Building AI models shouldn't require a spreadsheet degree to figure out where to buy compute, but current market fragmentation makes it nearly impossible to compare costs at a glance. A new tool called Taipeiledger aims to fix that by tracking daily GPU rental prices across four major providers: AWS, Azure, RunPod, and Vast.ai. The dashboard, which launched on Hacker News on November 11, 2026, provides a normalized view of per-GPU-hour costs for critical hardware like the NVIDIA B300, H100 SXM, and consumer-grade RTX 4090.

Normalizing the Chaos

The core value of this tool is its methodology. Cloud providers often quote prices for entire instances, which can contain anywhere from one to eight GPUs. Taipeiledger standardizes this by dividing the hourly instance price by the number of GPUs included, giving developers a true per-unit cost. For example, an AWS p5.48xlarge instance with 8 H100s is broken down to a single H100 hourly rate. This allows for an apples-to-apples comparison between AWS on-demand pricing, Azure pay-as-you-go rates, RunPodโ€™s secure and community clouds, and Vast.aiโ€™s marketplace median.

Significant Price Disparities

The data reveals stark differences depending on the hardware tier. For the high-end NVIDIA B300, prices range from $8.99 to $17.80 per hour, with Vast.ai offering the lowest entry point at $9.94. In contrast, the widely used H100 SXM shows a wider spread: AWS charges $6.88, while RunPod Community Cloud drops as low as $1.34. Even consumer cards show variance; the RTX 4090 is available for $0.34 on RunPod Community and Vast.ai, compared to $0.89 on RunPod Secure. The tracker notes that Vast.aiโ€™s marketplace data includes the number of active offers, which is crucial context since low offer counts can lead to volatile pricing.

Key Takeaways

  • The tool tracks 11 specific GPU models, including the new B300 and B200, providing rare visibility into next-gen hardware costs.
  • RunPod Community Cloud and Vast.ai consistently offer the lowest prices for H100 and RTX 4090 cards, often undercutting hyperscalers by 50% or more.
  • Azure pricing is notably inconsistent, with some entries marked as 'source not updated today,' highlighting the need for real-time verification tools.
  • The tracker includes a sanity check that holds prices at the previous value if a daily swing exceeds 50%, preventing misleading spikes from transient market errors.

The Bottom Line

If you are renting GPUs in 2026 and not comparing per-GPU-hour rates across at least three of these providers, you are burning money. Taipeiledger is a simple, necessary utility for anyone serious about optimizing inference or training costs.