Hugging Face Hardware Census: What GPUs the AI Community Actually Uses

Everyone talks about H100 clusters and USD 10,000 GPU rigs. But what does the AI community actually run?

Now we know. Hugging Face CTO Julien Chaumond just launched HF Hardware — the first real census of GPUs, CPUs, and Apple Silicon across the open-source AI community. The results are fascinating, and they tell a very different story than the one GPU marketing departments want you to hear.

The Numbers

NVIDIA: 45% — Dominant, But Not How You’d Think

NVIDIA owns nearly half of all reported GPUs. No surprise there. The surprise is which NVIDIA cards people actually use:

GPU	Users	VRAM	Street Price
RTX 3060	18,000	12GB	~USD 250 used
RTX 3090	16,000	24GB	~USD 700 used
RTX 4090	14,000	24GB	~USD 1,600
RTX 5090	8,000	32GB	~USD 2,000
RTX 5070 Ti	6,000	16GB	~USD 750
RTX 5060 Ti	6,000	16GB	~USD 400
RTX 4060	6,000	8GB	~USD 300
RTX 3080	—	10GB	~USD 350 used
RTX 4060 Ti	—	16GB	~USD 400

The most popular AI GPU in the world is a three-year-old, USD 250 card.

The RTX 3060 beats the 4090 by 4,000 users. It beats the brand-new 5090 by 10,000 users. Why? Because it has 12GB of VRAM at a price point that students, researchers, and hobbyists can actually afford. In AI inference, VRAM matters more than raw compute — and the 3060 hits the sweet spot.

The RTX 3090 at #2 reinforces this: it’s the cheapest way to get 24GB of VRAM. People aren’t buying the newest, fastest card. They’re buying the most VRAM per dollar.

NVIDIA generation breakdown:

• RTX 30xx: 27% — the workhorse generation, still dominant
• RTX 40xx: 27% — tied, but more expensive
• RTX 50xx: 19% — newest gen, still ramping
• Datacenter: 13% — A100s, H100s, the enterprise crowd
• GTX/RTX 20: 13% — still hanging on

AMD: 5% — The Uncomfortable Truth

AMD has spent billions marketing AI capabilities. ROCm has improved significantly. And yet:

GPU	Users	VRAM
RX 7900 XTX	4,000	24GB
RX 6700 XT	2,000	12GB
RX 6800	1,700	16GB
RX 6600	1,700	8GB
RX 7900 XT	1,500	20GB

AMD’s best card (7900 XTX) has fewer users than NVIDIA’s sixth most popular card. The entire AMD GPU ecosystem has fewer users than the RTX 3060 alone.

This isn’t a hardware problem — the RX 7900 XTX is excellent hardware with 24GB of VRAM at USD 800. It’s a software ecosystem problem. PyTorch on ROCm still has rough edges. Many models and frameworks are tested on CUDA first (or only). When a researcher hits a compatibility issue at 11 PM, they don’t file a bug report — they buy an NVIDIA card.

The VRAM Story

Look at which cards people actually chose, and a pattern emerges:

• 12GB (RTX 3060): 18,000 users — runs 7B-13B models
• 24GB (RTX 3090, 4090): 30,000 users combined — runs 13B-30B models
• 8GB (RTX 4060): 6,000 users — runs 7B quantized
• 32GB (RTX 5090): 8,000 users — runs 30B+ models

The community clusters around VRAM tiers, not compute tiers. Nobody’s buying a GPU for its TFLOPS rating. They’re buying it for how many parameters they can fit in memory.

This is the dirty secret of local AI: the bottleneck is almost always VRAM, not speed. A 7B model running at 20 tokens/second on a 3060 is perfectly usable. A 70B model that doesn’t fit in VRAM at all is useless regardless of how fast the card theoretically is.

What This Means

1. The “AI requires expensive hardware” narrative is wrong

The median AI practitioner is running a USD 300-700 GPU with 12-24GB of VRAM. Not an H100. Not even a 4090. The open-source AI revolution is happening on mid-range consumer hardware, and the models are adapting to fit.

Quantization (GGUF, AWQ, GPTQ) has made this possible. A 70B model that needed 140GB of memory two years ago now runs in 24GB quantized to 4-bit. The software caught up to the hardware people could actually afford.

2. AMD has a distribution problem, not a hardware problem

5% share with competitive hardware means something is broken in the software pipeline. Every time a model readme says “tested on A100 and RTX 4090” without mentioning ROCm, AMD loses another potential user. The ecosystem is self-reinforcing: developers test on NVIDIA because users have NVIDIA because developers test on NVIDIA.

AMD’s path forward isn’t better hardware — it’s paying for first-class ROCm support in the top 50 HuggingFace models, contributing PyTorch ROCm CI, and making “it just works” on AMD the default, not the exception.

3. Apple Silicon is the dark horse

The screenshot cuts off Apple Silicon, but it’s increasingly relevant. M-series chips with unified memory (up to 192GB on M4 Ultra) can run models that no consumer GPU can touch. A USD 4,000 Mac Studio with 192GB unified memory runs a full 70B model unquantized — something that would require multiple USD 1,600 GPUs otherwise.

The trade-off is speed (Apple Silicon is slower per-token than a 4090) but for many use cases — development, testing, private inference — “slow but fits in memory” beats “fast but doesn’t fit” every time.

4. Build software that runs on what people actually have

If you’re building AI tools, this data should inform your design:

• Target 12GB VRAM as the baseline. That’s where the largest single group of users sits.
• Optimize for inference, not training. Most of these are consumer GPUs running local models, not training clusters.
• Support quantized models first. The community has voted with their wallets: they’ll take quality loss for accessibility.
• Don’t require a GPU at all if you can avoid it. Tools like soul.py run entirely on API calls — no local GPU needed. Memory and identity shouldn’t require hardware investment.

The Bigger Picture

Hugging Face now hosts over 1 million models. This hardware census adds a crucial missing dimension: not just what models exist, but what hardware they run on in practice.

The gap between “state of the art” and “state of the community” is enormous. Papers benchmark on 8×H100 clusters. The community runs on a single RTX 3060. The models that win adoption aren’t the ones with the best benchmark scores — they’re the ones that fit on the hardware people actually own.

That’s the real insight here. And it’s one that every AI framework, model creator, and tool builder should internalize.

---

Data from Hugging Face Hardware Census, launched May 2026 by CTO Julien Chaumond. Self-reported by Hugging Face community members. Explore the live data at hf.co/hardware.