Who leads in AI depends on what you count: payments, agents or cloud

Ask who is winning business AI and you will get two recent answers. They come from different sources and they point in opposite directions. The lesson is not that one is wrong. The lesson is that "market share" in AI is no longer one number. It depends on whether you count who pays, who orchestrates, or who hosts.

Two studies, two leaders

Ramp tracks what U.S. businesses on its card and software platform pay for AI. In its September index, covering August, 43.8% of those businesses paid Anthropic for subscriptions or tokens. OpenAI reached 39.8%. Ramp also said overall adoption is still growing but slowing.

VentureBeat Intelligence surveyed companies about agent orchestration in September. Think of orchestration as the control layer that decides which AI agent gets a job, which tools it may use and how far its access extends. Of 107 companies using such a platform, 43% used OpenAI's, 22% used Anthropic's and 21% used Google's.

Why the numbers do not collide

Each study has limits, and VentureBeat names them. Ramp covers only businesses that use Ramp, which tend to be more tech-forward than average. Size does not matter in its tally: a firm with 50 staff and a firm with 50,000 each add one to the count. VentureBeat reported that small firms far outnumber large ones, so they shape the totals.

Production tells a third story. VentureBeat asked which provider carries most of their production AI. Google Cloud or Gemini led with 32%, ahead of OpenAI at 22%. Anthropic came in at 6%. VentureBeat cautions that some firms may have named the cloud that hosts a model instead of the model itself.

Real companies mix vendors

Home Depot shows why one label misleads. Angie Brown, its EVP and CIO, told VentureBeat the company is "a pretty hefty Google shop" with most of its data in Google's BigQuery warehouse. Even so, its teams use OpenAI models, hosted on Microsoft Azure, to grade answers from Google's models. For software, Claude drafts the code and OpenAI models check it.

That is one company, not a trend. But it shows how a single firm can appear in three vendors' counts for three different reasons.

The decision sits in the orchestration layer

The thesis here is simple. Winning a spot on the invoice and controlling the workflow are different positions. The orchestration layer decides which model gets a task, which tools an agent may call and what it is allowed to touch. That is where an organisation governs its agents. It is also where switching vendors gets hard or easy.

Ramp's economist says volume is rising, but may not offset lower prices. The growth comes from cheaper standard models such as GPT-5.6 Terra and Claude's Sonnet series. Frontier models drove 45% of token share, below an August peak of 53%. Ramp heard from businesses that set company-wide defaults to reduce frontier use, because standard models perform well and cost less.

One caution on spending. Median monthly AI spend per employee among the top 1% of firms fell 9.7% in August, to $7,205 from $7,976. Ramp notes that group is small and volatile, and that August is a vacation month. It also revised July upward after late transactions arrived, and says the figures are subject to change. Treat it as one month, not a direction.

Questions to put to your team

First, which layer of our AI stack do we own: the model contract, the agent platform or the cloud? Second, how many vendors run in production, and who approves the permissions each agent holds? Third, do routine tasks default to standard models, and is that setting written down? Fourth, are we paying the lower per-token prices now on offer, or the rates signed earlier?

A market-share chart answers one question. Your own architecture answers the one that decides your risk and your bill.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Ramp.