In most industries, a startup raises money and then spends it. In frontier AI, the scarce resource is often not money but compute: access to large clusters of modern accelerators.

Why compute is different

Training a competitive large model requires thousands of GPUs running for weeks. Those chips are expensive and, at times, simply unavailable. A startup that raises cash may still wait months for capacity. A partner that can deliver the hardware is therefore extremely valuable.

How the deals work

  • Strategic investment plus cloud commitment. Cloud providers invest in model developers, and the startup commits to running workloads on that cloud. Microsoft’s partnership with OpenAI and Amazon’s investments in Anthropic — which named AWS its primary cloud and training partner — follow this pattern.
  • Credits. Earlier-stage startups receive cloud credits as part of accelerator programs or investment deals.
  • Reserved capacity. Larger players sign multi-year contracts for capacity in specific data centers.
  • GPU-backed loans. Some infrastructure companies finance chip purchases with debt secured by the hardware itself.

Consequences for the market

  1. Concentration. Companies with privileged access to compute can move faster, which favors those allied with large clouds.
  2. Circular flows. Money invested by a cloud provider partly returns to it as cloud revenue, which complicates how we read both funding and revenue figures.
  3. Lock-in. Optimizing models for one provider’s hardware and services makes switching costly.

What to watch

When a funding announcement mentions compute, look for the length of the commitment, the provider, and whether the capacity is already built. In AI, the timing of hardware can decide who ships first.

A simple illustration

Imagine a startup that raises a large round in which part of the investment comes as credits for a specific cloud provider. The startup can train and serve models without paying cash for computing — but it is now tied to that provider’s infrastructure, chips and pricing. If it later wants to switch providers, it may face technical migration costs and lose remaining credits. Meanwhile, the cloud provider records revenue as the credits are used and gains a high-profile customer.

This structure explains why headline funding numbers can overstate how much flexible cash a startup actually has, and why cloud providers are among the most active investors in AI.

How startups reduce compute dependence

  • Use smaller, efficient models where they are good enough, reserving large models for hard tasks.
  • Fine-tune open-weight models instead of training from scratch.
  • Negotiate multi-cloud flexibility or portable infrastructure.
  • Optimize inference with quantization, caching and batching.
  • Focus on applications that build on existing models rather than competing at the frontier.

What it means for competition

When access to compute determines who can build frontier models, a few companies with the deepest pockets and infrastructure gain a structural advantage. Competition authorities have taken note, studying whether partnerships between cloud providers and AI developers limit competition. Open-weight models and falling inference costs push in the other direction, letting smaller companies build competitive products without frontier-scale budgets.

Questions to ask about a compute-heavy deal

How much of the funding is cash versus credits or hardware commitments? Are the credits tied to one provider, and for how long? What does the startup pay after credits run out? Who owns the infrastructure? The answers reveal how much strategic freedom the startup really has.

Related: how AI startup funding works and AI data center capex explained.

Key takeaways

  • Compute is now a core input for AI startups, often provided as cloud credits or chip access in funding deals.
  • Credits create dependence on specific providers and can inflate headline funding figures.
  • Cloud providers invest in AI startups partly to secure future customers.
  • Startups reduce dependence through efficient models, open weights and inference optimization.
  • Compute concentration raises competition concerns, while open models widen access.

Frequently asked questions

What are cloud credits?

Prepaid allowances to use a cloud provider’s computing services, sometimes given as part of investments or startup programs.

Why can’t startups just buy their own GPUs?

Some do, but large clusters require huge capital, data center space, power and expertise; cloud access is often faster and more flexible.

Does every AI startup need huge compute?

No. Many build products on existing models via APIs or fine-tune open models, needing far less compute than frontier labs.