The generative AI land rush has devoured billions in datacenter spending, and now vendors have found a way to pass the bill: imposing usage-based pricing that shifts infrastructure costs directly onto customers. Anthropic, OpenAI, and GitHub have already begun crafting plans where every API call, every token, every activated agent escalates the monthly cost. Microsoft adds the E7 license—a premium bundle that locks M365 Copilot, Agent 365, and security tools on top of the E5 subscription.

Research firm Forrester, after surveying more than 2,600 business and technology decision-makers, sounds the alarm: software budgets will swell in 2025, driven precisely by these pricing moves. The report notes that 80% of executives expect data and software spending to rise, not only because adoption is broadening but because of a perverse mechanism: the vendor offloads its own growing datacenter costs onto customers.

Bain & Company had already estimated that building AI datacenters will reach $2 trillion by 2030. Rather than absorbing that investment and recovering it over time through economies of scale, vendors are adopting pricing models that turn AI into a variable cost—almost an energy commodity—for the enterprise. The result is a predictability crisis: traditional FinOps was never designed for token-driven expenses, and KPMG found that nearly a third of business leaders struggle to understand and control operating costs when AI is deployed at scale.

Who really pays the hidden infrastructure

This shift has a structural effect that goes beyond accounting: it changes the incentives for companies evaluating where to run their Large Language Models. As long as vendors offered flat-rate subscriptions, the cloud appeared convenient and predictable. Now, with bills that can explode month after month, larger players are beginning to recalculate the Total Cost of Ownership of on-premise or colocation infrastructure, where hardware amortizes over time and inference can be optimized through techniques like quantization.

In this scenario, self-hosted deployments are no longer an ideological choice but a financial lever. Platform engineering and procurement teams already running containerized workloads are starting to evaluate runtimes like vLLM or TGI—tools that orchestrate inference on owned GPUs, providing full cost visibility and no end-of-month surprises.

The myth of AI replacing IT staff

The other myth the Forrester report dismantles is that AI is replacing IT jobs. “The AI-washing of layoffs will continue,” the analysts write, “as vendors trim for financial and restructuring reasons. Guard against inflated promises that AI can replace employees across the board.” The data is clear: technology staffing spending has not declined in recent years, despite high-profile cuts announced by Oracle, Microsoft, and Meta. In 2025, personnel already accounted for 35% of IT budgets, and for 2027, 67% of decision-makers expect to increase that share. Specific data and analytics roles are on the rise, with 68% of respondents anticipating a budget increase for that category.

This means AI is not eliminating work but transforming it: organizations need people who can manage complex pipelines, perform efficient fine-tuning, and implement semantic caching and usage guardrails against runaway spending. These are skills that in the cloud are paid for token by token, but in a controlled environment they become internal know-how.

Beyond FinOps: investing in foundations

Sharyn Leaver, Forrester’s chief research officer, sums up the strategic shift: “The organizations that outperform in 2027 won’t be those that spend the most on AI. They’ll be the ones that invest in the foundations that make AI effective: trusted data, strong governance, organizational readiness, and the ability to continuously adapt.” This re-framing touches the core of the on-premise approach: if data is well governed and the infrastructure is under control, AI becomes a calculable function, not a runaway variable.

Forrester still recommends strengthening FinOps practices with capabilities suited to the token era—dynamic model routing, semantic caching of responses, usage thresholds. But these are stopgaps. The real question is whether it makes sense to keep paying for every single token, perhaps on sensitive data, when an investment in inference hardware can slash unit costs and preserve data sovereignty. This isn’t about returning to the autarchic datacenter; it’s about placing workloads based on a transparent calculation of TCO, latency, and compliance.

Vendors have chosen the consumption path because infrastructure is expensive and shareholders demand profitability. Customers are now forewarned: the infrastructure bill is already on its way. The ability to read it, and to decide which line items to move outside the vendor’s perimeter, will define the winners of 2027.