| |
| ▲ | bleepblap 7 hours ago | parent | next [-] | | What I don't understand is how it's even a good low-margin business. Maybe I'm missing something but: Data centers (before recently) are low margin businesses because all the inputs are commodities: you buy power (joules), power (PDU), cooling hardware, physical racks, etc.. from the same vendors as everyone else. Worse, your biggest potential clients have the scale to just build it on their own and cut you out because of their scale and because you don't bring anything unique (outside of maybe physical proximity to an interesting market) xAI has all the same commodity inputs plus another huge upfront capital expense (GPU/storage/networking), and their customer base is exclusively the well-funded companies who would normally just build it on their own. I assume that they can't get better deals from nvidia than (e.g.) Microsoft because of their scale, so the unit cost of their inputs is the same or worse than their clients. So the whole game is hoping that they hope to charge more now because people can't build fast enough and try to recoup their upfront costs before either a) other capacity comes online and b) the installed hardware becomes obsolete. I'm being earnest -- it seems like they're trading one tiny margin service (datacenter) for another tiny margin service, with the added difficulty that there's an additional 10 figures of upfront expenditures and their viability depends solely on paying everything off before the price floor drops. Maybe it's staunching the bleeding, but it seems like not a great move | | |
| ▲ | cco 3 hours ago | parent | next [-] | | You're not wrong in the long term, either in general or for SpaceX. In the long term, hopefully the market stabilizes, new entrants can challenge Nvidia etc. But of course maybe not! However for SpaceX, this is a dead end move. They made a good decision on buying this compute when they did but they failed to use it to create a compelling model. So they're selling access to recoup some of their investment (maybe a profit?). But what's the plan as these chips age out over the next three to five years? Become a compute company? They claim they want to... in space! Regardless, they bought some valuable chips, failed to use them, but can now sell access and recoup over the next few years before they become outdated. | |
| ▲ | redox99 6 hours ago | parent | prev | next [-] | | It's like buying a ticket for a concert, realizing you can't go and that you can resell it for more than what you paid. You're right that long term it should stabilize into a low margin business. Elon is also much less risk averse than others, which helps to build stuff fast, possibly cheaper, pushing legality to the limit. Colossus was definitely built much faster than anything else. I think building datacenters suits him better than a pure software play, where "move fast break things" is already the norm. | | |
| ▲ | bleepblap 5 hours ago | parent [-] | | The concert analogy makes sense (I analogized it as "staunching the bleeding"). WRT SpaceX building data centers: I think there's a natural tension between a "low margin business" and "being risk adverse". SpaceX (the rocket business) did well because it was high risk and high reward. Building a 10b datacenter to hope to get a slice of a low-margin industry is high risk and low reward and just seems fundamentally like a losing strategy. | | |
| ▲ | redox99 5 hours ago | parent [-] | | It's not like Elon is a stranger to low margin. Making cars is low margin, and it's not like SpaceX has crazy margins now that we know the financials. Also I think stuff like Hetzner is a commodity. But are gigawatt scale data centers a commodity? You need those for AI training. Anyways their goal is datacenters in space, not traditional data centers. Although I think that's only viable for inference. |
|
| |
| ▲ | JumpCrisscross 5 hours ago | parent | prev | next [-] | | > because all the inputs are commodities AI compute hardware is not a commodity. And in a shortage, commodities can command high margins. xAI has lots of NVIDIA GPUs and HBM. It also has permits and power hook-ups, both things that are getting harder to come by day by day in the U.S. Natural gas is a commodity. Doesn't make having lots of right now bad business. > the whole game is hoping that they hope to charge more now because people can't build fast enough and try to recoup their upfront costs before either a) other capacity comes online and b) the installed hardware becomes obsolete Correct. But charging people now generates incumbency advantages that make beating (a) and (b) easier. (From what I can tell, (b) isn't an existential issue, at least for xAI, because they've basically already recouped their investment with commited contracts they'd have to fuck up on to lose.) | | |
| ▲ | bleepblap 5 hours ago | parent [-] | | > AI compute hardware is not a commodity. And in a shortage, commodities can command high margins. I don't see the distinction you're drawing about "commodity", but I'm happy to be wrong on that. My point was that spaceX's ai division is buying all their inputs from external vendors and can't meaningfully differentiate themselves from person Y who buys all the same hardware except for the fact they bought them first. Which... > Correct. But charging people now generates incumbency advantages I don't see now this is an "incumbency advantage". There's nothing that sticks their clients to stay there and sign up for the next data center. | | |
| ▲ | JumpCrisscross 5 hours ago | parent [-] | | > don't see the distinction you're drawing about "commodity" People pay markedly more for NVIDIA GPUs than they do for others. That opposes the fungibility requirement of a commodity. | | |
| ▲ | bleepblap 4 hours ago | parent [-] | | In the west, there's no actual competitor to NVIDIA hardware. Yes, people make other chips, but nothing is a serious drop-in replacement for the nv stack. Between the networking and software, they're truly a different "thing" of accelerator, and I don't consider them fungible at all. The US government tried to build 3 supercomputers with each of nvidia/amd/intel accelerators and you can see how it went | | |
| ▲ | JumpCrisscross 4 hours ago | parent [-] | | > there's no actual competitor to NVIDIA hardware...I don't consider them fungible at all Which is why nobody should claim NVIDIA makes a commodity. |
|
|
|
| |
| ▲ | bee_rider 6 hours ago | parent | prev [-] | | I wonder if they do have non-commodity AI capabilities, just, ones that don’t translate into a world-class frontier model. Like they might have hired really good AI infra folks, gotten really good uptimes on their nodes, gotten folks who really know how to configure Infiniband (or whatever). But then, didn’t find the folks who knew what to run on that infrastructure. Or maybe Grok just had too much political drama around it. | | |
| ▲ | bleepblap 5 hours ago | parent [-] | | Maybe they have something else im the books, I truly have no idea. But once you get down from the top rung of full-bandwidth cross section networking at the 100k node networking scale "AI" infra, theres no shortage of people who can do that. Most importantly, labor isn't the big chunk of the outlay. Even if they have 50 engineers clearing $1m/yr, that's pocket change for everything else EDIT: said 50 engineers at $50m/yr originally and meant 50 @ $1m/yr |
|
| |
| ▲ | lumost 5 hours ago | parent | prev [-] | | Why do we think frontier model vendors are high margin? |
|
