Use AI spending figures to understand the supplier market, then budget for your own workflow. Compare definitions and include the people and ongoing work behind the software.
What you'll learn
- Keep publisher, year and market scope attached.
- Distinguish spending totals from deals and deployment shares.
- Decide from a defined task and measured trial.
Gartner’s September forecast puts worldwide AI spending at about $2.7 trillion in 2026. IDC and Menlo track different markets and periods. Read each figure with its definition: infrastructure spending, an enterprise software estimate and the share of deals reaching production answer different questions.
Gartner’s September 2026 forecast is about $2.7 trillion in worldwide AI spending for the year, up 49.5%. It covers a much wider market than Menlo Ventures’ enterprise generative-AI estimate. A difference between those totals is not a measure of money that failed to reach production.[1][3]
For a business deciding what to buy, the useful distinction is between the market being counted and the work a proposed purchase will do. Keep the source, period, scope and unit together before comparing headline figures.
What does the global AI spending forecast cover?
Gartner’s September forecast covers worldwide AI spending across infrastructure, software, services and other named categories. Its table gives $2.67046 trillion for 2026, rounded to $2.7 trillion in the headline. The current model separates agents and assistants from AI software and adds consumer agents and assistants.[1]
The categories matter when comparing reports. Gartner’s agents-and-assistants line is $29.219 billion, while generative AI models are $28.266 billion. These are separate forecast rows; neither should be labelled all enterprise agent spending or treated as a completed-year result.[1]
IDC’s September 2026 analyst summary, based on its Worldwide AI and Generative AI Spending Guide 2026 V2, forecasts global AI IT investment from $695.4 billion in 2025 to $3.12 trillion in 2030. That is a different definition and horizon, not a competing estimate for Gartner’s 2026 endpoint.[2]
| Figure | Source and period | What it measures |
|---|---|---|
| About $2.7 trillion | Gartner, September 2026 forecast for 2026[1] | Worldwide AI spending across its defined markets |
| $29.219 billion / $28.266 billion | Same Gartner forecast[1] | Separate agents-and-assistants / generative-model rows |
| $3.12 trillion | IDC, September 2026 summary; forecast for 2030[2] | Global AI IT investment under the Spending Guide definition |
| $89.7 billion / $497 billion | IDC, July 2026 Infrastructure Tracker[4] | Observed Q1 spending / full-year 2026 forecast in that infrastructure market |
| $37 billion | Menlo Ventures, 2025 estimate[3] | Enterprise generative AI, including applications and infrastructure |
| $12.5 billion | Same Menlo estimate[3] | Foundation-model access within its infrastructure category |
Menlo excludes chips, inference/model-serving providers and AI features embedded in existing software from its market estimate. Its $37 billion therefore cannot be divided by Gartner’s global total to calculate a production success rate. The excluded categories are a scope choice, not evidence of failed investment.[3]
The earlier Gartner GenAI forecast of $644 billion concerned 2025 and included servers and devices. It is another distinct measure, rather than a software-and-services-only total. Keep the publication and market definition attached when it appears in an older presentation.[6]
Do not subtract one publisher’s enterprise software estimate from another’s global AI forecast and call the remainder waste. Compare the same market, year and unit before drawing a conclusion about investment or operational use.
Where did the $37 billion in enterprise GenAI go?
Menlo’s December 2025 report estimates $19 billion for applications and $18 billion for infrastructure, out of a $37 billion total. The total rises from a restated $11.5 billion in 2024, about 3.2 times as much. It combines a survey with a market-sizing model. These are estimated dollars across the included market, not a count of successful workflows or applications alone.[3]
Applications. The report distinguishes horizontal tools used across functions ($8.4 billion), departmental tools ($7.3 billion) and industry-specific tools ($3.5 billion). Coding’s $4 billion is inside departmental AI; healthcare’s $1.5 billion is inside industry-specific AI. Do not add those subcategories again to the application total.[3]
Infrastructure. Menlo allocates $12.5 billion to foundation-model access, $4 billion to training infrastructure and $1.5 billion to data-related infrastructure. This model’s infrastructure category differs from IDC’s server/storage tracker and Gartner’s global infrastructure definition.[3]
The approximate application/infrastructure shares are 51% and 49%, calculated from the rounded totals. The application subcategories total $19.2 billion because the published figures are rounded. Use them at the precision supplied, rather than claiming an exact reconciliation to the headline.[3]
Menlo estimates that 76% of enterprise AI use cases were purchased rather than built internally, compared with 53% in 2024. “Purchased” covers ready-made solutions; it is not a measure of work outsourced to implementation partners. A business still needs to decide who owns configuration, checking and operation after buying a tool.[3]
For your budget, separate purchased software, custom work and the people maintaining the process. Market totals can explain the supplier landscape, but they do not supply the cost breakdown of your own proposal. Ask for a scope that makes those responsibilities visible.
How did enterprise model market share change?
Menlo’s 2025 estimate gives Anthropic 40%, OpenAI 27% and Google 21% of enterprise model spending. Its method approximates dollars from weighted production model usage and available financial information. This is not a share of all enterprise AI spending, a company count or a current 2026 leaderboard.[3]
The comparisons span different years: Anthropic rises from 24% in 2024, while the cited OpenAI and Google baselines are 50% and 7% in 2023. Keep those years attached rather than describing every change as occurring in the same single year.[3]
For coding, Menlo estimates Anthropic at 54% and OpenAI at 21%. That is a category-specific share in the report, not a guarantee that the same provider will perform best on your documents, messages or approvals.[3]
Foundation-model access represents $12.5 billion of the $37 billion model, or roughly 34% by calculation. That share describes Menlo’s included spending; it does not mean every dollar purchases only inference or that the same fraction belongs in your budget.[3]
Price comparisons need a capability and period too. Stanford’s 2025 AI Index reports that the query cost at a GPT-3.5-equivalent benchmark level fell from $20 to $0.07 per million tokens between November 2022 and October 2024. That historical comparison is not a promise of falling prices for every provider or a proportional reduction in your whole operating bill.[5]
When assessing a supplier, ask for the actual model choice, expected usage, included review and the method for changing providers. Read what an AI agent is for the capability distinction before treating a provider’s market share as a workflow recommendation.
What does the budget-to-production conversion data show?
Menlo’s 2025 survey reports that 47% of AI deals reach production, compared with 25% for traditional SaaS deals. The denominator is deals once an organisation commits to exploring a solution. It is not the share of AI spending that produces a return, nor a success rate for all attempted projects.[3]
The method covers 495 US enterprise AI decision-makers at companies actively using AI tools. That selection matters when using its figures to plan for a small service business, a team new to AI or a country with different operating conditions.[3]
The report’s 16% enterprise and 27% startup figures concern deployments classified as true agents. They are not the percentage of enterprises or startups with at least one agent. The distinction is about architecture: planning and acting with feedback, rather than a fixed sequence around a model call.[3]
A spending total, a deal conversion rate and a deployment category need different denominators.
Menlo also estimates that 27% of AI application spend comes through product-led growth, compared with 7% in traditional software. Including its allowance for informal adoption, it says that buying motion may approach 40% of application spend. It does not say 40% of total AI usage is hidden from IT.[3]
For a local trial, record the starting scope and acceptance decision directly. If a case never enters production, note why; if it does, measure ongoing cost and completed work. A launch and a financial return are separate milestones, and both need a clear owner.
What does the spending curve mean for a business deciding now?
Use the spending figures to understand the supplier market, then decide from your workflow and budget. A rising global forecast does not set a deadline for your project, and a falling benchmark query price does not remove review, maintenance or connection costs.
Define the work. Name the input, required output and permitted action. For example, preparing a reply to an enquiry is a different purchase from authorising the agent to send it or change the customer record. Price the version of the task you are actually considering.
Compare ongoing responsibilities. Ask what the quoted scope includes after launch: corrections, changes to the connected systems, incident handling and a route back to people. A low usage charge can sit inside a much larger operating commitment.
Use a local decision point. Agree what a trial must show before expanding. Include staff time, completed outcomes and exception handling. If the task is not ready because records or permissions are unclear, resolve those dependencies before using a market-growth forecast to justify a launch.
The practical timing question is whether you have a definable problem and a way to judge the result. Say a team wants fewer unanswered enquiries: compare its current follow-up with a controlled draft-and-review trial, including missed cases and corrections. That is a proposed test, not a claim about a guaranteed return window.
Read AI agent ROI statistics for the distinction between measured task gains and financial returns. Start with what an AI agent is if you still need to distinguish an assistant, a fixed workflow and permission to act.
Frequently asked questions
How much is spent on AI globally in 2026? Gartner’s September 2026 forecast rounds worldwide AI spending to $2.7 trillion, up 49.5%. Its table lists agents and assistants separately from generative models. IDC’s broader spending-guide horizon and Menlo’s enterprise generative-AI estimate use different scopes and should be compared with their definitions attached.[1][2][3]
What percentage of AI spending reaches production? Menlo’s 47% figure is a deal-to-production conversion rate, compared with 25% for SaaS, not a percentage of dollars. Its 16% figure counts enterprise deployments classified as true agents, not the share of enterprises with agents. Neither supplies a universal return rate.[3]
How is enterprise AI spending allocated? Menlo estimates $37 billion in 2025: about $19 billion applications and $18 billion infrastructure. Horizontal tools, departmental tools and industry-specific tools are application categories. Coding is nested in departmental AI. The infrastructure total includes $12.5 billion of foundation-model access.[3]
Which AI models have the largest enterprise market share? Menlo’s 2025 estimate gives Anthropic 40%, OpenAI 27% and Google 21%, approximating enterprise model spending from weighted production usage and financial information. In coding it estimates Anthropic at 54% and OpenAI at 21%. These are report-specific spending estimates, not universal adoption shares or quality rankings.[3]
Notes
- Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026. September 16, 2026 Accessed October 15, 2026.
- IDC Worldwide AI and Generative AI Spending Guide 2026 V2 — analyst summary. September 9, 2026 Accessed October 15, 2026.
- 2025: The State of Generative AI in the Enterprise. December 9, 2025; survey November 7–25, 2025 Accessed October 15, 2026.
- AI Infrastructure Spending Holds Near $90 Billion in Q1 2026; Forecast Raised to $497 Billion. July 21, 2026 Accessed October 15, 2026.
- AI Index 2025: State of AI in 10 Charts. April 2025 Accessed October 15, 2026.
- Gartner Forecasts Worldwide GenAI Spending to Reach $644 Billion in 2025. March 31, 2025 Accessed October 15, 2026.
- Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026. May 19, 2026; superseded by September forecast Accessed October 15, 2026.
