Cloud credits and compute contracts: how AI startups protect their margin

Why the procurement details matter as much as the model

Most founders and investors now understand the basic tension in AI startups: running large models costs real money in GPU time, and that cost eats into gross margin in a way traditional software never did. What gets discussed less is that two startups with identical products and identical usage can end up with very different margins purely because of how they negotiated and structured their compute spend. The contract terms sitting behind a cloud bill are often as important to unit economics as the model architecture itself.

Credits are a runway tool, not a margin fix

Early on, many AI startups run largely on cloud credits handed out through accelerator programmes, cloud provider startup schemes, or as part of an investment. These credits extend runway and let a team experiment without immediately worrying about cost. But credits are temporary. When they expire, the startup is exposed to full list price, and if the product’s pricing was set while compute was effectively free, margin can collapse overnight. A useful discipline is to model gross margin as if credits did not exist from day one, so the true cost base is never hidden from the team or from investors reviewing the numbers.

Reserved capacity and commitment discounts

Cloud providers offer meaningfully lower prices in exchange for committing to a certain level of spend or capacity over a fixed term, rather than paying on-demand rates. This is standard practice for any company with predictable, sustained compute usage. The trade-off is flexibility: a commitment locks in a spending floor even if usage dips, and switching provider or region becomes harder. For a startup, the judgement call is about how confident the team is in its demand forecast. Commit too early, based on optimistic growth assumptions, and the fixed cost becomes a drag on the balance sheet. Commit too late, and the startup pays a persistent premium that a competitor with a locked-in rate does not.

Multi-cloud and provider leverage

Some startups deliberately keep workloads portable across more than one cloud or compute provider, even at some engineering cost, purely to retain negotiating leverage. Being able to credibly threaten to move a workload is one of the few tools a small company has when negotiating with a large cloud provider. This matters more for startups whose core cost is inference at scale than for those still in an experimental phase, where the overhead of multi-cloud architecture rarely pays for itself.

Inference cost is where margin is actually won or lost

Training cost is usually a one-off or periodic expense that can be amortised or, increasingly, avoided altogether by building on top of a third-party foundation model rather than training from scratch. Inference cost, the compute needed every time a customer actually uses the product, scales directly with usage and revenue. This is why serious operators focus heavily on cost per query or cost per generated token as a core metric, tracked with the same rigour as customer acquisition cost. Techniques that reduce this figure, such as using a smaller or fine-tuned model for routine tasks and reserving the most expensive model for cases that need it, routing requests intelligently, caching repeated outputs, and compressing or quantising models, tend to matter more for long-term gross margin than any single procurement negotiation.

Pricing has to reflect real marginal cost

A recurring mistake is pricing an AI product the way SaaS has always been priced, with a flat monthly fee regardless of usage, when the underlying cost structure is much closer to a utility. If heavy users cost the startup far more in compute than light users but pay the same subscription fee, the business is effectively subsidising its most expensive customers. Usage-based or hybrid pricing, where a base fee is combined with metered charges above a threshold, aligns revenue with the actual cost driver and protects gross margin as the customer base grows and usage patterns diverge.

What investors actually look for

When assessing an AI startup, investors increasingly ask for a clear breakdown of cost of goods sold that separates compute from other costs, and a gross margin trend line rather than a single snapshot. A startup whose margin is improving as it scales, because it is moving workloads to cheaper models, negotiating better rates, or shifting customers to usage-based pricing, tells a very different story from one whose margin is flat or worsening despite growth. Government-backed initiatives around AI compute access, including the UK’s approach to AI Growth Zones and compute capacity, are relevant here too, since where and how a startup can access subsidised or co-located compute can materially change its cost base. Founders should check current guidance directly with the relevant government department rather than relying on figures that may have moved on.

The practical takeaway

Compute cost versus gross margin is not just a modelling exercise done once at a funding round. It is an ongoing set of decisions about credits, commitments, provider leverage, model selection and pricing structure, each of which shifts the trade-off in a different direction. Startups that treat compute procurement as a core operational discipline, reviewed as regularly as sales pipeline, tend to protect margin far better than those that treat it as a fixed cost to be endured.