A transparent, well-regarded US GPU marketplace and RL-training platform with competitive observed rates and open-source tooling, but with no SLA, no compute attestation and GPU prices behind a login gate.
by the bestcloudmining editorial desk · review version 2 · prose updated 2026-09-24 · facts as of 2026-09-24 · how we verify
Prime Intellect is a US-based GPU marketplace and hosted AI training platform offering on-demand and reserved compute for research and inference. It caters to developers and AI teams needing high-end GPUs like H100s and H200s without long-term commitments. While it has a strong reputation in open-source AI circles and transparent operator details, it lacks service-level guarantees and verifiable compute attestation.
You can rent GPU instances (H100, H200, B200) via on-demand, reserved, or monthly contracts, with access through SSH, Jupyter, CLI, REST API, or Python SDK. The platform also offers token-based hosted training and inference for models like Qwen3.5 and Llama-3.2. You do not own the hardware or receive physical access; you only get remote compute time. Minimum purchase is $0, but actual pricing requires API access.
Contract types include on-demand, reserved, and monthly, but the exact term lengths in days are not specified. There is no published maintenance fee or uptime SLA, and payouts are not applicable since this is compute rental, not mining. Refunds or credits are handled case-by-case for provider-side failures. Minimum purchase is $0, but live rates require API access with authentication.
There is no verifiable proof that rented GPU hours execute as promised - no hardware attestation or confidential compute. While Prime Intellect publishes open-source tools (prime-rl, verifiers) and hosts a public status page, these do not confirm job integrity. Third-party research threads report real usage on 8xH200 setups, but this is not independent audit-level verification.
Prime Intellect Inc. is registered in Delaware with a San Francisco operational base, and its careers page shows active hiring. External Reddit threads from August 2026 confirm real multi-GPU workloads (8xH200 for up to 8 days) tied to its research. No billing disputes or exit scams have surfaced. However, the lack of an SLA and gated pricing API raise caution despite otherwise credible operator data.
GPU instances start at $0.47/hr for H200 variants, with H100 at $2.43/hr ($0.94/hr for Spot), B200 at $3.49/hr, and B300 at $4.99/hr. Hosted training starts at $0.02 per 1M input tokens. Pricing is in USD, with standard fiat payments; alternative digital payments require manual arrangement. Rates are behind an API key, so live availability must be queried directly. KYC status is unknown.
Prime Intellect is a credible option for researchers and developers needing flexible, high-end GPU access with transparent operator details and real-world usage evidence. However, the absence of an SLA and compute attestation means you must trust the provider. Consider starting with small jobs to verify performance before scaling.
| Documented price | $0.470 per GPU/hr · $0.060 per 1M tokens (output, smallest Hosted Training model) |
|---|---|
| Notes | Observed 2026-09-24 on the marketing home page (live rates gated: availability API returns 403 without key): H200 from $0.47/hr (variants $1.23/$1.80/$1.99), H100 $2.43/hr (Spot $0.94/hr), B300 $4.99/hr, B200 $3.49/hr, GH200/A100/A40/RTX Pro 6000 $3.14/hr. Reserved B300 3-year worked example $5.00/hr/GPU. Multi-node H100 cluster ~$52.80/hr premium / ~$40.80/hr value (FAQ). Hosted Training per 1M tokens: Qwen3.5-0.8B $0.02/$0.06/$0.06; Llama-3.2-1B $0.02/$0.06/$0.06; gpt-oss-20b $0.10/$0.30/$0.40; Qwen3.5-35B-A3B $0.25/$0.75/$1.00. |
Run the long math before paying: cloud mining calculator →
Ready to try Prime Intellect? Start small, verify one payout, and only scale up once the hashrate or compute actually matches what you paid for.
Based on our checks as of 2026-09-24: Prime Intellect scores 6.9/10 in our weighted review. We logged 1 red flag (listed above). User reports and operator background are documented with dated sources in the meta-review below.
You get remote access to GPU instances or hosted model training/inference via API, SSH, or Jupyter. You do not own hardware or receive physical control - only compute time on their marketplace.
Pricing starts at $0.47/hr for H200 GPUs and $0.02 per 1M input tokens for training. No maintenance fee is listed, but there’s no SLA, and full rates require API access.
The minimum purchase is $0, and you can access some pricing publicly, but live rates and availability require an API key. This allows small-scale testing before larger commitments.
Payouts don’t apply - this is compute rental, not mining. You pay for usage, and the service provides access. Usage is billed hourly or per token, depending on the service.
If operations cease, your access stops immediately. Since you don’t own hardware or have an SLA, contracts offer no recovery path. Always back up data and minimize lock-in.
Renting avoids upfront hardware costs and maintenance but lacks ownership. Unlike mining, you’re not earning crypto - you’re paying for compute. It’s cost-effective for short-term, high-intensity AI workloads.
Host-set GPU marketplace with per-second billing; on-demand, interruptible and reserved tiers, plus serverless and clusters.
Singapore-run GPU aggregator reselling capacity from 5+ neoclouds; H100 from $2.64/hr on-demand, per-minute billing.
Centralized open-model inference and GPU-cluster provider with published per-token and per-GPU rates; fiat-only and not decentralized.
No comments yet - be the first. Posting runs a short proof-of-work in your browser (anti-spam), no account needed.