Zero-Based Budgeting for Cloud and AI: Why 2026 Ended the Last-Year-Plus-Growth Era

The CFO approved the AI budget in January. By April it was 40% over, and no one in the room could say which workload caused it.

That scene is playing out across enterprises in 2026, and it exposes a budgeting model that’s quietly expired. For decades, technology budgets were built the way most budgets are built: take last year’s run rate, add a growth allowance, defend the delta. The base was assumed; only the increment got scrutinised. That worked when infrastructure cost behaved like rent — stable, predictable, contracted. It doesn’t work when a single prompt costs anywhere from a fraction of a cent to fifty cents, when a GPU cluster sits idle 95% of the time while the invoice looks perfectly steady, and when 98% of FinOps teams are now managing a cost category that barely existed three budget cycles ago.

Zero-based budgeting — justify every expense from a blank slate, each period, against the business outcome it delivers — was conceived in the 1970s. What’s new is that cloud and AI have made it the only budgeting discipline that survives contact with the 2026 technology estate.

80-85% of organisations miss AI cost forecasts by more than 25%. AI is now 19% of total cloud spending, up from 8% in 2023 — the fastest-growing cost category is the one the roll-forward model handles worst. As much as 95% of enterprise AI infrastructure spend is wasted at low GPU utilisation, and the underutilisation penalty on identical hardware can run 24x to 36x per token. Only around 29% of executives can confidently measure AI ROI, against roughly 79% who perceive productivity gains — the value is real, but the budget defence is broken.

What zero-based budgeting means for a technology estate

Zero-based budgeting rejects the single assumption underpinning most corporate budgets — that last year’s spend is a legitimate starting point. Every line begins at zero; every line earns its funding by demonstrating alignment with a strategic outcome. The principle is unglamorous and, in its original form, expensive to run — only about half of organisations implementing it realise the savings they expected, often because they treat it as a procedural change to the budgeting calendar rather than a principle for making better spending decisions in the first place. The failure is rarely the method. It’s applying the method to the wrong things, or applying it once a year to a cost structure that changes daily.

Technology spend is exactly that cost structure. When the estate was a fixed set of servers and a handful of SaaS contracts, an annual zero-base was feasible if painful. In 2026 the estate is multi-cloud infrastructure, Kubernetes clusters, token-metered inference, Databricks and Snowflake workloads, agentic pipelines, and data centre — and that scope has expanded faster than any organisation can realistically staff for by hand. Zero-basing it manually isn’t slow. It’s impossible. The discipline has to be automated into the platform governing the spend, or it simply doesn’t happen.

Why the roll-forward budget breaks in 2026

Roll-forward budgeting assumes a stable base. AI spend has no stable base — a single request can vary in cost by two orders of magnitude depending on model, context length, and output size, and that variability multiplies across thousands of daily requests and multiple vendors adopted independently by different teams. The result is a forecast wrong on arrival: 80-85% of organisations miss AI forecasts by more than 25%, and G1000 organisations are projected to face a 30% rise in underestimated AI infrastructure costs by 2027, driven not by reckless spending but by under-forecasting expenses unique to AI workloads specifically.

The scale of exposure is what makes this a board problem rather than a FinOps inconvenience. Global AI spending is forecast at $2.5 trillion in 2026, a roughly 44% year-over-year jump, with AI now representing 19% of total cloud spending against 8% in 2023. The fastest-growing category in the enterprise cost base is the one the incumbent budgeting model governs least well.

There’s a deeper structural reason the model fails, and it’s the one most forecasts miss entirely: the token invoice — the number finance actually sees — is only one of roughly nine distinct AI cost buckets identified in recent industry analysis. Inference infrastructure, idle GPUs, the KV cache the invoice never shows, evaluation and monitoring, model-churn integration cost, and more sit beneath the visible line. Organisations that can’t measure their AI ROI underestimate AI cost by 40-60% on average, because they count the licence and miss the iceberg underneath it. A budget built by rolling forward last year’s visible number therefore inherits last year’s blind spots and compounds them.

The GPU cluster that costs far more than the invoice suggests

Consider the most expensive thing in a modern AI estate that no roll-forward budget will ever flag: an idle GPU. At batch size one, a GPU spends most of its time waiting to load model weights for a single active request, and utilisation falls below 5%. The invoice for that GPU-hour looks identical whether the hardware is saturated or nearly asleep. What changes invisibly is the cost per useful token — on identical H100 hardware, effective cost can span $0.21 to $15.25 per million output tokens depending on offered load, an underutilisation penalty of 24 to 36 times. A GPU at 10% load turns a $13-per-million-token workload into a $130 one. Estimates place total wasted enterprise AI infrastructure spend as high as 95% at low utilisation.

The Idle GPU the Invoice Hides
──────────────────────────────────────────────────────────
What finance sees              What the workload actually costs
──────────────────────         ──────────────────────────────────
GPU-hour: stable               Batch size 1 -> utilisation < 5%
Monthly invoice: flat          Cost/M tokens at 10% load: $130
Line item: "AI compute"        Cost/M tokens at target load: $13
──────────────────────────────────────────────────────────
Same hardware. Same invoice. 10x the cost per useful token.
Roll-forward budgeting cannot see this. A zero base can.
──────────────────────────────────────────────────────────

This is where tokenomics enters the budget conversation — the unit economics of token-metered workloads: cost per million tokens, input/output ratio, cached-token ratio, GPU utilisation, and how all of it drifts over time. The optimisation levers are real and compounding — routing requests to right-sized models, quantisation, caching, continuous batching — but they only pay off if every layer is instrumented; miss one layer and the savings collapse at the others. The point for a budget owner is that none of this is legible from the top line. The token bill doesn’t reveal the cached-token ratio. The cloud invoice doesn’t reveal the idle cluster. You can’t zero-base what you can’t attribute, and you can’t attribute token, GPU, and data-platform cost to a named workload with a tool built to reconcile monthly cloud invoices. That gap between what the invoice shows and what the workload actually costs is precisely why so many 2026 AI budgets miss by 25% or more.

Zero is the method. Value is the target.

The most common way zero-based budgeting fails is subtle: teams treat it as a hunt to spend less. Unfocused zero-based exercises become opportunistic, seeking any opportunity to cut spend with no concern for business impact, and organisations that budget by business value instead of cost alone achieve over 30% more profitability or market-share growth. The name misleads. The goal was never zero. The goal is funding what returns value and defunding what doesn’t.

For technology, that reframing is the entire game in 2026. Zero-basing a GPU cluster tells you the spend is unjustified if it sits idle. Value-basing it tells you whether the workload it runs actually earns its keep — cost per inference against the revenue that inference generates, or cost per verified outcome, clearing the bar or not. Boards have stopped accepting efficiency projections as proof of value; the directive to finance in this budget cycle is to show measurable return, not promises. The uncomfortable truth is that most AI probably does generate value, and most organisations simply can’t prove it, because they never set up the measurement in the first place.

This is why zero-basing and value-basing are a sequence rather than a choice. Pressure-test each workload for strategic alignment first — does it map to an outcome anyone can actually name? Then pressure-test its cost. A workload with a named owner, a measurable outcome, and a defensible cost per unit survives the zero base. A workload that’s none of those things is exactly what the discipline exists to surface. The organisations keeping their AI budgets intact through the 2026 cycle aren’t the ones with better models. They’re the ones that baselined before they deployed, attributed every dollar to an owner, and can express the result in the P&L vocabulary a board actually accepts — cost per transaction, cost per customer, margin impact — rather than tokens and GPU-hours no director thinks in.

How to zero-base cloud and AI spend

Five moves, each a prerequisite for the next. Normalise the full estate first — a zero base is impossible if every data source is a separate conversation, so cloud, Kubernetes, AI workloads, Databricks, Snowflake, SaaS, and data centre all need to resolve into one specification before any of them can be justified against a common outcome. Attribute every dollar to an owner — the share of spend with a named owner is a core maturity marker, because unattributed spend can’t be justified from zero at all; for AI specifically this means per-chain, per-model, per-environment attribution, not a shared account where GPU and egress cost quietly disappears. Build decision packages with kill/continue criteria — each workload gets explicit thresholds (token budget caps, GPU idle limits, commitment coverage targets) enforced by policy rather than a quarterly spreadsheet review, funding releases through kill/continue gates rather than full transformations upfront. Rebuild from outcomes, not run rate — the budget assembles from cost per inference, cost per customer, and cost per verified outcome, rather than last year’s number plus a percentage. And govern continuously — consumption-based spend drifts within the period itself, so real-time anomaly detection and automated enforcement keep the zero base valid after planning ends, making the discipline a live control system rather than an annual event obsolete by February.

What this requires operationally

Every step above assumes two capabilities most cost tools simply don’t have: normalisation across the full technology surface, and attribution down to the individual AI workload. Because a FOCUS-native platform is native to the specification rather than adapted to it, it can normalise every cloud provider alongside Kubernetes, the major AI providers, data platforms, SaaS, and data centre into a single attributable view — the zero base covers the full estate, not the 60% a single-category tool reaches before Kubernetes and data platforms fall out of scope. Native AI governance — token budget caps, agentic kill-switches, GPU idle detection, per-chain attribution — gives finance the workload-level justification a zero base actually demands, and real-time anomaly detection surfaces cost drift before the invoice lands, so the base holds through the period rather than just at the planning meeting.

The commercial model matters here more than it might first appear. A platform priced as a percentage of the spend it governs has a structural incentive misalignment — it profits as the bill grows. Flat enterprise licensing that scales with the estate rather than the spend keeps vendor incentives aligned with the cost reduction a zero base is actually meant to produce. For regulated industries where billing data can’t leave the perimeter, BYOC deployment keeps the collector inside the customer’s own environment. A European energy utility used the same underlying capabilities to remove roughly €1M from its Databricks estate — a zero base applied to a data platform a cloud-only tool would never have reached in the first place.

Roll-forward budgeting assumes last year was right. Cloud and AI made that assumption expensive — the discipline of 2026 is to fund what returns value and defund what can’t prove it, every workload, every period, from zero.

See what a zero base looks like across your full estate. Book a 30-minute session and we’ll map real per-workload attribution against your actual technology footprint. Request a demo →

Frequently asked questions

What is zero-based budgeting for cloud and AI? It’s the practice of justifying every unit of technology spend from a zero base each budget period against a measurable business outcome, rather than carrying forward the prior period’s spend plus a growth allowance. Applied to technology, the principle is unchanged from its original definition, but the cost structure is harder — cloud and AI spend is consumption-based and non-linear, so a token-metered workload or an idle GPU can’t be modelled as a fixed line. It requires per-workload attribution, since unattributed spend can’t be justified from zero at all.

Why does roll-forward budgeting break specifically for AI spend in 2026? Because AI cost is consumption-based, non-linear, and largely invisible in the systems finance uses to plan. A single prompt can vary in cost by two orders of magnitude, and that variability across thousands of requests and multiple vendors makes a static forecast structurally unreliable — 80-85% of organisations miss AI forecasts by more than 25%. The token invoice is only one of roughly nine AI cost buckets, so a forecast built on it alone omits most of the actual cost.

What’s the difference between zero-based and value-based budgeting? Zero-based budgeting rebuilds the budget from a blank slate, requiring a business case for every expense. Value-based budgeting ranks every justified expense against the value it returns. Organisations budgeting by business value rather than cost alone achieve over 30% more profitability or market-share growth. For AI, zero-basing tells you a cluster is unjustified if it sits idle; value-basing tells you whether the workload it runs earns enough business outcome to deserve funding at all.

How much cloud and AI spend do enterprises waste on idle GPUs? Estimates place wasted enterprise AI infrastructure spend as high as 95% at low GPU utilisation. At batch size one, utilisation falls below 5%, and on identical H100 hardware effective cost can span $0.21 to $15.25 per million output tokens depending on load — a 24x to 36x penalty. This stays invisible to roll-forward budgeting because the invoice looks stable while cost per useful token multiplies underneath it.

How should an enterprise apply zero-based budgeting to its technology estate? Five steps: normalise every cost source into one specification, attribute every dollar to a named owner, build decision packages with kill/continue criteria, rebuild the budget from business outcomes rather than run rate, and govern continuously with real-time anomaly detection. Since scope has expanded faster than most teams can staff for, this genuinely can’t be a manual annual exercise — it needs automated normalisation, attribution, and enforcement built into the platform itself.

References

• FinOps Foundation — State of FinOps 2026
• Gartner — Use Zero-Based Budgeting to Rightsize Your Budget
• VentureBeat — Enterprise GPU Utilization: The AI Infrastructure Problem
• arXiv — Beyond Per-Token Pricing: A Concurrency-Aware Methodology for LLM Infrastructure Cost Estimation
• CIO Dive — FinOps Teams Gain Clout as AI Costs Climb

Related reading

• AI Unit Economics: Why Your AI Invoice Can No Longer Answer the Board’s Question — July 16, 2026
• Why Cloud and AI Budgets Fail Annual Planning — May 28, 2026
• Board-Level AI ROI: Why $600B in Investment Is Delivering Single-Digit Returns — and What Fixes It — May 21, 2026
• Cloud Cost Optimization Is Dead. Long Live Technology Value Management — May 5, 2026

© 2026 DigiUsher. All rights reserved. Privacy Policy · Terms of Service · Trust Center