$7.9 Trillion AI Hardware Boom Stalls at Data Center | Sam Ramji

Sam Ramji (0:00)

By 2030, about $7.9 trillion will have been spent collectively on AI compute.

SPEAKER_2 (0:06)

Wow.

Sam Ramji (0:07)

$7.9 trillion.

SPEAKER_2 (0:08)

Oh my gosh.

Sam Ramji (0:08)

Right.

$4.7 trillion will have gone into AI hardware alone by 2030 Wow. But the industry can't actually stand it up fast enough. It's deploying capital way faster than it's able to deploy the hardware, but it's left the people who do the difficult work in factories and data centers behind. They're still doing stuff, breakneck speed, it's challenging, it's physically difficult. They're in these high decibel environments that are tiring and dark, they're hot, all these things are going on. So we see an opportunity to make the entire system better by serving the people who are doing the work.

SPEAKER_2 (0:45)

Hello everyone and welcome back to MTS. Today I am joined by Sam Ramji, who is the co-founder and CEO of Sailplane. Thank you so much for joining me today.

Sam Ramji (0:53)

It's a privilege. Thank you for the time.

SPEAKER_2 (0:55)

Let's talk a little bit more about Sailplane. What is the core problem you guys are trying to solve?

Sam Ramji (0:59)

Well, the problem that we see is that there's an enormous amount of money going into AI hardware. By 2030, about $7.9 trillion will have been spent collectively on AI compute.

SPEAKER_2 (1:10)

Wow.

Sam Ramji (1:11)

$7.9 trillion.

SPEAKER_2 (1:12)

Oh my gosh.

Sam Ramji (1:12)

2017, we were all flabbergasted that all the hyperscalers together, we're going to spend $17 billion. I feel like Austin Powers. Yeah.

$4.7 trillion will have gone into AI hardware alone by 2030 Wow. But the industry can't actually stand it up fast enough. It's deploying capital way faster than it's able to deploy the hardware.

That's the fascinating problem that we see of all these different constraints of chips, and power, and people's expertise. We really think that the revolution here is serving people, but it's left the people who do the difficult work in factories and data centers behind. They're still doing stuff, breakneck speed, it's challenging, it's physically difficult. They're in these high-decibel environments that are tiring, the dark, they're hot, all these things are going on. We see an opportunity to make the entire system better by serving the people who are doing the work.

SPEAKER_2 (2:08)

What I'm hearing right now is people are deploying capital very fast but the actual hardware needed to make these things work isn't being innovated at the same speed. Where exactly are you guys coming in here?

Sam Ramji (2:21)

We're an agentic software stack company that we're sitting between the hardware and the people who want to use the hardware. Let's say Anthropic who's running a big workload on a company like CoreWeave. We're serving those who are manufacturing the hardware or validating it or then getting it into production where you're cabling the whole system together so you can run as a cluster so you can run these giant workloads. What's really interesting about these systems is NVIDIA has built these data centers in a rack. So I'll tell you a little bit about the racks, what makes this really interesting. These racks are called the NVL72, right? And they're like Jensen's sort of dream child and baby, and they're awesome. So it's called 72 because it's got 72 GPUs. So 72 Blackwell GPUs, that is a lot when you think about what does a B200, B300 cost.

These racks are about seven feet tall. They weigh three and a half thousand pounds. They take gallons and gallons of coolant. They've got 1,296 simultaneous cable connections, so that NVIDIA's NVLink will light the whole thing up as if it's one giant GPU. Yeah. It's so powerful that it's become the only game in town for the Frontier Labs and for the Neo Clouds. So these things are just flying off the shelves. Everybody wants them, but they're so Frontier level in pushing the boundaries of physics and what the hard work can handle, that they're finicky. So they require a lot of expertise just to get up and running the first time in a factory, let alone bringing them up, maintaining them and keeping them running. A lot of Neo Clouds that we've talked to are running 10-20 percent spares. Now, to put that in context, you've got a $10 billion data center.

Then you're spending another $2 billion on spares to make sure that you can keep meeting the service levels.

SPEAKER_2 (4:04)

That's interesting because what you're saying here is that actually getting to run these, especially these machines are really finicky. So there's all these things that happen behind the scenes. So maybe NVIDIA is not particularly working on, maybe the teams themselves who need this infrastructure aren't particularly working on. So you identified this as a market opportunity that needs new types of agentic software to come involved. So you guys come to play at this level. So you try to figure out what's going on, what's going wrong. How did you actually identify that this was a market opportunity worth solving?