Enterprise LLM Market Share: Why Four Trackers Name Three Different Leaders

Ask who leads AI market share in the enterprise and you get three answers depending on which tracker you open. Menlo Ventures puts Anthropic at 40% of enterprise LLM API spend. Ramp's AI Index says 43.8% of US businesses pay Anthropic and 40.6% pay OpenAI. OpenRouter's token data says the most used model on its network is from DeepSeek. None of them is wrong. They are measuring different things.

Enterprise LLM market share depends on what you count

Market share needs a numerator and a denominator. The three trackers people cite most often each pick a different pair, and that choice decides the leader before any data is collected.

Here is the short answer to the question most readers arrive with. If you want to know who earns the most enterprise API revenue, Menlo is the closest public proxy and it says Anthropic. If you want to know which vendor most businesses pay, Ramp says Anthropic by a narrowing margin. If you want to know which models developers route the most tokens through, OpenRouter says Chinese open weights, led by DeepSeek.

Spend: dollars flowing to a model provider

A spend measure weights every customer by how much it pays. One large customer running a coding agent all day counts for more than a thousand small firms with a chat subscription. This is the measure that maps to vendor revenue, and it is the hardest to observe from outside.

Paying customers: how many firms have a bill

A penetration measure counts each business once if it pays a vendor anything in the period. A $20 seat and a $2 million API contract count the same. It tells you about reach and sales motion, and almost nothing about revenue concentration.

Routed tokens: how much text a model processes

A volume measure counts tokens, the chunks of text a model reads and writes. Cheap models win on volume because developers send them bulk work. A model priced at a fraction of a cent can lead on tokens while earning a small share of the dollars.

Readings are the most recent each source published that this post could verify. Menlo's is 10 months old. Ramp's is monthly. OpenRouter's rankings page updates continuously, but its author share chart did not render for verification, so the June study is used.

Menlo Ventures measures enterprise LLM API spend

Menlo's December 2025 report is the source behind most headlines saying Anthropic overtook OpenAI. It surveyed 495 US enterprise AI decision makers between 7 and 25 November 2025, with an independent research firm running the fieldwork.

The headline is a three-way split. Anthropic held 40% of enterprise LLM API spend, OpenAI 27% and Google 21%, which leaves 12% for Meta's Llama, Cohere, Mistral and a long tail. Menlo's own history shows the shift: OpenAI held 50% in 2023 and Anthropic 12%.

Coding is the sharpest split. Menlo estimated Anthropic at 54% of enterprise coding share against 21% for OpenAI. Foundation model APIs were $12.5 billion of an estimated $37 billion in 2025 enterprise generative AI spend.

How Menlo turns a survey into dollars

Respondents reported the share of their AI workloads running on each model. Menlo then weighted those answers by the scale of each application and triangulated against public financials where available. The report calls the result "estimated dollars spent based on proportion of production API usage."

That definition has moved. Menlo's July 2025 mid-year update, built on 150 technical decision makers, said its share "reflects the proportion of production AI usage, not spend." Six months later, a methodologically similar number was described as estimated dollars. I do not think that makes the figure useless. It does mean the 32% mid-year reading and the 40% December reading are not a clean time series.

Two scope choices matter as much as the sample. The sizing excludes inference and model serving through AWS, Google Cloud, Azure and Fireworks as a spend category. It also excludes AI features built into existing software. Menlo also marks Anthropic as a portfolio company in its own charts, which is disclosed and worth keeping in mind. How Menlo's same dataset splits startups against incumbents is covered in the analysis of whether incumbents' distribution beats better AI.

Ramp AI Index counts businesses that pay, not dollars

Ramp is a corporate card and bill pay company. Its AI Index, written by economist Ara Kharazian, reads the transactions its customers run through the platform. Ramp states that it counts "corporate card and invoice-based payments in our measurements."

That gives Ramp something Menlo lacks: observed payments rather than recalled workload shares, published monthly. The latest release, dated 14 October 2026, reports September data. Anthropic was paid by 43.8% of businesses, OpenAI by 40.6%, and 55.9% paid for any AI at all.

Ramp first showed Anthropic ahead in its 13 May 2026 release. In April data, Anthropic reached 34.4% of businesses against 32.3% for OpenAI, the first time it had passed OpenAI on this measure.

Penetration is not market share

The Ramp figures sum to more than the total, and that is the tell. If 55.9% of businesses pay for any AI and 43.8% pay Anthropic and 40.6% pay OpenAI, at least 28.5% of all businesses on Ramp pay both. That is my arithmetic from Ramp's three published numbers, and it is a floor, not an estimate.

So Ramp's "lead" is a statement that slightly more firms have an Anthropic bill than an OpenAI bill. It says nothing about which bill is larger. Ramp itself has noted that 80% of spend on OpenAI and Anthropic is driven by 1% of firms, which is exactly why a count of firms and a share of dollars can diverge.

Ramp also has a sample frame. Its customers are US businesses that chose a modern spend platform, which skews toward startups and mid-market firms. Purchases made through a cloud provider's marketplace, or on an employee's personal card, are not visible in the pages I read. Ramp's own per-employee spend data is used in the post on why cheaper tokens produce bigger AI bills.

OpenRouter token share shows what developers route

OpenRouter is a single API that lets developers send requests to hundreds of models from dozens of providers. Its public rankings order models "by tokens processed through the OpenRouter API." That makes it the only one of the three trackers built entirely on observed machine traffic.

On OpenRouter the leader is neither Anthropic nor OpenAI. OpenRouter's own analysis covered over 450 trillion tokens from 1 January to 14 June 2026. It found DeepSeek "effectively doubled its token share over the period (from 9% to 18%)." DeepSeek had been the top model on the network since mid-May.

The same post says Chinese models passed American ones in token share as of early June 2026, after US models had served about three quarters of tokens in 2025. OpenRouter is explicit that these are "share of token volume (input and output combined), not share of spend."

Two features of the traffic explain why. Developers on a router are price sensitive and send bulk, automated work. OpenRouter also measured agentic workloads burning about 15 times more tokens per request than human use, so whichever model wins agent traffic wins the volume chart. OpenRouter's earlier study with a16z researchers states that the data "reflects OpenRouter's vantage point as an API service, not the entire ecosystem."

Enterprises mostly buy direct contracts or go through their cloud provider. So a large share of enterprise traffic never touches a router. What open weight models actually cost to run in production is set out in the comparison of Qwen, DeepSeek, Llama and GLM in production.

The fourth tracker: a16z's CIO survey says OpenAI

A fourth public instrument complicates the clean "Anthropic has overtaken OpenAI" story. Andreessen Horowitz published its third annual CIO survey on 30 January 2026, built on 100 verified VPs and C-level executives at Global 2000 companies.

Its findings run the other way. The survey found 78% of these CIOs using OpenAI models in production, directly or through cloud providers. It put OpenAI at a majority of wallet share, "at ~56%," with Anthropic and Google gaining. Anthropic was in production at 44% of firms, rising to over 63% including testing.

The disclosure is symmetrical with Menlo's, which is useful. a16z opens the section with "we are investors in OpenAI." Menlo marks Anthropic as a portfolio company. Each investor's survey shows its own portfolio company in the stronger position. I do not read that as bad faith, since both samples are disclosed. I read it as a reason never to cite either one alone.

The a16z sample also says it was "not randomly sampled" and "may over-represent organizations that are more advanced in AI adoption." Neither investor survey publishes its questionnaire in full. That pattern is examined in the breakdown of why AI ROI surveys reach opposite results.

The three biggest disagreements in AI market share

Put the four instruments side by side and three disagreements stand out. Each has a specific mechanical cause, and none requires anyone to be wrong.

1. Anthropic versus OpenAI depends on company size

Menlo says Anthropic leads enterprise API spend 40% to 27%. a16z says OpenAI holds about 56% of model wallet share. Both claim to measure enterprise spend on models, and the leader flips.

The cleanest explanation is the frame. a16z counts only the Global 2000, where Microsoft's Azure OpenAI channel and existing ChatGPT Enterprise contracts are strongest, and it explicitly includes models consumed through cloud providers. Menlo's sample is broader US enterprise, weighted by application scale, which favours heavy API users such as coding tools. My view is that Menlo better describes where new API dollars go, and a16z better describes where the largest incumbents' budgets sit today.

2. Open weights are 1% or the majority, depending on the door

Menlo found Chinese open source models at 1% of enterprise LLM API usage, with all open source at 11%. OpenRouter found Chinese models ahead of American ones on tokens by June 2026.

Both can be true at once. Enterprises buying through procurement and security review mostly choose US closed models. Developers paying per token on a router choose whatever is cheapest for the job. The gap is the distance between the procurement door and the developer door, and it is the most decision-relevant number in this post.

3. The direction of travel is disputed

Menlo's series shows a one-way trend: OpenAI from 50% in 2023 to 27% in late 2025. Ramp's September data shows OpenAI adding 1.12 points of business adoption against Anthropic's 0.25, the fastest OpenAI gain in the releases I checked.

The cause here is timing. Menlo's reading is from November 2025 fieldwork and has not been refreshed. Ramp's is from September 2026. Anyone quoting Menlo's 40% in October 2026 as current is quoting a number older than several model generations. I would treat the Ramp trend as the more current signal on reach, and treat the question of current spend share as unanswered in public data.

Which AI market share number fits which decision

The useful question is not which tracker is right. It is which tracker answers the question in front of you. Here is how I would map them.

The last row is the one most commentary skips. If you need to know who earns the most, the companies' own revenue disclosures are the better source, and the AI revenue leaderboard built from reported numbers does that work. A share estimate is a proxy for revenue, and the proxy should never outrank the thing itself.

For a founder deciding which models to build against, I would weight Ramp and OpenRouter over the two investor surveys. Both observe behaviour rather than recall it, and both update fast enough to catch a shift within a quarter. The practical cost of guessing wrong is lower than it looks, because the real cost of switching model providers sits mostly in prompts and evals, not in contracts.

Where this comparison is weakest

This post compares instruments rather than measuring the market, and that has limits I should state plainly.

First, the freshest reading for OpenRouter here is from June 2026. OpenRouter's live rankings page shows usage data through 15 October 2026, but its author market share chart did not render in a form I could verify. Secondary trackers report weekly author shares from that page. I have not used them, because a figure I could not open is a figure I should not print.

Second, Ramp's monthly figures are revised. Its September release gives Anthropic 43.8% for August, and its October release gives 43.8% for September after a stated 0.25-point gain. Those cannot both be unrevised. Small month-on-month moves in this series deserve less weight than the trend.

Third, I cannot see Ramp's sample size or its exclusions in the releases I read. Purchases through AWS, Azure or Google Cloud marketplaces are a large channel for enterprise model spend, and if they are invisible to card data, Ramp undercounts exactly the buyers that a16z oversamples.

Finally, the strongest counter-argument to this whole post is that the trackers agree more than they disagree. Three of four put Anthropic and OpenAI as the top two US vendors, and all show Anthropic gaining since 2023. If your question is simply "which two vendors matter," the answer is not contested. The disagreements only bite when someone quotes a precise percentage as the share of the market.

Frequently asked questions

Where to start

Pick the one AI market share figure you or your team quoted most recently, in a deck, a memo or a sales call. Write down which of the three units it was: dollars, paying firms or tokens. If you cannot tell, the figure should come out of the deck until you can.

Then replace the share figure with the measure that matches your decision from the table above, and add its date beside it. A dated, correctly labelled 43.8% of businesses is more persuasive to a sceptical buyer than an undated "40% of the market," and it survives the follow-up question.

Related reading