Entire Hydra team is excited to support El Salvador in its Sovereign compute initiative! Bitcoin + AI Accelerators is the new playbook.

1. Wat

Normally I’d agree, but Elon is the rare guy who can make disparate focus areas work at scale. He’s not just going to quit models and with his track record he shouldn’t. And ultimately models won’t be a sustainable business for anybody, it’s a loss leader to sell chips and compute hours. Perfect for SpaceX.

So you’re telling me we have a ways to go

Alright that sold me

Elon’s optimizing for owning as much compute and the power required for it as possible. Thats it, that’s what everything boils down to on a long enough time horizon. He correctly realizes it’s the most important resource in existence and is building a verticalized engine to secure it. Space is ultimately just about having access to continuous solar power and infinite room to house it all in. Starlink is primarily a useful bootstrapping mechanism, and secondarily a highly redundant comms network to connect all that compute. Tesla is autonomous logistics and robotics to operate the whole supply chain for a global network of Terafabs and Gigafabs. XAI/SI is leveraging those compute resources to generate higher margin returns and also can be used internally to solve bottlenecks and automate the supply chain using Musk proprietary training data, speeding up the flywheel. It’s nearly perfect really. The only thing I feel is missing is terrestrial modular nuclear plants to power the fabs and factories continuously. But he can easily acquire one (including the one that spun out of Boring company and he may already have significant stake in iirc). SpaceX/Tesla is becoming the actual self replicating autonomous super intelligence.

The large spread between GPU hours and inference revenue has been making the rounds. If you own the GPUs and are on the hook for the debt, inference is high risk high reward. Is that token pricing going to last long term as competition continues to increase? Are you going to be able to keep up with other token providers innovation? Maybe. Long term bare metal GPU contracts are conversely the lowest risk play. Lock in a contract or two, and comfortably pay back the hardware in 2-3 years. Exceptionally healthy returns, but upside is completely capped. The security is also arguably illusory. On-demand bare metal GPU hours are in the middle and being slept on. Risk is not materially increased from long term contracts unless you think a massive crash is coming (in which case your long term counterparty is just as at risk and why are you in the business anyway), and your upside is uncapped.

The SF group think issue is real. Hydra Host had zero SF engineers and was the first company to correctly identify bare metal as the most critical part of the Neocloud stack, especially when it comes to standardizing deployments for functional compute marketplaces. It just wasn’t hot or buzzy at first. Now all the SF neoclouds and compute marketplaces care, including customers and former customers of ours that originally dismissed it. Independent, first principle thinking always wins.

It’s pretty obvious Apple will meaningfully compete with NVIDIA before AMD and Google do.

Obvious when the biggest one already did, it’s called SpaceX. Lots of companies that imagined they will train capable models or run capable inference services will end up “just” renting bare metal.

And it looks like it'll be able to run on 4x M5 Ultras.

That's the inherent problem with Anthropic's paternalistic stance on blocking "offensive" capabilities. Offense is the best defense. You're either sitting around waiting for officially disclosed CVEs and patches, until which you are just completely helpless. Or you can pen test the hell out of your own systems and find the exploits first.

Been waiting for them to drop this for some time! Congrats to the team!

If the AI Infrastructure market has a collapse, it will be because we can't make GPUs efficient enough fast enough. Not the opposite. Few.

Hello @FoundationHQ