Older GPUs Are Not Obsolete, They Are Repriced: A 2026 Value-Tier Map

◆ THE DEPRECIATION ASSUMPTION BROKE

Scarcity upstream pushes demand backward

The assumption that older silicon gets cheaper as newer silicon arrives broke this year. Worth understanding why, because the mechanism decides what to do about it.

The price of an older GPU in 2026 is set by how constrained the current generation is, not by how capable the older part is. Blackwell supply is tight, HBM and GDDR7 are both constrained, and demand that cannot get current-generation capacity pushes backward into whatever will run the workload. That pushes older-tier prices up, not down. Even V100 pricing has seen an upward bump, and Volta launched in 2017.

So the useful question stopped being how old a card is. It became what the card can still do, and what the market is charging for that specific capability right now.

◆ TWO CLOCKS, NOT ONE

Capability and deprecation move at different speeds

Two clocks run separately here, and conflating them is where most planning goes wrong.

The capability clock is the precision floor. FP8 is the real line in the current generation of tooling. The L40S has it, at 724 TFLOPS. The A100 does not, and tops out at BF16 for training and INT8 for inference. That single gap matters more than the four years between the two architectures, because everything built after 2023 assumes low-precision throughput is available. An A100 and an L40S can rent for similar hourly rates and still not be the same tier.

The deprecation clock is the vendor support list, and it moves far more slowly than either marketing or resale pricing suggests. NVIDIA AI Enterprise Infra 8.1, released May 2026, removed V100 and Volta from its supported GPU list. Ampere was not removed. The Infra 4.10 LTSB branch reached end of life in July 2026. That is what deprecation actually looks like when it happens, and it is checkable on each release rather than inferred from vibes.

Jensen Huang described the A100 fleet publicly on 13 August 2026 as mission-capable from 2020 through 2029. CoreWeave has committed to renting A100 through 2029. A nine-year service life on a part discontinued in year four.

◆ THE 2026 VALUE-TIER MAP

◆ WHAT EACH TIER IS ACTUALLY FOR

Each owns a constraint, not a performance band

The A100 is the double-precision and multi-tenant option. FP64 Tensor Core performance of 19.5 TFLOPS makes it one of the few cloud-accessible GPUs suited to molecular dynamics, materials simulation, and computational fluid dynamics, where the newer inference-optimized parts do not compete. MIG partitions one card into seven isolated instances for multi-tenant serving. 80GB and roughly 2 TB/s of bandwidth cover models in the 30B to 70B range. What it lacks is FP8, permanently.

The L40S is the cost-per-token option. $0.023 per million against the H100's $0.026, achieved by being cheaper per hour by a wider margin than it is slower per token. Its strongest cases are the least discussed ones: image generation, where 48GB handles SDXL at any standard resolution and Flux.1 to 1080p at a rate comparable to H100 for a fraction of the hourly cost, and embedding endpoints, where a single card serves millions of embeddings per hour at low utilization. It has FP8. It has no NVLink and no MIG.

The RTX PRO 6000 is the single-card capacity option, and the one that requires no facility change. 96GB, more than an H100, at roughly half the bandwidth. Llama 3.3 70B fits at FP8 with room for KV cache. It runs in a standard tower on a normal wall outlet, works with standard drivers and consumer tooling, and needs none of the liquid cooling or 14 kW power budget a current-generation server system demands. It also has no NVLink, which ends any multi-card ambition.

◆ NOT A DISCOUNT RACK

Where the mismatch comes from

The pattern across all three is that each one owns a specific constraint rather than a general performance band. Double precision and partitioning. Cost per token. Single-card capacity without infrastructure.

None of them is a cheaper version of an H100. Treating the value tier as a discount rack is what produces the mismatch where a team ends up on an A100 for a workload that needed FP8, or on an L40S for something that needed tensor parallelism.

◆ BUYING VERSUS RENTING

The used discount is priced risk, and channel beats generation

The practical consequence for anyone buying rather than renting: used discounts are real but they price in support risk. Used A100 units trade at roughly 40 to 60% off with no warranty and limited driver support. That gap is not free money, it is the market pricing the possibility that the support list changes.

And for anyone renting, channel matters more than generation. The spread between the cheapest and most expensive A100 hour is wider than the gap between an A100 and an H100 at a single provider. Choosing the tier correctly and then renting it badly leaves more money on the table than choosing the wrong tier and renting it well.

◆ FAQ

Frequently asked questions