@sfcompute // @CloudNativeFdn TAG Infra Co-Chair // K8s SIG Infra Co-Chair // ex- @LambdaAPI @autodesk @plangrid @digitalocean // Tweets are my own.
USA
Joined August 2011
- Tweets3K
- Following1.4K
- Followers640
- Likes24.3K
Yes, broadly speaking. Our model supports both first-party and third-party GPU capacity within the same market order book.
sfcompute.com/requirements has our requirements
Close. We’re not an exchange operator. We operate a physically settled forwards market.
We physically verify and validate the compute provided through the market platform we built.
We have the engineering prowess to burn in and validation the physical GPUs, not just broker them to another without knowing the state of the hardware.
I think it’s turned into a negative stereotype to disparage a lot of the work it takes to bring GPUs online.
SemiAnalysis definitely plays a factor in contributing to this especially after the IREN fiasco
Correct, not shipping racks. We’ve built both the market/order book and the infrastructure required to deliver specific cluster capacity for a specific window.
We track ownership down to the node-hour. When that window arrives, the holder gets the actual compute on the cluster.
We operate the software and cluster underneath to make sure the buyer gets working GPUs, not just a contractual entitlement to hardware somewhere in a colo.
Dylan Page retweeted
we have gotten a flood of questions from startups & venture firms asking how to protect themselves from the debt markets freezing.
to coordinate, we're hosting a small off-the-record q&a in sf on thursday night
you can sign up here:
forms.gle/hg9RsJh2gKdNgkfW7
SFCompute is not a neocloud
SFCompute is not a neocloud
SFCompute is not a neocloud
It’s a compute market that also operates and handles physical settlement of clusters.
This is VERY HARD to do.
SF Compute exists to make compute accessible for start ups while supporting larger customers. We do this through our underlying orderbook, which allows larger tenants to resell their idle capacity for the little guys to sublease and use. We invented compute markets through this process. Our orderbook makes us a forwards market that’s physically settled, where physical settlement means we design and operate the clusters behind the orderbook. You can actually trade physically-settled block compute on our platform today.
We’ve been in the space since 2023. A lot of people have asked what we’ve been doing for the last 3 years and why it had taken us so long to emerge as a compute market. It’s actually extremely hard to do physical settlement. The plethora of compute markets you’ve seen popping up on your feed lately are not physically settled.
SF Compute is not a neocloud. We could have made so much money by being extremely early in the space but decided not to, in order to create a new type of business in the compute industry so that start ups can continue to exist. You can get GPUs on both long term and short term contracts, you can resell what you don’t need to lower your overall compute price, you can block trade compute on our platform, and if you wanted to own the hardware yourself for its residual value, we can operate it for you as a service. SF Compute is everything you could want in one company.
We’re kind, earnest people who reinvent the wheel. We’ll never try to sell you something we’re not proud of. We always try to give the very best for all the people we work with.
Dylan Page retweeted
You can literally do this on our platform. People who have resold their capacity with us have realized a 25% discount on their overall compute price.
Dylan Page retweeted
We're helping a number of funds explore this option.
If you'd like to secure compute for your portfolio, while you own hardware & its residual value, please reach out. (DMs open)
We're a full supercomputing cloud that will make the process turn key for you.
Good morning
045fa83b7c7a42dba1acb2c19794c1bf1091c448eb9c483503e8f8e110fcec42
github.com/kubernetes/kubern…
I’ve heard good things about Mill Valley. Looks like a northern Michigan fall if you hadn’t mentioned the location
Dylan Page retweeted
Experts in building and managing supercomputers are hard to find. Good thing we have a ton of them at @sfcompute. Plus we’ll manage the compute across your entire portfolio, and help you resell whatever you don’t use. Yes, you can have your cake and eat it too :)
Dylan Page retweeted
messaging my wife a link containing proof of her husband cheating on her from an anonymous instagram account and then when she clicks on it texting her to let her know she failed the family phishing test and she has to retake my cybersecurity course
Dylan Page retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
It’s not. I’ve seen it first hand. Usually smaller startups have lower utilization vs bigger companies that have experienced cluster operations teams. RMAs and GPU failures and improperly configured Infiniband partitions account for most of it.
I’ve seen utilization anywhere from 40-90% depending on cluster, customer, workload. It’s lower over time if it’s a cluster dedicated for training vs inference.
It’s obvious we are running into the physical limitations of power. No matter how many parallel datacenters you build or retrofit, power will be the stumbling block.
Demand for compute will not go down. Supply will continue to be out paced and prices will increase in the short to medium term.
One thing folks don’t see or hear is from the cluster operators about how many of these GPU clusters sit underutilized. They are being hoarded.
SFCompute provides both the ability to price the future, improve financing conditions, and unlock the ability for asset owners to resell their GPUs when they are not being used.