Consolidation Is Narrowing the Field
GenAI market consolidation is happening at extraordinary speed. Global AI spending approached $1.5 trillion in 2025 and is forecast to exceed $2 trillion in 2026, while AI funding doubled 2024’s record $108 billion. The capital is concentrating: the three largest Q3 2025 rounds went to foundation-model developers — Anthropic ($13 billion), OpenAI ($8.3 billion) and Mistral ($1.5 billion). High compute costs and compressed margins are pushing the market toward acquisitions and acqui-hires, with large model companies buying application-layer firms that have proven product-market fit.
For buyers, consolidation removes alternatives. Every acquisition narrows the field of credible competitors in a category and hands the acquirer licence to re-bundle and reprice — the same mechanic we examine for traditional software in IT vendor M&A impact on contracts. Tracking where the field is thinning is now core market intelligence, which is why it sits in the market intelligence pillar and runs through the broader AI vendor landscape.
The consolidation is happening at two layers at once, and they pull in opposite directions for buyers. At the model layer, a small number of foundation providers are accumulating the capital and compute to dominate — concentration that reduces choice and supports premium pricing. At the application layer, those same providers are absorbing the tools that sit on top, so the agent or copilot you buy this year may be owned by your model vendor next year. The risk is a vertically integrated stack where one supplier controls the model, the application and the pricing — exactly the lock-in enterprises spent the last decade trying to escape in traditional software.
Token Prices Fall, Bills Rise
The central paradox of AI pricing in 2026 is that unit costs are collapsing while total bills climb. Analysis of 2.4 billion enterprise API calls shows the blended cost of AI fell 67% year on year, from $18.40 to $6.07 per million tokens, and Gartner projects that by 2030 inference on a trillion-parameter model will cost around 90% less than in 2025. Yet aggregate spend is rising sharply, because agentic workloads could drive a 24-fold increase in token consumption by 2030.
| AI pricing signal | Figure | Buyer implication |
|---|---|---|
| Blended token cost YoY | −67% ($18.40 → $6.07/M) | Renegotiate rates as unit prices fall |
| Projected inference cost by 2030 | ~90% cheaper | Avoid long fixed-rate lock-ins |
| Agentic token consumption by 2030 | ~24x | Volume, not unit price, drives your bill |
| Microsoft Copilot enterprise seat | $30 → $60 | Challenge seat doubling against usage |
| Accidental single-month Claude bill | $500M | Usage caps are non-negotiable |
The risk is consumption running unchecked. One enterprise accidentally spent $500 million on Claude in a single month after rolling out AI access with no usage caps, and Uber reportedly burned through its entire 2026 AI budget in four months — roughly 5,000 engineers on coding assistants, with heavy users running $500 to $2,000 each per month. Falling token prices are no comfort if consumption is uncapped — the budgeting problem mirrors the consumption-pricing shift across the wider SaaS market.
This is the trap of the falling-price narrative. When a vendor points to a 67% drop in token cost as reassurance, it is describing the unit, not the bill. Agentic systems make many model calls per task — a single autonomous workflow can chain dozens of inferences — so consumption scales with capability, not headcount. The enterprises absorbing the worst surprises are those that budgeted as if AI were a per-seat licence and discovered it behaves like a metered utility with no upper bound. The unit-economics improvement is real and worth capturing, but only a consumption ceiling turns it into a predictable cost.
Per-Agent Meters and Bundling
Vendors are responding to consolidation and compute costs with a second meter. Salesforce Agentforce and Microsoft Copilot now charge for what AI agents do, not just how many people have access. Microsoft has doubled Copilot enterprise pricing from $30 to $60 per seat, while Salesforce has shipped three different Agentforce pricing models in roughly 18 months — from $2 per conversation, to Flex Credits at $0.10 per action ($500 per 100,000 credits), to tiered editions. Almost every major vendor is converging on a hybrid of seat licence plus metered agent activity.
This bundling pressure is the AI face of the broader repricing wave, and it features heavily in our Microsoft 2026 strategy analysis. The buyer danger is accepting an agent meter with no ceiling — a structure that converts a predictable seat cost into open-ended exposure.
The proliferation of models within a single vendor compounds the problem. Salesforce shipping three Agentforce pricing constructs in eighteen months is not indecision — it is a market testing what buyers will tolerate, and each iteration tends to extract a little more. When pricing changes that often, any multi-year commitment to a specific model risks being stranded above market within a quarter or two. The defensive posture is to treat the current pricing construct as temporary and to negotiate the right to move to whatever model the vendor offers next, rather than being locked to the one you signed.
Contracting for a Consolidating Market
In a market this volatile, contract flexibility beats price certainty. Avoid long fixed-rate token commitments when unit prices are falling 67% a year — you will be locked above market within months. Instead, negotiate the right to benchmark and reset rates periodically, hard usage caps and alerting thresholds, transparent unit definitions, and the ability to substitute models as the field consolidates. Open-weight models are an increasingly credible hedge, reinforcing the open source enterprise adoption case as protection against proprietary lock-in.
Procurement governance has to catch up with the technology. The enterprises that lost control of AI spend did so because access was granted without the budget guardrails that any other metered utility would carry. Before any broad rollout, set hard per-user and per-workload ceilings, real-time alerting at defined thresholds, and an automatic cut-off that requires sign-off to exceed. These are not constraints on innovation; they are the difference between a controlled pilot and a $500 million surprise.
Consolidation makes portability the other priority. As the field narrows to a few foundation providers, the buyer’s protection is the ability to switch — which means avoiding deep dependence on one vendor’s proprietary tooling, abstracting model calls behind a layer you control, and keeping at least one alternative model validated for your key use cases. The provider that knows you can move is the provider that keeps its pricing honest as the market consolidates around it.
Keep AI procurement on a short leash contractually. Twelve-month terms, not multi-year, suit a market where unit prices fall 67% a year and pricing models change every few quarters; the flexibility to renegotiate or switch is worth more than any volume discount a longer commitment might buy. In a consolidating, fast-repricing market, the shortest credible commitment is almost always the cheapest over time.
Treat every agent meter as a budget risk until it has a ceiling, and separate AI line items from your core software renewals so they remain optional and independently priced. Our AI contract red flags white paper catalogues the clauses that create runaway exposure, and you can request a confidential briefing on any AI agreement where consolidation has narrowed your alternatives.