Microsoft Imposes AI Token Budgets After $24.1 Billion OpenAI Spend Reveals Cost Crisis

Microsoft to Engineers: Don't Go Crazy With the AI Use. It's Not Cheap

As the cost of AI continues to skyrocket, Microsoft engineers are getting new limits on their AI use in the workplace. According to a company email obtained by 404 Media, Microsoft is making the more affordable OpenAI GPT-5.6 its default model for internal use to "get greater value from our token investment." Additionally, Microsoft divisions will soon have an "AI token budget target" that employees can individually track. Tokens are like AI credits; the more words you use when prompting an AI, the more tokens you eat up. For a time, tech employees were encouraged to use AI as much as possible, but that activity, or tokenmaxxing, resulted in huge bills for the companies using it. In the email seen by 404 Media, Microsoft EVP Jay Parikh said, "We all need to be aware of how we consume tokens. Tokenmaxxing is not what we are optimizing for." Instead, "I want all of us focused on maximizing outcomes that move the needle for our customers and our business," Parikh said. "As such, we are updating our internal guidance and managing token spend with the same discipline we apply to every other critical resource." Parikh's email links to internal Copilot guidelines, which say Microsoft divisions will have an "AI token budget target," but that document doesn't get into specifics. This comes about two months after GitHub Copilot moved to usage-based billing, and some users quickly hit their limits. Microsoft is reportedly also trying to cut down on AI costs by sending Microsoft 365 AI prompts to its internal MAI models rather than Anthropic and OpenAI. Companies like Amazon, Adobe, Atlassian, and Citi have also cracked down on their employees' token spending, 404 Media reported last month.

Microsoft tells engineers to curb their token-burning enthusiasm

Concerned about cost, Microsoft is reportedly warning employees against wasteful AI use. In an email seen by 404 Media, Microsoft Executive Vice President Jay Parikh warned that individual divisions would be given targets and could face restrictions. "Tokenmaxxing is not what we are optimizing for. I want all of us focused on maximizing outcomes that move the needle for our customers and our business," the email states. The Register asked Microsoft what it made of the report, and a spokesperson told us the company had "nothing to add." Microsoft is hardly short of cash and has continued to pour money into its AI investments, but its apparent efforts to curb wasteful AI spending bring it into line with other companies attempting to rethink how employees use AI. Tokenmaxxing is the practice of using as many AI tokens as possible, and treating the volume as proof of productivity. A token is the basic unit used to process and bill for the input and output of AI models. If a company is measuring employees' use of AI, burning through as many tokens as possible will make the figures look rosier. However, companies need to consider the cost of those tokens alongside other factors, such as productivity. As The Register observed in April: "Tokenmaxxing isn't an AI strategy". Microsoft does not appear to be seeking an overall reduction in token use - it wants to be an AI-first company, after all. It simply wants more bang for its buck. "As we accelerate our use of GitHub Copilot to deliver on our goals, we all need to be aware of how we consume tokens," the email reportedly said. GitHub moved to usage-based billing in June. Although Microsoft's code-hosting tentacle measures consumption in AI Credits rather than raw tokens, the change makes the cost of engineers burning through them harder to ignore. Not that this sends a particularly great message to customers. As one anonymous Microsoft staffer told 404 Media: "This really feels like the ultimate admission that we, as hosts of AI infra, can't afford our own AI products. And if that's even partially the case, how could the companies we sell it to manage?" That's a little simplistic. Microsoft can afford its AI products; the email suggests it wants engineers to use them with greater regard for cost and results. Which is, perhaps, the message customers should take away. AI tools and services carry a cost that needs to be measured, rather than treating "how many tokens did you consume?" as a measure of AI adoption. ®

Microsoft AI exec tells developers to default to OpenAI's top model as part of efficiency push

Microsoft is telling developers working on AI coding projects to rely on OpenAI's top-tier model over rival products as part of an effort to maximize efficiency. "Internally, shifting more workloads to OpenAI models helps us get greater value from our token investment," Jay Parikh, executive vice president of Microsoft's CoreAI engineering group, wrote this week in a memo to employees that was viewed by CNBC. Tokens measure the scale of AI processing, with one token equal to about three-quarters of a word. While Microsoft has built its own artificial intelligence programming model and gives cloud customers access to over 11,000 models, including from Anthropic, the company wants staffers to take advantage of valuable intellectual property rights that come from the software giant's early investment in OpenAI. Parikh, whose group includes GitHub, Visual Studio and Visual Studio Code, told staffers to default to OpenAI's flagship GPT-5.6 Sol when working in the GitHub Copilot coding tool, and use that model most of the time. OpenAI released GPT-5.6 Sol in July. Efficiency in AI spending is becoming increasingly important across corporate America after a brief era of so-called tokenmaxxing, when developers were encouraged to run up large token bills without worrying about their output. A slew of open-weight models, largely out of China, have gained popularity because they're cheaper to access than the frontier models and allow users to tweak them and host their work on the infrastructure of their choice. For the large hyperscalers, Wall Street is starting to demand more from their massive AI spending commitments, with capital expenditures from Microsoft, Amazon, Alphabet and Meta expected to top $700 billion collectively this year. Across the group, free cash flow dwindled in the latest quarter -- and even went negative for Amazon and Alphabet. Microsoft's cash generation fell by 23% from a year earlier, a mild decrease compared to its peers.

Microsoft's AI revenue is ~70% OpenAI, a filing shows

A filing shows Microsoft booked $24.1 billion from OpenAI last year, most of its AI business. The same day, an executive told staff to default to OpenAI's model internally too. Microsoft's AI business looks like a triumph. Its shares jumped after earnings last week, and the AI unit is growing fast. A new filing shows how much of that rests on a single partner. Microsoft booked $24.1 billion in revenue from OpenAI during the year to June, Bloomberg reported. The company disclosed the figure in a filing last week. That is most of its AI business. By Bloomberg's estimate, OpenAI is more than half of Microsoft's actual AI sales, and probably around 70%. It is the clearest picture yet, and the first time Microsoft has spelled the figure out. Investors have pressed the company on exactly this. One accounting researcher, Olga Usvyatsky, suggested the disclosure may be tied to OpenAI's plans for an eventual listing. How the 70% is built The 70% is an estimate, not a hard number. Microsoft has only ever disclosed its total AI business twice. It did not update the figure at its latest earnings. Bloomberg assumed the unit kept growing at the 123% rate Microsoft reported in March. That would put the AI business around $34 billion for the year. Set the disclosed $24.1 billion against it, and OpenAI is roughly two thirds. The money is not quite what it sounds. Under their agreement, OpenAI pays Microsoft for cloud computing and for the cost of building models. It also hands over a share of its revenue. Much of the $24.1 billion is OpenAI's own compute bill, routed through Azure and booked as Microsoft income. Against Microsoft's total revenue, OpenAI shrinks to under 10%, closer to 7%. The concentration only jumps out when you narrow the lens to the AI unit. Even the internal tools run on OpenAI On the same day the number circulated, Microsoft gave a smaller sign of the same dependence. Jay Parikh runs its CoreAI engineering group. In a memo, he told staff to default to OpenAI's flagship GPT-5.6 Sol in GitHub Copilot, CNBC reported. They should use it most of the time. The reason he gave was money. "Shifting more workloads to OpenAI models helps us get greater value from our token investment," Parikh wrote. Microsoft holds intellectual-property rights to OpenAI's models through 2032. Pointing its own engineers at Sol squeezes more from that stake. It is also part of a wider retreat from tokenmaxxing. That was the brief era when developers ran up big AI bills without watching the output. The memo was first reported by 404 Media. It lands as cheaper open-weight models, many from China, pull spending the other way. The diversification that hasn't landed None of this is for lack of trying. Microsoft has built its own MAI models and offers cloud customers more than 11,000 of them. It has reached for alternatives too, including a $5 billion bet on Anthropic. GitHub Copilot alone, now past 50 million users, serves models from Anthropic, Google, Moonshot, xAI, and Microsoft itself. And yet the filing says roughly 70% of the AI revenue still traces to OpenAI. The effort to spread the risk has not moved the number that matters. Engine or exposure Whether that reads as strength or fragility depends on what the money is. KeyBanc analyst Jackson Ader wants to know where it comes from. Is it services sold to OpenAI, or the benefit of Microsoft's own investment? "The more of that revenue comes from services to OpenAI," he said, "the more favorably I'm going to look at it." The bears are blunter. The writer Ed Zitron notes Microsoft has spent more than $260 billion on capital projects since 2022. He argues much of its AI revenue is OpenAI's compute bill dressed as growth. If OpenAI stumbles, Microsoft is left with the sunk cost and the empty data centres. There is no sign of a stumble. The AI unit is still growing fast, which is why the stock climbed. But the filing settles a long argument with an uncomfortable answer. After years of work and hundreds of billions in spending to become more than OpenAI's landlord, Microsoft, for now, mostly is one.

Microsoft tells employees to stop tokenmaxxing, sets division-level AI budgets

Microsoft's EVP told staff to curb AI token use, switched to a cheaper default model, and set division-level budgets Microsoft executive vice president Jay Parikh has told employees in an internal email that the company needs to rein in its AI token consumption, writing that "tokenmaxxing is not what we are optimizing for." The message, first reported by 404 Media, reveals that Microsoft has switched its default internal AI model to a cheaper OpenAI alternative and introduced formal AI token budget targets at the division level. As of July 2026, every Microsoft division now has an AI token spending cap. Employees can track their individual AI usage through an internal dashboard, and many engineers have been spending hundreds to a few thousand dollars a month in tokens. The shift marks a significant turn from the company's earlier posture of encouraging widespread AI adoption across its workforce. The move places Microsoft squarely in a pattern TNW has tracked since June, when companies including AT&T, Meta, Uber, Walmart, and Amazon began capping or throttling employee AI spending after discovering that token-priced tools behave nothing like the seat-based software licences finance teams know how to budget. Microsoft had already quietly cancelled most Claude Code licences inside its Experiences and Devices group in May, telling engineers to migrate to GitHub Copilot CLI by the end of its fiscal year. Parikh's email goes further. By making a cheaper model the default and creating division-level budgets, Microsoft is building the kind of metered infrastructure that treats AI tooling more like a utility bill than an enterprise software subscription. The anonymous Microsoft employee who shared the email with 404 Media described the budget caps as "the ultimate admission" that the company cannot afford to let its own staff use its AI products without limits. Microsoft is not alone in confronting the maths. Amazon, Adobe, Atlassian, and Citi have all introduced some form of AI usage throttling or spending visibility in recent months. Per-token prices have fallen roughly 98 percent since late 2022, but enterprise AI bills have still tripled because agentic tools consume dramatically more tokens per task than the autocomplete interactions that shaped the original pricing models. The broader financial picture at Microsoft remains strong. The company's most recent quarterly results showed revenue, operating income, and net income all beating Wall Street expectations. But the internal cost controls suggest that even a company generating record profits sees uncapped AI token spending as a line item that can spiral faster than the productivity gains it delivers. For employees, the message is clear: use AI, but know what it costs. That is a long way from the "AI for everyone" rhetoric of eighteen months ago, and it may be the most telling signal yet that the enterprise AI market is moving from an experimental phase into a procurement discipline where every token has a price tag and every division has a ceiling.

'Tokenmaxxing is not what we are optimizing for': Microsoft tells engineer to calm down on AI usage

* Microsoft aims to cut down on "tokenmaxxing" by employees * New guidelines will look to control AI token use to focus on ROI * This is despite Microsoft reporting record financial results recently Microsoft has apparently been forced to introduce limits on how much AI usage its engineers are allowed followed reports that some have been taking things to extreme. The software giant is looking to cut back on "tokenmaxxing" within the company - where employees use far more AI tokens that may be necessary. In an email seen by 404 Media, Microsoft warned employees that new limits on token usage would be introduced as it looked to focus on getting the most out of its AI platforms. New guidelines "As we accelerate our use of GitHub Copilot to deliver on our goals, we all need to be aware of how we consume tokens," the email to employees from Jay Parikh, an executive vice president at Microsoft said. "Tokenmaxxing is not what we are optimizing for," he continued. "I want all of us focused on maximizing outcomes that move the needle for our customers and our business." "As such, we are updating our internal guidance and managing token spend with the same discipline we apply to every other critical resource." In a bid to achieve "get greater value from our token investment", Parikh went on to say Microsoft is making access to the cheaper OpenAI GPT-5.6 model the default model for internal use. Employees were also reminded (via a link to updated internal Copilot guidelines) that as of July 2026 Microsoft divisions have an "AI token budget target," with employees also able to track their individual AI spending. "While there is no target spend value being shared at this time. The data shows that many engineers spend in the range of hundreds of dollars a month to a few thousand dollars in tokens," 404 Media reported the guidelines as saying. Parikh noted that Microsoft does not want to impair the company's progress towards becoming "AI-first," and that it will keep learning and adjusting its AI policies as models and products evolve. "We are not optimizing for fewer tokens," he said. "We are optimizing for more impact per token." The restrictions may come as a surprise to some Microsoft employees, given that the company recently reported yet another bumper financial quarter, and in that respect should have money to splash out on AI usage. However it is the latest step by Microsoft as it looks to focus internal AI usage. In May 2026, it was reported the company was reportedly canceling most of the Claude Code license it uses internally, with engineers being told to use GitHub Copilot CLI, with users given a June 30 2026 deadline to remove Claude Code from their workflows. However tokenmaxxing has proven to be an issue at other tech giants - perhaps most notoriously at Uber, which was forced to admit it had exhausted its entire annual 2026 AI coding token budget in just four months due to massive employee adoption of agentic tools like Anthropic PBC's Claude Code and Cursor. Amazon also recently revealed it had spent $1.8 million on an internal Claude Sonnet deployment intended for matching author details with product listings, after it ballooned far beyond its planned budget and was ultimately given menial tasks to do. Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.

Microsoft makes a controversial decision that changes its AI story

Microsoft spends billions building its own AI models every year. Internally, though, the company just told its own developers to reach for someone else's technology first. The instruction came from deep inside Microsoft's engineering organization, and it says as much about the economics of AI spending as it does about which chatbot writes better code. Why Microsoft defaulted GitHub Copilot to GPT-5.6 Sol Jay Parikh, executive vice president of Microsoft's CoreAI engineering group, told employees in a memo on August 4 to default to OpenAI's flagship GPT-5.6 Sol model when working inside the GitHub Copilot coding tool. "Internally, shifting more workloads to OpenAI models helps us get greater value from our token investment," Parikh wrote, according to CNBC. Parikh's group oversees GitHub, Visual Studio, and Visual Studio Code. Before the Sol switch, Microsoft's internal GitHub Copilot setup ran on an auto-router that defaulted primarily to Anthropic's Claude models. Microsoft engineers were, in effect, mostly coding with expensive Claude tokens on the company's own infrastructure. The memo told staff to lean on OpenAI specifically to capture value from Microsoft's early investment in the company. That is a notable instruction given that Microsoft has built its own AI programming model and gives cloud customers access to more than 11,000 models, including Anthropic's. As of July 2026, Microsoft divisions began operating under formal AI token budget targets for the first time. Parikh framed the shift as a break from what he called "tokenmaxxing," a period when developers were encouraged to run up large AI processing bills without much scrutiny of the output. "Tokenmaxxing is not what we are optimizing for," he wrote. "I want all of us focused on maximizing outcomes that move the needle for our customers and our business." 404 Media reported on the internal memo, which Parikh sent as employees were already logging monthly AI spending ranging from hundreds of dollars to several thousand per engineer. What the Microsoft OpenAI deal means for GitHub Copilot The guidance arrives nine months after OpenAI completed a corporate restructuring that extended Microsoft's intellectual property rights through 2032. Microsoft said separately in April it had stopped revenue-sharing payments to OpenAI, changing the financial shape of the partnership that once ran on shared profits rather than fixed licensing terms. Microsoft's fiscal year 2026 results illustrate why OpenAI stays at the top of the pecking order. The company reported full-year revenue of $331.8 billion, up 18%, with AI services accounting for a growing share. OpenAI-related investment gains contributed $4.963 billion to net income for the full year, reflecting the financial value Microsoft extracts from the partnership beyond direct revenue. The diversification play is real but different in character. Microsoft put up to $5 billion into Anthropic, and Anthropic agreed to spend on Azure in return. There are no public IP commitments in that relationship the way there are with OpenAI. Anthropic's Claude Code only runs on Anthropic's models. GitHub Copilot runs on everyone's, including Anthropic's, Google's, Moonshot AI's, xAI's, and Microsoft's own. That gives Microsoft the flexibility to point traffic wherever it wants. Right now, it's pointing at OpenAI. Microsoft stock and Azure $100 billion in context The memo drops during Microsoft's best stretch on Wall Street in years. July 29 earnings came in at $4.74 per share on $90.01 billion in revenue. Both beat the consensus of $4.24 and $87.62 billion, respectively. The stock jumped roughly 8% on the day, as TheStreet reported. Azure crossed $100 billion in annualized revenue and grew 43% in constant currency. Intelligent Cloud came in at $39.3 billion, up 32%. Worth noting though: a $3.2 billion gain from Microsoft's Anthropic investment added 27 cents to earnings per share. Part of the beat was investment accounting, not operations. That strength has stood out against rising skepticism toward hyperscaler AI spending generally. Alphabet raised its 2026 capex guidance to between $195 billion and $205 billion, pushing quarterly free cash flow negative for the first time, as TheStreet reported. Combined capital expenditures from Microsoft, Amazon, Alphabet, and Meta are expected to top $700 billion this year. What Microsoft's tokenmaxxing crackdown means for AI spending Microsoft isn't alone in doing this. Cheaper open-weight models, a lot of them from Chinese labs, have been quietly eating into the dominance of frontier models from OpenAI and Anthropic. They cost less to run. That gives budget-conscious teams a real alternative in a market that used to have only one direction: spend more. Parikh acknowledged CoreAI has not yet set formal token budgets for individual teams or employees, and encouraged staff to document both successful and unsuccessful uses of AI spending. GitHub Copilot itself has scaled to 50 million users, Microsoft said, even as newer entrants like Cursor have taken market share in the same space. Whether Microsoft's internal preference for OpenAI signals where enterprise customers should be looking remains an open question, but the memo makes one thing explicit: even Microsoft's own engineers are being told that more AI spending is not automatically better spending. The Arena Media Brands, LLC THESTREET is a registered trademark of TheStreet, Inc. This story was originally published August 7, 2026 at 2:03 AM.

Microsoft puts the brakes on AI: why is it limiting its use to its own employees?

Microsoft is rethinking part of its AI strategy when it comes to usage. According to 404 Media, the company will assign different token budgets to its divisions and invite each employee to check their spending to stay within them. The company's goal is to get more value out of each token, shifting the focus from indiscriminate consumption toward results. Better results from each token Jay Parikh, Microsoft's vice president, has asked engineers to measure the value of GitHub Copilot. The increasingly common term "tokenmaxxing" describes a pattern of usage in which more queries are made, more agents are used, and more models are tested, and it's this way of working that, according to Microsoft, has to give way to management like any other critical resource. The guidelines include budget targets starting in July 2026. Apparently, some engineers spend hundreds or thousands of dollars a month on tokens; which helps explain why Microsoft wants to tie each request to concrete results and adjust the limits according to usage. GPT-5.6 as the main option To improve efficiency, Microsoft has made GPT-5.6 the default model. The decision coincides with the drop in GPT-5.6 pricing and with Copilot's ability to work with different models for several years now. Microsoft is thus following a trend already set by AT&T, Meta, and Uber, with spending reaching as much as $7,500 a month per employee.

Microsoft tells employees to stop tokenmaxxing as it switches to OpenAI's GPT-5.6 Sol

Amazon, Meta, Adobe, Uber and other firms are also tightening AI usage as enterprise AI costs continue to rise. Microsoft is reportedly asking employees to use AI more efficiently instead of consuming the maximum possible computing resources. As per a 404 Media report, the company has internally urged staff to avoid tokenmaxxing even as it adopts OpenAI's GPT-5.6 Sol as the default AI model for employees. With this, the company wants to keep AI spending under control while continuing to expand the use of gen AI across their businesses. In an internal memo, Microsoft executive vice president Jay Parikh reportedly told employees that the company wants them to focus on the quality of outcomes rather than the number of AI tokens they consume through GitHub Copilot. For the unversed, tokens are the basic units AI models use to process prompts and generate responses, meaning more complex or frequent requests generally result in higher costs. Parikh also said that Microsoft's objective is not to reduce AI usage altogether but to generate greater value from every token spent. The memo also stated GPT 5.6 Sol was selected because it delivers better value for Microsoft's AI investment and is believed to be more cost-effective than some competing models, including Anthropic's Claude family. The report also claims that Microsoft has introduced updated internal Copilot guidelines that include AI token budget targets for business divisions. The employees are said to be able to monitor their individual AI usage, but no formal personal spending limits have reportedly been enforced yet. Internal data cited in the memo suggests some engineers consume AI tokens worth hundreds to thousands of dollars every month. Not only Microsoft, companies like Amazon, Uber, Adobe, Atlassian and Citi have reportedly introduced measures to reduce AI related expenses, while Meta previously implemented internal token budgets and discontinued its "Claudeonomics" leaderboard.