Saw the sync so CTO’s post asking for 256 H200s on a 1-year lease. Is GPU capacity really this tight, or is it normal for a video model lab? [ i will not promote ]

Saw a LinkedIn post from the CTO of sync labs (the lipsync/dubbing API, YC W24, did Wav2Lip back in the day). They want around 256 H200s or B300s, 32 nodes, 1-year minimum, and say big labs are taking all the capacity. Two things I can't comprehend. First, is it really this hard to find? Median on-demand H200 is about $5.15/hr right now and up roughly 47% over last year. 256 of them for a year is about $11M at that rate, less if reserved. But stock trackers show H200 rentable most weeks. Is the "can't find it" about 256 together on one fabric, not about single GPUs? Second, do lipsync and dubbing actually need this scale? The model is a face edit on existing video, not a full video generator. Is that mostly training cost, with inference cheap? sync's public price is around $0.05 per second of video for the base model, and HeyGen and Kling list lipsync at about $0.84/min. How does a company at that price pay for a cluster like that? They raised a $5.5M seed in 2024, I couldn't find anything later lmao