U.S. corporate AI bills are spiraling out of control—Uber burned through its entire annual token budget in four months, and Salesforce paid Anthropic approximately $300 million in one year. Right now, DeepSeek has topped the U.S. corporate subscription charts. Even more unusual, these companies aren’t downloading open-source weights to run locally—they’re paying directly and connecting to servers to transmit data. Even the chief economist admitted: “I didn’t expect U.S. companies to use DeepSeek.”

Article author and source: Wall Street Journal

At the end of May, Ara Kharazian, Chief Economist at Ramp, an U.S. corporate spending management platform, predicted that U.S. enterprises would become more cost-conscious in their AI spending, increasingly experimenting with open-source models or more affordable, lower-tier versions from OpenAI, Anthropic, and Google.

A week later, data from his own platform provided an answer that exceeded expectations.

On June 3, Ramp released its June Software Trends Ranking, with DeepSeek topping the list.

“I didn’t expect U.S. companies to use DeepSeek,” Karazian wrote in the report.

Ramp manages corporate cards and business bills for over 50,000 U.S. companies, processing billions of dollars in expenses each month. Its Software Trending Rank tracks which software vendors are receiving their first corporate payments at the fastest rates. Past regulars on this list have been Silicon Valley favorites like Figma and Fireworks AI—this is the first time a Chinese model company has topped it.

There’s another subtle detail in the payment records: these U.S. companies did not opt to download open-source weights and deploy models locally; instead, they paid DeepSeek directly.

To be clear: this is not just self-hosted open-source use. Enterprises are directly sending and receiving data through DeepSeek,” Karazian emphasized this point.

He added that this may be the clearest signal yet that U.S. companies are seeking lower-cost alternatives to OpenAI and Anthropic, with some firms willing to use cheaper Chinese models that "send U.S. data back and forth between servers hosted in China."

This is not the first time DeepSeek has appeared on the books of U.S. companies. In January 2025, R1 went viral globally, and DeepSeek’s usage briefly surpassed ChatGPT to top the free app chart on the U.S. App Store; according to Ramp’s data at the time, adoption among U.S. companies surged to 0.3% during that period, before quickly declining back to 0.1%.

If early 2025 resembled a collective trial by companies driven by media hype, this round of corporate adoption of DeepSeek is marked by a fundamental shift in motivation: the financial pressure of AI bills.

Uber burned through its entire annual token budget in four months.

How bad U.S. companies' AI bills are is no longer a secret known only within the industry.

Consulting firm Bain recently surveyed 951 global enterprises with annual revenues exceeding $100 million, delivering a blunt conclusion: after corporate AI spending surpassed $1 trillion, the actual cost savings generated by AI have generally fallen far short of expectations; even more strikingly, 44% of large enterprises are justifying their next round of AI investments using cost savings from the previous AI initiative that never materialized. Bain has labeled this practice “a structurally flawed cycle of betting.”

When it comes to specific companies, the numbers are even more striking: ride-hailing giant Uber revealed that it exhausted its entire annual token budget in just the first four months of 2026; Salesforce stated that its payments to Anthropic this year will reach approximately $300 million.

Tech giants are also tightening their purse strings: Amazon has halted its internal AI usage leaderboard after employees manipulated rankings by performing unnecessary tasks to increase token consumption; Microsoft plans to gradually phase out Claude Code subscriptions for employees in several key product teams by the end of June.

When even Amazon starts to care about tokens, cost is no longer just a concern for startups.

The U.S. companies that spend the most are precisely the two most expensive AI large model providers.

Ramp’s proprietary AI index shows that Anthropic and OpenAI have achieved paid adoption rates of 34.4% and 32.3% respectively among U.S. enterprises, with Anthropic surpassing OpenAI for the first time. Both companies’ flagship models have long been priced at the highest tier in the industry, meaning the bulk of corporate AI spending essentially goes toward their invoices. Karazian’s assessment of “cost discipline” refers precisely to this situation: money has already been spent, returns have been slow to materialize, and finance departments are now scrutinizing every line item of AI expenditure.

For businesses, there are only two ways to save money: use less or switch to something cheaper. Amazon and Microsoft chose the former. The companies on the trend list chose the latter.

Price dropped to a quarter; developers voted first.

Just as U.S. companies were struggling with exorbitant AI bills, DeepSeek delivered an offer that was hard to refuse in May.

On May 22, DeepSeek announced a permanent price reduction for its flagship model V4-Pro to one-quarter of the original rate: after the limited-time promotion ends, the input price will be as low as RMB 0.025 per million tokens (cache hit) and RMB 3 per million tokens (cache miss), with output priced at RMB 6, once again setting a new global low for mainstream large models.

According to third-party calculations, under the same task complexity, the average invocation cost of V4-Pro is approximately one-tenth that of GPT-5.5 and about one-eleventh that of Claude Opus 4.7.

Low pricing is not a new strategy for DeepSeek. As early as 2024, when the price cut of V2 triggered a domestic large model price war, founder Liang Wenheng stated in an interview the pricing principle: set prices after calculating costs—"neither selling at a loss nor seeking excessive profits." The confidence to lower prices comes from cost reductions brought by the next-generation model architecture: "APIs and AI should be accessible and affordable to everyone."

If those words sounded like an internal standard set by a Chinese startup at the time, two years later, DeepSeek’s inclusive pricing began to be adopted by U.S. companies’ business accounts.

Before the company invested, the developer community had already shown their trust in DeepSeek.

Data from OpenRouter, a global model aggregation platform, shows that during the week of May 18–24, DeepSeek V4-Flash ranked first globally with 3.43 trillion weekly token requests. The total weekly token requests for all DeepSeek models reached 5.74 trillion, surpassing Anthropic and Google, and marking the second consecutive week as the top global vendor. During the same period, the total weekly token requests for all Chinese models reached 9.22 trillion, outpacing U.S. models’ 4.93 trillion for the fourth consecutive week. OpenRouter’s user base is primarily composed of overseas developers, with Chinese developers accounting for only about 6%. Each request on this ranking reflects genuine choices made by overseas developers.

This week’s model usage leaderboard on OpenRouter places DeepSeek V4 Flash at number one; among the top ten are multiple Chinese models from Tencent, MiniMax, Xiaomi, and others. Source: OpenRouter

Developers vote with their feet first, and companies follow with their budgets. For enterprise software, this “developers use first, companies buy later” path is classic—over the past two decades, AWS, GitHub, and Slack all entered American enterprises this way. This time, however, it’s a Chinese model company walking this path.

Chinese models are entering U.S. companies—DeepSeek is neither the first nor the only one. As early as the end of 2025, Airbnb CEO Brian Chesky publicly stated that the company “heavily relies on Alibaba’s Qwen model, which is fast, high-quality, and cost-effective,” while OpenAI’s models were “not used much in production.” This remark drew inquiry from the U.S. House of Representatives this year.

In March this year, the AI programming tool Cursor, valued at nearly $30 billion, launched its new model, Composer 2. Within 24 hours of its release, developers discovered that its foundation was Moonshot AI's Kimi K2.5. Co-founder Aman Sanger later admitted, "Failing to disclose that it was based on Kimi in our initial blog post was a mistake."

These companies are using Chinese models differently than DeepSeek did this time. Cursor performed fine-tuning using Kimi’s open-source weights, and Airbnb’s Qwen was also deployed internally. Chekeski repeatedly emphasized, “We have not provided any data to any Chinese company.” They are leveraging the capabilities of Chinese open-source models, but the money is not going to Chinese companies.

Ramp, however, recorded something different: U.S. companies are paying DeepSeek directly, with data flowing directly in and out of DeepSeek. The real innovation in this month’s ranking is the shift from “using open-source weights to run independently” to “paying to connect directly to hosted services”—and this is precisely why Karazian emphasized, “This isn’t just self-deployment.”

Can this time be different?

Regarding the conclusion that DeepSeek has risen to the top of the U.S. enterprise subscription rankings, Karazian added a note in the report: “I wouldn’t overestimate the longevity of this trend. Enterprises face real competitive and security concerns when directly integrating DeepSeek.”

U.S. companies directly sending data to a Chinese company’s servers would inevitably raise concerns from compliance teams. This concern is real and may not be easily resolved by DeepSeek in the short term.

But the second half of Karazian’s statement carries equal weight: “American model companies should nonetheless take note of this competitive pressure and help businesses rein in their runaway AI spending by offering cheaper models or intelligent routing.”

Even if U.S. companies collectively withdraw from DeepSeek tomorrow due to security concerns, the cost pressures forcing them to this point will not disappear—OpenAI and Anthropic’s pricing models will still need to answer this question.

Moreover, the trend list features more than just DeepSeek. That same month, other上榜 platforms such as Fireworks AI, fal, and DeepInfra—open-source model inference platforms—also made the list. U.S. companies are collectively shifting toward open-source and low-cost solutions, with DeepSeek simply being the most prominent and surprising destination in this migration.

Cheap can also be the start of another problem. In the Agent era, a single task requires repeatedly calling models, writing code, running tests, and fixing errors on its own—consuming several times, even dozens of times, more tokens than the old one-answer-per-prompt model. Calculations show that when Agent adoption reaches 8%, its token consumption will match that of chatbots, with over five times more growth still ahead. When the price drops to a quarter, but usage surges tenfold, the bill only gets heavier. DeepSeek’s low pricing has brought more companies through the door—and pushed them deeper into this endless cycle of consumption.

One and a half years ago, the release of DeepSeek-R1 sent shockwaves around the world; the story of its “approximately $5.6 million training cost” caused panic across Silicon Valley, wiping nearly $600 billion off NVIDIA’s market value in a single day.

Although back then DeepSeek’s market panic could be calmed within a week, today the mounting pressure on AI budgets relentlessly confronts all U.S. companies every month.