A group buys it. The platform or ML engineering lead usually chooses the product. The CTO, VP of Engineering or a business unit head owns the budget. Security, procurement and finance can each stop the deal. Since 2025 the money sits in central IT and business unit budgets: innovation budgets fell from 25% to 7% of LLM spending (a16z, June 2025).
Which budget pays for AI infrastructure now?
Permanent budgets. In its survey of 100 CIOs across 15 industries, a16z found that innovation budgets had paid for a quarter of LLM spending a year earlier. By the 2025 survey that share was 7%, and the rest came from centralised IT and business unit budgets. The same CIOs expected their LLM spending to grow about 75% over the following year.
This changes who you sell to. An innovation lead could sign a pilot alone. A line item in an IT or business unit budget has an owner who must defend it at the next planning cycle. That owner wants a cost forecast, a security review and a name to call when the system fails.
The spend splits across several layers, and each layer has a different owner. Menlo Ventures estimated 2025 enterprise infrastructure spending at $18 billion (Menlo Ventures, December 2025):
| Layer (Menlo, 2025) | Spend | Typical owner inside the buyer |
|---|---|---|
| Foundation model APIs | $12.5 billion | CTO or central AI team, often through a cloud contract |
| Model training infrastructure | $4.0 billion | ML or research engineering lead |
| Data storage, retrieval and orchestration for AI | $1.5 billion | Data platform or data engineering lead |
The owner column is our reading of how these purchases run; Menlo's data covers spend only. Check it account by account.
Who sits in the buying group?
Plan for six roles in an AI infrastructure deal. Not every deal has all six, and in a company of a few hundred people one person can hold three of them.
| Role | Usual titles | What they decide | What they need from you |
|---|---|---|---|
| Champion | Head of ML platform, staff engineer, director of AI engineering | Whether the product works on their workload | A test on their own data with a pass mark |
| Economic buyer | CTO, VP of Engineering, business unit general manager | Whether to fund it from their budget | Cost per request now and after, and a forecast at their volumes |
| Security reviewer | CISO, security architect, AI governance lead | Whether data can leave their environment | Data flow diagram, deployment options, completed questionnaire |
| Procurement | Procurement manager, vendor management | Contract terms and vendor approval | Standard terms, insurance, pricing that fits their contract types |
| Finance | FinOps lead, finance business partner | Whether usage cost is predictable | Usage caps, alerts, a model of spend at two volume levels |
| User team lead | Head of the product or business team that ships the AI feature | Whether the feature can launch on your system | Launch date fit and a support contact |
Plan for the security review from the first call. a16z reported that security was heavily emphasised in its CIO interviews, and that security and cost had gained ground on accuracy as reasons to pick a model. Pricing matters too: most of the CIOs preferred usage-based pricing. That puts a finance reviewer in the room, because usage pricing moves with traffic.
Who signs, and who can stop the deal?
The economic buyer signs. Any reviewer can stop it. A champion can carry a deal through the technical test and still lose it at security or procurement, often months later. Ask the champion two questions on the first call: who signed the last infrastructure contract of this size, and what stopped the last one that failed.
Switching costs add a second reason to buy carefully. a16z found that agentic workflows were making it harder to change models, and that a switch could take a lot of engineering time. Buyers know your product will be hard to remove once their workflows depend on it, so they review it more before they sign.
Does the buyer change outside the engineering team?
Yes, and the share outside engineering is growing. Indeed Hiring Lab counted 822 US job titles touched by AI in the first quarter of 2026, 8.3% of all titles, up from 264 in the first quarter of 2022 (Indeed Hiring Lab, July 2026). Of those AI-touched titles, 63% sat outside tech occupations.
For an infrastructure seller, this means the user team often sits in operations, support, marketing or finance. That team wants the AI feature live, while engineering runs it. When a business unit pays, the general manager becomes the economic buyer and the platform team becomes a technical reviewer. Map both before you write the proposal.
How do you find the right names at an account?
Start from the system, then find the people. Read the company's engineering blog, conference talks and job posts to learn which team runs its models and its data pipelines. Then search for the leaders of that team.
Check every title before you write. Contact databases attach titles to the wrong company. In our own data we found a co-founder title that named one software firm sitting on the record of a different firm (our data, October 2026). We now drop any title that names a company other than the account. A wrong title sends your first message to someone with no say in the deal.
In what order should you contact the group?
Contact them in the order the deal will need them:
- Champion first. Confirm the workload, the metric and the pain, in their numbers.
- Economic buyer second, introduced by the champion, with a cost case at the buyer's volume.
- Security third, and early. Send the questionnaire and data flow before the technical test ends.
- Finance with the proposal. Bring the usage model and the caps.
- Procurement last, with standard terms ready, so legal review runs alongside the final test.
Each additional contact needs a reason tied to their role. A copy of the champion's email does not count. We cover how to price meetings with these buyers in how much to pay per qualified meeting.
A decision rule for multi-threading
Before you send a proposal, name one person for each role in the second table, or write down why the role does not apply at this account. If you cannot name the economic buyer and the security reviewer, the deal is still a pilot and should be forecast as one.
To map the buying group at your top accounts, start with The Diagnostic at $2,500.