Welcome to the GPU Pulse. I’m your host, Pete Paisley. If you’re holding Ampere or Hopper GPUs right now, we have some important updates. Three days ago, Jensen Huang took the stage at GTC 2026 and dropped some big news that’s impacting the GPU secondary market. The Vera Rubin platform is here — it’s in full production — and the cascading effect we predicted in our March market report has just shifted into overdrive.

First: what is Vera Rubin? It’s not just a new GPU. It’s a complete ecosystem. Seven chips working together:

• The Rubin GPU
• The Vera CPU
• NVLink 6
• ConnectX-9 SuperNIC
• BlueField-4 DPU
• Spectrum-6 Switch
• And the kicker: an integrated Groq 3 LPU for inference

The NVL 72 rack costs $3.5–$4 million — a 25% premium over Blackwell — and delivers an estimated 2–3x performance improvement in AI workloads. Jensen called it “the greatest infrastructure build-out in history for agentic AI.”

When hyperscalers start buying Vera Rubin racks, everything below it cascades down-market. Here’s the domino sequence:

It’s a chain reaction — and it’s happening in quarters, not years.

In our March report, we valued A100 80GB at $12,000–$20,000 with strong demand and said: hold for now. We’re changing that recommendation.

With Vera Rubin shipping later this year, we’re looking at a compressed timeline. We expect A100 prices to drop an additional 20–30% within six to nine months as H100 supply floods the secondary market.

For V100 and P100: our March report said liquidate immediately. The Vera Rubin announcement makes this even more urgent. V100s in the $1,500–$2,500 range and P100s in the $400–$800 range are about to become completely unsellable. Why would anyone buy a V100 when A100s will be $8,000–$10,000 in just 12 months? Accept 10–20% below market value right now to clear inventory — holding these is financially risky.

H100s are trickier — currently ranging $18,000–$22,000. The rental strategy from our March report is still valid, but the window is shrinking. If you’ve got H100s with an opportunity to rent them, rent them — but you’ve got maybe 12 to 18 months of strong demand left. If utilization is lower, sell before Vera Rubin volume production hits.

There’s a bright spot: consumer cards like the RTX 3090. That 24GB of VRAM is still gold for AI developers and content creators running local models. The 3090 at $800–$1,000 with very high demand is holding steady — because data center GPUs haven’t cascaded into prosumer pricing yet. But watch this space. If A100 40GB systems drop below $8,000, they’ll start competing with consumer cards on a price-per-gigabyte basis.

One more thing worth watching: the Groq 3 LPU integration. NVIDIA is acknowledging that training and inference need different hardware. If specialized inference accelerators take off, GPUs currently used for inference might get retired faster — but they could find new buyers in training and development markets. This bifurcation is worth watching.

GTC 2026 just pushed everything ahead. GPU lifecycles are compressing. The cascade is happening in quarters, not years. If you’re holding Ampere generation hardware, your window to maximize value is measured in months.

Our March report predicted this — but Vera Rubin moved the timeline up. Jensen Huang said the agentic AI inflection point has arrived. The question is: are you ready to act?