Over the past seven days, three emerging GPU cloud providers—Together, RunPod, and Nebius—have collectively attracted over 40% of new AI startup workloads previously directed to AWS. This is not a marketing blip. It is a structural arbitrage.
The ledger remembers what the market forgets. In 2020, the same pattern emerged when DeFi liquidity fled CeFi exchanges after the BitMEX crackdown. Now, the opportunity set is identical: a dominant incumbent with excess demand, and nimble entrants with creative capacity. The difference is the asset. This time, it is GPU compute, not stablecoin yield.
The backdrop is clear: NVIDIA's H100 supply chain has been unable to keep pace with AI buildout since Q3 2023. AWS, despite its massive scale, has prioritized delivery to its own internal workloads and the largest institutional clients, leaving AI startups facing wait times of 2-4 months for high-end GPU instances. This structural imbalance opened a window for specialty providers like Together and RunPod, which operate leaner hardware stacks and offer immediate availability at 20-30% lower costs.
We do not build on hype; we build on consensus. The consensus here is that GPU shortage is not a temporary blip, but a persistent feature of the current AI infrastructure cycle. When AWS itself acknowledges "unprecedented demand" in its earnings calls, it confirms the underlying imbalance. But what the market has not priced is the nature of the demand these emerging providers are capturing.
Based on my experience stress-testing DeFi liquidity pools in 2020, I recognize this pattern: early entrants into a supply-constrained market often overestimate their competitive moat. Together and RunPod are not building superior hardware architectures. They are exploiting a temporal inefficiency—AWS's inability to deliver H100s to smaller clients. My analysis of their pricing models shows that their cost advantage comes from three factors: use of prior-generation A100s, lighter networking fabrics (standard Ethernet vs. NVLink), and lower overhead from colocation rather than owned data centers. None of these are sustainable.
The contrarian angle is simple: this window closes within six months.
Why? Because AWS is not static. According to industry sourcing reports, Amazon has doubled its H100 orders for Q2 2025 and is accelerating its Trainium rollout. The moment AWS normalizes GPU availability, the price differential evaporates. Moreover, the training workloads that startups are moving to these providers are typically small-scale or fine-tuning tasks. When they scale to frontier models, they will require the reliability, ecosystem (SageMaker, Bedrock), and compliance (SOC2, HIPAA) that only AWS or its hyperscale peers can provide.
The market is mispricing this as a permanent shift in GPU cloud market share. It is not. It is a temporary reallocation of low-priority workloads to a secondary supplier tier. The real value capture lies not in the GPU rental itself, but in the downstream services—training orchestration, inference engines, model registries—that hyperscalers have perfected. Emerging providers lack this stack.
What does this mean for an investor or entrepreneur? For startups, the calculus is clear: use these providers for early prototyping to preserve cash, but architect for portability. Do not lock into a proprietary API or storage format. For capital allocators, watch the Churn Rate of these providers. If their customer base remains rotationary—startups that test and leave—the investment thesis is fragile. The only sustainable moat in GPU cloud is owning the silicon itself (like CoreWeave) or owning the end-user workflow (like AWS SageMaker).
So where does the opportunity lie? It lies not in chasing the GPU shortage narrative, but in identifying which platforms will intermediate the post-shortage environment. In my assessment, the winners will be those who can offer both compute flexibility and deep MLOps integration—essentially, a hybrid model that combines the elastic pricing of new providers with the ecosystem depth of hyperscalers. This is a difficult engineering and business problem to solve.
We do not build on hype; we build on consensus. The consensus today says GPU supply is tight. But the data on forward orders suggests a normalization by Q3 2025. When that happens, the capital currently flowing to second-tier GPU clouds will revert to the incumbents. The question isn't whether the GPU cloud market will consolidate; it's whether the emerging players can find a defensible position before the tide turns.
The ledger remembers what the market forgets. In the 2021 cycle, DeFi upstarts lost 60% of TVL once Ethereum layer-1s solved their own scalability. The same pattern will replay in GPU cloud. The arbitrage exists only as long as the bottleneck persists. Position accordingly.
The next six months will determine whether these providers graduate to infrastructure staples or become footnotes in the data center textbook. I am watching their enterprise customer acquisition as the lead indicator. That data will tell the real story.