Lambda Labs raised on-demand rates for its most popular training hardware this month. The single-GPU H100 SXM instance jumped from $4.29 to $4.89 per hour, a 14 percent increase that landed on November 8. The eight-GPU node version followed suit, moving from $31.92 to $36.72 per instance-hour. These changes affect the primary tier for large language model fine-tuning and distributed training workloads. I find it surprising that the company increased prices while competing providers like RunPod hold their H100 SXM secure cloud rate at $2.99 per hour. The price gap now widens significantly, making Lambda’s on-demand option less attractive for cost-sensitive teams.
New Rates and Configuration Costs
The pricing structure remains consistent across different memory tiers, but the absolute costs have shifted upward. For A100 80GB instances, the single-GPU rate climbed from $2.79 to $3.19 per hour, while the eight-GPU configuration stayed flat at $14.32 per instance-hour. This discrepancy suggests that volume discounts or bundled configurations may be protecting larger deployments from immediate hikes. The A10 24GB inference card held steady at $1.29 per hour, and the A100 40GB PCIe variant remained unchanged at $1.99 per hour.
| Plan | Previous Rate | Current Rate | Change |
|---|---|---|---|
| H100 80GB x1 | $4.29/hr | $4.89/hr | +$0.60 |
| H100 80GB x8 | $31.92/hr | $36.72/hr | +$4.80 |
| A100 80GB x1 | $2.79/hr | $3.19/hr | +$0.40 |
| A100 80GB x8 | $14.32/hr | $14.32/hr | No change |
These figures represent on-demand rates with no spot or reserved pricing available in the current public data. The B200 cluster pricing, which starts at $9.86 per GPU-hour for sixteen-GPU configurations, also reflects the broader trend of premium hardware commanding higher entry points. Users should note that these rates exclude storage and egress fees, which can add significant overhead for data-intensive pipelines.
Strategic Context and Alternatives
Lambda’s move aligns with its shift toward owning AI data centers rather than renting capacity. The company grew annualized revenue from $425 million at the end of 2024 to $760 million by the end of 2025, focusing on long-term contracts with hyperscalers and frontier labs. This strategy reduces reliance on hourly rentals, allowing them to raise prices for on-demand users who lack committed contracts. The infrastructure includes a 24MW facility in Kansas City and over 30MW through EdgeConneX in Chicago and Atlanta, ensuring stable power supply for dense clusters.
However, the price hike makes alternatives more compelling for smaller teams. Thunder Compute offers H100 PCIe access at $3.20 per hour, saving about 3 percent compared to Lambda’s new rate. For A100 workloads, the difference is starker: Thunder charges $1.09 per hour versus Lambda’s $2.79 per GPU-hour for the eight-GPU SXM node, a 61 percent savings. CoreWeave lists H100 SXM on-demand at $4.76 per hour, slightly below Lambda’s new $4.89 rate, while AWS remains far higher at $12.29 per hour.
I would recommend checking if your workload fits within the eight-GPU minimum for SXM nodes, as single-GPU access requires the more expensive PCIe variant. If you need flexible scaling without multi-month commitments, the current price gap favors competitors. Does the added stability of Lambda’s owned infrastructure justify the premium for short-term experiments?
