Future Scenarios

Consolidation

The neocloud landscape will likely consolidate over the next several years:

• Smaller players acquired by larger ones.
• Some failures as capital tightens.
• Hyperscaler acquisitions of strategic neoclouds.
• Cross-category consolidation (marketplace + inference + dedicated).

The current set of 8-12 named neoclouds may compress to 4-6 by 2030.

Compute futures impact

The CME and ICE compute futures markets launched in 2026 (covered in detail in the dedicated guide). Implications for neoclouds:

• Hedging instruments let neoclouds lock pricing on future capacity.
• Customer procurement gets more sophisticated.
• Capital structures evolve around compute as a financial commodity.
• Speculative trading enters the picture, potentially affecting prices.

GPU price normalization

The acute H100 shortage of 2023-2024 has been replaced with more normal supply-demand dynamics. As Blackwell and successor generations meet demand, prices may continue to normalize. The implications:

• Neocloud margins compress on the cost side.
• Customer pricing competition intensifies.
• Capital structures stress if pricing assumptions don't hold.

Geopolitics

Export controls, AI policy, and energy infrastructure decisions all affect neoclouds:

• Export controls on GPUs to certain countries affect supply chain.
• AI policy (EU AI Act, US executive orders) affects compliance burden.
• Energy infrastructure investment affects datacenter siting.
• Sovereign AI initiatives (UAE, Saudi, India, etc.) create new buyers.

Alternative chips

If AMD MI300X / MI325X / successor chips gain meaningful share, the NVIDIA-dependent neocloud strategy becomes more risky. Custom silicon (Microsoft Maia, Google TPU, AWS Trainium) also affects the picture.

So far NVIDIA's dominance has persisted. Whether it persists through the next several generations is one of the most consequential questions for neoclouds.

AGI-capability effects

Frontier AI capability trajectories affect demand:

• Continued capability scaling drives ever-larger training compute demand.
• Capability plateaus might shift weight from training to inference.
• Specific architectures (different attention, state-space models, etc.) might shift the optimal compute profile.

Neoclouds positioned for the current dominant architecture (transformer-heavy training and inference) face directional risk if the architecture shifts.

Takeaway

The forward picture is shaped by many forces. The final chapter examines the investment landscape that funds the industry.