Current Market Outlook

The market is pricing a 79% chance that NVIDIA H100 SXM compute will cost more than $3.83 per hour by December 31, 2026. That is a relatively strong bet. It suggests traders see a clear upward trajectory for GPU compute pricing, not a decline or stagnation. The $3.83 threshold matters because it sits well above current spot pricing on most cloud providers. As of mid-2025, H100 SXM instances on AWS and Azure run between $2.50 and $3.00 per hour for reserved terms. The market is essentially saying those discounts will disappear.

Key Factors Driving the Odds

Supply constraints are the primary driver. NVIDIA's H100 production has been strained since launch, and the company allocated most 2024 capacity to hyperscalers and sovereign AI projects. The remaining open market supply is tight. New fabs from TSMC won't meaningfully increase CoWoS packaging capacity until late 2026, which directly limits H100 output.

Demand is the other half. Enterprise AI adoption is accelerating, not slowing. Companies that missed the first wave of GPU procurement are now scrambling for capacity. The H100 remains the gold standard for training large models, and the B100 successor is priced even higher, pushing budget-conscious buyers toward the H100 instead of away from it.

What Could Change These Odds

A major demand collapse from AI model commoditization or a sudden shift to cheaper inference hardware could push prices below $3.83. The B100 ramp is the biggest wildcard. If NVIDIA delivers B100 volume earlier than expected, H100 secondary market prices could drop sharply as data centers upgrade. Watch for NVIDIA's Q4 2025 earnings in February 2026 for capacity guidance. Also monitor the Ornn index itself. If it shows sustained pricing below $3.50 through mid-2026, the 79% probability will look overconfident.