AI strategy is clearest in the invoices

Ask an executive about AI strategy and you get the announcement version. Ask the finance lead which invoices grew this year and you get the strategy. The two answers disagree more often than either party would like, and when they do, the invoices win.

NVIDIA is the clearest recent case. Model releases got the headlines in 2026. Network switches got the capital. NVIDIA's networking line is now a quarterly number larger than most software companies earn in a year, and it got there by selling the plumbing that sits next to its chips. Follow the invoices and the strategy is obvious. Follow the keynotes and you would have missed it.

Spending is a commitment, messaging is a hope

A budget line with an owner and a renewal date tells you what a company has decided to depend on. A slide tells you what it wants you to believe.

That distinction pays off in two places. In a portfolio review, a company's AI-related invoices tell you which capabilities the management team has bet on and which ones live only in the board deck. In your own budget, the same sort tells you if the program can feed the models it has bought.

The Monday-morning sort

Pull the last twelve months of AI-related spend and put each line into one of two piles.

1. Licenses: model access, seats, platform subscriptions, anything you could switch off without a person noticing for a week.
2. Plumbing: integration work, scheduled exports, field mapping between systems that disagree, credential and permission work, the contractor who connected the CRM to the ERP.

Then answer three questions.

• What is the ratio? A program that is all licenses and no plumbing has bought capability it has no way to feed.
• Which pile grew faster this year? Growth in the plumbing pile means somebody has met the real constraint and paid to remove it.
• Which invoices sit next to your product? Adjacent spend is where the next strategic move is hiding.

Ten minutes with the AP ledger. No roadmap required.

Where the capital went in Q1

Source facts, from IDC's 1Q26 Ethernet switch tracker and NVIDIA's Q1 FY27 results:

• The worldwide Ethernet switch market reached $15.4 billion, up 39.8% year over year. The datacenter segment inside it grew 61.0% to $10.0 billion.
• NVIDIA became the number one vendor by revenue in datacenter Ethernet switching for the first time. Its Ethernet switch revenue rose 192.7% to $2.1 billion, a 21.5% share of the datacenter segment.
• NVIDIA reported data center networking revenue of $14.8 billion for the quarter, up 199% from a year ago, inside total revenue of $81.6 billion.

My read: a chip company took the top spot in a switching market from incumbents built to sell switches. It did that by moving into the constraint next door, the network its own GPUs choke on. That is the adjacency move, and it showed up in invoices well before it reached a keynote slide.

The adjacency move works at company scale

The same move is available to a mid-market company, and it gets overlooked because it looks unglamorous.

A distributor whose customers keep calling about lead times is sitting next to a service. A manufacturer answering specification questions by email eleven times a week is sitting next to a data product. Read what customers ask for that you give away by hand. Structure it first. Charging for it is a separate conversation.

An illustrative example, not a client case: a $40M industrial distributor sorts its AI spend and finds $180,000 in licenses and $12,000 in integration. The models are fine. Nothing connected to them has current data. The next dollar belongs in plumbing, and the invoice that proves it will come from a contractor rather than a vendor.

The line item most proposals leave out

The same blindness shows up in AI proposals. They price the license and the build. Few price the connections: the export that has to run on a schedule, the field mapping between two systems that disagree, the permission that needs a person who left in March. Those weeks are where projects slip, and the slip stays silent because no plan named them.

Put integration on the plan as its own line with its own owner, before the model work. It makes the estimate look worse in the room and makes it true.

Licenses get announced. Plumbing gets paid for. The market priced that work at $14.8 billion last quarter. If a proposal on your desk prices it at zero, the work did not disappear. It moved into the part of the timeline no status report tracks.

Sort your own spend into the two piles. Which one grew faster in 2026? I am collecting answers, and the ratio you find will tell you more about your AI strategy than the roadmap does.