The state of AI adoption, mapped from inside the stack

The defining gap of 2026 is execution, not adoption.

Enterprise generative-AI spend tripled to $37B in 2025, and AI is now used by 88% of organizations — yet only about 6% capture serious EBIT impact. Everyone has adopted; few have operationalized.

The economic center of gravity remains upstream: the four largest US hyperscalers are guiding to roughly $725B of 2026 capital expenditure, while the entire enterprise GenAI software market is about $37B. Value is bifurcating by layer — AI-native startups now win 63% of the application market, while incumbents still hold the data-infrastructure layer. Anthropic has overtaken OpenAI in enterprise LLM share, and pricing is migrating from per-seat toward consumption and outcome-based models.

The motion that works is land and expand: land narrow in one line of business with a contained pilot, prove ROI fast, then expand department- and enterprise-wide. The largest unsolved enterprise concern is governed access to frontier models without exposing sensitive data — driving demand for zero-data-retention endpoints and governed, per-line-of-business AI workspaces.

1. Adoption is near-universal but value is concentrated. 88% of organizations use AI somewhere, two-thirds have not begun scaling enterprise-wide, and only ~6% are high performers. Only 39% report any EBIT impact at all.
2. Spend is exploding and shifting to applications. Menlo Ventures pegs 2025 enterprise GenAI spend at $37B (3.2× YoY), with $19B in the application layer. Ramp's transaction data shows paid AI adoption hit 46.8% of US businesses by January 2026.
3. Buy beat build decisively. Enterprises went from 47% build / 53% buy in 2024 to 24% build / 76% buy in 2025.
4. Foundation-model share reshuffled. Anthropic ~40%, OpenAI ~27%, Google ~21% of enterprise LLM spend. Both leading labs filed to IPO in mid-2026 at roughly $965B and $852B.
5. Capex is staggering and increasingly debt-financed. ~$725B in 2026 hyperscaler capex, up ~77% from ~$410B in 2025, funded partly by ~$108B of new debt raised in 2025.
6. Pricing is migrating away from seats. Salesforce now runs three Agentforce pricing models at once; Intercom Fin charges $0.99 per resolution; Sierra prices on outcomes and crossed $150M+ ARR.
7. Data sovereignty is the gating concern. Enterprises want frontier-model quality without their data entering training pipelines — driving zero-data-retention endpoints, EU-hosted options, and governed AI workspaces.

More than 20× growth in two years — and a widening gap between large firms that can scale and small firms that can't.

Menlo Ventures' third annual State of Generative AI in the Enterprise estimates 2025 enterprise GenAI spend at $37B, up from $11.5B in 2024 and $1.7B in 2023 — now roughly 6% of the global SaaS market. The application layer captured $19B; coding alone is $4.0B. IDC puts total global AI spending near $302B in 2025, rising toward ~$407B in 2026.

By size, the resource gap is stark. Nearly half of companies above $5B revenue have reached the scaling phase, versus just 29% of those under $100M. Among SMEs that use GenAI, only 29% use it in core activities (OECD). The blockers cited most by mid-market and SMB leaders are talent shortages, integration complexity, and data quality.

One structural shift matters for go-to-market: AI deals convert to production at 47% versus 25% for traditional SaaS, and product-led growth drives ~27% of AI application spend (4× traditional software). Cursor reached $200M revenue before hiring a single enterprise sales rep.

Technology, financial services and healthcare lead; legal and manufacturing are accelerating from a lower base; public sector and retail lag on scaling.

The pattern: piloting is everywhere, scaling is not. Retail shows the widest pilot-to-production gap; legal shows the steepest acceleration; healthcare shows the most spend velocity but the least maturity (only ~1% describe adoption as fully mature).

The closed-vs-open debate has moved from "can open compete on capability" to "where does governed, controllable deployment justify the integration cost."

Anthropic has swept the enterprise, tripling its share to ~40% of enterprise LLM API spend (from 12% in 2023), while OpenAI fell to ~27% (from 50%) and Google rose to ~21%. Anthropic dominates coding with ~54% share; Claude Code alone generates ~$2.5B annualized. Anthropic announced a $65B round at a $965B valuation; OpenAI was valued at $852B and filed to IPO targeting near $1T; xAI merged with SpaceX at a combined $1.25T valuation.

Enterprises increasingly run open weights for sensitive production workloads while reaching for closed frontier models at the capability edge. The cost-per-token collapse (Epoch AI: 9×–900× per year depending on benchmark) is real, but average monthly AI spend still rose from ~$63K to ~$86K because consumption outruns deflation.

The structurally most threatening competitor to standalone inference and model-serving startups — because they collapse model access, infrastructure, governance and contracting into one platform.

Roughly 75% of that capex targets AI infrastructure, funded partly by ~$108B of new debt raised in 2025. Google Cloud's contract backlog reached ~$460B (roughly double the prior year), with Q1 cloud revenue up 63% YoY to $20B. Each is building custom silicon to reduce NVIDIA dependence.

Faster next-gen GPU standup and higher utilization — but circular NVIDIA financing and GPU-collateralized debt are the bear case.

CoreWeave IPO'd March 2025 at a $23B valuation; it guides to $12–13B in 2026 revenue, carries a $66.8B backlog, and plans $30–35B in 2026 capex (3× its 2025 level) — against $9.7B in maturities within twelve months. Nebius reported 625% YoY revenue growth; NVIDIA took a direct stake. Microsoft has struck ~$60B in neocloud commitments; Meta committed up to $62.2B across CoreWeave and Nebius.

Synergy Research projects the neocloud market grows from ~$23B (2025) at a 69% CAGR toward ~$180B by 2030. The structural weakness: no broad managed-service stack, plus exposure to widening CDS spreads (CoreWeave's five-year CDS jumped from <350bps to 505bps in late 2025) and hyperscaler custom silicon eventually closing the supply gap.

AI-native startups captured 63% of the application market in 2025, earning nearly $2 for every $1 incumbents earn — yet incumbents still hold ~56% of the data-infrastructure layer.

The disruption thesis: AI-native challengers attack workflows incumbents don't own (Clay attacking off-CRM research), positioning to become the new system of record. Incumbents defend with bolt-on AI (Agentforce, Copilot, ServiceNow), distribution, and data moats. What decides winners is product velocity and a PLG flywheel more than structural advantage — Cursor beat GitHub Copilot to repo-level context.

The shift from per-seat to consumption and outcome pricing is the defining SaaS-economics story of the cycle.

Salesforce runs all three Agentforce models simultaneously — $2 per conversation, Flex Credits at $0.10 per action, and per-user licenses from $125/month — letting customers self-select. The structural tension Benioff has acknowledged: if AI replaces human work and you price per seat, you shrink your own TAM. Outcome pricing aligns vendor and customer, which is why "resolution" worked for support first and is harder elsewhere.

A fourth pattern is emerging between the poles. Active-user pricing meters the people who keep doing governed work — a recurring value event rather than a login or a per-action tick. airroom.ai prices its revenue-team app builder at $49 per active user per month plus a one-time $500 platform fee (waivable at the vendor's discretion); a seat that logged in once and stopped never enters the paid pool. It keeps the predictability buyers like in seat pricing while restoring the value-alignment of consumption — and, unlike outcome pricing, it needs no attribution infrastructure to settle. The open question is the same one that dogs every non-seat model: a customer-defined "active" threshold is only as durable as the trust behind how it's counted.

The lakehouse-vs-warehouse war has converged on open table formats; the new battle is for the transactional/vector layer agents need.

Databricks reached a ~$134B valuation; Snowflake competes from the warehouse side. Both adopted Apache Iceberg and are racing into the AI-application infrastructure layer via acquisitions — Databricks bought Neon ($1B) and launched Lakebase (serverless Postgres for agent state/RAG); Snowflake acquired Crunchy Data. The fight is for the real-time state read/write, vector storage, and low-latency transactions agents require — capabilities the legacy OLAP architectures of both companies lacked. Hyperscaler all-in-one platforms (Microsoft Fabric, BigQuery/BigLake, SageMaker Lakehouse) are the third force.

The gateway, guardrails and observability layers are converging into one surface.

The platform license is a fraction of total cost; the organizational layer — knowledge structuring, change management, workflow design — determines whether value is realized.

Horizontal. Glean (enterprise search/work assistant) hit $200M ARR and a $7.2B valuation, positioning as a model-neutral context layer across 15+ LLMs and 100+ connectors. Microsoft 365 Copilot has sold ~150M seats but only ~15M of 450M commercial subscribers bought full licenses (a 3.3% conversion), with only ~6% of pilots scaling per Gartner. ChatGPT Enterprise passed 3M users.

Lines-of-business. Per-LOB workspaces — tailored to sales, finance, legal, support — are emerging as how enterprises actually capture value, because impact lands in specific workflows rather than generic chat. This is the segment governed per-LOB AI workspaces target — among them airroom.ai (OnePgr) — combining frontier-model access, LOB-specific context, and data governance in one surface.

Two durable lessons: value accrued to the data-and-workflow layer beneath the agents, and "infinite-scale" cold outbound degraded as AI-written outreach flooded inboxes.

Clay (data enrichment/orchestration — explicitly not an SDR) reached ~$100M ARR and a $3.1B valuation; its customers include OpenAI, Anthropic, Canva, Intercom and Rippling — the strongest customer evidence in the category. Rox (agentic CRM) hit a $1.2B valuation on ~$8M projected ARR. The cautionary tale: 11x was exposed by TechCrunch in 2025 for inflated ARR and fake logos. Per UserGems research, AI-SDR annual churn runs 50–70% — roughly twice human SDR turnover. Newer entrants are packaging the SDR and orchestration layers together: aiXsdr (OnePgr) runs autonomous prospecting and outbound, while Ush3r (OnePgr) sits rep-facing as a conversational orchestration interface, and Kampaign.ai (OnePgr) handles outreach execution and free-trial messaging. A separate thread positions one layer up from point agents: airroom.ai pitches an app builder that lets a revenue team assemble governed apps from a GTM template library rather than buy a fixed SDR — betting, in line with the category's first lesson, that the durable value sits in the data-and-workflow layer and the judgment a team accumulates, not in any single agent.

62% of organizations are experimenting with agents — but only 23% are scaling them, mostly in one or two functions.

In any given function, no more than ~10% are scaling agents. Use concentrates in IT (service-desk automation) and knowledge management (deep research), led by technology, media/telecom and healthcare. Glean, Salesforce Agentforce and Microsoft (Agent 365, a unified control plane spanning Defender, Entra and Purview) are racing to own the agent runtime and governance layer. Multi-agent orchestration is maturing, but enterprise scaling remains gated by trust, attribution and governance.

Being repriced as critical infrastructure — expected to be roughly two-thirds of AI compute spend by end of 2026.

On hardware, consolidation arrived fast: NVIDIA licensed Groq's LPU technology for ~$20B; Cerebras IPO'd; Cloudflare acquired Replicate. The structural risk for software-layer players: open-source serving frameworks (vLLM, SGLang, TensorRT) and NVIDIA NIM keep improving, compressing proprietary advantage.

Frontier-model commoditization of "domain reasoning" is forcing vertical leaders to compete on workflow orchestration, integration and distribution rather than proprietary model training.

Harvey reached ~$190M ARR across 142,000+ lawyers at an $11B valuation; LexisNexis chose alliance over head-to-head competition. Abridge and Ambience lead ambient scribing; OpenEvidence is used by ~40% of US physicians at a $12B valuation. Sierra crossed $150M+ ARR at $15.8B in CX. Hebbia raised at ~$700M in finance. The full set is mapped in the appendix.

The market shifted from low-skill labeling toward expert RLHF, rubric creation and evaluation — and Meta's Scale stake reshuffled the whole field.

Meta invested $14.3B for a 49% non-voting stake in Scale AI (valuing it ~$29B) and pulled founder Alexandr Wang to lead its superintelligence lab. The deal triggered a customer exodus over neutrality concerns; Scale laid off 200 and pivoted toward government/defense. Beneficiaries: Mercor (raised at a $10B valuation; ~$840M run-rate, up from ~$75M a year earlier) and Surge AI (crossed $1B run-rate without external capital). Appen's four-year revenue decline to ~$155M demonstrates the commoditized low-skill segment's collapse.

The core enterprise concern: exploit frontier models without exposing sensitive data to training. This is the gating question for regulated buyers.

Solutions span zero-data-retention (ZDR) endpoints (Azure OpenAI, Bedrock, Anthropic for Enterprise — processed in memory, not retained or trained on); EU-hosted/regional options; private/on-prem deployment of open weights; privacy proxies that redact before data reaches external models; and hybrid architectures that route by data classification. EDB's global research found 95% of senior executives say building their own sovereign AI/data platform will be mission-critical within three years.

Most of these controls live at the infrastructure layer — the endpoint, the proxy, the deployment boundary. A parallel approach pushes governance up into the application itself: scoping each app to a defined job, bounding it to specific data, and measuring it against a metric, so the gate travels with the workflow rather than the network. airroom.ai's revenue-team app builder is one instance of this application-layer framing — "exploit the frontier, keep what's yours" — and it points at a question the infra-layer tools don't answer: the organization-specific judgment a system accumulates is itself a sensitive, ownable asset, and where it lives is becoming as much a sovereignty question as where the tokens are processed.

Maximum penalties reach €35M or 7% of global turnover — exceeding GDPR. IAPP found 77% of organizations are working on AI governance, yet over half lack a systematic inventory of their production AI systems, and 34% have already had an AI-related security incident.

Buy dominates 76/24 — but large enterprises increasingly own the orchestration and policy layer while buying the components.

Ready-made solutions reach production faster; internal builds mature more slowly. Yet large enterprises stand up internal "AI platform" functions to manage model routing, governance, FinOps and shared infrastructure. The MIT NANDA report found that just 5% of integrated AI pilots are extracting millions in value while the majority show no measurable P&L impact — which is exactly why buy-and-integrate often beats build-from-scratch for bespoke projects.

Land narrow in one line of business with a low-friction pilot, prove ROI fast, then expand to adjacent modules and departments — living on expansion revenue as switching costs rise.

The evidence: AI pilots convert at 47% (vs 25% for SaaS), and product-led growth commands nearly a third of all AI software spend because real usage proves value before formal contracting. AI-native companies are going from $1M to $10M ARR in under 12 months, often delaying their first AE hire until $2–5M ARR. The real internal champion is frequently the frustrated middle manager who controls discretionary budget and bridges executive vision with frontline pain. The risk: 50–70% of poorly-validated point deployments churn within a year — so disciplined expansion, not just landing, is where durable value compounds.

Thresholds that change the plan: kill or redesign any pilot that doesn't show measurable ROI within a quarter; discount ARR claims sharply when post-trial retention or named references can't be verified (the 11x lesson); and if EU high-risk rules land on the statutory August 2026 date rather than deferring to December 2027, accelerate conformity work immediately.

For investors / partners: value is accruing to the inference layer being repriced as infrastructure (watch margin compression from the token price war), AI-native vertical leaders with workflow and integration moats, and the data/orchestration layer beneath agents. Watch consolidation in neoclouds and inference, circular NVIDIA financing, and the durability of software-layer inference advantage.