Did you see the pricing fine print? Hailuo charges extra for 'premium' data types.

I've been conducting a detailed cost analysis of several emerging GPU-as-a-service providers, with Hailuo being a primary subject due to their advertised competitive rates for A100/H100 instances. While the base per-hour pricing appears attractive at a glance, a meticulous review of their documentation and pricing schedule reveals a significant and potentially costly nuance: supplemental charges for what they term "premium" data types.

Specifically, their standard FP16 and BF16 precision compute is covered by the listed instance rate. However, utilizing FP8, FP64, or even INT8 precision operations incurs an additional surcharge. This is not a trivial implementation detail; workloads leveraging mixed precision training, certain inference optimizations, or scientific computing may automatically engage these data types. The surcharge is applied as a multiplier on the base GPU cost for the duration the "premium" type is active.

Consider the following breakdown for their `H100-80G-SXM5` instance, assuming a listed base price of `$4.12`/hr per GPU:

* Base Compute (FP16/BF16): **$4.12 / GPU / hour**
* FP8 / FP64 / INT8 Surcharge: **+20%** (as per current documentation)
* Effective rate during FP8 operation: **$4.12 * 1.20 = $4.94 / GPU / hour**

This represents a **$0.82** per GPU hour premium. When projected over a sustained workload—such as a training job spanning multiple days—or scaled across a large cluster, the cost delta becomes substantial. For a 64-GPU cluster running an FP8-optimized training job for 7 days, the premium alone calculates to:

```python
premium_per_gpu_hour = 4.12 * 0.20
total_premium = premium_per_gpu_hour * 64 gpus * 168 hours
# total_premium = $0.824 * 64 * 168 = $8,859.65
```

This nearly **$8.9k** is an *additional* cost on top of the already ~$44k base compute cost, purely for the data type. This pricing model stands in contrast to providers who typically include all standard precision types within a single vCPU/GPU hour price.

My key questions for the community are:

* Has anyone encountered this in practice? Can you confirm how granularly this surcharge is applied—is it per-second billing only when the kernel uses FP8, or is it triggered for the entire instance if any process uses it?
* What monitoring or billing line-item details does Hailuo provide to track the activation of these premium data types?
* Are there any configuration or environment settings (e.g., specific container images, framework flags) that can inadvertently default to a "premium" type and thus inflate costs unexpectedly?

This underscores the critical importance of moving beyond headline instance rates. A true total cost of ownership (TCO) analysis must factor in data type requirements, network egress, storage I/O patterns, and instance startup latency. I am compiling a comparative spreadsheet and will share my findings, but empirical data from actual users would be invaluable.

-cc