What is H100?
NVIDIA H100 is a Hopper-based Tensor Core GPU introduced in 2022 for AI, machine learning, and deep learning workloads. As of mid-2026, its rental price has increased compared to a year earlier, and its availability has tightened following the collapse of B200 GPU supply, while older H100s remain useful for inference.
Release history
- Dec 2025 - Zain Asgar noted that newer chips like the B40 or B200 show significant performance gains over the H100, with 20-40% improvements in many cases.
- Apr 2026 - Jack Farley reported that GPU availability for B200s has collapsed, with H100s close behind.
- Apr 2026 - Philip Kiely stated that an H100 is more expensive now than it was a year ago in terms of rental basis.
- Apr 2026 - Nathan Labenz said that a frontier chip like an H100 uses about the same amount of energy when on as a microwave or electric teapot.
- Jul 2026 - Ben Pouladian said that H100s and A100s over three years old are still running and useful for inference, though not for training.
- Aug 2026 - Philip Johnston said that his team knows where an H100, B200, and H200 will fail if blasted with high-velocity protons and heavy ions.
In the discourse
Attributed discussion of H100.
Trinity Large (AI2) targets a 400B model deployable on a single rack of eight H100s at FP8, aiming for competitive open-model throughput at accessible hardware scale.
“You'll be able to get really good throughput, a little bit lower concurrency on a rack of eight H100s at FB8.”Nathan Lambert · 27 Jan 2026
Philip Johnston on StarCloud's unique radiation-failure data for flagship NVIDIA GPUs.
“I think we're the only people in the world now that know where both an H100, a B200, H200 will fail if you blast it with high velocity protons and heavy ions.”Philip Johnston · 5 Aug 2026
Older Nvidia GPUs (H100s and A100s, 3-plus years old) remain productive for inference even when obsolete for training, extending effective hardware lifespans beyond initial expectations.
“GPU are actually lasting longer than people expected. I think you have H100s and A100s that were like over 3 years old, still running, and they're still useful. Not so much for training, but you can use them for inference.”Ben Pouladian · 14 Jul 2026
Kernel-level optimization yields only single-digit percentage gains on H100s (already heavily optimized) but 20 to 40 percent improvements on B40/B200, and 2x or more on Mac, AMD, and Intel hardware.
“If you start looking at think is like a B40, or a B200, you'll start seeing like much more significant gains to be had over there. you know, we have seen 20 30 40% improvements in performance in many cases.”Zain Asgar · 2 Dec 2025
Jack Farley on GPU scarcity: B200 availability has already collapsed, H100s are close behind.
“GPU availability for B200s has collapsed H100s close behind.”Jack Farley · 10 Apr 2026
Running a frontier-class model like Kimi K3 requires approximately 80 H100-class GPUs, making it feasible for enterprises and sophisticated attackers but not ordinary end users.
“In practice, and I think it's like I saw somewhere because I was curious yesterday, 80 H100-class GPUs.”Steve Gibson · 29 Jul 2026
Philip Kiely on the unexpected direction of H100 rental pricing.
“An H100 is more expensive now than it was a year ago in terms of a rental basis.”Philip Kiely · 30 Apr 2026
A frontier AI chip (H100) consumes roughly the same power as a microwave or electric teapot when running.
“A frontier chip today like an H100 or it basically uses the same amount of energy when it's on as a microwave or as like an electric teapot.”Nathan Labenz · 1 Apr 2026
H100 rental prices have risen year-over-year, countering expectations of commodity GPU cost declines.
“An H100 is more expensive now than it was a year ago in terms of a rental basis.”Philip Kiely · 30 Apr 2026