Western companies are quietly switching to Chinese AI models as U.S. frontier AI prices rise

A year ago, the biggest question in enterprise AI was which frontier model delivered the best results. Today, a different question is driving boardroom conversations: Which model delivers the best results for the lowest cost?

That shift is pushing a growing number of Western companies to an answer that would have surprised many just months ago. Instead of relying exclusively on expensive AI models from OpenAI, Anthropic, and Google, businesses are turning to Chinese-developed models such as DeepSeek, Qwen, Kimi, and GLM.

The change is not about politics. It is about economics.

AI has become one of the fastest-growing operating expenses for software companies. As AI assistants take on longer conversations, larger datasets, coding tasks, customer support, and autonomous workflows, inference costs have climbed far beyond what many executives expected. For some companies, AI spending now rivals payroll.

We saw a striking example of that trend in May, when we reported on a company that accidentally spent $500 million on Claude AI in a single month after forgetting to set usage limits. Most businesses will never see bills anywhere near that size. Still, the story exposed a problem spreading across the industry. AI costs are rising much faster than many companies anticipated.

That financial pressure is changing how businesses think about AI infrastructure. Instead of asking who built the smartest model, more engineering teams are asking which model delivers the best value for each task.

Silicon Valley’s frontier models still lead many advanced benchmarks. Their strengths remain clear for complex reasoning, research, and sophisticated coding. Yet many everyday business tasks do not require the most expensive model available. Writing emails, summarizing documents, extracting data, classifying customer requests, and powering AI agents can often be handled by models that cost a fraction as much.

That realization is reshaping enterprise AI strategies.

Several technology leaders have publicly argued that companies should rely on premium frontier models only when the workload truly demands them.

Microsoft CEO Satya Nadella has repeatedly spoken about reducing inference costs across AI infrastructure. Palo Alto Networks CEO Nikesh Arora has highlighted the growing importance of smaller models for enterprise deployments. Coinbase CEO Brian Armstrong recently explained how the company is reducing AI spending by defaulting employees to lower-cost open-weight models for many internal tasks.

Their comments reflect a broader shift across the industry.

A recent post on X highlighted examples ranging from Lindy’s adoption of DeepSeek V4 to Cursor’s use of Moonshot AI’s Kimi models, Shopify’s deployment of Alibaba’s Qwen, Airbnb’s reliance on Qwen, and reports involving Coinbase, Uber Eats, Siemens, and Microsoft’s evaluation of DeepSeek. Social media lists should always be treated with caution, as some claims continue to evolve. Independent reporting and company statements, though, point to the same conclusion: Western companies are becoming far more willing to use Chinese AI models when the economics make sense.

The numbers help explain why.

Research firm Gartner projects AI coding costs will exceed the average developer’s salary by 2028. Its surveys show roughly three-quarters of executives expect technology budgets to increase this year, with nearly half anticipating double-digit spending growth.

Those rising expenses are pushing companies to rethink how they deploy AI.

Many organizations now use routing platforms such as OpenRouter to match each request with the most cost-effective model available. Premium models remain available for demanding workloads, including advanced coding and reasoning. Less expensive models handle routine requests that make up the majority of enterprise traffic.

The strategy is already changing usage patterns.

According to a Citi research note, open-source models processed through OpenRouter accounted for 65% of tokens in June, up from 34% in January. That increase points to growing confidence in open-weight alternatives as businesses seek to control AI spending.

“Chinese models charge as little as 18 cents per million tokens versus $4 average for top models,” Reuters reported, citing Citi

AI Costs Are Becoming an Infrastructure Problem

Inference has quietly become one of the highest operating costs in enterprise AI.

Training frontier models requires enormous computing resources, yet running them every day often proves even more expensive for companies serving millions of users.

Providers such as OpenAI, Anthropic, and Google charge premium prices for their most capable models. Those costs add up quickly for businesses generating billions of tokens each month.

Chinese AI labs have attacked that problem from a different direction.

Companies including DeepSeek, Alibaba, Moonshot AI, Zhipu AI, and MiniMax have focused heavily on lowering inference costs without sacrificing performance on everyday business tasks. Their models frequently cost between five and thirty times less than comparable frontier models, according to published pricing and industry benchmarks.

That gap has become difficult for engineering teams to ignore.

The lower prices are not the only attraction.

Most leading Chinese models are released with open weights, allowing companies to self-host them, fine-tune them with proprietary data, and avoid dependence on a single AI vendor. For many enterprises, that flexibility is almost as valuable as the lower operating costs.

Large cloud providers have made adoption easier.

Amazon Web Services, Microsoft Azure, and Google Vertex AI now offer access to several Chinese-developed models through managed cloud services. Companies can deploy those models with enterprise security controls, unified billing, compliance tools, and assurances that customer data remains isolated from model developers.

Usage data points to growing momentum.

OpenRouter reports DeepSeek and several other Chinese models among its most heavily used systems. Vercel recently disclosed that DeepSeek’s share of token usage jumped from less than 1% to 17% in a single month, making it one of the fastest-growing models on the platform.

Key Examples of Western Companies Switching to Chinese AI Models

Lindy Cuts Millions From Its AI Bill

Few companies illustrate the economic shift more clearly than AI assistant startup Lindy.

In June 2026, CEO Flo Crivello announced that Lindy had moved its AI agent traffic from Anthropic’s models to DeepSeek V4.

The decision was driven by simple business math.

Crivello said inference costs had grown larger than the company’s payroll. Moving to DeepSeek is expected to save millions of dollars each year, with no meaningful decline in quality for many of Lindy’s core workloads, including email triage and automated replies.

To address security concerns, Lindy runs the model through Atlas Cloud, a U.S.-based inference provider that keeps customer data inside domestic infrastructure.

Crivello summarized the decision with a line that quickly spread across the AI community.

“You don’t need God to write your email.”

His point was straightforward. Premium frontier models remain valuable for difficult reasoning tasks. Most business workflows simply do not require the industry’s most expensive AI every time a customer sends an email.

DeepSeek has continued pushing costs lower through technical improvements.

Its DSpark speculative decoding system combines lightweight draft generation with intelligent verification, allowing models to produce responses much faster under production workloads. Internal benchmarks show generation speeds improving by 57% to 85% with only minimal latency overhead.

For companies processing millions of requests each day, improvements like those translate directly into lower infrastructure bills.

Shopify Finds a 75x Cost Reduction

Shopify reached a similar conclusion after evaluating one of its internal AI systems.

Engineers replaced an OpenAI GPT-5 pipeline used for merchant data extraction with a self-hosted multi-agent system powered by Alibaba’s Qwen 3.

The results were striking.

Shopify reported a 75-fold reduction in per-unit language model costs alongside higher output quality. The company has since fine-tuned Qwen3-32B for natural language automation within Shopify Flow, resulting in faster responses and more accurate workflow generation.

The project reflects a growing trend across enterprise software.

Companies increasingly view frontier models as premium tools reserved for specialized tasks rather than as default engines for every AI request.

Airbnb Looks Beyond Silicon Valley

Airbnb has taken a similar approach.

CEO Brian Chesky has publicly praised Alibaba’s Qwen models, describing them as “very good,” “fast,” and “cheap” for powering customer service workloads.

The company uses multiple AI models across its products, yet Qwen has become an important part of its customer support systems. Airbnb says customer data is processed through domestic infrastructure or compliant channels rather than being sent directly to Chinese providers.

The company’s decision has attracted attention in Washington.

Lawmakers have questioned whether U.S. businesses should rely on Chinese-developed AI systems, particularly for products handling sensitive customer information. Those concerns have become part of a larger debate about national security, data sovereignty, and long-term dependence on foreign AI technology.

For Airbnb and many other companies, the calculation remains practical. Lower costs, faster responses, and competitive performance carry significant business value, provided security and compliance requirements can be met.

Cursor Builds on Kimi

Cursor, the AI coding platform from Anysphere, is another example of the shift. The company has used Moonshot AI’s Kimi models as part of its coding stack, according to industry chatter and model-tracking discussions. The appeal is clear: coding assistants generate enormous token volume, and even small price differences can turn into major savings at scale.

Coinbase Pushes Cheaper Defaults

Coinbase CEO Brian Armstrong has said the company is trying to keep AI spending under control by changing default model choices. Instead of sending every task to the most expensive frontier model, Coinbase is routing more internal usage to lower-cost open-weight models, including GLM and Kimi.

That does not mean Coinbase is abandoning U.S. frontier models. It means the company is treating AI more like cloud infrastructure: use the right system for the job, not the priciest one by default.

Uber Eats, Siemens, ChapsVision, and Microsoft Enter the Picture

The same pattern is showing up across larger companies. Uber Eats has been linked in industry discussions to the use of Qwen. Siemens has been associated with both DeepSeek and Qwen. ChapsVision has been tied to Qwen. Microsoft has tested DeepSeek V4 in some settings, according to the circulated industry list.

“Top executives such as Microsoft’s Satya Nadella, Palo Alto Networks’ (PANW.O), Nikesh ‌Arora, and Coinbase Global’s (COIN.O), Brian Armstrong, have said smaller, cheaper models can handle a big share of corporate needs,” Reuters noted.

Those examples should be framed carefully. Some are based on public posts, model-routing chatter, or reports that may change as companies update their AI stacks. Still, together they point in the same direction: Chinese AI models are moving from developer curiosity to real enterprise consideration.

For companies under pressure to cut AI bills, the question is no longer whether Chinese models can compete. The question is where they are good enough, cheap enough, and safe enough to replace more expensive U.S. models.

Challenges and Geopolitical Context

The growing adoption of Chinese AI models comes as Anthropic CEO Dario Amodei warns lawmakers that open-source AI is moving down a “very dangerous path.” Amodei has argued that widely available frontier models could make it harder to prevent misuse and protect national security, adding another dimension to the debate over enterprise AI adoption.

🚨ANTHROPIC CEO: OPEN SOURCE AI IS GETTING DANGEROUS

Anthropic CEO Dario Amodei told lawmakers that open-source AI is moving down a “very dangerous path.”

His warns that once powerful models are released openly, companies lose the ability to monitor misuse, revoke access, or… pic.twitter.com/099ptGEEP4

— Coin Bureau (@coinbureau) June 28, 2026

Adoption is not without friction. U.S. lawmakers have launched investigations into companies such as Airbnb and Cursor (Anysphere) over potential national and data security risks associated with Chinese AI models. Concerns include data sovereignty, possible backdoors, censorship alignment, and long-term technological dependence.

Many adopters mitigate those risks by self-hosting open-weight models or routing them through U.S. cloud providers that isolate customer data. Large regulated enterprises remain cautious. Startups and mid-sized companies have generally placed greater emphasis on economics, provided security and compliance requirements can be met. Chinese model developers still face challenges converting growing usage into enterprise trust and long-term revenue.