The Supply Landscape
NVIDIA's supply allocation for H2 2026 is heavily skewed toward Blackwell and Blackwell Ultra. CoWoS-L advanced packaging capacity, which was the binding constraint in 2024-2025, has expanded roughly 3x to approximately 60,000 wafers per month. This translates to roughly 1.8 million B200 and B300 GPUs in H2 2026, up from approximately 1.2 million in H1.
The catch is that hyperscalers AWS, Azure, GCP, and Oracle have pre-committed to approximately 75% of the H2 2026 Blackwell output through 12-36 month capacity reservation agreements. This leaves roughly 450,000 B200/B300 GPUs for the neocloud and enterprise market in H2 2026, which is insufficient to meet estimated demand of 700,000-800,000 units.
Pricing Projections by GPU Tier
Spot pricing for H100 will continue its downward trajectory, reaching $1.50-2.00/GPU/hr by Q4 2026 as secondary supply floods the market. H200 spot pricing will follow a shallower decline, settling at $2.20-2.80/GPU/hr. B200 remains supply-constrained and will hold at $4.00-5.00/GPU/hr through end of 2026 before declining as B300 ramps in Q1 2027.
Reserved contract pricing tells a different story. 12-month reserved contracts for B200 are currently commanding 25-35% premiums over spot, reflecting buyer desperation for guaranteed capacity. We expect this premium to compress to 10-15% by Q2 2027 as supply catches up with demand.
| GPU Model | H2 2026 Spot Range | H1 2027 Forecast |
|---|---|---|
| H100 SXM | $1.50 – $2.00/hr | $1.00 – $1.50/hr |
| H200 SXM | $2.20 – $2.80/hr | $1.80 – $2.40/hr |
| B200 SXM | $4.00 – $5.00/hr | $3.00 – $4.00/hr |
| B300 NVL | $5.50 – $7.00/hr | $4.50 – $5.50/hr |
| A100 SXM | $0.80 – $1.20/hr | $0.60 – $0.90/hr |
Hyperscaler Demand Dynamics
The three largest GPU buyers in 2026 are Microsoft, Google, and Meta, collectively accounting for an estimated 65% of all Blackwell output. Microsoft alone has committed to approximately 350,000 B200/B300 GPUs for 2026 delivery, primarily for OpenAI inference workloads and internal Copilot infrastructure. Google's allocation is split between B200 for external cloud customers and TPU v6 for internal workloads.
The critical dynamic for non-hyperscaler buyers is the hyperscaler's appetite for secondary disposal. As hyperscalers transition capacity from H100 to Blackwell, an estimated 200,000-300,000 H100 units will enter the secondary market in H2 2026, predominantly from Microsoft and Oracle. This secondary wave will depress H100 pricing faster than current spot markets reflect.
Blackwell Supply Constraints
B200 supply in H2 2026 remains the most constrained segment. NVIDIA's yield on the B200 die has improved to approximately 70%, but the complex packaging required for the B200 NVL72 rack-scale system creates a separate bottleneck. Each NVL72 system consumes 72 B200 GPUs plus 36 NVLink switches, and production is limited to roughly 1,200 systems per quarter across all OEM partners.
B300 supply will begin ramping in Q1 2027 with initial volumes of 200,000-300,000 units, ramping to 500,000+ per quarter by Q3 2027. Early allocation will again favor hyperscalers. Enterprise buyers seeking B300 should expect 12-18 month lead times for new orders placed in H2 2026.
Budgeting Recommendations by Team Type
For AI startups at Series A-B: budget based on H200 spot pricing of $2.50-3.00/GPU/hr for H2 2026 workloads, with the expectation that H100 availability will create a spot market floor at $1.50. Avoid long-term B200 reservation contracts unless your workload specifically requires Blackwell's FP4 capability. The flexibility to arbitrage across H100 and H200 spot markets will save 30-40% versus fixed-price reserved instances.
For enterprise teams: the window to lock in favorable B200 reserved pricing is Q3-Q4 2026. By Q2 2027, the combination of B300 ramp and increased CoWoS capacity will create a buyers' market for B200. Negotiate 12-month contracts now with Q1 2027 break clauses. For B300, expect to pay a 40-60% premium over B200 through H1 2027.
Upside and Downside Risks
Downside risks are concentrated in geopolitics. Further export controls on advanced GPU shipments could bifurcate the market, with prices in restricted markets (China) rising while unrestricted markets see oversupply. NVIDIA's silicon strategy pivots around this uncertainty, with both the B200 and B300 having export-compliant variants with reduced interconnect bandwidth.
The upside scenario sees CoWoS-L capacity expanding faster than expected, potentially reaching 80,000 wpm by Q1 2027 and allowing NVIDIA to ship 2.5+ million Blackwell GPUs in 2027. In this scenario, B200 spot pricing could drop to $2.50-3.00/hr by mid-2027, significantly earlier than our baseline forecast.
Our Recommendation
Do not overpay for B200 reserved capacity in H2 2026. The premium over spot is inflated by fear of supply constraints. Use H200 spot instances for workloads that do not strictly require Blackwell FP4 or large HBM capacity. Reserve B200 only for production inference serving with committed SLAs that justify the premium.
Plan B300 procurement now for Q2-Q3 2027 availability. Get in the queue with multiple providers to increase the probability of allocation. ClusterBid's marketplace provides real-time pricing across 40+ providers, enabling teams to optimize their GPU mix dynamically without multi-year commitments.