Companies Are Spending 5x More on AI Now

Context

Mainstream coverage of enterprise AI this week was about restriction, such as Apple cutting how much its engineers can spend on AI per day. NLW argues the real picture is the opposite: business AI spend is not just growing, its growth rate is speeding up. He walks through the 10 most interesting numbers from Menlo Ventures' fourth annual State of Enterprise AI report, based on a mid-September Morning Consult survey of about 400 US enterprise decision makers who control roughly $4 billion in AI spend. He flags that Menlo is an investor in Anthropic and many other AI companies, so it is not a neutral observer. For PMs, the report is a map of how buyers actually choose models, vendors and budgets, which shapes pricing, positioning and build-versus-buy decisions for any AI product.

The Big Idea

Agents changed AI from something companies buy per seat into something they buy per token, so spending now scales with the amount of work done rather than the number of people.

That shift is why enterprises overshot their own AI budgets by about 2.5x this year, and why the money flowed to frontier models and AI-native vendors in the categories where agents do new kinds of work.

Key Insights

1. AI spend growth is accelerating

Menlo's estimates of annual enterprise AI spend show growth that slowed and then sped up again:

• 2023: $1.7 billion.
• 2024: $8.5 billion, a 5x jump.
• 2025: $33 billion, a slower 3.9x jump, which is what you would expect as a market matures.
• 2026: almost $177 billion, a jump bigger than even the 2023 to 2024 multiple.

NLW's explanation is agents. From the release of Claude Opus 4.5 in November 2025 onward, AI spend "stopped being about seats and started being about tokens," which reset expectations of how big the market can get.

2. Depth of use grew, not breadth

• The share of enterprises using AI rose only from 55% to 65% between 2025 and 2026.
• That still leaves 35% of enterprises reporting no AI use, which NLW reads as a sign of how much inertia the business world has.
• Average AI spend per enterprise that does use AI rose 4.7x over 2025.
• The reason, again, is agents: they work independently, run in parallel and keep working after people step away, so spend scales with work, not headcount.

3. Companies overshot budgets by 2.5x

The average company planned to grow AI spend about 1.9x from 2025 to 2026. A company that spent $10 million in 2025 expected to spend $19 million in 2026. Most firms built those plans on per-seat adoption across more teams. Actual spend grew 4.7x, about 2.5 times more growth than budgeted. NLW stresses this is not necessarily out-of-control spend. It is people discovering what they can do with raw intelligence. The swings between "token maxing" and token restrictions inside companies are a symptom of planning for a per-seat world and getting a per-token one.

4. Buyers chose capability over cost

Despite all the talk about cost efficiency, enterprise buyers moved to more expensive models:

• A year ago the most popular models were Claude Sonnet 4.5 ($3 input, $15 output per million tokens) and GPT-5 ($1.25 input, $10 output).
• Today the most popular are Claude Opus 5.5 ($4 input, $20 output) and GPT-6 Astra ($10 input, $60 output).
• Frontier models' share of spend rose from 35% to 63%, while mid-tier, small and older models all lost share.
• This happened even though labs cut small-model prices by 80% or more.

NLW does not think this kills the cost story. His view is that each frontier jump (Opus 4.5, and now Opus 5.5) unlocks use cases that do not translate easily to other models, so early adopters concentrate on the state of the art. As adoption and the stack mature, he expects cheaper, more efficient models to fill in a more complex stack.

5. Token share tells another story

Open models look small by spend but larger by usage:

• Open source fell from 6% to 4% of enterprise spend between 2025 and 2026.
• Over the same period its share of tokens rose from 11% to 19%.
• Kimi accounted for 61% of open-source spending this year.
• Near-frontier models such as Grok 4.7, Kimi K3, DeepSeek V4 and GLM 5.3 together still take under 5% of API spend.

NLW argues token share matters more over the medium to long term. One common Wall Street view is that open models may end up handling most tokens while closed frontier models keep most of the spend on the highest-value use cases.

6. A two-horse race for enterprise

Enterprise API spend share by provider, per Menlo:

• Anthropic: 12% in 2023, 24% in 2024, 40% in 2025, and 55% in 2026. NLW credits the Opus models becoming the default for long-horizon agentic coding.
• OpenAI: 50% in 2023, 34% in 2024, 27% in 2025 (the first year Anthropic passed it), then a recovery to 31% in 2026, driven by the GPT-5.6 series, GPT-6 and a heavy enterprise focus.
• Google: down from 21% to 9%, as companies followed the best model capabilities. NLW says Gemini 4 Carbon looks good early but needs more battle testing.

7. Lab coding agents exploded

• Revenue for the labs' own coding agents, Codex and Claude Code, rose from $1 billion at the end of 2025 to $24 billion at the end of 2026, a 24x increase.
• Other coding platforms such as Cursor, Cognition, Factory and Lovable still grew almost 3x, from $2.6 billion to $7.8 billion.
• So the labs' growth did not come at the cost of independent coding tools. The labs' apps were simply far bigger.

8. Verifiable domains grew fastest

Growth in AI app spending by category:

• Financial services: 8x (but only about $800 million total).
• Agentic coding: 7.7x (about $34 billion total).
• Customer success: 4.3x.
• Sales: 3.6x.
• Legal: 2.8x.
• Marketing: 1.8x.

NLW's reading is that domains where answers can be checked, where agentic-coding-style power applies even outside app building, benefited most in 2026.

9. New capabilities favor AI natives

Whether AI-native vendors (like Anthropic and OpenAI) or incumbents (like ServiceNow) win depends on whether the job is new:

• AI natives take 96% of coding app spend, 95% of finance and operations, and 81% of design.
• Sales is close to even, at 54% AI native and 46% incumbent.
• Incumbents hold over 50% in marketing, customer success, HR, data science and IT. In IT, only 9% goes to AI natives.

Spend tends to flow to tools already embedded in, or acting as, the company's system of record. Agentic coding is a net-new capability with no plugged-in incumbent, so new vendors had room to enter.

10. AI money is mostly new

The early-2026 "SaaS apocalypse" assumed that letting companies build their own apps would gut SaaS. Menlo found something milder. SaaS application market growth slowed from 13% in 2025 to 11% in 2026. Of the AI spend Menlo could trace, only 21% came from existing software budgets and 55% came from net-new budgets. NLW cautions that the next threat to SaaS is barely starting: agents using software directly and "headless" software with no human interface.

Mental Models & Frameworks

Seats versus tokens

• Per-seat model: spend grows with the number of people given access, so budgets are planned by headcount and rollout scope.
• Per-token model: spend grows with how much work AI does, so a single team running agents in parallel can multiply cost without adding a single user.
• When to use it: any time you price an AI product, forecast its cost or explain a usage spike. NLW says AI budgeting looks more like cloud compute budgeting, but even that analogy breaks down, so it needs its own approach.

Concentration: risk or leading indicator

When a small number of customers drive most AI spend, there are two readings:

• Bearish: if one big spender pulls back, the outsized hit can drag the whole market down.
• Bullish: the heaviest users are not a natural ceiling but a preview of where everyone else will go, so their spend can be projected across all companies that could spend that much.

Menlo takes the bullish view and reads the 4.7x rise in per-enterprise spend as a sign the long tail has barely started.

Spend share versus token share

Spend share shows where the money goes today. Token share shows where the work goes. Open-source models losing spend share while gaining token share is a sign of a market where cheap models carry volume and expensive models carry the highest-value tasks. Watching both numbers avoids misreading which models are actually winning.

Net new versus incumbent work

A quick test for whether an AI-native product can win a category: is the job net new, or does it sit on an existing system of record? Net-new jobs like agentic coding go mostly to AI natives (96% of coding spend). Jobs tied to long-standing vendors, like IT (91% incumbent), mostly stay with incumbents adding AI.

Trade-offs & Nuance

Frontier spend versus cost efficiency

Buyers paying more for frontier models and labs cutting small-model prices by 80% are both true at once. NLW frames it as a phase: early adoption concentrates on the state of the art because new capabilities only exist there. Later, as the stack matures, cheaper models take on more of the work. A PM who designs only for the cheapest model may miss today's demand, and one who designs only for the frontier may miss where volume is heading.

Survey source and bias

The data comes from Menlo Ventures, which is heavily invested in Anthropic and other AI companies. NLW still finds it useful because it directionally matches broader sentiment about business AI spend. Treat the exact numbers with care, but the direction as credible.

Narratives run ahead of reality

NLW admits that the conversation in heavily AI-focused circles can get ahead of what the average business is doing. A third of enterprises still report no AI use, and incumbents still win most categories. What feels obvious inside the AI bubble may not yet describe your customer.

Common Mistakes

Mistake: Budgeting AI by seat count

• Why it happens: most firms planned 2026 AI spend by extending per-seat adoption to more parts of the company.
• Why it hurts: agent-driven usage grew spend 4.7x against a planned 1.9x, which caused whiplash between encouraging heavy token use and then restricting it.
• Better approach: forecast AI cost from expected work and token volume, and treat it as its own category rather than forcing it into a past bucket.

Mistake: Assuming AI just eats SaaS

The early-2026 SaaS selloff assumed AI spend would replace software budgets. Menlo found 55% of traceable AI spend came from net-new budgets versus 21% from existing software budgets. Incumbents that own the system of record still win most non-coding categories. The more real long-term risk NLW points to is agents operating software directly, not companies rebuilding every tool themselves.

Practical Application

Price AI features on work

If your AI product bills per seat, check whether heavy users running agents cost you far more than light users. Model a usage-based or token-linked component so revenue rises with the work your product actually does, the way enterprise AI spend now scales with work rather than headcount.

Design for a mixed model stack

• Do: build the highest-value workflows on the best available frontier model, since enterprise buyers are choosing capability over cost.
• Then: route high-volume, lower-stakes steps to cheaper or open models, which already handle 19% of enterprise tokens.
• Why it works: it matches the direction NLW expects, with frontier models carrying hard tasks and efficient models carrying volume.

Test whether the job is new

Before positioning an AI-native product against an incumbent, ask whether the job is new work or sits on an existing system of record. In categories like IT, where incumbents take 91% of AI app spend, plan to integrate with the incumbent rather than replace it. In net-new categories like agentic coding, there is room to own the workflow outright.

Track your heaviest AI users

Identify the customers or internal teams with the highest token use and study what they do. Under the bullish reading of concentration, their behavior shows where the rest of your users are heading, so their workflows are a roadmap input, not just a cost line.

Questions to Consider

• If our AI product's cost scales with tokens but our price scales with seats, which customer segment is quietly unprofitable?
• Are we choosing a cheaper model for a workflow where our buyers would actually pay more for frontier-model capability?
• Is the job our AI product does net new, or does it compete with a system of record that customers have used for years?
• What would our usage data look like if we tracked share of tokens by model, not just share of spend?
• Is our roadmap built on what AI-focused circles talk about, or on what our average customer, who may not use AI at all yet, actually does?

Bottom Line

Agents moved enterprise AI from a per-seat purchase to a per-token one, so spend now follows the amount of work AI does. That is why companies overshot their AI budgets, why buyers paid up for frontier models, and why AI natives win where the work is new while incumbents hold the systems of record.

Case Studies Mentioned

Anthropic's climb to 55% share

Anthropic's share of enterprise API spend went from 12% in 2023 to 55% in 2026, overtaking OpenAI for the first time in 2025. NLW ties it to the Opus models becoming the default for long-horizon agentic coding from Opus 4.5 onward. The lesson is that in a capability-driven market, owning the best answer for one fast-growing workload can reshape overall share.

Google's slide to 9%

Google fell from 21% of enterprise spend in 2025 to 9% in 2026. NLW says Gemini 3 Pro was at the frontier in November 2025, just before Opus 4.5 arrived and agentic coding took over. Since then Google has not been in a real three-way race at the top. Enterprise buyers moved quickly to wherever the best capability was.

Claude for personal finance

In the headlines, Anthropic relaunched Claude for personal finance with read-only bank connections through Plaid. Claude will not give investment advice and instead helps with budgeting by analyzing spending. Anthropic framed it as a first step, saying bigger money decisions are best served by Claude working alongside licensed professionals and regulated firms. It is a clear example of scoping an AI product to reduce trust and regulatory risk in a sensitive domain.

Tools & Products

Resources Mentioned

Notable Quotes

"Spending then scales with work, not with headcount." (NLW)

"We planned for one paradigm, a per-seat paradigm, and instead got a per-token paradigm which needs a totally different type of strategy." (NLW)

"Buyers still, despite all of the narrative about cost and cost efficiency, are choosing capability over cost." (NLW)

"Perhaps our narratives and storytelling are getting out ahead of where businesses actually are." (NLW)