How much are we actually spending on AI?

The question is: how much are we actually spending on AI?

The honest answer in most organisations is that nobody knows. AI spend is spread across SaaS contracts, individual subscriptions, Copilot licences, embedded features, shadow tools, and pilot projects that quietly turned into production workloads. It is not a line item. It is fifty line items pretending to be background noise.

I want to be direct about this. Most leaders reading this are about to get caught flat-footed. The AI bill is about to spike, the spike is going to be larger than anyone is forecasting, and the reasons are sitting in plain sight in the earnings reports nobody is reading carefully.

This is the article I would want to read if I were running AI strategy for a mid-market or enterprise business right now. Not because the news is exciting. Because the math is not subtle, and the leaders who do not pay attention this quarter are going to spend 2027 explaining variance to their boards.

The number nobody is sitting with

Last week, four companies confirmed they will spend a combined $725 billion on AI infrastructure in 2026. That figure is 77% higher than 2025. Analysts expect 2027 to exceed $1 trillion across the same four companies.

If you are reading this and your reaction is "interesting macro news, but it doesn't apply to me," you are exactly the audience I'm trying to reach.

These companies are not running charities. That $725 billion has to come back, with margin attached. The path from their capex to your line items is so short that I do not understand why almost nobody is connecting the two halves of the equation publicly.

Let me connect them.

How they will get the money back

Hyperscalers have three obvious paths to recoup capex. All three are already in motion. Most enterprise leaders are watching one of them and missing the other two.

The first one is loud and easy to forecast. The second one is silent and consequential. The third one is happening right now in front of you and most leaders are interpreting it as a procurement optimisation rather than a revenue strategy.

Here is the uncomfortable truth. Microsoft did not launch E7 because customers were asking for a bundle. They launched it because bundles consolidate procurement, raise average revenue per user, and lock in long-term contracts. AWS and Google Cloud will launch their own versions of E7 in the next 12 months. They have to. The capex math demands it.

What I'm watching most leaders get wrong

Three patterns I'm seeing in the conversations I'm having, written down because someone has to say them.

Pattern one: leaders are treating AI spend as a technology line item when it's becoming a strategic line item. AI spending in 2024 was small enough to ignore. By 2027 it will be a top-five operating expense for most knowledge-economy businesses. Treating it as IT-managed and IT-reported is going to fail badly. It needs the same scrutiny finance gives to headcount, real estate, and marketing.

Pattern two: leaders are over-indexing on the AI tools they bought and under-indexing on the AI features that came embedded in things they already pay for. Your sanctioned vendors quietly became AI vendors without renegotiating the contracts. Those features were free during the customer acquisition phase. They are getting repriced now. Your real AI spend is much higher than your AI line item suggests.

Pattern three: leaders are assuming vendor concentration is fine because the vendor is reliable. This is the mistake that will hurt the most. If 70% of your AI spend is locked to one hyperscaler, you have a pricing power problem dressed up as a partnership. When that hyperscaler raises prices in 2027, your options are pay or move. Moving will take 18 months. You won't have 18 months.

The conversation you should be having this week

Three questions to bring to your next leadership meeting. They are uncomfortable on purpose.

Most leadership teams cannot answer any of these three questions in the room. That is the problem. Not because the answers are hard to find. Because nobody has been asking.

What I would do if I were running AI strategy in your seat right now

Three actions, in order of priority.

Run a real audit this month, not next quarter. Pull every SaaS contract, every Copilot licence, every API key with billing attached, every team subscription. Add it up. The first version of the number will be embarrassing. That is the point. You cannot manage what you cannot see, and most enterprises are operating with deliberate blindness because the number is uncomfortable.

Tier your workloads before someone else tiers them for you. Identify which AI use cases produce real business outcomes and which ones are productivity theatre. The hyperscalers are about to tier their own pricing, and the unit economics of your low-value workloads are going to break. Better to retire those workloads on your terms than have the math retire them for you.

Build vendor diversification into your 2026 roadmap. Not as a price negotiation tactic. As a long-term resilience play. The companies that get to 2028 with concentrated vendor exposure are going to be the ones explaining bad quarters. The ones with diversified AI vendor relationships will have meaningful pricing leverage.

The honest bottom line

I am not writing this to predict doom. AI is going to keep getting better. Productive uses are going to keep producing returns. The companies that use AI well are going to keep pulling ahead.

What I am saying is this. The era of cheap, loosely tracked, vendor-subsidised AI is ending. Anyone who tells you otherwise is selling you something or hasn't read this week's earnings reports carefully.

The leaders who do the audit this quarter, who have honest conversations about vendor concentration, who tier their workloads before the market forces them to, are going to spend 2027 with options. The leaders who don't are going to spend 2027 with surprises.