one RTX 5090 and its AI video making capacity and bandwidth needed to serve

 This is the math that matters. RTX 5090 can make video faster than you can upload it.

1. Max limits of RTX 5090

• VRAM: 32GB GDDR7, 512-bit bus 
• Memory Bandwidth: ∼1.79 TB/s / 1,792 GB/s 
• CUDA Cores: 21,760 
• AI Power: 3352 TOPs 
• Power: ∼575W - will thermal throttle if you run 24/7 non-stop in a closed case

Video limit because of VRAM: You CANNOT generate 10 min in one go. Current max is ∼10 seconds per generation. For longer videos, you need to chain clips. 

2. Highest card available today

• Consumer: RTX 5090 - 32GB - best for you
• Workstation: RTX PRO 6000 Blackwell Server Edition - 96GB GDDR7, 24,064 CUDA cores, 752 Tensor Cores, 600W TDP, Bandwidth up to 1597 GB/s 
• Datacenter: Nvidia B200 - 192GB HBM3e, ∼8 TB/s bandwidth. 4-5x faster than 5090, but $30k and not rentable on Vast.

For LTX, PRO 6000 can load full 40GB+ model without quantization. A single PRO 6000 is as fast as 4x RTX 5090s.

3. If you run RTX 5090 non-stop, how much video will it make?

LTX has 2 versions:

LTX-Video 13B [fastest, what you will use]:
Can generate a 5-second clip in roughly four seconds on an RTX 4090. RTX 5090 cuts that time roughly in half.
So on 5090: 5 sec video in ∼2 seconds. 

That is 2.5x faster than real-time.

• Per hour: 3600 sec / 2 sec = 1800 clips x 5 sec = 9000 sec = 2.5 hours of new video per 1 hour of GPU time
• Per day non-stop: 60 hours of video per day per 5090

LTX-2 / LTX-2.3 [4K quality]:
RTX 5090 generates 4K 10-second videos in ∼3 minutes, 720p 4-sec clips in ∼25 seconds. With NVFP4, generation is 3x faster. 

So 4K: 10 sec video in 180 sec = 0.055x real-time

• Per hour: 20 clips = 3.3 minutes of new 4K video per hour
• Per day: 80 minutes of 4K video per day per 5090

4. Bandwidth needed to SEND all that to users

Assume final delivery bitrate:
720p = 8 Mbps = 1 MB/s = 3.6 GB per hour of video
4K = 35 Mbps = 4.37 MB/s = 15.7 GB per hour of video

If you run ONE RTX 5090 non-stop and ONE user downloads everything as it's made:

Model	Video Made Per Day	File Size Per Day	Upload Speed Needed to Keep Up
LTX-Video 720p on 5090	60 hours	216 GB	20 Mbps
LTX-2 4K on 5090	1.3 hours	20.5 GB	2 Mbps

[sustained]

Shocking right? 20 Mbps is enough for non-stop 720p factory. This is why GPU rental sites can offer cheap bandwidth.

But if 100 users want to download the same day's output at same time:

Multiply by 100.

• 720p factory: 20 Mbps x 100 = 2 Gbps upload needed
• 4K factory: 2 Mbps x 100 = 200 Mbps upload needed

If 1000 users: 20 Gbps and 2 Gbps.

If you use the highest card - B200:
B200 is ∼4-5x faster than 5090. So it makes ∼300 hours of 720p video per day = ∼1 TB file size per day.
Single user needs 100 Mbps sustained to receive it. 100 users need 10 Gbps.

Bottom line for your startup:

One RTX 5090 will never saturate a 1 Gbps fiber if you serve 1-10 users. The bottleneck is NOT bandwidth, it's generation speed for 4K.

If you want to serve 100+ concurrent users downloading final 10 min movies, don't serve from the GPU machine. Let the GPU upload once to R2 at 10 Gbps, then R2 + Bunny CDN serves 10k users at 100 Gbps edge.

That's how RunPod can offer FREE egress - because the real bandwidth cost for non-stop video is tiny compared to GPU cost.

Want me to calculate how many 5090s you need to support your 10-min movie workflow for 100 daily users?