Cheaper Chinese AI Models Steal Enterprise AI Spend

Enterprise AI budgets are colliding with token-based billing, and the first winners may be cheaper Chinese AI models, not the biggest U.S. labs.

Large AI labs including Anthropic and OpenAI could see growth pressured as companies rein in AI spending and route more work to lower-cost alternatives, according to PYMNTS, citing Financial Times reporting from Thursday, June 18.

The shift hits the companies that assumed enterprise AI adoption would keep expanding in a straight line. It won’t. Buyers are moving from excitement to control. The question inside enterprises is no longer “Which model is best?” It’s “Which model is good enough for this job, at a cost finance can tolerate?”

Enterprise Buyers Are Turning Chinese AI Models Into a Budget Weapon

The core signal is blunt: enterprises are no longer treating frontier models as the default for every workflow.

PYMNTS says rising costs are being driven by two changes. First, companies are moving from simple chatbot use to AI agents, which consume more computing power. Second, AI labs are shifting from flat subscriptions to token-based billing. That changes procurement behavior fast.

A chatbot answer is one thing. An agentic workflow can involve multiple model calls, tool use, retrieval, follow-up reasoning, and background execution. Each step can add tokens. Each token carries cost.

That creates an opening for Chinese AI models, which PYMNTS says can charge less than U.S. companies because of more efficient models and China’s lower energy costs. OpenRouter data cited in the report shows Chinese AI models now have greater token consumption than U.S. ones, a reversal from the start of the year.

For buyers, that is leverage. If a cheaper model can summarize documents, classify tickets, translate text, draft code, or answer internal search queries well enough, the premium model loses automatic priority.

For related XOOMAR coverage on how model choice changes technical work, see Long Docs Split ChatGPT vs Claude Technical Writing Race.

Builders Face the New AI Cloud Bill: Tokens, Agents, and Usage Caps

Developers and internal AI teams are seeing the same pattern cloud teams saw years ago: small unit costs look harmless until usage spreads across thousands of employees or production systems.

PYMNTS reported in February that enterprises moving from pilots to production found traditional SaaS billing does not map cleanly to AI. Instead of charging per employee, AI can charge per token, API call, generated image, inference cycle, autonomous workflow, or several of those at once.

That matters because AI usage does not stay neatly inside one team. PYMNTS says companies that once encouraged employees to use AI tools while costs were lower are now cutting back.

Reported controls include:

• Usage caps: Limiting how much employees can consume.
• Tool routing: Telling staff to use the right model for the task.
• Cheaper models: Switching to older models when frontier performance is unnecessary.
• Open-source models: Adopting alternatives that reduce dependence on premium APIs.

The clearest corporate examples are already concrete. PYMNTS says Uber exhausted its full-year 2026 AI budget by April, leading executives to say the company was “back to the drawing board” and that the productivity case had not closed.

“back to the drawing board”

PYMNTS also reported that Walmart limited employee AI usage on June 1, moving workers from unlimited tokens to a set token allocation for its in-house AI agent Code Puppy.

The question for builders is simple: can they preserve productivity gains while cutting the most expensive model calls?

CFOs and CIOs Are Splitting Workloads by Price and Risk

Enterprise buyers are not abandoning AI. They are sorting it.

Frontier models still matter where reasoning quality, reliability, and sensitivity justify higher costs. But not every task needs the best available model. A routine summarization job has different economics from high-stakes analysis.

That is where cheaper Chinese AI models are gaining attention. The additional source packet cites Rest of World reporting on U.S. developers using DeepSeek to reduce costs, including one user who said an hour of coding cost about $10 on Claude versus under $0.50 on DeepSeek. That is anecdotal, not a universal price table, but it explains the buyer psychology.

Fortune also reported that Shanghai-based MiniMax debuted its M1 model on June 16 and claimed it spent $534,700 renting data center compute to train it. Fortune said that is nearly 200-fold cheaper than estimates that ChatGPT-4o likely exceeded $100 million to train, though MiniMax’s claim has not been independently verified.

The appeal is obvious. Lower cost means more experimentation, broader internal access, and less pressure to ration AI only to senior staff or high-value workflows.

But the limits are real. Fortune noted geopolitical and national security concerns around Chinese-developed models, including censorship requirements and a bipartisan House Select Committee report from April that said DeepSeek’s responses are “manipulated to suppress content related to democracy, Taiwan, Hong Kong, and human rights.”

For CIOs and legal teams, the question becomes harder: when is the savings worth the governance burden?

OpenAI and Anthropic Face a Margin Test as Customers Shop Around

The pressure on OpenAI and Anthropic is not only competitive. It is structural.

Premium labs need revenue growth to support heavy spending on chips, data centers, research talent, and energy. But their enterprise customers are learning to manage AI consumption like a variable operating cost. That weakens easy pricing power.

Fortune reported that OpenAI recently cut the cost of using its o3 reasoning model by 80%. That kind of move can defend usage, but it also shows the pricing fight is already active.

A simple enterprise split is emerging:

This is where internal platforms matter. If a company can route tasks across multiple models, OpenAI or Anthropic becomes one supplier, not the default layer for everything.

For XOOMAR readers tracking pressure around frontier AI companies, our coverage of Five-Month Exit Jolts Barret Zoph's OpenAI Comeback offers related context on how closely the market watches execution inside major AI labs.

The question for premium labs: can they prove enough incremental value to remain the expensive default?

Cloud Providers, Legal Teams, and Employees See Different Risks

Not every stakeholder wants the same outcome.

CFOs want predictability. Token budgets, usage caps, and cheaper models give them control.

CIOs want performance without vendor sprawl. Multi-model procurement can reduce cost, but it adds evaluation work, routing rules, monitoring, and support questions.

Legal and security teams care about data handling, auditability, and exposure to geopolitical risk. Chinese AI models may be cheaper, but the governance review can be heavier.

Employees mostly want tools that work. They care less about model brand than speed, accuracy, and whether approvals block them from using AI during actual work.

Cloud providers also sit in an awkward position. More AI usage means more compute demand, but cheaper and more efficient models could change where that demand lands. Fortune noted MiniMax’s claimed efficiency, if accurate, could pressure cloud providers such as Amazon’s AWS, Microsoft’s Azure, and Google Cloud Platform, and could affect demand for Nvidia chips.

The question across the stack is not whether AI usage keeps growing. It is who captures the spend.

The AI Pricing Fight Now Runs on Evidence, Not Hype

The next phase of enterprise AI will be measured by cost per useful output.

That is XOOMAR’s read of the PYMNTS report: cheaper Chinese AI models are not winning because every buyer suddenly prefers China-made AI. They are winning the workloads where price, availability, and acceptable quality beat brand prestige.

Evidence that would strengthen this thesis includes more OpenRouter share gains for Chinese models, broader enterprise token caps, more public price cuts from premium labs, and more companies reporting budget overruns like Uber or usage limits like Walmart.

Evidence that would weaken it would be clear enterprise pushback on Chinese model governance, poor independent validation of low-cost model claims, or premium labs proving that higher-priced models deliver enough productivity to survive finance scrutiny.

The winners won’t simply be the labs with the most famous models. They’ll be the vendors that make AI affordable, governable, and useful at scale.

The Bottom Line

• Enterprise buyers are shifting from chasing the best AI model to finding models that are cost-effective for each job.
• Token-based billing makes agentic AI workflows more expensive and harder for finance teams to approve without controls.
• Cheaper Chinese AI models could gain share if they handle common business tasks well enough at lower cost.