I’m saying that in the medium term (next year or two), opencode will struggle to get their next clusters to scale, and their largest bottleneck will be credit. If i had to guess what @thdxr’s biggest problems in getting new clusters are, i suspect they would be the following: - everyone is sold out - all the new clusters have long lead times and no matter what dax wants to pay he can’t get it earlier - everyone asks him for a large prepayment, say 30% to 100% on a long contract. If he asks for a shorter contract, the prepayment % goes up, not down. - any cluster he wants is likely to be delayed - he’s asked to bet his demand over the next several years, and if he’s wrong in one direction, the business either doesnt grow or blows up. all of those problems are the same credit problem i am describing. OpenCode’s incentives, regardless if it’s what they want to do, is to sign larger, long term contracts with credit worthy counterparties. This is the same problem that @ivanburazin just hit with Daytona. Daytona isn’t IG, and the cloud they signed up with couldn’t get the debt, so that cloud came back and asked them for a larger prepayment, because the terms of the clouds debt wasn’t expected.

okay public speaking is hard, apologies for the folks who showed up to watch me overheat and bomb, so here’s all the stuff i meant to explain right now you are being asked for a near 100% prepayment because neither you nor the cloud is considered credit worthy, and so nobody can get debt. As ai scales, we are asking the credit markets to take on a massive single sector deployment of capital. More money deployed, more scared everyone gets. this problem will flow down into the token layer. Every inference provider or model lab struggles to be credit worthy enough to get compute, and their lever to improve their situation is to lock in big prepayments from their customers or long term contracts from credit worthy customers. So this problem doesn’t just effect compute buyers, it effects token buyers. We’re entering a world where fortune 500s can very quickly deploy huge token spend ON CREDIT where the startup will need to raise capital. Startups got to exist because human capital could leave a firm and get ahead of the big firm by exploiting a combination of incentive misalignment, speed, and taste. You probably have the hidden assumption that the fortune 500 is slow and incapable of competing with your ultra fast startup. This belief does not necessarily hold in the case where the agent does the thing. the “market” is working on solving it. We know, because we are the market, and we’re working on it. sfcompute is working on ways of firming up the collateral of the loans, which makes it safer to lend to a startup. this is basically the whole reason we exist, the problem everyone is experiencing now is the problem junelark, our little ai lab years ago, hit however, we won’t succeed in time. so, in the short term, startups and venture firms should learn to protect themselves. we are mostly watching startups and investors make unforced error after unforced error, because this is not a simple subject. first, venture firms need to understand that they are in the compute business now and deploying capital to pay for a 100% prepayment for 10 startups where 6 of them die and 4 of them make 50% margins on that capital is basically locking in non-venture returns. many funds, via low understanding of the space, are wrecking their funds. there’s an economic solution for vc’s and it is owning the compute and trading equity for gpu/hrs. If the company dies, you keep the residual value. Each of you are looking to capture the next anthropic or openai, but the only way these companies scaled was by trading compute for equity. meanwhile, startups have three options in a credit crunch. The first is to raise a wild amount of capital, enough to pay for your deployments with far larger prepayment or to survive until things chill out. theres a reason we’re seeing a number of multi-billion dollar rounds simultaneously. The second is to design your business model to support variable priced tokens and compute. As the financial ecosystem builds around compute, you will be able to trade “lower up front payments” for “variably priced compute or tokens”. Your business will get tailwinds from the infra under you if you setup your business model such that YOUR customers have variable prices and you’re just adding a % take rate on top of the underlying floating price. Every year, you will become more and more capital efficient. The economic winners in ai & inference are firms that make lower and lower margin every year, but become more capital efficient & scale faster every year, making their gross profit increase. The third option is asynchronous or interrupted products. If you have a variable cogs, but a fixed rate card for your customers, it can’t have 100% availability or you’ll go bankrupt. But it is totally possible to design your product such that customers like this.

Starting soon, an investment grade company will be able to deploy hundreds of millions of dollars very quickly, to do things that in theory, a startup could do. I would like to live in the world in which those 2 people get "unlimited AI tokens" to pursue their own thing. Instead, I think we're heading to a world where that startup will need to pay a large prepayment, which they will not be able to raise from investors. The way to fix this is by unlocking lending against collateral, rather than lending against credit quality of the offtakers. We must make this work like cars or houses, where a wide variety of folks can get a loan, rather than like semiconductors, where starting a new fab is a geopolitically difficult feat to accomplish.

Givemeanode is now SF Autoresearch, a new product from SFC. We built the scalable, resilient platform for agent-driven machine learning research. The current moment is important and we want to help you take it seriously.

right now, buying compute is just an absolutely awful experience, regardless of who you buy from. it's backroom deals, backed up supply chains, broken hardware, and large, risky purchases in a high stress environment. san francisco compute exists to try and solve these problems. our team has grown from the tiny ai lab that was desperately just trying to sublease our cluster, to a force to be reckoned with: real veterans of the industry, the same folks that built & financed the other neoclouds. our team is incredible & has produced some remarkable work that i'm really excited to celebrate, but we haven't solved the problem yet. we have solved a few things, but our jobs not done. i often don't talk about sfc's success because even though we have grown ridiculously fast, there is still just so much more ahead of us. we see the hard technical problems that accompany supercomputing and the nasty financial problems that accompany its financing as linked. the path to scaling ai is through safety. the bottleneck of ai is risk and our mission is to reduce that risk. it's a tricky, nuanced problem that requires bright minds. if you're one of those, we'd love for you to join us at the san francisco compute company. we have a long way to go

if there ever was a time to take a big breath and take things seriously, now would be the time put down the rage brain, that which is moloch, and think if we have but one shot at making this go right, we should make that shot wisely that extends beyond the labs, but to the peanut gallery that watches from the streets of san francisco and around the world you, the reader of this post, should take stock of the situation, tune your beliefs towards what you ACTUALLY believe is true, not what that monkey brain feeling thinks is true. gun to your head what do you believe? lock in. get some sleep and forgive the people you normally fight with. we, the people in this industry, are likely more "in it together" right now than we have ever been before. the world will look to us to explain what is going on. the world will misinterpret, they may ridicule, they may criticize, but it is our job to do the right thing anyway. it is our job to lead, because there is no one else who can.

this is basically what i'm worried about a populist backlash towards this seems like it could fly off in directions that make alignment & other problems harder to solve, not easier, because the average person will misunderstand the problem

the average on-platform customer on sfcompute saves about 25% it's the difference between $4.5/hr and $3.3/hr we wrote about why and how you can scale faster with reduced risk

i remember when the general theory was sf was going to die, because the big, greedy, tech companies moved out of the city and took the tax rev with them the only people who stayed were these small unknown companies like openai and a bunch of random startups like anthropic

strongly agree, people who are saying we should build infinite ai-optimized tiktok seem like they hate their children and want them to live in a worse world This is real-world simulator, for heavens sake, use it to build robots!

hi! please meet it's a new autoresearch product from the san francisco compute company. try asking it to "download insert-model-here off huggingface and optimize it 4x and don't stop until it's done." let it run for days. you know those massive grab bag of black magic tricks others do? we want to let you do that too, for cheap. lets let anyone train models. it's backed by a big, fair queue that makes sure nobody can just buy out the whole thing, so there's always space for experiments. struggling to get gpus? try here <3 it's not just a business promise, it's how the product works! it doesn't require you to sign some big long reserved contract for years, where you're betting your balance sheet on your gpu bill, because it's backed by the liquidity of sfcompute. it's built for ridiculous redundancy & reliability. jobs can die and restart from the same point & the scheduler keeps headroom for you. even the platform itself going down does not stop your agent, because we've built ticketing straight into the mcp that lets your agent recover from outages. we've taken the thing offline, started it up again, and agent-supervised training runs just keep on going. it's got platform-native support for grant programs, meaning you can buy capacity, and sponsor other accounts. it's got s3 compatible storage, hugging face integrations, scratch disks & persistent volumes, clock locking, ncu support, metrics, logs, traces, checkpointing, and a whole grab bag of other stuff. but you shouldn't worry about that. let your agent sweat those details. we'd love for you to try it, DM me for some GPU credits if you'd like!