GPU Market Outlook H2 2026: Supply, Demand, and Pricing Trends

The GPU market in June 2026 is fundamentally bifurcated. H100 supply is abundant, with data centre operators reporting 4-8 week lead times for new orders and spot pricing stabilising at $1.20-1.80/GPU-hour, down from $4.00+ in early 2025. The H100 buildout that began in 2024 has reached peak delivery, and the secondary market is active with GPU capacity trading on open exchanges.

At the same time, B200 remains severely constrained. NVIDIA allocated approximately 80% of initial B200 production to hyperscalers (AWS, Azure, GCP, Oracle), leaving only 20% for the broader market. Lead times stretch 36-52 weeks, and spot pricing remains above $7/GPU-hour. The B200 shortage is expected to persist through Q1 2027, when NVIDIA's additional fab allocation ramps.

This bifurcation creates a complex market where buyers must choose between abundant, cost-effective H100 capacity and scarce, premium-priced B200 capacity. This post analyses supply-demand dynamics, pricing forecasts, and procurement strategies across both segments.

H100 supply has reached equilibrium in mid-2026. NVIDIA's Hopper production lines, combined with TSMC CoWoS capacity expansion, have delivered an estimated 3.5-4 million H100 units since launch. Data centre operators report 85-90% utilisation on H100 fleets, with available capacity for new workloads at competitive pricing.

The H100 secondary market has matured significantly. GPU capacity exchanges (including ClusterBid's marketplace) facilitate spot and short-term contracts for H100 compute, with prices transparently quoted by GPU generation, location, and contract duration. The secondary market has added approximately 15-20% liquidity to the total H100 available capacity, creating more pricing options for buyers.

The key trend for H2 2026 is that H100 pricing is expected to decline further. New H100 shipments continue at approximately 500-700K units per quarter, while demand growth from AI startups has moderated (more startups moving to inference-only models that can run on smaller GPUs). Our forecast: H100 spot pricing falls to $0.80-1.20/GPU-hour by Q4 2026, and reserved 12-month pricing reaches $1.50-1.80/GPU-hour.

B200 supply is the dominant story in H2 2026. NVIDIA's Blackwell architecture uses a new TSMC 4NP (4nm) process with CoWoS-L advanced packaging, and the yield ramp has been slower than expected. Estimated B200 shipments in H1 2026: 800,000-1,000,000 units. Demand: 2.5-3 million units as estimated. The supply-demand gap means B200 will remain constrained through at least Q1 2027.

The B200 pricing outlook: reserved 12-month contracts at $4.50-6.00/GPU-hour, spot pricing at $7-9/GPU-hour, and secondary market premiums of 20-40% for immediate availability. Multi-year contracts (24-36 months) with volume commitments can achieve $3.50-4.50/GPU-hour from some providers, reflecting the longer commitment offsetting scarcity risk.

The H200 fills the gap between H100 and B200. With 141 GB HBM3e and 4.8 TB/s bandwidth, H200 offers approximately 1.2-1.4x the performance of H100 for memory-bandwidth-bound inference workloads. H200 pricing at $2.00-3.00/GPU-hour reserved makes it an attractive middle option. Lead times are 8-12 weeks, substantially better than B200.

The table below provides our supply forecast for each GPU generation through H1 2027. These estimates are based on TSMC CoWoS capacity allocation, NVIDIA wafer starts, and hyperscaler deployment schedules.

The critical metric for buyers is the supply-demand ratio. Below 0.8 (demand exceeds supply), pricing remains elevated. Above 1.2 (supply exceeds demand), pricing declines. B200's ratio of 0.35 signals persistent scarcity. H200's ratio of 1.1 indicates near-equilibrium with moderate pricing pressure. H100's ratio of 1.4 confirms a buyer's market.

GPU supply and pricing vary significantly by region. The US market (particularly US West) has the highest GPU density and most competitive pricing, benefiting from proximity to hyperscaler data centre clusters. European GPU capacity is growing but pricing is 15-25% higher than US, driven by higher power costs and longer supply chains.

APAC markets are bifurcated. Japan and South Korea have strong GPU availability through government AI initiatives. Southeast Asia (Singapore, Malaysia) is emerging as a GPU hub with competitive pricing ($0.10-0.20/GPU-hour premium over US). China operates in a separate market with restricted GPU access. Chinese AI labs rely on domestic alternatives (Huawei Ascend, Biren) and restricted NVIDIA H20 GPUs, creating a fundamentally different supply landscape.

The regional arbitrage opportunity in H2 2026: deploy latency-tolerant training workloads in regions with lowest GPU pricing (US Midwest, Southeast Asia, parts of Europe), keep latency-sensitive inference geographically close to end users. Multi-region GPU deployment reduces effective GPU costs by 15-25% for teams with workload portability.

The bifurcated GPU market demands a nuanced procurement strategy. For H100 capacity: do not commit to long-term contracts at current pricing. The market is entering a buyer's phase with declining prices. Use 6-month contracts or spot instances for H100 workloads, and negotiate aggressively on volume. Target $1.20-1.50/GPU-hour for 12-month H100 reserved contracts signed in Q3 2026.

For B200 capacity: commit early with 12-24 month contracts to secure allocation. B200 scarcity will persist, and waiting only increases the risk of not having capacity when needed. Negotiate upgrade rights to B200 Ultra (due early 2027) in your contract. Target $4.00-5.00/GPU-hour for 12-month B200 contracts, with a volume commitment of 100+ GPUs.

The hybrid strategy: use H100 for the majority of training workloads (where the 1.5-2x per-GPU performance gap does not justify the 3-4x price premium), reserve B200 for workloads that truly need the 192 GB memory (large model inference, long-context training), and supplement with H200 as a mid-range option for memory-bandwidth-sensitive applications.

NVIDIA's 2027 product cycle will reshape the market. Blackwell Ultra (B300) is expected in Q1 2027 with 288 GB HBM4 and 12 TB/s memory bandwidth, priced at $6-8/GPU-hour initially. The Rubin architecture (XR100, due late 2027) will introduce a new GPU platform, ecosystem transitions, and 1 TB+ memory capacity per GPU.

Market normalization is expected by mid-2027, when B200 supply finally catches up with demand and multiple GPU vendors (AMD, Intel, and custom ASICs) provide competitive alternatives. At that point, the GPU market will resemble a more traditional compute market with multiple vendors, standardised interfaces, and competitive pricing.

The strategic implication for buyers in H2 2026: structure contracts with flexibility for the 2027 transition. Include GPU swap rights (H100 -> B200 or B200 -> B300 at predetermined rates), keep 20-30% of capacity on spot/on-demand for maximum flexibility, and develop workload portability to easily move between GPU generations and providers. The teams that maintain procurement flexibility will capture the most value from the evolving GPU market.