LLM Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: August 17, 2026 | License: CC BY 4.0
The worldwide AI models and platforms market reaches $64.25 billion in 2026, and the language-model layer inside it more than doubles to $28.27 billion. But there is no single LLM market leader. Anthropic holds 40% of enterprise LLM API spend while OpenAI holds 76.85% of consumer AI referrals. Both figures are correct.
Quick Answer
Worldwide end-user spending on AI models and platforms is forecast at $64.25 billion in 2026, up 63.4% from $39.31 billion in 2025, per Gartner’s July 2026 forecast. Anthropic leads enterprise LLM API spend at 40%, per Menlo Ventures. OpenAI leads consumer AI referrals at 76.85%, per Statcounter. According to Axis Intelligence Research, the top three providers control 88.0% of the enterprise market and 88.51% of the consumer market — a 0.51-point difference in concentration with the leaderboard completely reversed.
Key Findings
- According to Axis Intelligence Research, the LLM Market Divergence Index (LMDI) reads +37.34 points for Anthropic, −49.85 for OpenAI and +12.00 for Google as of the August 2026 snapshot, measuring how far each provider’s enterprise share sits from its consumer share.
- Gartner forecasts worldwide spending on foundation generative AI models at $23.36 billion in 2026, up 104.2% from $11.44 billion in 2025.
- Menlo Ventures estimates Anthropic earned 40% of enterprise LLM spend in 2025, up from 12% in 2023, while OpenAI fell to 27% from 50% over the same period.
- Statcounter recorded ChatGPT at 76.85% of worldwide AI chatbot referrals in April 2026, its lowest level on record, down 7.36 percentage points from 84.21% twelve months earlier.
- According to Axis Intelligence Research, the enterprise and consumer LLM league tables have a Spearman rank correlation of −0.50 across the three leading providers, meaning position in one market carries no predictive value for position in the other.
Source: Axis Intelligence Research — CC BY 4.0
How Big Is the LLM Market in 2026?
The honest answer starts with a definition problem. Spending on “AI” in 2026 runs to $2.59 trillion by Gartner’s May 2026 forecast, a number that includes semiconductors, servers, network fabric and devices. Spending on the models themselves is a fraction of that, and it is the only figure that describes the LLM market as a buyer of language models would recognize it.
Worldwide AI Models and Platforms Spending, 2025-2026
| Segment | 2025 ($M) | 2026 ($M) | Growth | Source |
|---|---|---|---|---|
| Foundation generative AI models | 11,438 | 23,356 | 104.2% | Gartner, July 2026 |
| Domain-specific and specialized GenAI models | 1,583 | 4,910 | 210.0% | Gartner, July 2026 |
| AI application development platforms | 6,885 | 9,541 | 38.6% | Gartner, July 2026 |
| AI platforms for data science and ML | 19,405 | 26,444 | 36.3% | Gartner, July 2026 |
| Total market | 39,311 | 64,252 | 63.4% | Gartner, July 2026 |
Source: Gartner, “Forecasts Worldwide AI Platforms and Models Market to Grow 63% in 2026,” July 20, 2026. Gartner states a 2026 total of $64,252M; the published segment rows sum to $64,251M. Axis Intelligence Research reports both rather than silently reconciling the one-million-dollar rounding difference.
Isolating the two model segments gives the cleanest available read on the LLM market proper. According to Axis Intelligence Research, foundation and specialized generative AI models together account for $28.27 billion in 2026 against $13.02 billion in 2025 — growth of 117.1%, and a rise from 33.1% to 44.0% of the total AI models and platforms market. The language-model layer is not merely growing; it is taking share from the classical data-science tooling that used to define this category.
Enterprise Generative AI Spending and Where the Model Layer Sits
Gartner measures worldwide end-user spending. Menlo Ventures measures something narrower and, for competitive questions, more useful: US enterprise dollars, built from a survey of 495 enterprise AI decision-makers fielded November 7-25, 2025, combined with a bottom-up market model.
Menlo puts total enterprise generative AI spending at $37 billion in 2025, up 3.2x from $11.5 billion in 2024. Of that, $18 billion went to infrastructure and $19 billion to applications. Inside infrastructure, foundation model APIs took $12.5 billion. According to Axis Intelligence Research, that places the model layer at 33.78% of all enterprise generative AI spending — meaning roughly two of every three enterprise AI dollars are spent above the model, on the software that calls it.
These two figures are not interchangeable and Axis Intelligence Research does not average them. Gartner’s $23.36 billion is worldwide end-user spending on foundation models across all buyer types. Menlo’s $12.5 billion is US enterprise API spending, explicitly excluding chips, model serving on hyperscaler infrastructure, and AI features embedded in existing software. Combining a worldwide figure with a US enterprise figure would produce a number that means nothing. They are published here side by side, with their scopes stated.
Sarah Mitchell: The $2.59 trillion number and the $28 billion number get used in the same sentence constantly, usually by someone arguing the market is either enormous or a bubble. They are measuring objects two orders of magnitude apart. If you want to know what enterprises pay for tokens, the model rows are the only ones that answer the question — everything else in the $2.59 trillion is the building the tokens are served from.
What Is the LLM Market Share in 2026?
Two independent measurement systems, both credible, both published, produce opposite league tables.
Enterprise LLM API Market Share by Provider
| Provider | 2023 | 2024 | 2025 | Source |
|---|---|---|---|---|
| Anthropic | 12% | 24% | 40% | Menlo Ventures, Dec 2025 |
| OpenAI | 50% | Not published | 27% | Menlo Ventures, Dec 2025 |
| 7% | Not published | 21% | Menlo Ventures, Dec 2025 | |
| All other providers | — | — | 12% | Menlo Ventures, Dec 2025 |
Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise, published December 9, 2025. Shares approximate dollars spent, derived from reported proportion of production API usage, weighted by application scale and triangulated against publicly reported financials.
Menlo’s own framing of the shift is direct: Anthropic tripled its share over two years while OpenAI ceded nearly half of its own. Google’s climb from 7% to 21% is the least-discussed movement in the table and, at 3.0x over two years, the second-fastest.
Consumer AI Assistant Referral Share, April 2026
| Assistant | Apr 2025 | Apr 2026 | 12-month change | Source |
|---|---|---|---|---|
| ChatGPT | 84.21% | 76.85% | −7.36 pp | Statcounter |
| Google Gemini | 2.31% | 9.00% | +6.69 pp | Statcounter |
| Perplexity | 12.07% | 7.73% | −4.34 pp | Statcounter |
| Microsoft Copilot | 0.23% | 3.76% | +3.53 pp | Statcounter |
| Claude | 0.30% | 2.66% | +2.36 pp | Statcounter |
| DeepSeek | 0.88% | 0.01% | −0.87 pp | Statcounter |
Source: Statcounter Global Stats, AI chatbot referral share worldwide, monthly series, data through April 2026, released May 7, 2026. Based on over 3 billion monthly page views across more than 1 million websites. Grok is absent from the series because it does not pass referral data in its header. April 2026 rows sum to 100.01% because of publisher rounding.
The LLM Market Divergence Index (LMDI)
LMDI — LLM Market Divergence Index. A signed measure, in percentage points, of how far a provider’s enterprise position sits from its consumer position.
Formula: LMDI = enterprise LLM API spend share − consumer AI referral share
Inputs: Enterprise share from Menlo Ventures (December 2025). Consumer share from Statcounter (April 2026). Provider-to-product mapping: Anthropic to Claude, OpenAI to ChatGPT, Google to Gemini.
| Provider | Enterprise share | Consumer share | LMDI | Reading |
|---|---|---|---|---|
| Anthropic | 40.00% | 2.66% | +37.34 | Enterprise-weighted |
| 21.00% | 9.00% | +12.00 | Moderately enterprise-weighted | |
| OpenAI | 27.00% | 76.85% | −49.85 | Consumer-weighted |
Source: Axis Intelligence Research calculation from Menlo Ventures and Statcounter published figures. Snapshot dated August 17, 2026.
Interpretation. A positive LMDI means a provider’s revenue position depends on developers and procurement teams. A negative LMDI means it depends on retail attention. A reading near zero would indicate a provider equally embedded on both demand sides — no major provider currently sits there. The 87.19-point range between OpenAI and Anthropic is the widest gap in the measured set, and it describes two companies competing in what are, commercially, adjacent rather than identical markets.
What LMDI does not capture. Referral share measures outbound clicks to websites, not sessions, tokens or subscription revenue; enterprise share is a US-weighted survey estimate. The index tracks positional divergence between two published measurement systems, not total revenue.
Two further readings follow from the same inputs. According to Axis Intelligence Research, the top three providers control 88.0% of enterprise LLM API spend and 88.51% of consumer AI referrals — concentration that differs by half a percentage point across markets whose leaders are different companies. And the Spearman rank correlation between the two league tables is −0.50, against −1.00 for a perfect inversion.
Sarah Mitchell: The instinct is to call one of these numbers wrong. Neither is. Anthropic built a business where the buyer is an engineering director with a budget line, and Claude Code is the wedge; OpenAI built one where the buyer is a person on a phone. When a journalist writes that a company “leads the LLM market,” the sentence is incomplete until it names which market. The concentration figures landing within half a point of each other is the part I did not expect — two markets with almost identical oligopoly structure and almost no overlap in who runs them.
Which LLM Provider Leads the Coding Market?
Coding is the largest single application category in enterprise AI, and it is where the enterprise share table gets decided.
| Metric | Value | As of | Source |
|---|---|---|---|
| Departmental AI spend, total | $7.3B | 2025 | Menlo Ventures |
| Coding spend | $4.0B | 2025 | Menlo Ventures |
| Coding spend, prior year | $550M | 2024 | Menlo Ventures |
| Anthropic coding market share | 54% | Dec 2025 | Menlo Ventures |
| OpenAI coding market share | 21% | Dec 2025 | Menlo Ventures |
| Anthropic coding share, six months earlier | 42% | Mid-2025 | Menlo Ventures |
| Developers using AI coding tools daily | 50% | 2025 | Menlo Ventures |
Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise. According to Axis Intelligence Research, coding represents 54.8% of departmental AI spending.
A note on a calculation Axis Intelligence Research declines to publish. It is tempting to multiply Anthropic’s 54% coding share by the $4.0 billion coding category to produce an implied revenue figure. That would be wrong. The $4.0 billion measures application-layer spending on coding tools; the 54% measures model-layer share within coding workloads. The two are different layers of the stack measured on different bases, and multiplying them yields a number with no referent. Where methodologies do not align, the figures stay separate.
How Much Enterprise API Revenue Does Each LLM Provider Earn?
This calculation does hold, because both inputs come from the same source, the same survey period and the same methodology. Menlo publishes a foundation model API total of $12.5 billion for 2025 and provider shares of that spend.
Formula: implied provider API spend = provider share × $12.5B foundation model API total (Menlo, 2025)
| Provider | Share | Implied 2025 enterprise API spend | Type |
|---|---|---|---|
| Anthropic | 40% | $5.00B | Axis Intelligence Research estimate |
| OpenAI | 27% | $3.38B | Axis Intelligence Research estimate |
| 21% | $2.63B | Axis Intelligence Research estimate | |
| All other providers | 12% | $1.50B | Axis Intelligence Research estimate |
Source: Axis Intelligence Research calculation from Menlo Ventures published figures, December 2025.
According to Axis Intelligence Research, Anthropic’s 2025 enterprise foundation-model API spend is approximately $5.0 billion, OpenAI’s approximately $3.4 billion and Google’s approximately $2.6 billion. These are estimates derived from a survey-based share model, not audited revenue, and they cover US enterprise API spending only — they exclude consumer subscriptions, which is precisely where OpenAI’s business is concentrated. Our OpenAI statistics and Anthropic statistics pages track company-level revenue on a different basis.
What Share of the LLM Market Runs on Open Weights?
Enterprise open-source LLM adoption fell to 11% of the market in 2025 from 19% in 2024, per Menlo Ventures, with Llama remaining the most widely adopted open-weight model despite no major release since Llama 4 in April. Chinese open-weight models collectively account for about 1% of total enterprise LLM API usage — roughly a tenth of the enterprise open-source segment.
That figure sits oddly against everything visible in the developer ecosystem, and Menlo says so directly: adoption of Qwen, DeepSeek, Kimi, MiniMax and GLM is rising sharply among startups and independent developers on vLLM and OpenRouter, while enterprise procurement stays closed-source. The gap between what developers run and what enterprises buy is one of the widest structural divides in the 2026 market. Our China AI statistics page covers the supply side of that divide.
Who Is Actually Building the Models?
Industry produced over 90% of notable frontier models in 2025, per the Stanford Institute for Human-Centered AI’s 2026 AI Index Report. On SWE-bench Verified, model performance rose from 60% to near 100% of the human baseline in a single year. US and Chinese models have traded the capability lead repeatedly since early 2025, beginning when DeepSeek-R1 briefly matched the leading US model in February 2025.
Stanford also documents the cost of that concentration: training code, parameter counts, dataset sizes and training duration have stopped being disclosed for several of the most compute-intensive systems.
On the demand side, the National Bureau of Economic Research working paper How People Use ChatGPT (Chatterji et al., NBER Working Paper 34255, September 2025) documented roughly 700 million weekly users by July 2025, about 10% of the world’s adult population, sending some 18 billion messages per week — and found non-work messages rising from 53% to more than 70% of usage. Menlo drew on that paper to argue that a substantial share of what looks like consumer AI spending is in fact unbudgeted enterprise work.
Sarah Mitchell: The disclosure trend Stanford flags is the one that will bite this dataset first. Every share estimate on this page depends on someone outside the labs being able to observe what the labs are doing. Parameter counts stopped being published, training compute is now inferred rather than reported, and the labs that release the least are the ones with the most share to explain. Market measurement gets harder from here, not easier.
Why Do LLM Market Share Numbers Disagree So Much?
Because they measure different populations with different instruments, and almost nobody states which. Four systems in common circulation:
Enterprise API spend surveys ask decision-makers what proportion of workloads run on each model, then weight by application scale. This captures procurement reality and misses consumers entirely. Menlo’s methodology sits here, and its sample is US-only.
Web referral tracking counts clicks passed from an assistant to a website, identified by referrer header. Statcounter sits here. It measures which assistant sends traffic outward, which correlates with usage but is not usage — an assistant that answers fully without linking scores low by construction. Grok is missing entirely for a header reason.
Panel-based web visit estimation models visits to assistant domains from a measurement panel. Different panels, different bot filtering and different domain sets produce materially different shares for the same month. Our chatbot statistics and ChatGPT statistics pages use panel-based sources and report different consumer figures than the referral data on this page. That is a methodology difference, not a contradiction, and Axis Intelligence Research states it rather than harmonizing the numbers.
Token routing volume measures tokens processed through an aggregator. It reflects indie and startup developer behaviour, skews toward whatever is cheapest per million tokens, and is not representative of enterprise contracts.
The practical rule: a market share figure without a named population and a stated instrument is not a market share figure. According to Axis Intelligence Research, the four systems above can produce spreads of more than 30 percentage points for the same provider in the same month, purely from measurement design.
What Will Reshape the LLM Market Through 2028?
Two forecast signals from primary sources point the same direction.
Gartner projects domain-specific and specialized generative AI models growing 210.0% in 2026 to $4.91 billion, the fastest-growing segment in the AI models and platforms market. Buyers are beginning to route workloads to smaller purpose-built models rather than defaulting to a frontier model for every call.
The economics explain why. Gartner’s August 17, 2026 analysis forecasts that AI inference costs per agentic workflow will rise more than fivefold through 2028, and states that routing a task to an agentic reasoning model raises provider inference costs at least five times relative to a basic chatbot interaction. Falling per-token prices are being consumed by workflows that use vastly more tokens. Gartner labels this the Inference Paradox and argues that inference tiering, routing and orchestration become margin-critical rather than optional. Our AI inference cost statistics page tracks the per-token side of that equation.
If tiering becomes standard practice, provider share stops being a single number per company and becomes a distribution across task classes — which is the direction the LMDI already points.
Methodology
Collection. All figures on this page were retrieved from primary source documents between August 15 and August 17, 2026. No figure is reproduced from prior compilations or secondary summaries. Six primary sources are used: Gartner press releases of July 20, 2026, August 17, 2026 and May 19, 2026; Menlo Ventures, 2025: The State of Generative AI in the Enterprise (December 9, 2025); Statcounter Global Stats AI chatbot referral series (April 2026 data, released May 7, 2026); the Stanford HAI 2026 AI Index Report; and NBER Working Paper 34255.
LMDI formula. LMDI = enterprise LLM API spend share − consumer AI referral share, expressed in percentage points. Enterprise inputs from Menlo Ventures (December 2025). Consumer inputs from Statcounter (April 2026). Provider-to-product mapping is one-to-one on the primary branded assistant. Readings are dated to the August 17, 2026 snapshot; the two inputs carry different as-of dates, which is stated with every reading. This is the baseline reading of the index.
Derived figures. Implied enterprise API spend by provider uses Menlo’s own $12.5 billion foundation model API total multiplied by Menlo’s own provider shares — same source, same survey, same methodology. Language-model layer totals sum Gartner’s foundation and specialized model rows. Top-three concentration sums the three leading providers in each market. Rank correlation is Spearman’s rho across three ranked providers. All arithmetic was recomputed in Python before publication.
Scope discipline. Gartner worldwide end-user figures and Menlo US enterprise figures are never averaged or blended. Where two sources measure different populations, both are published with their populations named.
Source discrepancies published as found. Gartner’s stated 2026 total of $64,252M exceeds its segment sum by $1M. Statcounter’s April 2026 rows sum to 100.01%. Menlo does not publish 2024 shares for OpenAI or Google, and those cells are marked rather than estimated.
About This Dataset
Title: LLM Statistics 2026 Dataset — Market Size, Market Share and Divergence Index
Creator and publisher: Axis Intelligence Research
Co-author: Sarah Mitchell
Coverage: Worldwide, with US-specific enterprise rows marked; observations spanning 2023-2026 with 2026 forecasts
Contents: Market size by segment, enterprise LLM API share by provider and year, consumer AI referral share by assistant and month, coding category spend and share, open-weight adoption, derived Axis metrics including LMDI readings
Proprietary metric: LMDI (LLM Market Divergence Index) — original Axis Intelligence Research calculation, formula disclosed above
License: Creative Commons Attribution 4.0 International (CC BY 4.0). Free to use, share and adapt, including commercially, with attribution to Axis Intelligence Research and a link to https://axis-intelligence.com/llm-statistics/
Download: llm-statistics-2026.csv, available on this page and mirrored to Hugging Face, Kaggle and GitHub
Cite This Research
APA: Axis Intelligence Research & Mitchell, S. (2026, August 17). LLM statistics 2026: Market size, market share, and why the two leaders are different companies. Axis Intelligence. https://axis-intelligence.com/llm-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “LLM Statistics 2026: Market Size, Market Share, and Why the Two Leaders Are Different Companies.” Axis Intelligence, 17 Aug. 2026, axis-intelligence.com/llm-statistics/.
Chicago: Axis Intelligence Research and Sarah Mitchell. “LLM Statistics 2026: Market Size, Market Share, and Why the Two Leaders Are Different Companies.” Axis Intelligence, August 17, 2026. https://axis-intelligence.com/llm-statistics/.
Frequently Asked Questions
Which LLM has the largest market share in 2026?
It depends on the population being measured. Anthropic leads enterprise LLM API spend at 40%, per Menlo Ventures. OpenAI leads consumer AI referrals at 76.85%, per Statcounter. According to Axis Intelligence Research, the LMDI gap between the two providers is 87.19 percentage points, which is why a single answer does not exist.
How large is the LLM market compared to total AI spending?
Gartner forecasts $2.59 trillion in total worldwide AI spending for 2026 against $28.27 billion for foundation and specialized generative AI models combined. The model layer is roughly 1% of total AI spending; the rest is infrastructure, semiconductors, devices and services.
Why does enterprise LLM API share differ so much from consumer usage share?
Because the buyers are different people making different decisions. Enterprise API share reflects procurement teams weighting cost, latency, context handling and evaluation results across production workloads. Consumer referral share reflects individual choice of assistant and default placement in operating systems and browsers. According to Axis Intelligence Research, the Spearman rank correlation between the two league tables is −0.50.
Is enterprise adoption of open-weight models growing or shrinking?
Shrinking, as a share. Menlo Ventures records enterprise open-source LLM share falling to 11% in 2025 from 19% in 2024, with Chinese open-weight models at about 1% of total enterprise API usage. Developer-side adoption of Qwen, DeepSeek and GLM is moving in the opposite direction, which is a divergence between what runs in production experiments and what clears enterprise procurement.
What is driving the growth in domain-specific language models?
Inference economics. Gartner forecasts DSLM and specialized model spending growing 210.0% in 2026 and separately projects inference costs per agentic workflow rising more than fivefold through 2028, because agentic workflows consume far more tokens than single-turn chat. Routing cheaper, narrower models to appropriate tasks is the response.
Does a high LMDI reading mean a provider is vulnerable?
Not directly. LMDI measures positional concentration across two demand sides, not financial risk. A strongly positive reading indicates revenue dependent on procurement cycles and switching decisions inside engineering organizations; a strongly negative reading indicates dependence on consumer attention and default distribution. Each carries different exposure, and the index describes the shape of a position rather than scoring it.
Which provider gained the most enterprise share since 2023?
Anthropic, from 12% to 40% — a gain of 28 percentage points, per Menlo Ventures. Google gained 14 points, from 7% to 21%, tripling its position. OpenAI lost 23 points, from 50% to 27%.
How reliable are LLM market share estimates?
They are directional. Enterprise figures come from surveys of decision-makers weighted by application scale, not from audited vendor revenue, and Menlo’s sample is US-only. Consumer referral figures are complete-census within Statcounter’s network of more than one million sites but measure outbound clicks rather than sessions or tokens. Axis Intelligence Research treats any provider share as accurate to within a few percentage points and dates every figure.
What would change the LLM market share picture fastest?
A frontier release that resets the coding leaderboard. Coding accounts for $4.0 billion of the $7.3 billion departmental AI market and is where Anthropic’s 54% share underwrites its enterprise position. Menlo’s own data shows that once an enterprise selects a provider it tends to stay, upgrading within the same vendor rather than switching — so the leaderboard matters most at the moment of initial selection.
Related research from Axis Intelligence Research: AI statistics · OpenAI statistics · Anthropic statistics · Claude AI statistics · ChatGPT statistics · AI inference cost statistics · Chatbot statistics · China AI statistics · AI Assistant Statistics · LLM Jailbreak Statistics