Compute @FlumeFabric, Co-founder @Supertraceai, @flumeinternet. Proud Cal 🐻
New York, NY
Joined March 2009
- Tweets907
- Following932
- Followers461
- Likes2K
👀 hmm
Exclusive: Amid a server shortage, Anthropic has been asking where to find 5 to 10 MW of power in older buildings, as demand revives interest in decades-old telecom facilities.
Read more: thein.fo/4AjtXQm
1
153
It’s almost impossible to get 1-3MW colo right now. We’re building dark fiber and powered shell to extend Colo sites out to buildings with power and cooling in cities.
1
2
175
Small distributed sites are def the way. The challenge with Brightspeed sites that we’ve seen in the CLEC (old telephone sites) we’ve surveyed is that they don’t have the chilled water capacity you need for 7MW of liquid cooled chips.
1 - as soon as you need to roll in a new air cooled chiller unit you run into environmental permitting
2 - if you drop in air cooled B300s your PUE goes through the roof and you are probably at 2MW of critical IT
Still definitely the way forward vs large centralized sites.
Exclusive: Nvidia has submitted a proposal to Brightspeed, an Apollo-owned broadband provider reportedly nearing bankruptcy, to convert some of its 130-plus telecom buildings into AI data centers.
Nvidia estimates the sites have roughly 1 gigawatt of available power, potentially enough to run tens of billions of dollars of Vera Rubin chip systems.
Read more: thein.fo/4AjtXQm
2
143
The second order effect of the craziness in the compute market is that the lending markets have become insane and sometimes shady.
Seeing 35% IRR equipment lease deals or waterfall JVs get slung around by boutique banks who are sometimes hiding behind two levels of syndication. No transparency on fees with term sheets being offered in 24 hours and sometimes pulled just as quickly. If you're financing compute clusters in the 1-4MW range stay safe out there.
2
124
Talking about how many tokens your platform serves is like an ISP in 2005 talking about how many bytes of data it passed on its network.
1 - Will be abstracted away by newer pricing models that are all you can eat, with token caps that are too high to keep track (similar to ISPs throttling at high data caps)
2 - the volume of tokens is not as important as the type of models and workloads it was spent by
Standing up 18GW in two years and making compute 82% of cash flow would be epic. Wouldn’t bet against them but:
1 - They’d have to have all of that in construction now at sites not yet made known to the public 🧐
2 - Also curious where the debt line item sits in this
Over the last week I built a bottoms up SpaceX financial model through the end of 2028
The numbers that came out are insane
There are wide error bars given how quickly things are changing but I wanted to distill something complicated into something that can at least be debated
* This is not investment advice!!!
** Sequoia is a major SpaceX shareholder
*** Past performance is not indicative of future results
1
6
386
GPU clouds building 1-3MW clusters that can be turned up quickly with debt that amortizes in under 3 years can handle higher credit costs than the clouds building 30MW+ sites mainly due to time to RFS from drawing down on their loan. People will soon realize these are two different categories.
The middle market for compute financing ($10M to $300M) is alive and well.
This is what we’re seeing from our OEMs who are saying hey can ship B300s to us quickly. Priority is given to those who can install fast
Microsoft said that "some third-party GPU-as-a-service contracts have been delayed to the second half" of FY27, so calendar quarters Q1 and Q2 2027..
I wonder which "neocloud" contracted to deliver capacity to Microsoft in Q4-26 is running behind schedule..
216
How does one make a futures market / derivatives market for compute when the majority of it is sold B2B? Is there any other commodities market that has this much price opacity? Everything else is a consumer good, I don't get it.
118
Yup - I track our compute sales folks progress by the number of shared slack channels they set up every week
buying huge amounts of compute is still weirdly informal.
as much as I like thinking about indexes and markets, they exist on top of how compute is actually bought today (slack messages, negotiations via calls and brokers, price changes, gpus being given to someone else last minute, etc)
you could make a gpu price index this weekend by collecting prices from providers, deciding how to weigh them, and then coming up with a number. frankly the barrier to making an index is low.
just look at the different prices already for an h100 (using different sources, measuring different parts of the market):
- $1.55/hour (GetDeploying, cheapest available)
- $2.53/hour (Silicon Data, standardized neocloud index)
- $2.96/hour (Ornn, completed on-demand rentals)
- $3.27/hour (CF Benchmarks, approved provider panel)
- $3.40/hour (GetDeploying, median across providers)
the idea that everyone's pushing is "standardization". indexes are just one way to do that for the price.
making contracts easier to compare and streamlining the process could be a different adjacent approach to standardizing compute!
2
181
It helps the financial underwriting if you can reduce the time it takes from putting the downpayment on the chips to first token produced by 75% 😉 @FlumeFabric
The biggest bottleneck in the GPU industry at the moment is not chips, CoWoS, or even power. It's credit.
Sharing some thoughts on compute markets, residual value, and where compute financing is heading.
Substack link in the comments.
2
222
Just got an email for a batch of 200x 3 yr old Dell PowerEdge XE9680 - 8x NVIDIA H200 (141 GB) / 2 TB DDR5 servers, on the market for $450k per… what
107
