新川サーバー | エンタープライズ サーバー ハードウェア サプライヤー
The H100 Lead Time in January 2025 Was 36 Weeks. Today It Is 12. Here Is Why That Matters for Your 2026 Budget.
In January 2025, ordering 100 H100 GPUs from any server vendor returned a lead time of 30-40 weeks. By January 2026, the same order returns 10-14 weeks. The supply chain has not magically improved — NVIDIA has doubled H100 production capacity, AMD has shipped MI300X in volume for the first time, and the initial wave of hyperscaler GPU purchases has moved from procurement to deployment. The GPUs that were stuck in allocation queues in 2024-2025 are now flowing to the enterprise market. For procurement teams that shelved AI infrastructure plans because the lead times were impossible, 2026 is the window to act.
GPU Pricing Trends: H100, A100, L40S, and T4 in Mid-2026
| GPU | Mid-2025 Street Price (Single Unit) | Mid-2026 Street Price (Single Unit) | 12-Month Change | Volume Availability (100+ Units) |
|---|---|---|---|---|
| NVIDIA H100 80GB SXM5 | $28,000-32,000 | $22,000-26,000 | -20 percent | Available. Lead time 10-14 weeks. |
| NVIDIA H100 80GB PCIe | $25,000-28,000 | $19,000-23,000 | -18 percent | Available. Lead time 8-12 weeks. |
| NVIDIA A100 80GB SXM4 | $14,000-17,000 | $11,000-14,000 | -20 percent | Available. Lead time 4-8 weeks (secondary market). |
| NVIDIA A100 80GB PCIe | $12,000-15,000 | $9,500-12,500 | -18 percent | Readily available. Lead time 2-4 weeks. |
| NVIDIA L40S 48GB PCIe | $8,000-10,000 | $7,000-9,000 | -12 percent | Readily available. Lead time 2-4 weeks. |
| NVIDIA T4 16GB PCIe | $2,500-3,500 | $2,200-2,800 | -15 percent | Immediately available from distribution. |
| NVIDIA H200 141GB SXM5 | 該当なし (not yet GA) | $35,000-40,000 | New product | Limited. Lead time 20-30 weeks. Allocated to hyperscalers first. |
The key insight for procurement teams: H100 pricing has dropped roughly 20 percent year-over-year as supply catches up with demand. A100 pricing has similarly fallen as the market shifts attention to H100 and H200 — making A100 the best price-performance GPU for mixed HPC-plus-AI workloads in mid-2026. L40S and T4 pricing is stable and availability is strong. The GPU you could not buy 18 months ago is the GPU you can negotiate on price for today.
AMD MI300X and Intel Gaudi 3 — The Alternatives That Change the Negotiation
NVIDIA still commands roughly 80-85 percent of the AI GPU market, but AMD MI300X and Intel Gaudi 3 are shipping in volume as of early 2026. The MI300X delivers 192 GB of HBM3 with 5.3 TB/s memory bandwidth — competitive with H200 on memory capacity and exceeding H100. The Gaudi 3 delivers competitive training throughput at roughly 65-75 percent of equivalent H100 pricing. Neither platform has the CUDA software ecosystem maturity of NVIDIA, but both are viable for organizations with the engineering resources to support non-CUDA AI stacks (PyTorch with ROCm for AMD, PyTorch with Habana SynapseAI for Intel).
The procurement implication is straightforward: even if you have no intention of buying MI300X or Gaudi 3, their existence in the market gives you leverage in NVIDIA pricing negotiations. A multi-vendor GPU RFP that includes AMD and Intel alternatives typically yields 5-10 percent better NVIDIA pricing than a single-vendor NVIDIA-only RFP. The threat of a competitor is worth more than the competitor itself.
The GPU Server Total Cost Equation: Beyond the GPU Purchase Price
| Configuration | GPU Cost (8x Unit Price) | Server Cost (8-GPU Chassis, XE9680 or G5500 V7) | 3-Year Colocation Power at $0.12/kWh | ネットワーキング (400GbE InfiniBand, 2 アダプター + switch port) | 3-Year TCO |
|---|---|---|---|---|---|
| 8x H100 SXM5 | $200,000 | $120,000 | $52,000 | $35,000 | $407,000 |
| 8x A100 SXM4 | $100,000 | $110,000 | $43,000 | $28,000 | $281,000 |
| 8x L40S PCIe | $68,000 | $80,000 | $29,000 | $14,000 (200GbE Ethernet) | $191,000 |
The GPU itself accounts for roughly 50 percent of the 3-year TCO for H100 clusters, but only 35 percent for L40S clusters — because the server, networking, and power costs scale more slowly than the GPU cost. For inference workloads where L40S delivers competitive throughput per dollar, the TCO advantage over H100 is approximately 2.1x. For training workloads where H100 FP8 and Transformer Engine deliver unique throughput, the H100 premium is justified by time-to-model. The right GPU is a function of your workload, not just the GPU price tag.
Three GPU Procurement Strategies for 2026
- Buy H100 now, deploy in Q3. With lead times at 10-14 weeks and pricing down 20 percent year-over-year, H100 is more available and more affordable than at any point since its launch. If you have budget approved for AI training infrastructure, Q2 2026 is the optimal procurement window. H200 will become more available in 2027, but H100 is not obsolete — it will be the workhorse AI training GPU for the next 3-4 年.
- Deploy L40S for inference, lease H100 for training. L40S is available in 2-4 weeks, costs roughly 30 percent of H100, and handles 90 percent of enterprise inference workloads without breaking a sweat. For organizations that train models quarterly but serve inference continuously, buying L40S servers for the inference tier and using GPU cloud (Lambda Labs, CoreWeave, AWS P5) for the episodic training bursts is often cheaper than owning an H100 cluster that runs at 15 percent utilization.
- A100 for mixed HPC-plus-AI. If your workload mix includes CFD, molecular dynamics, or financial simulation alongside AI training, the A100 at current mid-2026 pricing is the best price-performance GPU on the market. 80 GB HBM2e, FP64 Tensor Cores, NVLink multi-GPU, and availability in 2-4 weeks. The A100 will be productive for the next 5 years in mixed-workload data centers.
Source Your AI GPU Infrastructure Through Xincuan
We configure GPU servers — Dell XE9680, Dell R760xa, xFusion G5500 V7 — with your choice of NVIDIA H100, A100, L40S, or T4 GPUs. 工場直送価格, 3-年保証, global shipping, and free GPU architecture consultation to match your workload to the right GPU.
Request a GPU server configuration and quote
Explore GPU products: A100 | L40S | T4 | AI Servers