxiaoyin) May 27, 2026

But how are Chinese AI models offering such low prices when many US AI companies continue to struggle with the economics of large-scale AI deployment?

“The low cost of Chinese AI APIs reflects a combination of genuine technical efficiency and strategic ecosystem economics,” observes Marc Einstein, Research Director at Counterpoint Research.

“Advances in model architecture and infrastructure optimisation have reduced underlying costs, while major platforms are willing to operate at thinner margins because AI services help drive broader cloud and enterprise adoption,” he tells AIM.

The Economics of ‘Cheap API’

Chinese models succeed in three areas: software optimisation, lower operating costs, and hardware strategies.

Technical efficiency has become a defining characteristic of the Chinese AI ecosystem, driven in part by export-control-related compute constraints. Models from DeepSeek, Qwen, Zhipu, and Kimi increasingly rely on techniques that reduce both training and inference costs.

For example, DeepSeek’s R1 family introduced Group Relative Policy Optimisation (GRPO), which removes the need for a separate critic model during reinforcement learning. The approach reduces memory requirements by up to 50% and lowers training costs substantially compared with traditional PPO-based methods.

More recently, Fuli Luo, head of Xiaomi’s MiMo family of models, explained in a post on X the rationale behind the company’s latest API price cuts. According to Luo, infrastructure and software optimisations increased cacheable token capacity fivefold and reduced caching costs by roughly 80%, allowing the company to operate near break-even despite the aggressive pricing.

The notion that low-cost APIs are loss-making is also challenged in disclosures from Chinese AI firms. 

According to its IPO prospectus, MiniMax’s Open Platform business generated a 69.4% gross margin during the first nine months of 2025. Zhipu also reported an 18.9% gross profit margin in its Open Platform and API business, which provides cloud-based services built around its general-purpose models.

The Base Cost of Compute

Beyond software, Chinese AI labs benefit from structurally lower operating costs.

“Chinese AI labs spend far less on virtually every aspect of the model development and deployment pipeline: compute infrastructure, R&D, training, and inference,” notes Kyle Chan, fellow at the John L Thornton China Center at Brookings Institution.

Meanwhile, NVIDIA CEO Jensen Huang has estimated that while a 1 GW data centre costs roughly $40–60 billion to build in the West, comparable Chinese facilities may cost closer to $26 billion.

Engineering talent is also significantly cheaper than in Silicon Valley, while industrial-scale electricity and permitting costs remain lower. 

These factors allow Chinese firms to stretch capital much further than many of their Western counterparts.

However, Chan notes that China’s cheap inputs are partially “offset by more energy-efficient chips and hardware systems in the US.” 

The US ecosystem relies on unthrottled GPUs to maximise performance per watt and token throughput, thereby improving deployment economics despite higher operating costs. Yet American AI companies continue to shoulder far larger infrastructure and frontier-model investment programmes, suggesting that lower operating costs may still give Chinese providers an advantage in inference and model deployment.

Giants like Alibaba are further lowering costs by reducing their dependence on third-party hardware vendors like NVIDIA. 

In its Q4 FY25 earnings, Alibaba revealed that it is actively leveraging its proprietary T-Head silicon to bypass the severe 60–80% margin premiums typically charged by global merchant chip vendors. It disclosed that T-Head chips have achieved scaled mass production, with cumulative shipments surpassing 470,000 units, and now power more than 60% of Alibaba Cloud’s external AI workloads. 

While acknowledging that domestic silicon may still lag behind frontier US GPUs in raw peak performance, CEO Eddie Wu emphasised during an earnings call that this closed-loop strategy guarantees superior value for money.

Chinese AI models are currently running on a mix of infrastructure, including NVIDIA GPUs available within China, Huawei's Ascend chips, and, in some cases, overseas compute resources.

One notable example came in early 2025, when DeepSeek made its first confirmed international deployment at the Aramco Digital Data Center in Dammam, Saudi Arabia, as part of the kingdom's $14.9 billion push into AI and digital infrastructure. By hosting its models on Saudi infrastructure, which could provide access to more advanced hardware and alternative architectures, such as Groq accelerators, DeepSeek may gain access to compute resources that are difficult or impossible to deploy domestically.

Why a Large-Scale Shift is Unlikely

Given the revenue US AI labs earn from APIs, Chinese models do pose a genuine threat. 

“OpenAI derives an estimated 73% of its revenue from consumer subscriptions and 27% from APIs, while Anthropic remains more heavily dependent on API and enterprise customers, with an estimated 85% of revenue coming from APIs,” Einstein notes.

However, can Chinese models eat into OpenAI’s and Anthropic’s API revenue on a large scale? 

“While lower-cost Chinese models may attract price-sensitive developers, the more likely outcome is a segmented market, where lower-cost providers capture experimentation and cost-sensitive workloads, while frontier labs compete on performance, trust, security, and enterprise value,” Einstein adds. 

Several developers across Reddit, Hacker News, and Chinese forums increasingly describe model selection as a routing problem rather than a winner-takes-all contest, reserving frontier models for complex coding and agentic workflows while directing bulk inference workloads toward lower-cost providers.

One contributor on the Chinese internet forum Zhihu suggested using DeepSeek, particularly for workloads involving large outputs, repeated inputs that benefit from caching, and applications with lower multimodal and latency requirements.

“DeepSeek’s price cut is not about stealing GPT-5.5 workloads. A rational user almost never moves GPT-5.5-level tasks to DeepSeek,” added the contributor. “If that happens, either the task was overestimated before, or the quality requirement was never that high.”

Even Crivello observes that Lindy’s migration required significant engineering work. “I think these things are stickier than they sound—it's not easy to switch models,” he tells AIM. 

He also points to the trust that US-based frontier AI models have earned and the brand effects conveyed by the popular phrase that does the rounds among developers: ‘no one ever got fired for picking Anthropic.’ “And for a lot of companies, this spend [to switch to Chinese AI models] isn't worth optimising,” adds Crivello. 

Pricing advantages aside, data sovereignty remains a structural barrier that enterprise adoption cannot route around. Sending sensitive workloads to API servers outside the United States creates real compliance exposure for US government contractors, financial institutions, and healthcare companies operating under sector-specific data regulations.

Can the Economics Hold?

For companies such as Alibaba, low API prices are part of a broader strategy to drive consumption, expand market share, and strengthen adjacent cloud businesses. Its broader presence across China’s technology ecosystem helps absorb these losses and supports the strategy economically.

But for pure-play Chinese AI labs, can this model support the enormous costs of AI development over the long term?

“You can see in some of their financial statements that many are struggling to recoup the costs of computing and training,” Chan says.

Despite healthy API economics, overall profitability remains elusive. Zhipu reported a net loss of RMB 4.72 billion ($657 million) in 2025, driven largely by R&D spending and rising compute costs. 

MiniMax likewise posted a loss of $1.87 billion for the year, with R&D expenditure of $252.8 million far outpacing its $79 million in revenue. 

DeepSeek operated for nearly two years without external funding and was financed by its parent company, the hedge fund High-Flyer Capital. “However, even DeepSeek is feeling pressure to raise capital to expand compute infrastructure,” says Chan, pointing to the company's first external fundraising effort to raise about RMB 50 billion ($7.4 billion) at a valuation of RMB 350–400 billion ($52–59 billion). 

Aggressive pricing may be winning Chinese AI labs market share today, but whether they emerge as a durable force in the industry will ultimately depend on how effectively they balance low-cost offerings with the immense investment required to remain competitive at the frontier.