While some companies are capping employee AI usage to hold down ballooning costs, many others across the technology industry are still going all-in on AI spending.
Big cloud providers such as Amazon Web Services, along with software companies like Palantir, Snowflake and HubSpot, have not set hard spending limits for engineers, who use AI coding tools such as Anthropic's Claude Code and OpenAI's Codex to build applications and features. The leading frontier model labs have also avoided caps on themselves. Venture capitalists say many of the startups they back still encourage employees to use AI as much as possible.
Meanwhile, Salesforce has in recent months begun setting team-level spending caps for forward-deployed engineers, the staff who build AI agents for customers, according to people familiar with the matter. Not all employees are aware of the limit: one Salesforce forward-deployed engineer told The Information he had assumed his budget was uncapped.
More than half a year has passed since surging AI costs first triggered industrywide concern, and the differing approaches show how sharply the industry's leading players diverge on how to govern AI usage. In April, The Information reported that Uber found Claude Code easily burned through its AI budget, while Meta employees embraced "tokenmaxxing," a push to consume as many AI tokens, the units that large models use to process information, as possible. Around the same time, many companies began looking for ways to control their AI tool bills. Meta told employees it had started limiting engineers' AI spending, and Uber built its own coding orchestration software called harness that automatically switches between different large models to cut costs.
One engineer at Workday said the enterprise software company gives each engineer a very low initial budget, typically $100 to $300 a month. A Workday spokesperson said the budget is only a "starting point," not a hard cap, and Chief Technology Officer Gabe Monroy said in an interview that additional budget can be requested for special projects. Startups backed by Salt Lake City venture firm Oquirrh Ventures actually have higher budgets, even though they are far smaller than Workday, with annual revenue of just $15 million to $80 million. Oquirrh founder Ron Heinz said the eight technology companies it backs give engineers roughly $500 a month each.
Other companies actively encourage engineers to use coding tools. HubSpot Chief Product and Technology Officer Duncan Lennox said the company pushes engineers to use its in-house coding tool, which can switch between Anthropic, OpenAI and open-source models. At the start of the year, only about 15% of its more than 2,000 engineers used the tool frequently in their daily work; by the end of September, that share had soared above 90%. Lennox said cost is "not a big issue" because the tool has greatly boosted engineer productivity. Much the same is true at roughly 50 portfolio companies of S3 Ventures and more than 20 early-stage projects at Bain Capital Ventures, which generally set no hard token limits. "As an early-stage, venture-backed company, the first goal is to be at the forefront of technology, both in what you deliver as a product and in the tools you give employees. So we see very little of that kind of spending control, and I think that is the right strategy," said Alyssa Ko, a partner at Bain Capital Ventures. S3 Ventures partner Eric Engineer said part of the reason portfolio companies are insensitive to cost pressure is that they are cutting staff or slowing hiring. Compared with the labor cost of adding an engineer, spending heavily on AI coding tools "is nothing; as long as they get the value, companies feel there is no need to set a cap."
Technology companies such as Rippling and Uber do set token budgets, but the amounts are relatively generous, and engineers can request increases that are generally approved. The same is true at software company Autodesk, where the total team budget works out to $2,500 per person per month, allocated at the team leader's discretion.
Of course, some companies are tightening budgets. One employee said Walmart recently cut engineers' token budgets; Bloomberg reported that Apple reduced its daily budget for engineers from $300 to $200; and The Information reported last month that Microsoft recently trimmed engineer AI budgets by a tenth. A software engineer at Disney said the company has paused employee applications for tiered AI coding tool budgets, which range from $50 to $25,000 a month. Some companies that previously had no limits are now preparing to impose them.
People familiar with the matter said Blackstone plans to use a new AWS feature to set caps on engineers' AI spending, allowing AI expenditure to be controlled on an individual employee basis. Over the next few quarters, the private equity firm plans budgets of about $5,000 a month for heavy users of coding tools and large models, a floor of $250 a month for low-usage employees, and the ability for low-budget employees to request more. Last week, AWS CEO Matt Garman said at a Blackstone technology conference in San Jose that the hyperscaler lets employees experiment freely with AI tools to improve efficiency, while reminding them not to waste company money. He gave an example: an employee used company AI tools to write a sitcom script. "I thought, 'That's funny, but is that the best use of company money? Probably not.'" Garman said Amazon currently has no hard spending limits, but relies on orchestration tools to automatically route engineer requests across different frontier and open-source models to control costs; in some engineering scenarios, that approach can cut spending by as much as 40%. Amazon and other AI service providers are also planning or have already launched features that let customers track spending. Anthropic's Claude enterprise customers can set companywide or individual spending limits and use a dashboard launched this summer to view AI usage; OpenAI offers similar capabilities.
Not surprisingly, several companies at the center of the AI wave let their own employees use tokens as needed. An OpenAI spokesperson said the lack of a cap is because arbitrary limits would restrict innovative experimentation, though the company has an internal dashboard tracking each employee's usage and does not encourage "tokenmaxxing." One Nvidia employee said engineers' AI budgets can reach as much as $50,000 a month, while non-engineers can get up to $10,000 a month. Another employee said he knew of only one person at the company who had hit a limit, an employee who used open-source AI to generate training data.
Comments