Welcome back to the GPU Pulse, your weekly deep dive into the graphics processing market where we track the heartbeat of the hardware that powers AI, gaming, and everything in between. I’m your host, Pete Paisley, and today is March 14th, 2026. If you’ve been watching the GPU market lately, you know things are interesting. We’re seeing a massive migration happening right now.

Enterprises retiring Ampere generation hardware, the RTX 3090 becoming the darling of the secondary market, and some legacy GPUs that are basically on life support. Today we’re breaking down the top 10 high-end GPUs approaching retirement age, their current market values, and what this means to you if you’re buying, selling, or managing GPU assets.

Let’s start with the elephant in the room — the data center workhorse that’s finally getting put out to pasture: the NVIDIA A100 80GB. Launched in 2020, the A100 is now five to six years old, putting it right at typical retirement age for enterprise hardware. And here’s what’s fascinating: despite being old by tech standards, these cards are still commanding $12,000–$20,000 in the secondary market. That’s 40–50% depreciation from peak pricing, but we’re talking about a GPU that originally cost $10,000–$15,000.

Why is it holding value so well? Three reasons: memory, memory, and memory. That’s 80GB of HBM, and it’s still incredibly valuable for mid-sized AI training workloads — think models up to 70 billion parameters. Plus there’s an HBM shortage right now, keeping prices elevated across the board. Research institutions, startups building multi-GPU clusters, and organizations that need serious compute but can’t justify H100 pricing — the A100 is a sweet spot for a lot of these use cases.

The RTX 3090 launched in September 2020 at $1,499. Today — five and a half years later — it’s selling for $800–$1,000. That’s 47–53% depreciation, which honestly isn’t bad for a consumer GPU that’s been through a crypto boom, a pandemic, and multiple generational cycles.

Here’s why the 3090 is the card to watch: 24GB of VRAM at a consumer price point. You can’t get that anywhere else without spending significantly more. The RTX 4070 and 5070 have 12GB. The RTX 4080 has 16GB. If you’re running local AI models — Llama 270B, Stable Diffusion XL, anything that needs serious memory — the 3090 is your best bang for the buck.

Demand is very high right now. AI hobbyists, content creators, small businesses, and professionals who need the VRAM but don’t want to drop $3,000 on an RTX A6000. The market is calling this the sweet spot GPU. And we agree.

The RTX A6000 launched in 2020 and is now selling for $3,500–$5,000 in the aftermarket — a 25–35% depreciation from its original MSRP of ~$4,650. Compare that to consumer cards and you can see why professional GPUs hold better value: they’re built for longer life cycles, certified for professional applications, and carry enterprise support.

The A6000 is still in demand for CAD/CAM workstations, medical imaging, scientific visualization, and professional 3D rendering. But the RTX 6000 ADA is now out with better performance at the same 48GB memory capacity, triggering workstation refresh cycles.

The A40 is in a similar position: $3,000–$4,000 in the secondary market, down from ~$5,000 at launch. The A40 was always more of a data center virtualization play — VDI deployments, multi-tenant GPU sharing, rendering farms. Its successor, the L40S, is now available, pushing A40s into retirement.

The V100 launched in 2017 — it’s now eight to nine years old and selling for $1,500–$2,500. That’s 70–80% depreciation from its original $8,000–$10,000 price tag. Demand is low to moderate, mostly from organizations with V100-optimized code or extreme budget constraints.

The problems are real: it’s not power efficient compared to modern GPUs, it lacks modern AI features like FP8 and Transformer Engine, and it’s simply old. If you’re running a data center, you’re not buying V100s in 2026 unless you have a very specific legacy use case.

The P100 is even worse. Launched in 2016 — now eight to ten years old — selling for $400–$800. That’s 85–90% depreciation. Demand is low: educational projects, extreme-budget AI experimentation, hobbyist clusters. That’s it.

Depreciation patterns to know:

• Consumer GPUs (RTX 30/40 series): 40–50% within 12 months of next-gen launch
• Professional GPUs (RTX A series): 25–35% over four to five years — much more stable
• Data center GPUs (A100, H100): 40–55% as the next architecture launches
• Legacy GPUs: 70–90% after seven-plus years

The memory premium effect is real. GPUs with high VRAM are commanding premium pricing on the secondary market: RTX 3090/3090 Ti (24GB) holding value; A6000/A40 (48GB) strong demand; A100/H100 (80GB) enterprise-grade pricing.

Optimal selling windows: Consumer GPUs — sell three to six months before next-gen launch. Professional GPUs — sell at the three to four year mark. Data center GPUs — sell when the next architecture enters volume production.