Nearly half of enterprises are over budget on AI, and the money is coming out of your services line

Week of 21 September 2026 | Compiled by Team Paringa

The AI overrun is no longer an anomaly to explain; it is the operating norm. This week's data shows where the gap is being funded, what the model providers are doing about it, and why the FinOps community is converging on Amsterdam to argue about tokens.

What Happened This Week

🔴 Futurum: 46.9% of enterprises over budget on AI, and only one in six slows down

The Futurum Group published its 2H 2026 CIO and Technology Buyers survey on 10 September, drawn from 1,636 enterprise technology decision makers worldwide. The headline: 46.9% of organisations report actual AI spend above plan (35.6% moderately, 11.3% substantially), against 31.8% roughly on plan and just 5.6% under. A further 10.0% have no formal AI budget to measure against at all.

The more useful finding is what happens next. Among the 767 over-plan organisations, 47.6% ask for more budget, 43.3% absorb the overrun and reconcile later, and only 17.2% pause or reduce the initiative. When the gap is closed by reallocating within IT, external labour is cut first: 60.9% of reallocating organisations reduce contractors and consultants. Where AI-driven headcount cuts land, IT support and helpdesk (60.2%) and software and application development (58.4%) are most affected. Notably, 23.1% of over-plan organisations fund the overrun from a non-IT business unit budget, evidence that AI funding authority is moving outside the CIO's control.

Futurum's Mitch Ashley summarised it bluntly: when nearly half of organisations are over budget and only one in six responds by slowing down, "the budget has stopped being a control". His advice to buyers is to "fund the governance layer explicitly rather than let it surface as an overrun".

The figures line up with earlier 2026 data. The FinOps Foundation's State of FinOps 2026 reported 73% of organisations spending more on AI than planned, and DoiT's survey of 500 finance leaders in the US and UK found 79% experienced AI cost overruns in the prior 12 months, with the most FinOps-mature organisations reporting the highest measured overspend (a mean 30.9%) because they are the ones actually instrumented to see it.

Sources: Futurum Group, DoiT, FinOps Foundation State of FinOps 2026

The CFO conversation has changed. It is no longer "why is AI over budget" but "which line did you raid to pay for it". Cutting contractors and support to fund unmetered token consumption is a bad trade, because the overrun recurs every month and the capability you cut does not come back. The organisations getting this right are doing three things: putting an owner on every AI cost line, measuring cost per unit of output rather than total spend, and setting consumption caps before procurement, not after the first shock invoice. Every one of those is a governance change that costs far less than the overrun it prevents.

🟡 Model providers ship cost controls: effort caps, cache diagnostics, gateway pricing sync, pay-as-you-go

FinOps Weekly's 11 September provider round-up captures a fortnight in which the major AI vendors released tooling aimed squarely at spend governance. Anthropic's Claude Code v2.1.267 added a maxEffortLevel administrator setting to cap reasoning effort across providers, plus fixes to prompt-cache invalidation that reduce repeated transmission of tool definitions and system prompts. v2.1.260 now shows likely causes of prompt-cache misses in /cost telemetry, and v2.1.268 lets enterprise gateways define pricing in configuration so that client-side cost reporting matches the gateway spend meter.

Google Cloud made Gemini Enterprise pay-as-you-go broadly available to all projects on invoiced billing accounts from 10 September, removing previous eligibility restrictions and widening the door to usage-based AI consumption. OpenAI's GPT-6 Astra, released via the Responses API, adds adjustable reasoning effort and preservation of cached prompt prefixes, while its new Agents API offers managed orchestration for long-running agent workloads.

The pattern is consistent: reasoning effort, cache hit rate and agent runtime are now the levers that decide the bill, and vendors are exposing them because customers are demanding it. Each control is also a new thing to configure, monitor and govern.

Sources: FinOps Weekly

Every one of these features is a saving only if someone owns it. An effort cap that is never set, a cache-miss diagnostic nobody reads, and a pay-as-you-go tier with no budget alert are how overruns happen with the vendor's blessing. CIOs should ask their platform teams a simple question this week: for each AI provider we use, who sets the caps, who watches cache efficiency, and what happens at 80% of budget? If the answer is a name and a dashboard, good. If it is "the engineers know", that is the 46.9%.

🟡 Tokenomicon + FinOps X Amsterdam opens 22 September, sold out; GPU price gap between hyperscalers and neoclouds widens

The FinOps Foundation's Tokenomicon + FinOps X event runs 22 to 23 September at the Muziekgebouw in Amsterdam, with 500-plus attendees and registration sold out. The event, backed by the Linux Foundation, the new Tokenomics Foundation, the ITAM Forum and FOCUS, is explicitly framed around the gap between two disciplines: FinOps, which asks whether the AI bill matches the plan, and tokenomics, which asks whether the plan can get cheaper without losing quality. Speakers include infrastructure and FinOps leads from Deutsche Bank, Booking.com, Lundbeck, Adevinta, Transavia and UNICC, alongside Microsoft, Google Cloud and AWS.

On the compute side, Thunder Compute's September market analysis puts the GPU rental market at roughly $52bn in 2026, heading for $199bn by 2031. Median on-demand H100 pricing is about $4.17 per hour on dedicated GPU clouds versus $7.89 on AWS, Azure and Google Cloud, an 89% premium for identical silicon, rising to 132% on H200. Reserved one-year commitments save around 25% and three-year around 45%, while spot capacity runs at roughly half of on-demand but can be reclaimed mid-job. Blackwell allocation remains largely reserved for hyperscalers through 2027, and the memory shortage driven by HBM demand continues to push component costs up.

Sources: Tokenomicon, Thunder Compute, Linux Foundation

A sold-out practitioner conference on token economics tells you where the discipline is going: unit cost per token, per inference and per agent run is becoming the core metric, and FOCUS-normalised billing data is the plumbing that makes it comparable across providers. On GPUs, the near-doubling premium for hyperscaler capacity is a procurement decision hiding in plain sight. Not every workload needs to sit next to the rest of your cloud estate, and a hybrid placement strategy with committed spend where utilisation is proven is the fastest hard saving available on AI infrastructure this year.

What CIOs, CFOs and FinOps Leads Are Dealing With

The AI spend problem has stopped being a forecasting problem and become a control problem. The Futurum data shows budgets are exceeded as a matter of course, the response is to find more money rather than manage consumption, and the funding is increasingly coming from outside IT. At the same time, model providers are handing customers more knobs than ever: effort levels, cache behaviour, gateway pricing, agent runtimes, pay-as-you-go tiers. The gap between what can be controlled and what is actually controlled is where the overrun lives.

FinOps teams built for reserved instances and rightsizing are being asked to govern token flows, agent loops and GPU placement across hyperscalers and neoclouds, often with no single owner for the AI line and no agreed definition of success between finance and engineering. The organisations that close that gap first will be the ones that can still fund the next initiative without cutting the people who deliver it.

Three Things Worth Watching Next Week

1. Tokenomicon + FinOps X Amsterdam, 22 to 23 September - Watch for announcements on the Tokenomics Foundation, FOCUS updates for token-level billing data, and the practitioner benchmarks on cost per inference that will set the vocabulary for 2027 budgets.

2. Q3 quarter-end AI invoices - With 46.9% of enterprises already over plan, September close is when finance teams see whether the summer's agent pilots have become a recurring line. Expect budget reallocation requests to spike in the first fortnight of October.

3. Broadcom's vSphere Standard refresh and VMware renewal cycles - Broadcom has signalled a return to lower-end virtualisation bundles after admitting it pushed VCF too hard. Any renewal landing this quarter is a leverage point to reprice, and to test alternatives, before committing to multi-year subscription terms.

Questions? hello@paringa.ai

Paringa helps CIOs and CFOs put an owner, a unit cost and a cap on every AI and cloud line, with partner Atmoz providing the continuous optimisation engine inside a managed service. The goal is simple: fund the next initiative from savings, not from the services budget. | 21 September 2026 | © Paringa Ltd