Due diligence on an AI infrastructure and GPU services business

Diligence on a GPU hosting or compute infrastructure business follows the same broad shape as any acquisition — financials, contracts, corporate standing — but a handful of checks specific to this category matter more here than in most small-business purchases, and skipping them is where buyers get hurt. A generic diligence checklist covers the basics; this category needs a few additions built specifically around the hardware and the agreements that keep it running.

Search for security interests on the hardware

Compute hardware is commonly financed or leased, and a lender or lessor with a registered security interest has a claim on that equipment that survives a change of business ownership unless it’s specifically dealt with. In Ontario, this means a search of the registry maintained under the Personal Property Security Act; other provinces run similarly structured registries under their own equivalent legislation. Where a security interest turns up, the purchase agreement needs to address whether it will be paid out at closing, assumed by the buyer, or otherwise resolved — silence on the point is not a resolution.

Confirm what’s actually assignable

Pull the data-centre lease, the colocation contract and any power agreement and read the assignment provisions directly rather than relying on a summary. Many of these agreements require the landlord’s or provider’s consent before they can be assigned to a new owner, and a buyer needs to know, before closing, whether that consent has been sought, whether it’s been granted, and what happens to the deal if it isn’t. The same review applies to customer compute contracts — confirm they’re actually assignable to a new owner and not personal to the current one.

Test how firm the customer commitments really are

Read the customer contracts for minimum commitment language, term length and cancellation notice, and compare that language against what the seller has represented about revenue stability. A contract with no minimum commitment and a short cancellation notice period is a different asset than a multi-year reserved-capacity agreement, even if both are currently generating similar revenue, and a buyer should weight recent utilization trends more heavily where commitment language is weak.

Assess the hardware’s real remaining competitive life

A technical review of the hardware fleet — its generation, condition and age relative to what’s currently available in the market — tells a buyer more about future earning power than the seller’s description of it as being in good working order. Equipment that still functions perfectly can still be approaching the point where customers prefer to buy capacity elsewhere, and a buyer needs a realistic view of how many years of competitive service the fleet has left, not just how many years it will physically keep running.

Verify data-residency claims and check the software IP chain

Where the business markets itself on data-residency or sovereignty commitments to regulated or government customers, don’t take the claim at face value — ask for the technical documentation showing where data actually resides and confirm it matches what customer contracts promise, since a gap between the two is a real liability a buyer inherits the moment the deal closes. Review the business’s documented role and obligations as a data processor under federal privacy law, and where any customer relationships cross an international border, confirm whether the hardware or the relationship has been assessed against any applicable regulatory screening. Separately, where the business has built any proprietary orchestration or scheduling software, trace the ownership chain the same way a buyer would for any AI product — pull every contractor and vendor agreement that touched the code and confirm each one actually assigned the resulting intellectual property to the company, rather than simply describing the work as done.

Confirm no licensing gap has been assumed away

No AI-specific licence currently governs GPU hosting or compute resale in Canada, and it’s tempting to read that as meaning there’s nothing to check on the regulatory side. A buyer should still confirm, directly, that the business’s specific customer mix and hardware sourcing haven’t triggered an adjacent registration or reporting obligation the seller hasn’t flagged — the absence of a dedicated licence is not the same as the absence of any regulatory footprint at all, and it’s a cheap question to ask relative to the cost of discovering the answer after closing.

What a finding actually means for the deal

A registered security interest on the hardware doesn’t necessarily kill a deal — it usually means structuring the purchase price to pay it out at closing, or getting the lender’s written consent to an assumption. A power or colocation agreement that turns out not to be assignable without consent means that consent needs to be obtained, or built into the deal as a closing condition, before the buyer commits to a price. And thin customer commitments don’t automatically mean walking away — they mean pricing the business on its demonstrated, not its promised, revenue, and possibly structuring part of the price as an earnout tied to retained utilization after closing.