GPU Pricing · OYASS · Open Models

Pricing comparison across 4 GPU providers, how it connects to OYASS, which models you can actually own, and hardware recommendations — compiled July 14, 2026.

Sources: Vast.ai, Salad, Terra Compute, Hashrate, freeintelligence.ai, velab.org/oyass

1. GPU Pricing Comparison

Four tabs open in the browser — here's how they stack up across consumer and datacenter GPUs.

Per-Hour Pricing Matrix

GPU	VRAM	Vast.ai	Salad	Terra Compute	Hashrate
RTX 5090	32GB	$0.42–$0.48	$0.25	—	Listed
RTX 4090	24GB	$0.35–$0.39	$0.11–$0.18	—	Listed
RTX 3090	24GB	$0.20–$0.30	$0.09–$0.10	—	Listed
RTX 3080	10GB	~$0.12	$0.06	—	Listed
RTX 3060	12GB	~$0.08	$0.05–$0.08	—	Listed
H100 SXM	80GB	$0.90–$2.27	—	Multi-GPU	Listed
B200	—	$4.75–$5.08	—	—	Listed
A100 80GB	80GB	$0.75–$4.00	—	Multi-GPU	Listed

Best price highlighted in green. Hashrate is a comparison index, not a provider. Terra Compute sells bare-metal server hosting at $0.075/kWh.

🥬 Salad — Cheapest Consumer GPUs

60,000+ daily active GPUs. Distributed network (gamers' machines). Best for batch inference on consumer cards.

RTX 5090: $0.25/hr · RTX 4090: $0.11–$0.18/hr · RTX 3090: $0.09–$0.10/hr

🌐 Vast.ai — Widest Selection

20,000+ GPUs, RTX 3060 → B200. Marketplace model — prices vary by host. Best for datacenter GPUs (H100, B200).

H100: $0.90–$2.27/hr · B200: $4.75–$5.08/hr

🏗️ Terra Compute — Bare-Metal

You buy the server, they host it at $0.075/kWh. Multi-GPU builds. Different model: you own the hardware, they provide power/cooling/colo.

Server builds: $4,000–$8,000 + $0.075/kWh

💡 Recommendation: For OYASS workloads running 24/7 inference on 27B-class models, Salad is cheapest for consumer GPUs. For H100/B200 datacenter models, Vast.ai has the best selection and pricing. For sustained high-utilization, buy a used 3090 ($700) — it pays for itself in 2–4 months vs. renting.

2. OYASS — Own Your AI Software Service

OYASS (velab.org/oyass/) is a VE LAB product that builds branded, owned AI workers for operators. The core insight: rent a unicorn vs. own a donkey 🫏.

$3k–$10k

One-time build — scope workflow, build worker, brand it, hand over the fork

Pass-through

Compute is your keys, your accounts — OYASS never meters it

$500–$1.5k/mo

Optional upkeep — tuning, model updates. Cancel anytime, your 🫏 keeps running

Yours

The fork — code, workflow, data, brand. Runs where you run it. Doesn't disappear.

OYASS vs. Subscription — The Math

	Rent (SaaS)	OYASS (Own)
5-year cost	$20/mo × 60 = $1,200	$5k build + $0 compute = $5,000
10-year cost	$20/mo × 120 = $2,400	$5k build + $0 compute = $5,000
You own it?	❌ No	✅ Yes — fork, data, brand
Can it disappear?	✅ Yes — deprecation, pivot, shutdown	❌ No — runs where you run it
Compute cost	Bundled (opaque)	Pass-through (transparent)

The real comparison isn't $20/mo vs. $5k — it's renting forever vs. owning once. Different posture, different math.

OYASS builds worker

→

Needs compute

→

GPU pricing comparison

→

Customer picks: rent or buy

OYASS Compute Strategy

Since OYASS passes compute through to the customer, the GPU pricing comparison directly answers the customer's question: "What will it cost me to run this?"

Light usage (<100 hrs/mo): Rent on Salad/Vast.ai — $10–$50/mo for a 27B model
Heavy usage (500+ hrs/mo): Buy a used 3090 ($700) — pays for itself in 2–4 months
Datacenter models (70B+): Rent H100 on Vast.ai at $0.90–$2.27/hr, or buy a Mac Studio M3 Ultra 192GB ($4,199+)
Multi-tenant / multiple workers: Terra Compute bare-metal at $0.075/kWh — own the server, host it remotely

3. Open-Model Transparency Matrix — Which Models Can You Actually Own?

From freeintelligence.ai/open-models/ — 12 labs scored across 7 transparency dimensions. The key question for OYASS: can you truly own and control this model?

🏆 Fully Open (7/7) — You Can Actually Own These

These release weights + training data + code + checkpoints + recipes + permissive license + data tracing. The full stack, reproducible from scratch.

Lab	Model	Size	License	Score	Best For
Ai2	OLMo 3.1	32B	Apache 2.0	7/7	Gold standard. Full data (Dolma, 6T tokens), every checkpoint, training code, OlmoTrace. Roadmap: Olmo-MoE 2026.
Hugging Face	SmolLM3	3B	Apache 2.0	7/7	Small, fast, fully open. 11.2T tokens. Great for fine-tuning and edge deployment.
EleutherAI	Pythia	12B	Apache 2.0	7/7	Original open-data lab. The Pile dataset. New: Common Pile v0.1 (openly-licensed data only).
LLM360	K2	—	Apache 2.0	7/7	Community-owned AGI. All artifacts: code, data, checkpoints, intermediate results.
BigCode	StarCoder 2	—	Apache 2.0	7/7	Code model. Training code + The Stack v2 dataset released. Rare fully-open code model.

⚠️ Open Weights + Reports (Partial)

Weights are downloadable, training data is not. You can run and fine-tune, but you cannot fully reproduce or audit.

Lab	Model	License	Score	Notes
DeepSeek	V3 / R1	MIT	~3/7	Open weights + detailed technical reports. R1's RL pipeline is documented enough to be reproducible in principle. No dataset release.
Mistral	Mixtral	Apache 2.0	~3/7	Open weights + papers describing recipes. Training data not released. Member of NVIDIA Nemotron Coalition.
Cohere For AI	Aya	Custom	~3/7	Open weights + multilingual data cards documenting composition. Raw dataset not released.

🔬 Frontier-Scale Open Outlier: NVIDIA Nemotron

NVIDIA builds frontier-scale open models and releases the data — distinct from "frontier-class" capability.

Nemotron 3 (Dec 2025): Training datasets, recipes, ~10T tokens of open data
Nemotron Coalition (GTC March 2026): Pools data and compute with Mistral, Perplexity, Cursor, LangChain
Nemotron 4 (roadmap): Coalition-built base model, intended to be open-sourced

❌ Closed Frontier — You Cannot Own These

Lab	Model	Score	Why
Meta	Llama 3/4	~1/7	Open weights only. Restrictive community license. No data, no code, no checkpoints.
Google	Gemma	~1/7	Open weights only. Custom terms. No data.
Alibaba	Qwen	~1/7	Open weights only. MIT license (good), but no training data.
OpenAI	GPT-5	0/7	Fully closed. API-only. No weights, no data, no code.
Anthropic	Claude 4	0/7	Fully closed. API-only.

💡 OYASS Recommendation — Which Models to Build On:
For full ownership (data + weights + code): Ai2 OLMo 3.1 (32B) — the gold standard. Apache 2.0, full data release, OlmoTrace for auditability. Best for regulated industries or customers who need complete control.
For maximum capability in an open package: DeepSeek V3/R1 (MIT) or Qwen 3 (MIT) — open weights, permissive license, frontier-competitive. Accept that training data is private.
For small/fast/edge: SmolLM3 (3B, Apache 2.0) — fully open, runs on a laptop, great for fine-tuning.
For frontier-scale with data: NVIDIA Nemotron 3/4 — the only frontier-scale model that releases training data. Coalition-backed.

4. Hardware Recommendations — The Local AI Box

From freeintelligence.ai/local-ai-box/ — the definitive guide to picking local-AI hardware in 2026.

The Default Pick: Used RTX 3090 24GB ($700)

Three things make the 3090 the consensus pick in 2026:

24GB is the right VRAM — big enough for 27B-class models (Qwen 3, GPT-OSS 20B, DeepSeek V3 Q4) at usable quantizations. 16GB cards cap out at 8B-class.
First-class software support — Ampere architecture, CUDA 12, PyTorch BF16, TensorRT-LLM, vLLM, llama.cpp, Ollama, LM Studio. Every tutorial assumes you have this card.
Actually a GPU — drives a display, runs Blender, plays games. Not a server card needing SXM2 adapters or custom cooling.

Budget Tiers

Tier	Cost	Hardware	Runs
Entry	~$300	Tesla P40 24GB ($180) + DIY	8B–13B models, slow
Sweet spot	~$700	Used RTX 3090 24GB	27B Q4, 70B Q3, 20B fast
From scratch	~$1,500	New build + used 3090	Same as sweet spot + reliability
Silent	~$4,200	Mac Studio M3 Ultra 96GB	70B Q4, near-silent, 30W
Frontier	$7,500+	Mac Studio 192GB / multi-GPU	Anything that fits in VRAM

What Runs on 24GB (Q4/Q5)

GPT-OSS 20B — 50+ tok/sec, covers ~80% of daily work
Qwen 3 27B — new consensus for local coding, matches GPT-4-class
Llama 3.3 70B Q3 — fits, ~15 tok/sec, previous-gen "big local" benchmark
Mistral, Gemma, Phi, Qwen Coder — 7B–14B class, fly on a 3090

What doesn't fit: Anything >32B at Q4+. Upgrade to V100 32GB (~$700, CUDA 11 lock-in) or Mac Studio.

💡 OYASS Hardware Strategy:
For most OYASS customers, the answer is: start renting on Salad/Vast.ai, buy a 3090 when utilization justifies it. The $700 3090 pays for itself in 2–4 months of 24/7 inference vs. renting at $0.20–$0.30/hr. For customers who need silence or 70B+ models, the Mac Studio M3 Ultra is the right answer despite the $4,200+ price tag.

5. Open-Models Builder Profiles

From freeintelligence.ai/open-models-report/ — detailed profiles of every organization building truly open AI models.

🥇 The Fully-Open Builders

Builder	Flagship	Max Size	Data	License	Why It Matters
Ai2	OLMo 3.1	32B	Dolma (6T tokens)	Apache 2.0	Gold standard. Every checkpoint, training code, OlmoTrace. Roadmap: Olmo-MoE 2026.
Hugging Face	SmolLM3	3B	FineWeb-Edu (11.2T)	Apache 2.0	Data stewards. Maintain FineWeb corpus used by many open models. Progressive checkpoint releases.
EleutherAI	Pythia	12B	The Pile	Apache 2.0	Original open-data lab. Common Pile v0.1 built entirely from openly-licensed text. Proving unlicensed data isn't necessary.
LLM360	K2	—	Full	Apache 2.0	"Community-owned AGI." All artifacts open: code, data, checkpoints, intermediate results. With MBZUAI and Petuum.
BigCode	StarCoder 2	—	The Stack v2	Apache 2.0	Hugging Face + ServiceNow. Rare fully-open code model. Training code + dataset released.
NVIDIA	Nemotron 3	Frontier	~10T tokens	Open	Only frontier-scale builder releasing data. Nemotron Coalition pools data + compute across Mistral, Perplexity, Cursor, LangChain.

🔬 Smaller / Newer Fully-Open Entrants

Builder	Model	Notes
Stanford	Marin	Academic fully-open model
ETH Zürich / EPFL	Apertus 70B	Swiss fully-open 70B model
AMD	Instella	AMD's fully-open entry
Zyphra	Zamba	Zamba models + Zyda dataset. Fully open.
BLOOM / T5	—	Older but fully open. Pioneers.

💡 For OYASS: Which Builder to Partner With:
Ai2 is the safest bet — gold-standard transparency, Apache 2.0, active roadmap (Olmo-MoE 2026). Build OYASS workers on OLMo 3.1 for customers who need full auditability.
NVIDIA Nemotron is the frontier-scale bet — if OYASS needs to offer frontier-class capability with data transparency, the Nemotron Coalition is the only game in town.
DeepSeek (MIT) for maximum capability in an open-weight package — accept the data transparency tradeoff.
SmolLM3 for lightweight, edge-deployed OYASS workers that need to run on customer hardware.

6. The Strategic Picture

How It All Fits Together

OYASS builds owned AI workers

→

Needs models customers can own

→

Open-Model Matrix picks the right model

→

GPU pricing tells compute cost

→

Hardware guide picks rent vs. buy

The OYASS Customer Decision Tree

What model do you need? → Check the Transparency Matrix. Need full ownership? OLMo 3.1. Need max capability? DeepSeek V3. Need edge deployment? SmolLM3.
What hardware does it need? → 27B models run on a $700 3090. 70B+ models need H100 rental or Mac Studio. 8B models run on a laptop.
Rent or buy? → <100 hrs/mo: rent on Salad/Vast.ai. 500+ hrs/mo: buy a 3090. Multi-tenant: Terra Compute bare-metal.
What's the total cost? → OYASS build ($3k–$10k) + compute (pass-through) + optional upkeep ($500–$1.5k/mo).

Key Numbers Summary

Item	Cost	Notes
OYASS build (one-time)	$3,000–$10,000	Scope, build, brand, handover
OYASS upkeep (optional)	$500–$1,500/mo	Tuning, model updates
RTX 3090 (used, buy)	$700	Runs 27B models, pays for itself in 2–4 months
H100 rental (Vast.ai)	$0.90–$2.27/hr	For 70B+ models
Consumer GPU rental (Salad)	$0.05–$0.25/hr	Cheapest for 8B–27B models
Mac Studio M3 Ultra 96GB	$4,199	Silent, 30W, runs 70B Q4
Ai2 OLMo 3.1 (fully open)	Free (Apache 2.0)	Gold standard for ownership
DeepSeek V3 (open weights)	Free (MIT)	Max capability, no data transparency
Terra Compute hosting	$0.075/kWh	Bare-metal, you own the server