A data or AI consultancy’s largest recurring cost after payroll is frequently cloud infrastructure and model-compute spend, AWS or Azure bills, GPU rental, and API usage against hosted large language models. The expense treatment itself is usually simple. What takes more attention is the billing decision sitting next to it: whether that cost is passed through to the client or absorbed into the fee, because that choice is what actually determines the engagement’s margin.
The Expense Treatment Is Usually Straightforward
Cloud infrastructure and model API costs are consumed as used, not purchased and owned, so they are almost always current expenses, deducted in full in the period incurred, the same current-expense category covered generally in software, SaaS, cloud, AI tools, domains, and subscriptions. There is no capital cost allowance question here the way there is for owned servers or lab hardware, since the consultancy is not acquiring a depreciable asset, only paying for consumption.
The one situation worth flagging separately is a prepaid, multi-year committed-use contract, a cloud provider discount program requiring an upfront multi-year spend commitment, which may need to be recognized over the committed period rather than expensed entirely in the year paid, depending on the contract structure. Most usage-based monthly billing does not raise this question at all.
The Question That Actually Matters: Pass-Through or Absorbed
Where a data or AI consultancy’s compute costs diverge from an ordinary software subscription is scale and variability. A client-facing model deployment or a data pipeline running against a large dataset can generate compute costs that are a meaningful fraction of, or in some cases exceed, the consulting fee itself for that engagement. How that cost is billed changes both the firm’s margin visibility and its GST/HST position.
Passed through as a disbursement. The client is billed directly for actual compute usage, itemized separately from the advisory or development fee, often at cost or with a modest markup. This keeps compute cost swings from compressing the firm’s own margin, since the client absorbs usage variability directly. It also keeps the compute reimbursement outside the firm’s own margin calculation, since it is a cost recovery rather than firm revenue in substance, though it still needs to be reported and, where applicable, taxed correctly, covered in client reimbursements, pass-through costs, and disbursements.
Absorbed into a flat or blended fee. The consulting fee is priced to include an estimate of compute cost, and the client pays one number regardless of actual usage. This is simpler for the client and can work well for predictable workloads, but it shifts usage risk onto the firm. A project that was priced assuming moderate model usage and then requires significantly more inference or training compute than estimated eats directly into the firm’s margin, sometimes turning a profitable engagement unprofitable partway through.
Neither approach is universally correct. The decision should be made consciously per engagement type, reflected clearly in the contract and the invoice structure, and applied consistently rather than shifting silently between projects in a way that makes margin comparisons across clients meaningless.
Tracking Compute Cost by Client or Project
Cloud providers generally support cost allocation tags or labels at the resource level, and using them, rather than letting all compute spend land in one undifferentiated expense account, is what makes engagement-level margin visible. A consultancy running workloads for three clients on a shared cloud account, without per-client cost tagging, can only see total compute spend for the month, not which client relationship is actually profitable once compute is factored in.
This matters more for data and AI engagements than for most consulting work because compute cost as a share of total project cost is unusually high and unusually variable, tied directly to data volume, model size, and client usage patterns that the firm does not fully control once a client-facing deployment is live. A monthly close that reconciles cloud spend to the general ledger without breaking it down by client is missing the number that actually drives engagement-level decisions about pricing and scope.
GST/HST on Passed-Through Compute Costs
Where compute costs are billed to the client as a disbursement, the GST/HST treatment generally follows the underlying supply. Compute billed to a Canadian client is subject to GST/HST the same as the advisory fee itself. Compute billed to a non-resident U.S. client as part of a broader zero-rated engagement, covered in AI consulting revenue, USD invoicing, and GST/HST zero-rating, generally follows the same zero-rating treatment as the rest of the engagement, provided the pass-through documentation supports that the cost is part of a single zero-rated supply of services rather than a separate taxable transaction.
SR&ED and Compute Costs
Compute spend tied to a genuinely eligible SR&ED technological investigation, as opposed to routine model inference for a standard client deliverable, can factor into an SR&ED claim, but this should not be assumed automatically because AI infrastructure is involved. The distinction between eligible experimental work and ordinary client delivery, covered in SR&ED for software and AI contractors, applies to compute costs the same way it applies to development labour: the technological uncertainty has to be real, not just the technology being used.
Related Guides
- Software, SaaS, cloud, AI tools, domains, and subscriptions covers the general current-expense treatment this guide applies specifically to compute-heavy client billing.
- Client reimbursements, pass-through costs, and disbursements covers the mechanics of billing a cost through to a client as a disbursement rather than absorbing it into the fee.
- SR&ED for software and AI contractors covers when development and compute costs may qualify for SR&ED versus ordinary client work.
- Project profitability for a small consulting firm covers the broader margin-tracking framework compute-cost tagging feeds into.
Scope of This Guide
This guide covers the expense treatment and client-billing structure for cloud infrastructure and AI model-compute costs incurred by Canadian data and AI consultancies. It does not cover:
- SR&ED claim preparation itself, which requires its own technical and financial documentation
- Data residency, privacy, or vendor contract terms for cloud and model providers, which are separate from the tax and billing treatment
- QST-specific treatment for Quebec-based engagements, which generally mirrors the federal GST treatment but should be confirmed separately
This is general information, not advice for a specific engagement. A CPA reviewing your actual client contracts and cloud billing structure can confirm the correct treatment for your file.