Generative AI is reshaping where companies spend and how vendors earn. Upfront licensing is no longer the dominant story. Early enterprise budgets are skewing toward compute and integration — GPUs, cloud time, and the engineering needed to weave models into real workflows. Investors are increasingly backing scale players that can drive down per-request serving costs, while corporate finance teams track model-refresh cadence, utilization, and labeling expenses. Below I break those drivers into actionable numbers and outline scenarios that matter for investors and market watchers.
Adoption patterns — three clear clusters
Adoption of generative AI falls into three practical groups once a market matures:
– Early adopters: pilots and proofs-of-concept, about 15% of firms.
– Scaled adopters: departmental, production deployments, roughly 35%.
– Platform integrators: company-wide embedding, around 50%.
Using a working universe of ~175,000 global firms with 100+ employees, a 35% scaled-adoption rate implies about 61,000 firms running generative AI in production for specific workflows. That cohort underpins the market-sizing and revenue estimates that follow.
Where the money flows
Enterprise spending concentrates in three buckets:
1. Vendor licensing and API fees.
2. Professional services: integration, customization, compliance.
3. Infrastructure: cloud GPU/TPU time, storage, and egress.
A reasonable annual range per firm for vendor + services is USD 200k–1.2M; the midpoint sits near USD 700k. Multiply that midpoint by 61,000 scaled firms and you arrive at roughly USD 42.7 billion in vendor and services revenue. Infrastructure typically adds another 20%–60% of vendor fees, which pushes the combined addressable market into the USD 51–68 billion band under current assumptions.
Unit economics and pricing dynamics
Vendor margins rest on three core costs: capitalized model development (training), marginal inference, and customer acquisition/servicing. For example, a notional training capex of USD 300M amortized over 10 trillion inference tokens yields a training cost of about 0.00003 USD per token. Marginal inference for efficient large-scale serving often translates to roughly 0.5–5 cents per thousand tokens. Together, blended per-thousand-token costs land in the low single-digit cents; list prices typically apply 2x–6x markups, explaining the observed range of roughly 1–30 cents per thousand tokens across vendors and use cases.
Two dynamics to watch:
– Falling inference costs (better chips, quantization, optimized serving stacks) tend to drive per-call prices down and volumes up.
– Rising per-seat API consumption (deeper automation, richer outputs) expands the vendor + services pool even as unit prices compress.
The multimodal cost drag
Workloads that include images or video add bandwidth and egress expenses. For instance, one million multimodal API calls averaging ~2 MB outbound equals roughly 2 TB of egress — a nontrivial cloud bill today. These costs can lengthen payback periods for vendors. With a customer acquisition cost (CAC) of USD 80k and an annual contribution margin of USD 220k, payback is under a year; cut that margin by 30% and the payback window stretches, tightening cash flows and putting pressure on valuations.
Who captures value
– Cloud providers and GPU vendors capture the majority of infrastructure spend.
– Systems integrators and consultancies benefit from rollouts and bespoke integration work.
– Industries with standardized workflows — professional services, finance, parts of healthcare — move faster from pilot to production and therefore account for disproportionate spend.
– Verticals that demand auditability and accuracy (legal drafting, clinical summarization) tolerate higher per-token prices; commodity tasks push prices lower through volume.
Productivity and macro effects
Micro-level studies show automation gains in the 10–25% range for specific software tasks. If you conservatively model an 8% uplift applied to 20% of white-collar hours in affected sectors, the aggregate comes to roughly a 1.6% sector-level labor productivity improvement. Scaled, economy-wide adoption could therefore translate into approximately 0.2–0.5 percentage points of measured GDP growth, depending on penetration and whether firms channel gains into capital investment and further innovation. The market is already sizable and will reallocate value across cloud providers, model owners, integrators, and vertical specialists. Watch inference-cost trends, egress/bandwidth for multimodal workloads, and per-seat consumption — those variables will determine who wins the most economic value as adoption climbs.
