10+ hours of podcasts, in 5 minutes.
21 quotes from 1 episode on 20VC with Harry Stebbings, each with a timestamped link to the source.
Vertically integrating AI data center infrastructure allows companies to bypass severe supply chain delays and capture shifting market margins. Chase Lochmiller notes that Crusoe bypassed external vendor wait times by manufacturing power hardware in 28 weeks.
A 140-megawatt data center consumes only as much water per year as 10 single-family houses.
“One of our giant buildings in Abalene, Texas that consume, you know, it's a call it 140 megawatts of power is like what's budgeted for each of those buildings on the on the first eight buildings. They use about the same water annually as about 10 single family homes.”
Building large data centers in local communities typically drives retail energy prices down for residents.
“When you look at markets where data centers have made investments and have built big data centers, typically energy prices for communities have come down…”
Expanding key-value caches rapidly overflow on-chip high-bandwidth memory during neural network inference.
“The KV cache can get quite large and so it can expand well beyond the amount of memory you have on chip in the HBM…”
Selling across multiple layers protects infrastructure builders from price cycles. Margins naturally move between electrical power, physical data centers, raw chips, and managed services.
“What Crusoe is focused on doing is what we believe the opportunity is, to build an AI super major…”
“Our margins are going to move around across electrical, data centers, chips, and services…”
“There's really three products that we ultimately sell to customers where we're making money…”
Relying on external supply chains causes extreme delays for large compute facilities. Producing electrical distribution centers internally allowed builders to reduce wait times from 100 weeks down to 28 weeks.
“The supply chain to support large scale AI data centers is sort of like a game of whack-a-mole…”
From Crusoe Cut Power Center Lead Times from 100 Weeks to 28
“When we set out to build the first two buildings in Abilene, the first it was a little over 200 megawatts of compute capacity, we had sort of committed to doing this in one year and the next closest bid was 2 and 1/2 years…”
From Crusoe Cut Power Center Lead Times from 100 Weeks to 28
“We said how quickly could we make this ourselves…”
From Crusoe Cut Power Center Lead Times from 100 Weeks to 28
Securing large load interconnection agreements with local utilities and obtaining air permits for new power generation create multi-year hurdles. These friction points will prevent an estimated 50% of announced facilities from opening.
“We've seen very prominent people say, 'Oh, we expect 50% of data centers that are planned not actually to be built and ready to go.' What percent of data centers do you think that are planned will not actually go online? 50% seems reasonable.”
“…getting a large load interconnection agreement done with the local utility…”
“…getting an air permit if you're bringing online new generation.”
“Elon has this very interesting framing of this…”
From Crusoe Cut Power Center Lead Times from 100 Weeks to 28
“Energy is definitely a key constraint so power that's available for compute is a key bottleneck…”
“It's one of these things where, you know, kind of coming back to this notion of thinking like a mountaineer when you're going through these planning processes. A lot of things can go wrong.”
“Exxon very famously doesn't hedge their oil exposure…”
“The way we depreciate the assets today is we use a six six-year depreciation cycle…”
“When you look at that business where it abstracts away the actual underlying chip, the underlying compute from the service that people are receiving, it opens up new monetization engines that can persist for a much much longer period of time…”
“Here we are 3 years later and the prices being charged for utilizing hoppers is higher than the rates that were being charged 3 years ago when they were when they were brand new…”
“I think people underestimate the ingenuity of applications and application developers of turning compute capacity into value and into valuable services for the economy.”
“When tokens are fed into a large neural network you can compute the output tokens by basically running this feed forward process and running all these matrix multiplications. That takes time. You may also already know the answer of that output token.”
How we attribute quotes. Every quote was matched against the episode transcript, so the words and the timestamp are real (we trim filler words like "um", nothing else). The name comes from our written summary of the episode, and we use it only when a separate check of the captions finds that person on the episode. YouTube gives us no voice-by-voice transcript, so open the timestamp to hear who is talking. See a wrong name? Tell us and we fix or remove it.
10+ hours of podcasts, distilled into one 5-minute read. Free, every Sunday.
For now, every subscriber gets both newsletters. No spam. Unsubscribe with one click.