NVIDIA H100

What hackers pick instead

• NVIDIA Tesla V100People recommend the used V100 as cheap VRAM and FP64 compute for local AI and physics work. Caveats: no bfloat16, loud fans, high idle power.16recommend
• AMD Instinct MI355XStrong memory and performance per dollar for large AI inference and training; critics say its software stack lags and it trails Nvidia's B300 on some benchmarks.14recommend
• Nvidia A100Long-lived datacenter GPU that people still run for AI and compute work years after release; the most common caveat is that it can be unreliable under long runs.10recommend
• Intel D270 GPU AI AcceleratorRecommended for holding many resident AI models in 480 GB of LPDDR5X memory. Critics say its low memory bandwidth makes it slow, and it is very expensive for individuals.10recommend
• NVIDIA B200Recommended for top raw AI throughput, a maturing software stack and low cost per token. Main caveat: rented B200 instances are scarce or pricey.9recommend

Every quote

The 34 most detailed of 39 comments. Each links to the original.

1. A single H100 is so OP, so hosting thousands in a data centre it's expected that it should remain unrivalled.

   Recommends · Mar 2026 · in “Anthropic takes legal action against OpenCode”

2. The machine is expensive, yes, but the customers also pay a lot to get it's advantages.

   Mixed · Jun 2026 · in “Show HN: Can Europe train a frontier AI model on the compute it owns?”

3. the easiest way to run them is by renting H100s or using a managed service offering

   Recommends · Jun 2026 · in “For Most of the World, Open-Source AI Is the Only Way Forward”

4. H100 are cheap enough but the power requirements make it a long term net negative

   Mixed · Jun 2026 · in “AI Coding at Home Without Going Broke”

5. Self-hosted 8xH100 is ~$250k, depreciated across three years

   Mixed · Mar 2026 · in “Intel's make-or-break 18A process node debuts for data center with 288-core Xeon”

6. SXM was the most expensive per hour but cheapest to finish

   Recommends · Mar 2026 · in “Which H100 Instance to Train Nanochat – Benchmarking PCIe, SXM, and NVL”

7. At that price point I would rather go for a single H100.

   Recommends · Sep 2026 · in “Thelio Mira AI Linux Workstation: 192 GB GPU Memory”

8. being able to rent an H100 for 2$ an hour was very nice

   Recommends · Nov 2025 · in “A $1k AWS mistake”

9. incredible that you can get an H100 for about $2/hr

   Recommends · Oct 2026 · in “Show HN: Using prediction margins to decide when to trust your classifier”

10. H100 to GB200 saw a 50x increase in efficiency

    Mixed · Apr 2026 · in “Hyperscalers have already outspent most famous US megaprojects”

11. a private H100SXM or B200 can go a long way

    Mixed · Sep 2026 · in “The current balance of power in open models”

12. For that price you get the H100.

    Mixed · Sep 2026 · in “Thelio Mira AI Linux Workstation: 192 GB GPU Memory”

13. Nvidia GPUs, particularly H100s, have a failure rate orders of magnitudes higher than traditional CPU-only hardware

    Criticizes · Feb 2026 · in “Data centers in space makes no sense”

14. Their failure rate is so high, I dont understand why anybody will even consider it feasible

    Criticizes · Dec 2025 · in “Economics of Orbital vs. Terrestrial Data Centers”

15. 8 x h100 80's don't give you enough to run the latest 1tn + parameter models

    Criticizes · Apr 2026 · in “Claude.ai Is Unavailable”

16. If you're still running H100s they draw >10x kWh/Mtoken as new designs

    Criticizes · Aug 2026 · in “OpenAI Jalapeño: Better than Nvidia Blackwell”

17. it's too small to run even the frontier open source models

    Criticizes · Jul 2026 · in “Nobody knows what a used GPU cluster is worth”

18. PCIe: 1,412ms/step - ~$112 (runpod)

    Criticizes · Mar 2026 · in “Which H100 Instance to Train Nanochat – Benchmarking PCIe, SXM, and NVL”

19. Sohu uses 144GB of HBM3E per chip with approximately 1.8x the memory bandwidth of an H100 SXM5.

    Mixed · Aug 2026 · in “Etched Sohu vs. Nvidia: Transformer ASIC vs. GPU (2026)”

20. high high end solutions like H100, H200, etc that have significantly more memory

    Mixed · May 2026 · in “Local AI needs to be the norm”

21. 16xH100 (or 16xB200) may be a better config for cost optimization

    Recommends · Jan 2026 · in “Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model”

22. the one winning, and continuing to win for some time, is Nvidia

    Recommends · Jul 2026 · in “American AI is locked down and proprietary. It's losing”

23. they are just waiting for the h100 to reach its peak lifecycle

    Mixed · Nov 2025 · in “Nvidia says its GPUs are a 'generation ahead' of Google's AI chips”

24. the last good fp64 gpu was the H100

    Recommends · Oct 2026 · in “GPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards”

25. a card that costs over $30,000 performs worse than integrated GPUs in such benchmarks as 3DMark and Red Dead Redemption 2

    Criticizes · Feb 2026 · in “The AI boom is causing shortages everywhere else”

26. NVIDIA will charge you a kidney to provide driver/hardware support if anything goes wrong.

    Criticizes · Apr 2026 · in “Claude.ai Is Unavailable”

27. Almost completely damn useless unless you really have no other alternative.

    Criticizes · Jul 2026 · in “Nobody knows what a used GPU cluster is worth”

28. has less shader compute capacity than the iGPU of AMD's Ryzen CPUs

    Criticizes · Dec 2025 · in “Framework Raises DDR5 Memory Prices by 50% for DIY Laptops”

29. H100 looked worse than expected in this specific setup.

    Criticizes · Feb 2026

30. A100 40GB SXM was nearly 2x as efficient as H100 SXM

    Criticizes · Dec 2025 · in “A100s are more cost-efficient than H100s for small training tasks”

31. NVIDIA H100 critical error† Hardware 0.15 (50d)

    Criticizes · Jan 2026 · in “Keeping 20k GPUs Healthy”

32. It won't last long because of space radiation.

    Criticizes · Nov 2025 · in “Nvidia's H100 GPU Takes AI Processing to Space”

33. my models were collapsing quickly

    Criticizes · Jan 2026 · in “Show HN: Floating-point drift between Apple M1 and H100 is real”

34. they are getting only 341 tok/s

    Criticizes · Apr 2026 · in “MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU”