The Fixed Cost That Nobody Priced

Anthropic's IPO prospectus landed on 29 September 2026, and the number that stopped the room was $518 billion. That is the compute and infrastructure obligation the company has committed to over the coming years, against roughly $4.6 billion of revenue in 2025. The filing lists $111.1 billion with Google, $110 billion with Amazon, $31.4 billion with Microsoft, and $161.2 billion of non-cancelable equipment lease obligations tied to Broadcom, generally spanning seven to ten years. One clause makes clear the shape of the risk: if actual spend falls short, Anthropic must pay the difference.

Most of the coverage treated this as an Anthropic problem, a capital structure curiosity for a company seeking a $2 trillion valuation. It is not. It is the finance function's problem at every institution deploying agentic systems this year, because the same spend-whether-you-use-it structure is showing up in enterprise AI contracts, and 93% of enterprises have already blown past their AI budgets before most of them have run a production agent for six months.

I have sat in enough of these reviews to see the pattern. 60% of agentic AI costs sit in refinement loops rather than primary inference, and most budget frameworks were built around inference costs alone. The finance owner signs off on a pilot costed at eighteen pence per invoice, and month one lands at three times the forecast. Nothing is broken. Attachments were larger than the sample used for costing, ambiguous invoices triggered a clarification loop with no iteration limit, and roughly 6% of runs exceeded twenty steps. The vendor did not lie. The model performed as specified. The bill still tripled.

That variance is the tell. Traditional enterprise software commits you to a per-seat or per-module fee. If adoption is slower than forecast, you carry underutilised licences, but the bill does not surprise you. Cloud infrastructure scales with usage, and finance teams spent the last decade learning how to forecast and govern that. Agentic AI is neither. Much of the spending is locked in, so investors cannot treat it like a typical pay-as-you-go software business where costs flex down when demand cools. If model usage or prices wobble, the company may still owe most of its compute and lease payments. That is the vendor's financing structure. It is becoming the enterprise buyer's obligation structure as well, because the capacity commitments large deployments require come with minimum guarantees.

The regulatory gap makes it harder to price. SR 26-2, the Federal Reserve's revised model risk management guidance issued on 17 April 2026, explicitly states that generative AI and agentic AI models are outside the scope of the guidance. The revised framework focuses most heavily on banking organizations with over $30 billion in assets, as well as institutions with high-risk profiles regardless of size. That scoping decision was deliberate. The guidance states that generative AI and agentic AI models are novel and rapidly evolving, and as such they are not within the scope.

The result is not that banks are free to deploy agentic systems without governance. It is that the systems driving the largest budget overruns live outside the model inventory, validation cadence, and materiality frameworks that boards and audit committees spent fifteen years learning to read under SR 11-7. A credit model that drifts 4% triggers a formal review. An agentic reconciliation system that triples its API call count because somebody uploaded quarterly statements instead of monthly ones does not, because banks now need to build a separate program for the generative and agentic AI systems SR 26-2 excludes from model risk scope. Most institutions have not built that program yet. The agent is in production. The bill is on the CFO's desk.

The shape of the financing makes the exposure systemic. Brookings projects total U.S. AI infrastructure investment at $10.3 trillion from 2025 through 2032, roughly 3.6% of annual GDP, and argues the financing is increasingly structured to stay off corporate balance sheets. Morgan Stanley calculates that more than half of the $2.9 trillion that hyperscalers will spend between 2025 and 2028 to build AI data centers will be financed with external capital. A lot of financial institutions, directly or indirectly, are exposed to these data centers either as lenders, or as guarantors of some of the debt, or as backers of the private credit funds funding these data centers.

That chain of obligation is not theoretical. Broadcom agreed to lend up to $42 billion in convertible notes to Anthropic, and a $60 billion syndicate is forming: a $42 billion senior secured tranche from Bank of America, Citi and Morgan Stanley, and an $18 billion junior tranche led by Blackstone, which plans to commit $9 billion itself. These are the same institutions selling agentic finance operations to their own clients. The credit risk on the infrastructure supplier and the operational dependency on the agent platform are starting to sit in the same portfolio.

The macro spend lines confirm the scale. The Big Five hyperscalers will spend over $600 billion on infrastructure in 2026, a 36% increase from 2025, with approximately 75% targeting AI infrastructure. Major tech companies were estimated to spend $650 billion on AI data centers in 2026. Hyperscalers are spending $13 on AI infrastructure for every $1 of current AI revenue, requiring either massive revenue acceleration or acceptance of multi-year payback periods. That gap is a bet. Every enterprise that commits to a three-year minimum on an agentic workflow platform is taking the other side of it, because if the vendor's utilisation falls short, the capacity cost does not disappear. It gets reallocated, repriced, or passed through.

The FinOps data shows the adjustment is already underway. The FinOps Foundation's 2026 survey of 1,192 organizations managing $83 billion in annual cloud and technology spend found that 98% of organizations now manage AI spend, and 73% report budget overruns. Actual per-token costs on major models fell 50-94% year-over-year by July 2026, but usage exploded: agentic workflows make 10-20 API calls per task instead of one. The unit economics improved. The bill went up. That is the signature of fixed infrastructure cost being amortized over variable demand that turned out higher than modelled.

The spend is no longer an innovation line. In 2024, AI spend sat in innovation and R&D funds, evaluated with the loose ROI expectations of an experiment. In 2026, that spend moved into operational technology capital, judged with the same rigor as an ERP investment or a headcount decision. The 2026 agentic AI market hit $12.4 billion, with 76% of CFOs funding autonomous finance agents that execute, not copilots that suggest. When the budget moves from the CTO's experimental envelope to the CFO's operational capital plan, the variance tolerance moves with it. A 3x overrun on a pilot is a learning. A 3x overrun on a production system with a three-year commit is a forecast failure that follows the finance owner into every quarterly review.

The vendor response has been to build commitment pricing that looks like reserved instances but does not behave like them. A reserved cloud instance is underutilised capacity you already paid for. A minimum agentic platform commitment is underutilised capacity you still owe for, plus the variable cost of the workload you actually ran, if it exceeded the minimum. The contract structure is starting to resemble a take-or-pay agreement, which is how you finance a pipeline or a power plant, not how you buy software. Gartner's 2026 AI Hype Cycle report forecasts 40% of AI agent projects will be cancelled by 2027 due to cost overruns alone, not technical failure, not market fit. When the reason is economics rather than capability, the problem is not the agent. It is the obligation nobody modelled when the ROI case was built on a per-task cost.

Uber's CTO confirmed that the full annual AI budget was exhausted by mid-April 2026. Uber is not a bank. It does not carry the model risk governance overhead or the regulatory validation burden or the data residency constraints that multiply AI deployment costs in financial services. AI deployment in financial services costs 20-40% more than in less regulated industries due to compliance overhead, talent premiums, and legacy system integration, and European banks spend an average of EUR 2.3 million per production AI use case, compared to EUR 1.4 million in retail and EUR 1.1 million in professional services. If a consumer platform with no prudential regulator ran out of money in April, the finance owner at a Tier 1 bank has a structural problem, not a tuning problem.

The boards I brief want to know when the cost curve flattens. The answer depends on whether you are measuring cost per output or cost per commitment. Per-output cost is falling. McKinsey projects AI could reduce certain cost categories by as much as 70% across the banking industry, but the net effect is expected to be 15-20%, or $700 billion to $800 billion, due to rising AI technology costs. That 50-percentage-point gap is infrastructure, which is exactly the line showing up as a minimum annual commitment in enterprise contracts this year. The efficiency is real. The offset is also real. The finance owner who receives the bill sees both, and the bill line is larger, more predictable, and less governable under the existing frameworks than the efficiency line.

Three things would change my view. One is a wave of contract amendments converting fixed minimums into true variable pricing with no floor, which would signal that vendor utilisation is high enough that they no longer need demand guarantees. I am not seeing that. Two is a regulatory update that brings agentic systems into a formal inventory and validation cadence, so the enterprise can govern the usage that drives the cost with the same rigor it applies to a credit model. SR 26-2 went the other direction. Three is a FinOps telemetry standard that attributes agentic cost to the business outcome it produced, so the variance can be defended as ROI rather than explained as surprise. 98% of organizations now actively manage AI spending, up from just 31% two years ago, but tracking spend is not the same as tying it to value, and the CFOs I work with can see every API call and still cannot answer whether the agent paid for itself.

The Anthropic figure is not an outlier. It is the cost structure becoming visible. The finance owner who receives the bill has two numbers to reconcile: the per-task ROI in the business case, and the minimum annual commitment in the contract. When those numbers were built in different quarters by different teams with different assumptions about scale and usage, the bill nobody forecast is the one that arrives in month four and does not go away.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.