Hacker News Reader

Comment by dgellow | Hacker News Reader

> A year from now, who knows what the situation is going to be like.

That cuts both ways, we are building datacenters for an immature technology that is quickly evolving. We have no idea what AI will look like in the next 5-10y. Everything that is planned to be built is based on the demand we see right now, not what it will be in the future. That means different GPUs that require different cooling systems, different power supplies, etc. NVIDIA already broke backward compatibility with their new cards, which requires a different infrastructure.

What is unserious is the opposite position: believing that we already know what will be valuable in the future and bet the entire economy on it, without any proof of positive ROI.

View on HN

Directionally we are seeing demand for more compute.

It's a reasonable assumption that data centers that are set up for large power usage and cooling will be valuable.

Claiming the opposite based on, well, nothing at all, in order to forecast a correction, seems less reasonable.

"Compute" is not one generic commodity. Filling your datacentres with ASICs that do nothing but compute SHA256 for Bitcoin mining was a smart move in 2015 but today that hardware is worthless e-waste.

What does the depreciation curve look like for nvidia cards purchased today? How long will it take to recoup the investment on this buildout? Will those datacenters pay for themselves before they're scrapped?

It's an interesting question. Some napkin math:

Let's say we serve a Fable class model on 8x B300.

From Kimi K3 metrics, with 8x concurrent streams, we would achieve 55-60 tok/s per stream, matching Fable 5.1 throughput.

432 tok/s x 3600 => 1.555M output tokens/h x 50$/M API price = $77.76 revenue per hour.

Assuming total API billing at 2.06x output token bill = $160.2 / hour or ~$20 per B300.

A server with 8x B300 could be $461.5k.

At an obviously unrealistic 100% utilization we would look at 4 months of revenue to match the cost of the server.

About how model serving works at scale and actual utilization I know little.

And for all we know Anthropic could serve their model with 64 streams on the same hardware instead of the 8 we assumed here.

Put succinctly, if a mismatch between the timelines of the bonds in the bond markets and the revenue streams occurs, it's a bubble. This increases the risk to current and future AI investments and buildouts. If the mismatch is strong enough, contagion in other sectors/areas of the world can cause a run to security and pop the bubble.

Just because the bond markets 1/2/3/5-year bonds are bloated does that mean the bubble will pop. It just increases the risk.

Reply on news.ycombinator.com