The Token Bill Is Coming. Most Companies Aren’t Ready.

How the AI inference cost crisis snuck up on enterprise and what to do about it.

There’s a conversation we keep having with clients.

It usually starts around month six of a successful AI deployment. The product works. Users love it. Then the finance team sends a message, and the number is three to five times larger than anyone expected.

The AI didn’t fail. The math did.

Why the Bill Arrives as a Surprise

In 2023 and 2024, most AI deployments were contained pilots. Costs were manageable because volume was small. Then the pilots moved to production — agentic workflows came online, chains of AI calls running continuously and the economics shifted.

Inference spending crossed 55% of all AI-optimized cloud infrastructure costs in early 2026, surpassing training for the first time. Inference accounts for 80–90% of total compute costs over a model’s production lifecycle.

In traditional software, the marginal cost of one more user request is essentially zero. In AI inference, every request costs tokens. Every token costs money. In an agentic workflow that chains five model calls per interaction, you’re paying five times what a single call would cost.

Some enterprises are now seeing monthly AI bills in the tens of millions of dollars — driven primarily by agentic AI with continuous inference.

The Subsidy Nobody Talks About

The API prices most companies are building on today aren’t real prices. They’re subsidized.

OpenAI generated approximately $3.7 billion in revenue in 2025 while losing an estimated $5 billion. Current API pricing is subsidized by venture capital and hyperscaler cross-subsidies, with price normalization expected within 12–24 months.

The pricing that CFOs have approved AI budgets around is being held artificially low while the biggest players fight for market share. When that fight settles, prices will adjust. Organizations that have built inflexible AI workflows around today’s pricing will feel that adjustment directly.

The Most Expensive Mistake in Enterprise AI

There’s a name for the architectural decision driving more unnecessary AI spend than anything else: the Big Model Fallacy.

It’s the assumption that frontier models should be used for everything because they’re impressive in demos.

A workflow using a frontier model to extract a date from a document — a task a model costing 1/200th as much handles with identical accuracy — isn’t a minor inefficiency. At enterprise scale processing millions of documents, it’s a structural cost problem that compounds daily.

Pairing model routing with semantic caching reduces API call volume by 30–50% for typical enterprise deployments.

The solution is routing each task to the cheapest model that can handle it reliably — frontier models only when the task genuinely requires them.

What Cloud 3.0 Actually Means

The market’s response has produced a three-tier hybrid architecture that routes workloads based on economics rather than defaulting everything to cloud.

When cloud AI costs reach 60–70% of what equivalent on-premises hardware would cost over a comparable period, the economics of on-premises begin to compete — even after accounting for CapEx and operational overhead.

Cloud handles variable, experimental workloads where you need flexibility without committing capital.

On-premises handles high-volume predictable production inference. On-premises delivers up to 8x lower cost per million tokens compared to cloud IaaS and up to 18x lower compared to commercial GenAI APIs, with breakeven in as little as four months at sustained utilization.

Edge handles latency-sensitive workloads where sub-50ms response is required.

83% of enterprises now plan to repatriate at least some workloads to on-premises or private cloud due to cost pressure.

The Question Worth Asking Now

If your organization is still in planning or early deployment, this isn’t a problem you’re facing yet — it’s one you can avoid.

The organizations suffering most right now deployed quickly without modeling production-scale economics. The opportunity is to design for those economics from the beginning: audit what each workflow costs at scale before building it, choose model tiers deliberately, and build routing and caching infrastructure that makes optimization possible.

The question isn’t “can we afford to deploy AI?” Almost every organization can, at pilot scale. The question is “have we modeled what this costs at 10x, 100x, production scale and built an architecture that handles that economically?”

The organizations that ask this early build AI programs that stay affordable as they scale. Those that don’t will eventually have the conversation with their CFO that so many others are having right now.

Wamisoftware works with startups and enterprise clients on AI integration and architecture. We design for production economics from the start — not as a retrofit when the first large invoice arrives.