[tadasv]

I hope that open models will dominate. The difficult part to reconcile for me is the amount of compute that's required to create and run such models. Small models are fine, I run local llms 27b param on a gpu, but it's not even close to frontier in capability. Who wants to drop $40k+ on hardware to run these things. Companies, maybe/perhapts. On the other hand, to run a DB I can get a server for $3k and handle tons of traffic on it and other things too.

[Avicebron]

Has anyone tried to run a data center as a Co-Op?

[fc417fc802]

A data center or a cloud? It's not difficult to find a good data center to colo at. The problem is then you have to bring your own hardware, technicians, and sysadmins.

However, if you don't trust cloud providers or inference providers for whatever reason then you probably aren't going to be excited to enter a co-op model where you're still effectively renting access to hardware that you don't directly own. There are already reasonably priced options to rent bare metal from a cloud provider.

The only way I see it working is if it's a bunch of medium to large sized businesses getting together to be able to rent out the spare capacity on hardware that they physically control. So an AWS equivalent where each rack is owned by a different company and retail VMs migrate between them transparently. But I question the overall economics of such an arrangement.

[embedding-shape]

> enter a co-op model where you're still effectively renting access to hardware that you don't directly own

If we're imagining a co-op, then the participants should all be equal owners in an organisation that owns the hardware itself, otherwise it's not much of a co-op really.

[fc417fc802]

But at that point you don't have the same sort of security guarantees about the hardware. This isn't like a farming co-op where a single expensive piece of equipment gets passed around to each participant one week at a time. There needs to be either an economic, security, or other advantage to entering into this arrangement for a medium to large sized player that would otherwise be colocating multiple racks in a privately owned datacenter or else renting bare metal instances in bulk from one of the hyperscalers.

[mattlutze]

Participants as equal owners doesn't preclude security guarantees.

The advantage is having a voice in setting direction, investment and ultimate shared incremental profit recovery over the rent capture from hyperscalers.

It's an interesting idea really. But it also comes with cooperation, which means trust, and a lot of for-profit enterprises are bad at operating cooperatively outside of 2- or 3- way engagements.

[orbifold]

I was thinking that would be great, too. What would be the equivalent for the property developer: one gpu server is 450k.

[rzerowan]

Would actually be a good biz model for the Colo facilities that keep shutting down as everyone moves to the big cloud providers.Now if they can get their hands on enough GPUs and RAM.

[eru]

Why would it be a good business model?

[danielheath]

I think it aligns incentives way better (but is almost impossible to set up).

[eru]

Incentives are pretty easy to align, if you have repeat customers and repetition. McDonald's sells me burgers I expect at a reasonable price, because they want me to come back tomorrow and not go across the street to Burger King.

[orbifold]

can you give examples of colo facilities in SF / New York that are shutting down?

[rasur]

coopcloud.tech ?

[rzerowan]

I believe until the hardware designs catch up to be more commodized ala cryto mining evolution from GPUs to ASICS for specfic algos. Designs (like Google TPUs equivalent) would also need to evolve to be more memory dense to be able to handle them. Untill then it seems will be system time shares for the larger models , probably with a bring your own model and pay as you go.

[Dylan16807]

> ala cryto mining evolution from GPUs to ASICS for specfic algos

I don't see it happening. A current gen GPU with a huge and fast block of memory isn't a perfect fit for these algorithms but it's relatively close. With cryptocurrency, mass small sha256 hashing was a totally different kind of computation.

[fc417fc802]

> isn't a perfect fit for these algorithms but it's relatively close

I don't think that's true. The best fit out of what's presently available perhaps. Inference is almost entirely memory bandwidth bound at present, to the extent that GPUs with HBM have a massive advantage over those with GDDR. TPUs appear to be a much better overall design.

I expect that a hypothetical advance in fabrication enabling processing elements to be placed directly adjacent to dense RAM on the same silicon (not merely in the same package) would be superior in all regards.

[Dylan16807]

> I expect that a hypothetical advance in fabrication enabling processing elements to be placed directly adjacent to dense RAM on the same silicon (not merely in the same package) would be superior in all regards.

Processing scales better than DRAM does. I think an HBM-like stack where the bottom layer has the math units is probably the ultimate form of that.

And it's possible that flash instead of DRAM is actually the better play, as long as you can hook up enough in parallel. RIP Optane.

[wcoenen]

> I don't see it happening.

Isn't it already happening with Cerebras? It's mentioned at the end of OpenAI's GPT 5.6 announcement:

"We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July"

[bluecalm]

There might be a community effort at some point. This happened in chess where the community recreated and then improved on Alpha Zero. You could run small training chunks on your machine. Some people donated thousands of hours of server time.

[tsss]

They never will. The only reason China releases open weights is because they can't compete on frontier models. Whoever has the frontier model has no incentive to give it away for free.