FOR PEOPLE BUILDING WHAT'S NEXT✳
VOL. 1 · NO. 1 NOVEMBER 20, 2026
THE TREND TIMES
Signals, Not Noise.
Join ↗
MARKET SIGNALS

Enterprise AI Spend at $177B: Cost Traps and Pricing Shifts

Enterprise AI spend hit $177B in 2026, averaging 2.5x plan as 89% redesign for token costs. See the benchmark data and how pricing models adapt.

BY THE TREND TIMES EDITORS

Replit used to charge 25¢ every time one of its agents hit a checkpoint. That worked while agents ran for a few minutes. As runtimes grew, some jobs cost Replit several dollars while the user still paid 25¢, so the company moved to usage-based billing (Stripe, Oct 2026).

The same mismatch now shows up on the buyer side. According to Menlo Ventures, enterprises planned to grow AI spend 1.9x this year and ended up at 4.7x on average, roughly 2.5x their plan (Menlo Ventures, 2026). Agents are the reason in both cases: they bill by work done, and nobody forecast how much work they would do.

If you sell an AI product or run an AI budget, you are on one side of that gap. This piece reads five primary sources from October and November 2026, keeps each survey’s numbers with its own sample, and ends with a checklist for this week and this month.

What the $177B Number Actually Measures

Menlo Ventures puts enterprise AI deployment spend at $176.8B by September 2026, up from $33B a year earlier (Menlo Ventures). That is 5.4x growth, faster than last year’s 3.9x. Deployment means using AI: $99.5B of infrastructure such as model inference and agent platforms, plus $77.3B of applications.

Read the figure with its method attached. Menlo built it from a bottoms-up market model and shows revenue as annualized run rate, so $177B is the spending pace as of September, not a closed calendar-year total. The survey behind it covers nearly 400 U.S. enterprise decision-makers who manage more than $4B in AI spend, fielded by Morning Consult from September 8 to 26 (Menlo methodology).

Two comparisons give the number scale. Building models drew $102.5B, so each $1 of development spend now meets about $1.70 of deployment spend. Deployment also equals more than 30% of public cloud revenue, up from 6% a year ago (Menlo Ventures).

Why Budgets Broke: Plans at 1.9x, Actuals at 4.7x

The average enterprise in Menlo’s survey planned to nearly double AI spend in 2026. Reported spend averaged 4.7x last year’s level against a planned 1.9x (Menlo Ventures). The median tells a calmer story: 2.2x actual against 1.8x planned.

The gap between mean and median matters for your own forecast. Menlo attributes it to a “fat head” of power users whose spend grew far faster than everyone else’s. If your company is not one of them, the 2.2x median is the closer benchmark.

Adoption did not drive the jump. The share of enterprises using AI rose only from 55% to 65%, while spend per enterprise nearly quintupled (Menlo Ventures). Menlo’s explanation is agents: they run in parallel, keep working after people step away, and so scale spend with work instead of headcount.

Agents Are Now the Biggest Token Consumers

OpenRouter, a gateway that routes developer traffic to hundreds of models, shows the shift in raw tokens. Agent traffic passed human traffic in early February 2026, after the launches of Opus 4.6 and GPT-5.3-Codex. Since then agent traffic has grown 37x, against 5x for human-driven use (Menlo Ventures).

Agents now generate 80% of OpenRouter tokens, about 130.9 trillion per week, or roughly 7x human usage (Menlo Ventures). Keep the sample in mind. Menlo notes that OpenRouter users skew toward individual developers and startups, so this is a leading indicator for enterprises, not a measurement of them.

Builder surveys point the same way. According to Menlo, 74% of AI builders now use agents to manage workflows, and 32% run multiple agents together (Menlo Ventures). The prior-year comparison of 27% measured startup deployments, so treat the jump as direction rather than an exact like-for-like rise. Code sandbox adoption went from 11% to 40%, search from 5% to 30%, and browser infrastructure from 4% to 14%.

The Cost Traps Nobody Budgeted For

Cloudera’s survey shows what overspending looks like inside operations. It is a different sample: 1,500 global IT leaders, fielded by Wakefield Research in September 2026 (Cloudera, Oct 2026). Its percentages should not be stacked on Menlo’s U.S. figures.

The headline: 89% of respondents have already limited or redesigned AI use cases because of rising token consumption. And 98% have pulled back an AI deployment because of failures, hallucinations or unexpected behavior (Cloudera). That second number matters for cost: a rollback usually means tokens spent without a finished result.

Governance adds its own line item. According to Cloudera, 75% name governance and compliance as a constraint on scaling AI, and 55% say those costs are growing faster than expected (Cloudera). Asked what they would do if they could restart, 49% said they would strengthen governance from the beginning.

Moving workloads is the next trap. Cloudera found 45% have moved AI workloads because of higher or more volatile costs. But 84% say data residency or sovereignty rules shape where AI runs, and 58% say a move needs moderate or major redesign (Cloudera). Switching to a cheaper host is rarely a config change.

Where the Money Actually Went

Spend is concentrated at the model layer. Menlo estimates that Anthropic, OpenAI and Google take $122.7B, or 69% of deployment dollars, up from 50% last year (Menlo Ventures). Their spend grew 7.5x, against 3.3x for everyone else.

Within enterprise LLM API spend, Anthropic’s share rose from 40% to 55%, OpenAI’s from 27% to 31%, and Google’s fell from 21% to 9%. Anthropic and OpenAI together hold 86%, up from 67% (Menlo Ventures). Buyers paid for capability: the most popular models moved from Claude Sonnet 4.5 and GPT-5 a year ago to Claude Opus 5.5 ($4/$20 per million input/output tokens) and GPT-6 Astra ($10/$60).

Open-source models show the opposite pattern. Their token share rose from 11% to 19%, about 2.9 quadrillion tokens a month, yet their spend share fell from 6% to 4% (Menlo Ventures). The average open-source token costs $0.13 per million, against a $0.70 blended price for closed models.

The application layer reached $77.3B, up from $19B. Coding ($34.0B, up 7.7x) and enterprise chatbots ($25.0B, up 3.0x) make up 76% of it (Menlo Ventures). Claude Code and Codex alone account for $24.0B, or 70% of coding spend, and AI-native vendors hold 96% of that market.

Most of the money is new. Menlo found 55% of incremental AI spend came from net new budgets and 22% from software budgets. Of that 22%, 13% stayed with incumbents converting SaaS licenses to AI, and only 9% ($12.6B) moved to AI-native vendors (Menlo Ventures).

How Pricing Models Are Being Rewritten

Sellers are changing how they bill because their costs now move with agent work. According to Stripe, two-thirds of the Forbes AI 50 use some form of usage-based pricing (Stripe, Oct 13). The pattern from its Monetize 2026 conference, with 300+ leaders from companies like Replit, Fin, Snowflake and Clay, is layered: subscriptions for predictability, usage pricing for agent work, and credits to bundle several billing units.

Replit shows the cycle. Usage billing closed its margin gap but brought unpredictable bills and abrupt paywalls. As inference got cheaper, Replit added a rate-limited Free Mode on a plain subscription, and it now changes parts of its pricing several times a month (Stripe).

Clay ran into the same problem from another side. Its pricing memo, cited by a16z, found that fixed AI prices overcharged simple workloads and made complex ones unprofitable (a16z, Nov 2026). Clay split Data Credits from Actions for orchestration, kept fixed credit prices for bounded AI work, and passed through actual token costs for volatile reasoning models with no markup.

Predictability has become a feature you ship. Snowflake gives customers forecasting tools and scenario builders, and Stripe now treats a usage control pane with alerts and spending limits as standard (Stripe). Annual usage buckets and discounted overages do the same job on the contract side.

Pricing Pages Are Now Read by Machines

Kyle Poyar of Growth Unhinged tested how six AI engines describe pricing for companies on the Cloud 100. He analyzed 7,600 AI-generated responses (Stripe). For one unnamed company with no accessible pricing, the models pulled from informal sources like Reddit and returned prices from $100 to $2,400 per user per month, a 24x spread.

Negative pricing language appeared in 88% of responses about that same company (Stripe). That figure comes from one case, not the whole Cloud 100, but the mechanism applies to anyone. If crawlers cannot read your pricing page, or it renders in a way agents cannot parse, a model will fill the gap with whatever it finds.

Your pricing page now has two readers. The buyer still signs, but an assistant is often summarizing the options first.

Seat Pricing Is Not Dead, but It Is Shrinking

Stripe’s platform survey covers 200 C-suite and director-level leaders at North American and European SaaS platforms, fielded in March 2026 (Stripe, Oct 19). Half use Stripe, and the sample splits evenly between vertical and horizontal platforms. New platform businesses on Stripe grew more than 180% year over year, and 57% of leaders worry about losing customers to AI-driven competitors.

Seats remain the base, not the headline. The share expecting seats to be their primary model falls from 16% in 2026 to 11% in 2027 (Stripe). Horizontal platforms move faster, with 7% expecting seats to stay primary against 14% of vertical SaaS. And 44% plan multiple pricing changes in the next 12 months.

The hybrids are taking shape. One in four plan to keep per-seat fees and bill agent work separately through credits, usage or outcomes; Zendesk already charges for resolutions alongside subscriptions. Another 7% plan to price agents like seats, with a base fee plus a license per agent (Stripe).

The 2027 revenue outlook splits by cohort. Among horizontal platforms serving midsize and enterprise customers, 60% expect usage or hybrid pricing to lead revenue, against 22% of vertical platforms serving micro and small businesses (Stripe). For embedded finance the split reverses: 34% of that vertical cohort and 3% of the horizontal one.

What the Benchmarks Do Not Tell You

Menlo’s spend data says agents push token use up, but its own price math says they can push the average unit price down. Cache reads can make up 95% or more of an agent’s tokens. Opus 5.5 charges $0.20 per million for them, against $4 for fresh input and $20 for output (Menlo Ventures). At 3% fresh input and 2% output, a session averages about $0.70 per million tokens.

So the per-token price falls while total spend rises. Neither number tells you what one finished task costs. That is the measure a16z asks sellers to prove: the full cost of a completed unit, including retrieval, retries, evaluation, human review and failed attempts (a16z).

Buyers expect bills to keep growing. In a16z’s August 2026 survey of 100 enterprise LLM decision-makers, 94% expected total LLM spend to rise and 37% expected it to at least double next year (a16z). That is a third sample again, separate from Menlo and Cloudera.

The case studies show what proof looks like. Harvey beat Microsoft Copilot and Repsol’s internal prototype in an eight-week pilot with 50 lawyers across 12 jurisdictions; Harvey reports four to six hours saved per lawyer per week and 96% adoption across the legal department. Decagon’s published Notion case reports up to a 34% improvement in ticket resolution time (a16z). Both are vendor-reported, which is why a16z pairs them with a cost-per-completed-unit test.

Primary Sources Behind This Analysis

Each source answers a different question and uses a different sample:

Action Steps for AI Builders and SaaS Teams

This week

  • Pick one agent workflow and measure its real token cost per completed unit (a reviewed contract, a resolved ticket), counting retries, cache reads, failed runs and human review time (a16z).
  • Fetch your pricing page with curl and with an AI assistant, then check that both return your actual plan prices in plain text.

This month

  • Ship in-app spend limits and usage alerts, or move to a hybrid model that bills orchestration separately from pass-through model cost, as Clay did.
  • Turn on prompt caching for agent sessions and route low-risk steps to cheaper models, then track average cost per million tokens against the roughly $0.70 Menlo benchmark (Menlo Ventures).

Sources

THE WEEKLY SIGNAL / EVERY THURSDAY

Signals, Not Noise.

One primary-sourced deep dive into AI agent standards and micro-SaaS every Thursday. Zero spam, ever.

Free. One email a week. Unsubscribe anytime.

ONE EMAIL A WEEKUNSUBSCRIBE ANYTIME