app/src/content/article.mdx

---
title: "AI Hardware Choices are Highly Variable and Sparsely Disclosed"
# subtitle: "What the top 4,000 models on Hugging Face reveal about hardware transparency - and why disclosure should be a release norm"
description: "A systematic audit of training- and inference-hardware disclosure across the top 4,000 most-downloaded models on Hugging Face, and the case for treating hardware as standard model-release documentation."
authors:
  - name: "Francisco Ríos"
    affiliations: [2]
    equalContribution: true
  - name: "Ian Reynolds"
    affiliations: [1]
    equalContribution: true
  - name: "Robert Praas"
    affiliations: [2]
  - name: "Avijit Ghosh"
    affiliations: [1]
    advisor: true
  - name: "Irene Solaiman"
    affiliations: [1]
    advisor: true
equalContributionNote: "These authors contributed equally."
advisorNote: "Advisors."
affiliations:
  - name: "Hugging Face"
    url: "https://huggingface.co"
    logo: "/image/hf-logo.png"
  - name: "Centre for European Policy Studies"
    url: "https://www.ceps.eu"
    logo: "/image/ceps-logo.png"
published: "Jun. 04, 2026"
licence: >
  Text and figures are licensed under <a href="https://creativecommons.org/licenses/by/4.0/" target="_blank" rel="noopener noreferrer">CC BY 4.0</a>, unless noted otherwise.
  Figures reused from other sources are excluded and marked in their captions (“Figure from …”).
tags:
  - AI hardware
  - compute
  - transparency
  - policy
  - open weights
tableOfContentsAutoCollapse: true
showPdf: true
---

import ExecutiveSummary from "./chapters/executive-summary.mdx";
import Introduction from "./chapters/introduction.mdx";
import Training from "./chapters/training.mdx";
import Inference from "./chapters/inference.mdx";
import Conclusion from "./chapters/conclusion.mdx";
import Appendix from "./chapters/appendix.mdx";
import AuthorContributions from "../components/AuthorContributions.astro";

<ExecutiveSummary />

<Introduction />

<Training />

<Inference />

<Conclusion />

<Appendix />

export const authorContributions = [
  { role: "Conceptualization",          people: "R. Praas, A. Ghosh" },
  { role: "Data curation",              people: "F. Ríos, R. Praas" },
  { role: "Investigation",              people: "I. Reynolds, A. Ghosh, R. Praas" },
  { role: "Methodology",                people: "F. Ríos, R. Praas, A. Ghosh" },
  { role: "Software",                   people: "F. Ríos" },
  { role: "Validation",                 people: "F. Ríos, I. Reynolds, R. Praas, A. Ghosh" },
  { role: "Formal analysis",            people: "I. Reynolds, A. Ghosh, R. Praas" },
  { role: "Writing (original draft)",   people: "F. Ríos, I. Reynolds, A. Ghosh, R. Praas" },
  { role: "Writing (review & editing)", people: "F. Ríos, I. Reynolds, A. Ghosh, R. Praas, I. Solaiman" },
  { role: "Visualization",              people: "F. Ríos" },
  { role: "Supervision",                people: "A. Ghosh, I. Solaiman" },
];

<AuthorContributions contributions={authorContributions} />

<div id="references" class="references">

1. Lawler, R. (2026). *Anthropic is in talks to use Microsoft's AI chips too*. [The Verge](https://www.theverge.com/ai-artificial-intelligence/935688/anthropic-is-in-talks-to-use-microsofts-ai-chips-too).
2. Emberson, L., & Sevilla, J. (2026). *Is a compute crunch coming?* [Epoch AI](https://epoch.ai/gradient-updates/is-a-compute-crunch-coming).
3. Fedasiuk, R. (2026). *America's Technology Siege Is Working as Intended*. [Asia Society, Center for China Analysis](https://centerforchinaanalysis.asiasociety.org/p/americas-technology-siege-is-working).
4. Qu, T. (2026). *Alibaba Ramps Up AI Push With New Chip, Model Upgrades*. [The Wall Street Journal](https://www.wsj.com/tech/ai/alibaba-unveils-new-ai-chip-upgrades-ai-model-421b196d).
5. Financial Times. (2026). *Huawei’s AI chip sales surge as Nvidia stalls in China*. [Financial Times](https://www.ft.com/content/b82fa156-d1db-40e5-bce5-3c5f8f54069b).
6. Gerut, A. (2026). *Encrypted texts reveal how Nvidia chips and U.S. tech are being smuggled to China and Russia*. [Fortune](https://fortune.com/2026/05/13/nvidia-chip-smuggling-china-russia-iran-export-controls-supermicro/).
7. Google. (2026). *Our eighth generation TPUs: two chips for the agentic era*. [Google](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/).
8. Los Angeles Times. (2026). *Inside the AI compute crunch driving Google researchers to quit*. [Los Angeles Times](https://www.latimes.com/business/story/2026-05-18/inside-ai-compute-crunch-driving-google-researchers-to-quit).
9. Shilov, A. (2025). *DeepSeek reportedly urged by Chinese authorities to train new model on Huawei hardware*. [Tom's Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/deepseek-reportedly-urged-by-chinese-authorities-to-train-new-model-on-huawei-hardware-after-multiple-failures-r2-training-to-switch-back-to-nvidia-hardware-while-ascend-gpus-handle-inference).
10. Reuters. (2026). *DeepSeek unveils new AI model tailored for Huawei chips as China pushes for tech autonomy*. [Reuters](https://www.reuters.com/technology/chinas-deepseek-returns-with-new-model-year-after-viral-rise-2026-04-24/).
11. Poolside. (n.d.). *Models*. [Poolside](https://poolside.ai/models).
12. CoreWeave. (2024). *Mistral AI and CoreWeave Demonstrate Partnership at NVIDIA GTC, Mistral AI Hackathon*. [CoreWeave](https://www.coreweave.com/blog/mistral-ai-and-coreweave-partnership-at-nvidia-gtc).
13. r/LocalLLaMA. (2026). *Buried lede: Deepseek v4 Flash is incredibly inexpensive from the official API for its weight category* [Online forum post]. [Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1su5gj5/buried_lede_deepseek_v4_flash_is_incredibly/).
14. Brand, F., & Denain, J.-S. (2025). *Why benchmarking is hard*. [Epoch AI](https://epochai.substack.com/p/why-benchmarking-is-hard).
15. Yuan, J., et al. (2025). *Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference*. [Hugging Face Papers: 2506.09501](https://huggingface.co/papers/2506.09501).
16. Qi, P., et al. (2025). *Defeating the Training-Inference Mismatch via FP16*. [Hugging Face Papers: 2510.26788](https://huggingface.co/papers/2510.26788).
17. Gond, R., et al. (2026). *LLM-42: Enabling Determinism in LLM Inference with Verified Speculation*. [Hugging Face Papers: 2601.17768](https://huggingface.co/papers/2601.17768).
18. Pape, D., Evertz, J., & Schönherr, L. (2026). *The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility*. [Hugging Face Papers: 2605.19537](https://huggingface.co/papers/2605.19537).
19. Angelo, J. (2026). *Microsoft reports are exposing AI's real cost problem: using the tech is more expensive than paying human employees*. [Fortune](https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/).
20. Midha, A. [@anjneymidha]. (2026). *Post on X* [Post]. [X](https://x.com/anjneymidha/status/2058611711867801989).
21. Hugging Face. (n.d.). *Hardware*. [Hugging Face](https://huggingface.co/hardware).
22. SambaNova. (2023). *Introducing BLOOMChat-176B: the multilingual chat-based LLM*. [SambaNova](https://sambanova.ai/blog/introducing-bloomchat-176b-the-multilingual-chat-based-llm).

</div>