Not owning GPUs has become the default, deliberate strategy for building an AI company — not a compromise founders accept reluctantly. H100 rental rates fell 64-75% in fifteen months, a dense ecosystem of neoclouds and inference-as-a-service providers now lets startups skip infrastructure entirely, and credit programs can fund a company’s first year before a founder writes a check. But the same flexibility that makes renting attractive also makes it fragile — and 2026 produced sharp, well-documented examples of exactly how.
The Market
Neocloud revenue tripled to roughly $23 billion in 2025, en route to $180 billion by 2030, because hyperscalers simply cannot build data centers fast enough — a neocloud can rack GPUs in months where a new hyperscale campus takes three to five years. CoreWeave (43 data centers, $66.8B contracted backlog, Nasdaq-listed at $23B), Lambda Labs, Nebius, and Crusoe anchor the enterprise tier; RunPod and Vast.ai’s peer-to-peer marketplace anchor the budget tier. Beneath that sits an inference-as-a-service layer — Together AI, Fireworks, Groq, Cerebras — where the same open-weight model can price 6x apart depending on provider, and where 2026 alone saw a $5.5B Cerebras IPO, a ~$20B Nvidia licensing deal for Groq’s chip technology, and Cloudflare’s acquisition of Replicate.
The Proof Point
Perplexity built a reported $14 billion company entirely on rented AWS infrastructure — never negotiating a GPU purchase order, never owning a data center. Cursor and Character.AI followed similar paths. None of them made the rent-vs-own decision once and stopped revisiting it.
The Risks Founders Underestimate
Three documented, specific risks anchor the piece: price volatility at renewal (Krea’s GPUs jumped 32% at contract renewal, from $2.80 to $3.70/hour, with providers demanding longer terms); vendor concentration and circular financing (Nvidia’s equity stakes and revenue-share arrangements with the neoclouds renting you capacity shape the price you pay, whether you know it or not); and expanding regulatory exposure (the Remote Access Security Act now extends export controls to cloud-based GPU access itself). Layered on top: a structural GPU shortage with 36-52 week lead times, and Microsoft reportedly requiring 1,000-GPU minimum commitments just to guarantee smaller customers access at all.
When Owning Makes Sense
For the substantial majority of startups, the crossover point never arrives during the venture-backed phase. The exceptions: usage that becomes large, predictable, and sustained for years; defensive ownership when rental capacity is being rationed away from smaller customers; and hard data-sovereignty requirements with no viable rental path.
The Takeaway for Decision-Makers
Rent by default, but architect for portability from day one — containerized workloads, provider-neutral storage, and an API abstraction layer that turns a provider switch into a configuration change, not a re-engineering project. The analysis closes with a staged checklist from formation through growth stage, and the throughline: not owning GPUs is the right strategy, but it’s no longer one a founding team can adopt passively.
