Source: raw/Companies_Are_Spending_5x_More_on_AI_Now.md (The AI Daily Brief, Nathaniel Whittemore, reading Menlo Ventures’ State of Enterprise AI report, https://www.youtube.com/watch?v=INC89yHLexM); Apple’s spending caps from raw/How_a_Data_Center_Slowdown_Benefits_Big_Tech.md (The AI Daily Brief headlines, https://www.youtube.com/watch?v=ssEWGHjKHTM); Amazon’s token memo and model strategy from raw/Ep._249_-_Anthropic_Files_for_IPO_AI_and_the_Meaning_of_Work_Microsoft_s_60K-Employee_Buyout_Offer.md (The Artificial Intelligence Show ep. 249, Paul Roetzer and Mike Kaput, recorded 2026-11-03, https://www.youtube.com/watch?v=_OODHaVlMH8). All fetched 2026-11-05 to 2026-11-09. Platform: YouTube.
Menlo Ventures’ fourth State of Enterprise AI report, read on air by NLW, says enterprise AI spending grew faster in 2026 than in 2025, on a larger base. The average company using AI now spends 4.7× what it did last year, about 2.5× what it budgeted. Over the same weeks, Apple and Amazon made news by restricting how their own engineers spend on AI. Both stories describe the same shift: AI spend moved from per-seat licences to per-token consumption, and budgets have not caught up. Every figure here is a podcast host reading a report. Menlo is an investor in Anthropic and other AI companies.
Key Takeaways
- Total spend: Menlo estimates enterprise AI spend at 8.5B (2024), 177B (2026). Growth was 5× and then 3.9×, and in 2026 it rose again, to about 5.4× (177 ÷ 33).
- Depth, not breadth. The share of enterprises using AI rose only from 55% to 65%, but average spend per using enterprise rose 4.7×.
- Budgets missed by about 2.5×. Companies planned to grow AI spend about 1.9× and actually grew it 4.7×. NLW’s explanation: they budgeted per seat and got per-token consumption.
- Buyers chose capability over price. Frontier models’ share of spend rose from 35% to 63%, even as labs cut small-model prices by more than 80%.
- Vendor share of enterprise API spend: Anthropic 55% (up from 40% in 2025), OpenAI 31% (up from 27%), Google 9% (down from 21%). Grok 4.7, Kimi K3, DeepSeek V4 and GLM 5.3 together are under 5%.
- Open weights: 4% of spend, 19% of tokens. Spend share fell from 6%; token share rose from 11%. Kimi takes 61% of open-source spending.
- Coding is the biggest category. Revenue from the labs’ own coding agents (Claude Code, Codex) went from 24B at end-2026. Other coding tools (Cursor, Cognition, Factory, Lovable and others) went from 7.8B.
- Most AI spend is new money. Of the spend Menlo could trace, 55% came from net-new budgets and 21% from existing software budgets.
- The other side: caps. Apple cut engineers’ AI spending caps from 200 a day (Bloomberg). Amazon told engineers not to treat token usage as a target, and had already shut an internal AI-usage leaderboard in May (CNBC).
The Menlo survey
- Sample: about 400 US enterprise decision-makers, surveyed by Morning Consult in mid-September 2026. Together they are responsible for about $4B of AI spend. This is Menlo’s fourth annual edition.
- Bias: Menlo is “heavily invested in the AI field, including in Anthropic”. NLW says the data is directionally consistent with what he hears elsewhere.
Growth
| Year | Estimated enterprise AI spend | Growth on prior year |
|---|---|---|
| 2023 | $1.7B | — |
| 2024 | $8.5B | 5× |
| 2025 | $33B | 3.9× |
| 2026 | almost $177B | about 5.4× ^[inferred] |
- Average spend per enterprise: 4.7× the 2025 level. NLW: “Spending then scales with work, not with head count.”
- Planned vs actual: planned growth about 1.9× (a company that spent 19M), actual 4.7×.
- Still not using AI: 35% of enterprises say they do not use AI.
Which models get the money
- Most popular models, a year ago and now: Claude Sonnet 4.5 (15 per million input/output tokens) and GPT-5 (10) then; Claude Opus 5.5 (20) and GPT-6 Astra (60) now.
- Spend by tier: frontier 35% → 63%. Mid-tier, small and older models all lost share.
- NLW’s reading: since Opus 4.5 (November 2025), businesses have been catching up to a differentiated set of agentic capabilities, so early adoption concentrates on the frontier. He expects smaller, cheaper models to take a larger role as stacks mature.
| Enterprise API spend share | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|
| Anthropic | 12% | 24% | 40% | 55% |
| OpenAI | 50% | 34% | 27% | 31% |
| — | — | 21% | 9% |
- Menlo reads OpenAI’s rise as a second-half recovery driven by the GPT-5.6 series (July) and GPT-6 (September) and its enterprise focus. Google’s fall tracks Gemini losing frontier status after Gemini 3 Pro; NLW says Gemini 4 Carbon may change that.
- Open weights: 6% → 4% of spend, 11% → 19% of tokens. NLW thinks token share matters more over time.
Where the money goes
- Coding apps: about **1B → 2.6B → $7.8B.
- Category growth: financial services 8× (but only about $800M), agentic coding 7.7×, customer success 4.3×, sales 3.6×, legal 2.8×, marketing 1.8×.
- AI-native vendors vs incumbents (share of app spend going to AI-native companies): coding 96%, finance and operations 95%, design 81%, sales 54%. In marketing, customer success, HR, data science and IT, incumbents take more than half; in IT only 9% goes to AI-native vendors.
- SaaS: Menlo found SaaS application market growth slowing from 13% to 11%, but most AI spend comes from new budgets rather than replacing software (55% net new vs 21% from software budgets).
Spending caps and token targets
- Apple (Bloomberg’s Mark Gurman, via NLW): engineers’ AI spending caps were lowered last month from 200 a day; another cut to 100** limits, which could fall to 300 a day is more than 1,500 a month limit in June. Some Apple insiders said few engineers ever came close to $300; others read the cuts as a sign of wider cost-cutting. Apple’s earnings call the same week mentioned AI 45 times, and its CFO called AI features a reason for customers to upgrade.
- Peter Steinberger’s question (via NLW): “do individual limits still make sense?” If agents run loops on a project and humans curate them, budgets may belong to the project or team, not the person.
- Amazon’s memo (The Artificial Intelligence Show): a memo to software engineers and managers, circulated in September and reported in late October, says Amazon has used token usage, AI tool adoption, code changes and reviews as proxies for AI impact. It calls those proxies flawed and warns that treating them as targets creates the wrong incentives, because token use can rise without the work improving. CNBC reported that Amazon shut an internal AI-use leaderboard in May after employees ran up tokens to climb it. Roetzer’s opinion: “Token maxing is what happens when the technical people are the ones leading the AI strategy.”
- Amazon’s model bets. Business Insider reported in July that Amazon deprioritised several of its Nova models (Premier, Omni, and the Reel video and Canvas image models) and moved engineers and compute to fewer projects, including a new frontier model expected at re:Invent (November 30 to December 4). It also cut some AI jobs and closed a research lab. Roetzer’s takeaway: build so you can switch models, for example through Bedrock.
- Consumption is the business model. About 80% of Anthropic’s revenue is consumption-based, per Mike Kaput’s reading of the S-1 (see Anthropic’s Public S-1).
- Security budgets (a16z survey of 100 security chiefs at companies with more than 1,000 employees, via The Artificial Intelligence Show): 87% expect cybersecurity spending to rise in 2027, nearly half expect their teams to grow, and 45% plan to buy or upgrade tools within a year to secure their internal AI agents.
Added 2026-11-11: more Menlo figures, and how companies are cutting the bill
Source: The Artificial Intelligence Show ep. 250 (raw/Ep._250_-_Midterm_Voters_Reject_Data_Centers_AI_Speeds_Up_Anthropic_s_Research_Nvidia_s_Spy_Probe.md, https://www.youtube.com/watch?v=2XY3q56sD04), Mike Kaput reading Menlo’s report and three cost stories. Same report, second reader.
Figures not given by NLW:
- Agents are the biggest consumers of models: 80% of all tokens on OpenRouter, about 7× human-driven usage on that platform (one proxy, Kaput notes).
- The top three labs (Anthropic, OpenAI, Google) take 69% of deployment dollars, up from 50%, or 2.3× the rest of the market combined.
- Anthropic and OpenAI together take 86% of enterprise LLM API spend, up from 67%. This matches the vendor table above (55% + 31%; 40% + 27%).
- Coding and enterprise chatbots are 76% of AI application spend.
- Budget source: 55% net new, 22% redirected from software budgets (NLW read 21%).
- Little replacement yet. Menlo found limited evidence of companies replacing core systems such as Salesforce or Workday with AI. Kaput notes Menlo invests in software companies.
Three cost stories (Kaput):
- LinkedIn “flattened the cost” of its AI hiring assistant while usage grew, by replacing OpenAI models with smaller models it trained in-house (The Information).
- Zapier’s Wade Foster argues for using AI only where judgement is needed: “your agent builds the automation and Zapier runs it”, on a trigger with the same steps every time, and an agent steps back in only to fix a broken step. He claims costs can fall by up to 71%; he is selling the approach.
- HubSpot, after a weak earnings report, is considering a “headless” strategy in which customers use ChatGPT or Claude to work with their HubSpot data, and is testing pricing that replaces separate product subscriptions with user seats plus usage credits (for sending emails, resolving tickets, enriching records).
Roetzer’s view (opinion): he now asks ChatGPT, connected to HubSpot, for webinar registration counts and scheduled a daily report in about 30 seconds instead of opening HubSpot’s dashboards. He expects software vendors to meter AI access to their data, notes that a computer-use agent logging in as one user could do the work of many seats, and proposes an “AI value engineer” role to decide which steps run on an LLM and which on cheaper deterministic automation.
How this fits the other October data
- Consistent with Ramp: Ramp’s economist said open-weight models are “less than 5% of business spend”; Menlo says 4%. See AI Adoption Data, October 2026.
- Not the same measure as The Information’s survey, which found OpenAI back on top by reader usage. Menlo measures API spend share among large US enterprises.
- Concentration, read two ways. NLW lays out both: concentrated spend is a market risk if a few big buyers pull back, or a leading indicator of where the rest of the market will go. Menlo takes the bullish reading. The bear case is in AI Bubble Signals.
Try It
- Budget AI like metered infrastructure, not seats. The average company here overspent its plan by about 2.5×. Set a monthly token budget per project with alerts, and review it against output each month. ^[inferred] See Admin Spend Visibility and Control.
- Do not reward token volume. Amazon’s memo and leaderboard are the example. Measure cost per completed task or outcome instead. See Token Economics Primer.
- Put caps on teams or projects, not only individuals, if agents run long loops on shared work (Steinberger’s point).
- Split model choice by token share, not just spend share. Open weights carry 19% of tokens for 4% of spend: route high-volume, low-stakes work to them and keep the frontier model for work that needs it. ^[inferred] See Cost & Intelligence Levers.
- For marketing, customer success, HR and IT, check your incumbent vendor’s AI features before buying a new tool. That is where Menlo finds most spend going. In coding, AI-native tools dominate.
- Keep a model-switching path. Amazon dropped models customers had built on; use a gateway or a multi-model platform.
Open Questions
- The report itself was not fetched. All Menlo figures are NLW’s on-air reading; the 2023 and 2024 vendor shares for Google are not given.
- **“40B globally in 2026) is in AI Adoption Data.
- The budget-source shares (55% new, 21% software) leave 24% unexplained on air. Kaput reads the software share as 22% (2026-11-11).
- Amazon memo date: the show gives both “released … on October 27th” and “circulated … in September”. The September date is taken as the circulation date.
- Apple’s CEO is named inconsistently in the caption (the episode names John Giannandrea as “newly installed CEO”, while Moonshots ep. 308 names John Ternus as CEO). Not used here.
- Names from captions: the Amazon executive leading the reorganisation (“Peter Dantis”) and the frontier-model lead (“Peter Aiel”) are not normalised here.
Related
- AI Adoption Data, October 2026 — KPMG, Ramp and The Information data from the month before
- AI Bubble Signals — why concentrated spend worries investors
- Anthropic’s Public S-1 — the seller’s side of the same spend
- AI and Work, Early November 2026 — how slowly large companies are actually deploying agents
- Token Economics Primer — cost per task as the unit
- Admin Spend Visibility and Control — how to set the budgets and dashboards
- Cost & Intelligence Levers — routing and caps in practice