|
As AI spending has eaten into corporate budgets, generating concern over rising costs at Fortune 500 firms from Uber to ServiceNow and Snowflake, some companies are beginning to throttle employees’ usage. The latest is AT&T, which has begun limiting some employees’ access to Microsoft’s Github Copilot, according to a person at the company. In a more striking example, Meta is also limiting employees’ AI spending on services from Anthropic and other providers, a sharp reversal from a few months ago when employees competed with each other to use as much AI as possible, known as tokenmaxxing, my colleague Jyoti reported Friday. Uber and Walmart similarly capped employee usage of AI coding tools, Bloomberg recently reported. And the Financial Times reported that Amazon has scrapped its “leader board” that ranked employees by the amount of AI they used. Some companies are effectively saying, “I told you so.” “We never celebrated tokenmaxing,” Box CEO Aaron Levie told me. “We never had leaderboards, so we didn't get ahead of our skis on sort of incentivizing the wrong thing.” To be sure, not everyone is aboard the token-minimizing trend. At Databricks, which sells database software and AI tools, “the AI budget is unlimited for engineers, so tokenmaxxing still exists,” Nikita Shamgunov, an engineering leader at the company, said at an event hosted by Nebius last week. That could reflect the relative sophistication of engineers at Databricks, which have long focused on using AI as efficiently as possible. And while “tokenminimizing” may be one way to cut down on costs, swapping out different models such as open source ones for different tasks could be a way to ease the growing pressure of AI bills. Executives of software companies such as Palantir and Box, which sell apps that let enterprise customers to access a variety of models, told me there’s a growing interest from customers in using cheaper or open source models to cut costs. As Microsoft’s Satya Nadella put it over the weekend in an essay on X, “The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see.” Nadella wants AI models to be like an easily swappable commodity, which makes sense because his company owns a dominant suite of productivity applications that are under threat from Anthropic and OpenAI. But Nadella’s post may seem like wishful thinking, given the technical advances we’ve seen so far this year at the leading AI model developers. Still, as I’ve previously written, cost concerns are likely a bullish indicator for software vendors like Palantir and Snowflake and model routing companies that allow customers to easily switch between different models so that simpler tasks can be handled by cheaper models. That path could be even more important as the U.S. government impacts who gets access to advanced models. Those companies are seizing the moment. Microsoft and Databricks, for example, have launched tools called gateways that can help companies control staffers’ AI access and spending. And the Nvidia-backed, $1.5 billion-valuation AI software development upstart Factory launched a new model router at the start of the month. Microsoft Echoes Anthropic’s Pricing Model Meanwhile, even as it tries to court cost-conscious customers, Microsoft on Tuesday revealed it will charge customers extra to use some of its flashiest new AI features. Months after it previewed new Copilot Cowork AI features that resemble Anthropic’s popular Cowork product, Microsoft on Tuesday finally disclosed the product’s pricing model, which appears to take a page from Anthropic’s book by combining per-seat and usage pricing. Microsoft’s Copilot Cowork primarily relies on Anthropic models to automate complex multi-step tasks within Microsoft’s Office 365 software. For instance, users can send Copilot Cowork a file full of screenshots of receipts and direct it to create a spreadsheet listing the expenses shown in the photos. That goes beyond more rudimentary tasks that Microsoft’s existing 365 Copilot was meant to tackle, such as summarizing emails or creating financial models in Excel. And like Anthropic’s enterprise software, Microsoft’s Copilot Cowork won’t be cheap. Microsoft said Tuesday that to access Copilot Cowork users will need a 365 Copilot license, which starts at $30 per user per month—and will be billed an additional charge on top of that based on how much they use Copilot Cowork. That’s similar to Anthropic’s per-seat-plus-consumption billing that it started rolling out to enterprise customers earlier this year. Microsoft appears sensitive to companies’ anxiety around spiking AI costs, though—executive vice president Charles Lamanna said in a blog post Tuesday that customers “have options to manage cost,” such as setting a ceiling on how much Copilot Cowork their employees can use. And Microsoft is previewing the ability for customers to swap out Anthropic’s models in Copilot Cowork for other models from OpenAI or Microsoft itself, which it says can deliver similar results at a lower price point. The company is also testing open source models that could swap out for Anthropic’s in some cases, according to an employee with knowledge of the situation. That might help it win over companies that want to try Copilot Cowork while “tokenminimizing,” especially as Microsoft aims to compete with Anthropic for those same customers. KPMG’s Hallucination If you want some comic relief, check out this piece from the Financial Times about an October report from consulting giant KPMG on enterprise AI adoption. Evidently, it included fake case studies that purported to show how businesses such as UBS and Swiss Federal Railways were making gains with AI. As all the chatbots tell you when they generate answers: double check their work! Or do your own research. KPMG told The Information, as it told FT, “We expect all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content and verify independent sources.” The firm said it “takes the accuracy and integrity of its published content seriously” and had removed the report from its website.
|
Gift Subscription
