Maria has been a contracting officer for the Department for six years, and this is her first AI infrastructure acquisition. Her program needs compute capacity to support a fine-tuning workload for a document analysis tool the agency has been piloting for eight months. The pilot worked. Now leadership wants to scale it, and Maria has been asked to make that happen by the end of the fiscal year.
Here’s what that actually looks like, week by week.
Maria’s first instinct, like most people new to AI procurement, is to start pulling vendor quotes. Her program’s technical lead, a data scientist named James, stops her before she gets too far.
“We need to actually define the workload first,” he tells her. “Is this training, fine-tuning, or just running inference on the model we already have? Because those are three completely different hardware conversations.”
It’s fine-tuning, it turns out, updating the existing document analysis model with new agency-specific data every quarter, not training something from scratch. That distinction alone rules out the most expensive tier of infrastructure vendors keep suggesting. James also flags something Maria hadn’t considered: their current model runs comfortably on a mid-range GPU configuration, but their planned expansion into processing scanned handwritten documents will need substantially more VRAM. Undersizing here, James warns, is the single most common and most expensive mistake he’s seen other agencies make.
By end of day Monday, they have a real answer: a fine-tuning workload, a defined VRAM floor, and a three-year growth horizon that assumes two more document types get added to the pipeline.
Maria loops in her agency’s security team Tuesday morning, later than she should have, she admits to herself. The system will process Controlled Unclassified Information once it’s fully deployed, which means it needs to go through the Risk Management Framework and eventually receive an Authority to Operate.
Her security lead asks a question that catches her off guard: “Does the hardware you’re looking at support TPM-based attestation? Because if it doesn’t, we’re going to hit a control implementation gap during assessment, and that’s going to delay your ATO by months, not weeks.”
Maria makes a note to add this to every vendor conversation going forward. She also learns, somewhat uncomfortably, that if this had come up after she’d already selected hardware, they might have been looking at a mid-program hardware swap instead of a straightforward specification requirement.
Maria talks to three hardware vendors Wednesday. Two of them pitch her a standard configuration within the first ten minutes of the call, before she’s finished describing the workload.
The third vendor asks questions first. What’s the actual VRAM requirement James identified? What’s the compliance categorization, and has her security team weighed in on hardware-level control requirements yet? What does the three-year growth horizon actually look like in terms of additional document types and expected data volume?
It’s the same conversation Maria had with James on Monday and her security lead on Tuesday, except this vendor is asking it back to her, which tells her something. This is a vendor who’s going to specify a configuration based on her actual requirement, not whatever configuration is sitting in inventory and easiest to sell.
Thursday brings a question from Maria’s contracting specialist that nearly derails the timeline: has TAA compliance documentation been requested from the shortlisted vendors yet?
It hadn’t. Maria assumed, incorrectly, that this would be handled during the formal proposal stage rather than something she needed to request upfront. Her contracting specialist explains that waiting until after a purchase order is drafted to verify compliance is exactly the kind of gap that can stall an acquisition for weeks, and that requesting a Manufacturer’s Certificate of Compliance now, while vendors are still competing for the business, gets a faster, more complete answer than requesting it later when there’s less incentive to respond quickly.
Maria sends the request Thursday afternoon. Two vendors respond within a day with full documentation. One takes over a week and eventually admits the specific configuration they proposed uses a component sourced from a country that isn’t TAA-compliant, information that would have been far more disruptive to learn after an award decision than before one.
By Friday, Maria has what she needs to make a real recommendation: a workload assessment that ruled out an expensive, unnecessary hardware tier, a compliance conversation that surfaced a hardware-level security requirement before it became a mid-program problem, a vendor who asked the right questions instead of leading with a pitch, and TAA documentation that eliminated one vendor from consideration for a reason that would have been far more painful to discover after signing a contract.
None of this happened because Maria followed a complicated, novel process. It happened because she asked the right questions in roughly the right order, workload first, compliance early, vendor evaluation grounded in both, and TAA verification before a purchase order rather than after one. The programs that struggle with AI infrastructure procurement aren’t usually failing at something exotic. They’re skipping a step that felt like it could wait, and discovering later that it couldn’t.
Maria’s recommendation memo goes to her contracting officer Monday morning. It’s a smaller, more specific configuration than any of the vendors initially proposed, and it clears every compliance question her security team is going to ask months from now, not just the ones on this week’s checklist.
Every procurement officer’s actual week looks different from Maria’s, different workload, different compliance category, different vendor conversations. But the underlying sequence, define the workload, surface compliance requirements early, evaluate vendors against both, verify TAA compliance before a purchase order, holds up across nearly every federal AI infrastructure acquisition, regardless of program size or complexity.
If this week feels familiar, or if you’re staring down a similar acquisition and want a second set of eyes on where you are in that sequence, our federal engineering team is available to talk through it.
→ Contact Ace Computers Federal Sales Team
→ View Federal and Government IT Solutions
→ Read the Complete Guide to Federal AI Infrastructure Procurement
