Written by: Agata Wojtas, Chief Commercial Officer, Digital Colliers
Meta's internal projection that it could spend up to $10B a year on Anthropic models is the kind of number that reframes a boardroom conversation. Not because most banks will ever spend that much. Because a hyperscaler with Meta's own model stack is still modelling worst-case vendor concentration on a single AI supplier. If they're doing that math, your CFO should be able to pull the same view for your bank on demand.
Most can't. And that's the gap worth closing before the next budget cycle, not after.
The question your board will ask in 2026
It's some version of this: what are we spending on AI, per vendor, per business line, per use case, and what happens to the P&L if one of those vendors doubles its prices or goes down for a week?
Right now, at most mid-market banks, answering that question is a two-week project involving three spreadsheets and a Slack thread. The Meta number matters because it forces the question into the open. Vendor concentration on a foundation model provider is now a real line item, not a footnote. Boards that have watched cloud spend balloon over the last decade know exactly how this movie ends if nobody's measuring.
And the regulator is already in the room. DORA has been in force since 17 January 2025, and it treats critical ICT third-party providers as a supervised concern. A foundation model vendor sitting inside your credit, KYC, or customer-service stack is exactly that.
Why your current close can't answer it
Here's the uncomfortable part. Month-end close at mid-market finance teams typically runs 8 to 10 days, and most of that work still happens in spreadsheets, with people pulling numbers across systems by hand. Best-in-class teams close in under 5 days, and even they aren't usually cutting AI vendor spend by use case.
So when the board asks the question in the March meeting, the honest answer is often: we'll get back to you in April. By which point the number has moved.
The plumbing problem is concrete:
- AI vendor invoices land as a single line in AP, tagged to IT or to a cost centre that owns the contract, not to the business unit consuming the tokens.
- Token usage lives in the vendor's console, in a format nobody in finance logs into.
- The mapping between a prompt call and a revenue-generating product is buried in application logs.
- Nobody owns the reconciliation between what engineering thinks it's spending and what AP actually paid.
Until those four things join up, you don't have an AI cost view. You have an AI cost guess.
What a live dashboard actually needs
The operators getting ahead of this in 2026 tend to do four things in parallel. None of them are exotic.
- Tag every API call at the source. Business unit, product, use case, environment. If it's not tagged when the call is made, no downstream pipeline will save you.
- Pull vendor usage data daily, not monthly. Anthropic, OpenAI, Azure OpenAI, Bedrock all expose usage APIs. Pipe them into the same warehouse your finance team already trusts.
- Reconcile usage to invoice weekly. Small deltas early are cheap. A 20% delta discovered at year-end is a board paper.
- Model the concentration scenario explicitly. What's the cost impact if your primary model vendor raises prices 40%? What's the operational impact if they're unavailable for 72 hours? Have the number ready before someone asks.
This is not a data science project. It's finance plumbing with an AI label on it.
The cost of not having the number
Around 95% of enterprise AI projects fail to reach production or ROI. A meaningful share of that failure is invisible spend on pilots that quietly kept running. If you can't see per-use-case cost, you can't kill the losers, and you can't scale the winners with any confidence.
The board question is coming. DORA is already live. The Meta signal tells you the sophisticated buyers are already modelling concentration risk on foundation models as a first-class financial exposure.
The teams that walk into the next audit with a live per-vendor, per-use-case view spend the meeting talking about strategy. The teams that don't spend it explaining why they can't answer the question.