Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Anthropic has toppled OpenAI in share of enterprise GenAI spending.
- Enterprise spending represents an AI boom, not a bubble.
- Agentic AI is still a niche compared to prompt engineering.
The most popular single use of generative artificial intelligence in the enterprise is coding tools for programmers. That tool’s popularity has helped Anthropic trounce OpenAI in the enterprise in 2025, according to a new survey of company spending.
“The foundation model landscape shifted decisively this year when Anthropic surprised industry watchers by unseating OpenAI as the enterprise leader,” according to the report, the third annual State of Generative AI in the enterprise, from venture capital firm Menlo Ventures.
Also: How Anthropic’s enterprise dominance fueled its monster $183B valuation
The report estimated that Anthropic now earns 40% of enterprise LLM spend. That’s up from 24% last year and 12% in 2023. Meanwhile, OpenAI lost almost half of its enterprise share, dipping to 27% from 50% in 2023, wrote authors Tim Tully, Joff Redfern, Deedy Das, and Derek Xiao.
The team based their market share on the dollar amount spent by enterprises on production API usage, as weighted by the scale of each customer. The report estimated that US enterprises spent a total of $37 billion, up more than threefold from last year’s $11.5 billion.
The report is specifically based on US spending, so it’s not clear what the data suggest about global trends. The team surveyed executives in the C-suite, IT, and engineering at 495 US companies in November, with the help of a third-party survey firm, to put together an estimate for total spending.
Anthropic dominating code automation
The report confirms earlier findings by Menlo Ventures on the popularity of Anthropic’s Claude large language models, as detailed back in August. It is largely being fueled by the automation of code-writing.
“Anthropic’s ascent has been driven by its remarkably durable dominance in the coding market, where it now commands an estimated 54% market share, compared to 21% for OpenAI,” the report said.
Also: Anthropic beats OpenAI as the top LLM provider for business – and it’s not even close
And coding tools, represented by startups such as Replit, Cursor, Harness, Windsurf, Augment Code, and All Hands AI, are now a $4 billion annual business, the Menlo Ventures team noted, “making it the largest category across the entire application layer,” and “Generative AI’s first ‘killer use case.'”
Anthropic’s success thus is being helped by those startups in coding tools. That is consistent with anecdotal observations, such as those by software entrepreneur Jeremy Burton in my October interview.
Also: Why AI coding tools like Cursor and Replit are doomed – and what comes next
Burton pointed out that most coding tools, such as Cursor and Replit, depend on Anthropic’s Claude technology.
“Most of those startups depend on Anthropic’s model,” Burton told me.
Menlo has a direct financial interest in the findings, as it funds many startups in the area, including Anthropic.
An AI boom versus a bubble
The report has a distinctly more positive tone than last year’s report by Menlo Ventures, which detailed the struggles of enterprises to get AI off the ground. It appears that a big shift in the intervening year is more and more packaged applications emerging versus do-it-yourself enterprise projects.
Also: Enterprises are struggling with what to do with Gen AI, say venture capitalists
The authors noted that “For a while, the prevailing wisdom was that enterprises would build most AI solutions themselves.” Last year, 47% of AI solutions were built internally versus 53% purchased; this year, that’s reversed.
“Today, 76% of AI use cases are purchased rather than built internally. Despite continued strong investments in internal builds, ready-made AI solutions are reaching production more quickly and demonstrating immediate value while enterprise tech stacks continue to mature.”
The rise of coding tools, but also vertical-market tools for customer experience and finance, prompted the venture capitalists to sound a singularly positive note.
“This year’s findings make clear that the shift is no longer speculative,” they wrote. “Enterprise AI is now a $37 billion market — the fastest-scaling category in software history. Across industries, AI has become core to how work gets done. Enterprises, seeing real returns, are doubling down.”
Noting the widespread fears currently of a funding bubble in artificial intelligence,” the authors replied that while “the concerns aren’t unfounded,” the “demand side tells a different story,” and the “real revenue” means that it’s more “a boom versus a bubble.”
The total estimated spending by enterprises this year, $37 billion by Menlo Ventures’s reckoning, is up more than threefold from last year’s number.
Agents are still a niche
One area that so far hasn’t played out is a category that Menlo Ventures has been quite optimistic about: agentic AI, where large language models can have hooks into other enterprise programs and data stores to perform more advanced tasks.
So far, the agents are not swarming the enterprise, neither in the packaged applications nor in the AI infrastructure spending, they noted.
In the largest and fastest-growing part of AI applications, horizontal AI applications, almost all spending, 86%, is on simpler co-pilot programs such as ChatGPT Enterprise, Claude for Work, and Microsoft Copilot, versus agentic AI such as Salesforce Agentforce, Writer, and Glean.
Also: The journey to fully autonomous AI agents and the venture capitalists funding them
In the $1.5 billion AI infrastructure category of spending, most of what’s being put into production is not agentic; it’s plain-old-fashioned prompt engineering, noted Tully and team.
“For all the talk of ‘agents,’ real production architectures remain surprisingly simple,” they wrote. “Only 16% of enterprise and 27% of startup deployments qualify as true agents — systems where an LLM plans and executes actions, observes feedback, and adapts its behavior — while most are still built around fixed-sequence or routing-based workflows wrapped around a single model call.”
The authors laid out several predictions at the end of the report, including the prospect that coding tools will prove superior to human coders in some mundane tasks.
“AI will exceed human performance in daily practical programming tasks,” they said. “There is no plateauing of LLM skill sets, especially in verifiable domains such as math and programming, where the best models will continue to get better and better.”
The report may be too bullish
It’s worth asking whether the positive attitude of the report is really supported by the data that Tully and team have gathered.
A three-fold increase in industry revenue in one year’s time is indeed striking, but it seems less impressive as you dig into the numbers.
While $37 billion in total annual generative AI sales in the US may sound like a lot, it is relatively small compared to, for example, amounts spent on cloud computing. This year, for just the top three cloud vendors, Alphabet’s Google Cloud, Amazon’s AWS, and Microsoft’s Azure, combined revenue is projected to reach $288 billion.
And if you examine the various categories of spending compiled by Tully and team, it’s clear there’s very little momentum for categories outside of the most predictable ones.
Within the $37 billion, $18 billion for infrastructure is the stuff that has been selling steadily for three years now, the usage of frontier model APIs, such as Claude and GPT, and the attendant infrastructure equipment. Within applications, $8.4 billion is for the earliest tools offered, the horizontal applications such as Microsoft’s Copilot offerings, and $4.2 billion is for the coding tools that are a big hit, such as Cursor.
What that means is that almost all of the spending by enterprises, 83%, is simply the most obvious use cases, namely, renting API usage, running co-pilots, and using coding tools. It’s heavily tilted toward the technical users most inclined to pick up Gen AI, the coders and product developers.
In almost every other category of the remaining $6.4 billion, the initial amounts spent seem minuscule.
For example, within departmental AI, a total of $360 million has been spent on AI-driven human resources applications, they noted, and $100 million on finance and operations applications using AI. Those numbers are tiny compared to the annual sales of, for example, Workday, a vendor of both HR and finance applications, which is projected to make $9.5 billion in annual sales this year.
Also: ChatGPT saves the average worker nearly an hour each day, says OpenAI – here’s how
And despite lots of talk about marketing teams in enterprises using Gen AI to replace creatives, the total spend by marketing on Gen AI was just $660 million, they estimated — a fraction of the revenue of companies such as Adobe and Figma.
The authors may have been a little too charmed by that overall three-fold rate of growth — the fastest rate of growth of any software category in history, as they said. When an entire industry consisting of every established software vendor on the planet, plus hundreds of venture-backed startups, goes after a single market at the same time, of course, there will be fast growth.
But when all of those companies manage to sell only $37 billion in twelve months, it’s not as big a triumph as it seems at first blush.