|
Welcome back! Microsoft CEO Satya Nadella said Wednesday that a lot more people were heavily using the company’s AI-powered applications. But his company’s first quarter financial results showed the usage was dragging down profit margins in its cloud unit, prompting Microsoft to effectively boost prices. The conundrum should sound familiar to every company that sells AI-powered apps. Look no further than Anthropic, which recently revamped its Claude Enterprise pricing to charge customers based on how much they used the tools on top of a flat monthly fee, rather than just charge a subscription price. The shift from seat-based prices is a hot trend, as Laura recently reported, though customers aren’t exactly loving it. In Microsoft’s case, the company said Wednesday that increased usage of GitHub Copilot, which relies in part on Anthropic models to write and edit code, dragged down gross profit margins in its cloud unit, which also includes the all important Azure server rental business. That margin fell about 5 percentage points to 56% in the first quarter compared to a year earlier. (Shareholders were unimpressed by Microsoft’s overall results, pushing the stock down 4.5% so far today, though Microsoft has recently set more aggressive pricing for customers renting servers equipped with Nvidia chips, which could boost Azure margins as demand has far exceeded supply.) Microsoft has already copied some of Anthropic’s pricing changes to stop some of the bleeding. The company on Monday debuted new pricing to charge GitHub Copilot customers additional fees based on how much they use it. And that’s on top of a price hike GitHub Copilot announced just last week! Meanwhile, gross margins increased slightly in the business unit that includes Office 365 AI products, which the company attributed to increased efficiency “even with” growing usage of those products. CFO Amy Hood told analysts that new consumption based pricing boosted margins. Nadella all but said Microsoft will follow the GitHub Copilot playbook for its other AI products, including 365 Copilot, its flagship AI tool that automates tasks in Outlook, Teams, and other Office apps. “Any per-user business of ours, whether it's productivity or coding or security, will become a per-user and usage business,” he told analysts in a conference call Wednesday. “That’s obviously already happening with [GitHub Copilot] coding [with] some of the business model changes we made this quarter … but it also speaks to the intensity of usage.” Microsoft is already moving towards consumption-based pricing for its core enterprise AI apps. Microsoft charges a $30 per month flat fee for enterprise users of 365 Copilot—and it said Wednesday that paying subscribers grew 33% quarter over quarter to 20 million. But the company recently added additional features to persuade customers to pay extra for custom “agents” that can work on tasks around the clock. Those tasks could include routinely updating a sales database or browsing the internet to create daily research reports, which are billed based on how much customers run the agents. Newer products also seem poised to usher customers towards more consumption pricing: the company is about to roll out a new software bundle, E7, which combines Office software, security tools, and 365 Copilot for $99 per user per month. That bundle includes a light version of the consumption-based agent features, but customers have to pay an additional amount when they exceed a usage cap for the agents. (Microsoft is also plotting new Copilot features similar to OpenClaw agents, which would likely rely on consumption based pricing.) ‘False Sense of Security’ In Nadella’s view, businesses are willing to stomach paying more for consumption because their use of Copilot will “compress workflows,” leading to eventual cost savings or faster revenue growth. That’s a hopeful statement, given the hand-wringing we heard from corporate chief information officers about Anthropic’s price hikes. And rivals like Google have been pitching the cost predictability of seat-based AI as a way to win big deals with large enterprises. Pricing models that mix consumption and seat based pricing can create a “false sense of security” that leads to unexpected costs for businesses, said Adam Mansfield, a consultant at UpperEdge who negotiates deals with Microsoft and other software firms on behalf of large customers. “The seat-based part is relatively easy to determine and forecast for,” Mansfield said. “The problem is the consumption-based pricing component … customers are not going to have an appetite to increase their AI spending when they weren't expecting it to happen.” Of course, the prices that AI app makers like Microsoft pay for the underlying models that power them fluctuate. Last summer, for instance, AI customers saw prices creep up thanks to an increased reliance on “thinking” AI models that gobbled up more tokens than earlier models. But Microsoft also said on Wednesday that it’s reduced the cost of running Copilot by making models run more efficiently, so Microsoft could find ways to pass along some future cost savings to its customers. To convince CIOs that the new pricing is great, Microsoft may try to point to its own experience using AI to save on labor costs. The company said Wednesday that overall headcount is down year over year “as we focus on building high performing teams that operate with pace and agility,” CFO Amy Hood said. CIOs and CTOs: let us know what you think! Everyone Has Already Launched ‘Stateful’ AI In February, Amazon and OpenAI introduced a jargony term for a new way to build AI agents involving a “stateful runtime environment.” We wrote a whole column about how stateful AI differs from the current paradigm of so-called stateless models that most businesses use in that they are customized to specific businesses so they understand how the business operates and remember past conversations and actions they took. On Wednesday, Amazon CEO Andy Jassy said stateful AI will give Amazon Web Services an advantage over cloud rivals because it essentially automates much of the work that corporate software engineers typically do when developing customized agents today. This work includes managing what types of data an agent is permitted to access, what kinds of tools it can access to complete tasks, and enabling it to recall past actions. “I think that‘s the future of how these agents are going to be built [and] it’s something that nobody else has,” Jassy said Wednesday as his company launched a preview of a new stateful product, Amazon Bedrock Agents, which OpenAI models power. We’re not so sure. Microsoft and Google also offer stateful AI development tools to customers, as does Anthropic. They just haven’t made a big deal out of it. The next question is whether and when stateful AI products turn out to be useful to enterprises.
| 
