Cloud services have deprived AI startups of GPU access, and their resources were absorbed by Anthropic and OpenAI

Short Summary

Cloud giants Microsoft, Amazon and CoreWeave have taken up most of the available graphics processors (GPUs), leaving startups in line. This has affected companies backed by venture funds such as Sequoia Capital, Founders Fund, General Catalyst and Andreessen Horowitz. The cost to rent GPUs over the last six months has risen by more than 25 %, and according to Azure the shortage could continue until the end of 2026.

1. What’s happening
Provider | How resources are allocated | Microsoft | Three‑tier scheme:
• Priority for ~1,000 large clients (annual contracts).
• Mid‑size companies get managers.
• Small and micro‑businesses via partners. | Amazon & CoreWeave | Similar model but with stricter quotas:
• Blackwell chips require a minimum of 1,000 units and an annual contract worth tens of millions of dollars.
• Older NVIDIA models are only available after waiting several weeks/months.

2. Why startups are in line
1. Priority deals
Microsoft, Amazon and CoreWeave have signed multi‑billion contracts with Anthropic and OpenAI, guaranteeing these companies first‑in‑line access to GPUs.

2. Reduced quotas for small business
The emergence of new AI tools has sharply increased demand for compute resources, while providers have cut limits for smaller clients.

3. Expiration of old contracts
Many startups signed 2–3 year agreements last year; now they are expiring, and providers use this as an excuse to reallocate capacity in favor of more financially capable customers.

3. Venture funds’ reaction
- General Catalyst
Managing partner Hemant Taneja sent a survey to founders of portfolio companies to gauge real GPU access. In response he announced the creation of shared compute pools and plans to negotiate with providers on behalf of his investments.

- Andreessen Horowitz & Index Ventures
After large providers cut public access in early 2023, these funds began building their own GPU pools for startups from their portfolios.

4. Consequences for specific companies
Company | Investments | Current situation
Anthropic – Despite a large contract, it is experiencing capacity shortages; Azure reports a deficit until the end of 2026. | Lightning AI – Manages ~40,000 GPUs, but about 40 companies are queued with a need for 400,000 chips. Hourly cost rose from $1.60 to >$2 (and higher on some configurations). | Krea – $83 million from Andreessen Ventures and Bain Capital Ventures. Six months ago they signed a contract for several hundred Blackwell chips at $2.80/hour. When trying to scale up capacity for new model training, provider reps stopped responding or cited resource shortages.

5. Numbers
- Price increase: over 25 % rise in GPU rental cost in six months.
- Quotas: Blackwell chips require a minimum of 1,000 units and an annual contract with a budget of tens of millions of dollars.
- Availability of older chips: wait times range from weeks to months, sometimes longer.

6. What’s next
1. Funds continue building their own GPU pools to reduce dependence on large providers.
2. Providers tighten control over access: if a client does not use allocated chips in full, they may permanently lose GPU access.
3. Long‑term contracts with major AI companies still give them priority, but even they face limits due to rising demand.

Conclusion
GPU shortages in the cloud are becoming a serious barrier for startups. Venture funds respond by creating their own pools and actively negotiating with providers, but large players continue to receive priority access. Rental costs rise and wait times lengthen – a situation worth monitoring closely in the coming years.