Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Higher usage limits for Claude and a compute deal with SpaceX

anthropic.com

Post reply on HN
Print view View original

371–380 of 519 posts

Re: Higher usage limits for Claude and a compute deal with SpaceX

#371

Post by lijok »

ceejayoz wrote: post #263
lijok wrote: post #259

Earlier quoted context omitted.

Why are you comparing the output of a datacenter to the output of a single sat? How much power do starlink sats draw and how does it compare to say 8x H200s?

> This gives us access to more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month. 27,500 satellites need launching - fast! - just for Claude to meet a demand spike?

There are over 10k starlink sats in orbit already. They’re obtaining approval for another ~30k.

So clearly not a problem for them.

Re: Higher usage limits for Claude and a compute deal with SpaceX

#372

Post by alsetmusic »

timmmmmmay wrote: post #243
0xbadcafebee wrote: post #77

Colossus 1 datacenter is the one using illegal power, is poisoning the air for poor communities near Memphis, and is potentially poisoning the water. It's likely the additional demand on the grid will cause massive blackouts during extreme weather events, putting residents at further risk. https://en.wikipedia.org/wiki/Colossus_(supercomputer)#Envir... So you can put Anthropic on your list of companies that like to t…

it's in a former appliance factory that's right next to two pre-existing TVA power plants, a Nucor steel mill, and a sewage treatment facility. you've been lied to about how close it is to a residential area, just look at a map

Air pollution travels.

Re: Higher usage limits for Claude and a compute deal with SpaceX

#373

Post by lijok »

gambiting wrote: post #269
lijok wrote: post #256

Earlier quoted context omitted.

Which is the same lifetime as a starlink sat

So what exactly is the benefit of having that thing in orbit then, where it costs you millions of dollars to put it there?

Millions of dollars? Where did you get that number from?

Re: Higher usage limits for Claude and a compute deal with SpaceX

#374

Post by jordanb »

floatrock wrote: post #168
brokencode wrote: post #44

Earlier quoted context omitted.

Pretty smart for SpaceX though. They’re turning an asset they made for a money-pit (Grok) into probably a major source of revenue ahead of their IPO.

We all remember 2 weeks ago when SpaceX bought $10B of Cursor services. https://news.ycombinator.com/item?id=47855293 Since Cursor often relies on Claude models, some of those services will flow back to their own datacenter compute. Especially if there's, lets call it, "customer demand loadbalancing optimization agreements" that makes those Cursor services prioritize Claude models using the app keys that get load-bal…

Big wheel keeps on turning

Re: Higher usage limits for Claude and a compute deal with SpaceX

#375

Post by exabrial »

kcb wrote: post #253
exabrial wrote: post #237

This is where I see the economy of AI going: * Inference becomes cheap - speciality accelerators hit the market and race to the bottom begins * Training remains expensive - This works out for Anthropic/OpenAI, they go into the business of training * Models become rental units or purchasable assets, you run on inference hardware - Rent or own inference hardware * Or you pay someone to do all of the above for you, at a…

There's no magic bullet for inference on cheap accelerators. Any accelerator will still require large amounts of high bandwidth memory.

The way to do it _today_ requires enormous amounts of HBM! However, we've never designed inference accelerators, which is actually a quite "trivial" problem, but we've just never had a need.

Groq (acqui-hired by NVidia) came up with a different processor architecture: metric shit-tons of SRAM attached to a modest single core deterministic processor. No HBM needed on this card, and 32x faster inference than today's best GPUs at inference!

These LPUs are pretty useless for training though, which is useful for companies training models! Training is expensive, inference is cheap (someday, not now).

There's also a Canadian company that _literally burned the model as a silicon mask_ on a chip. It's unbelievably (1000x) fast, but not flexible of course: https://chatjimmy.ai

Re: Higher usage limits for Claude and a compute deal with SpaceX

#376

Post by lijok »

jsnell wrote: post #335
Footnote7341 wrote: post #294

Earlier quoted context omitted.

The current bottleneck on compute is power and zoning. Solar panels are 5x more efficient in space, and there is no zoning in space.

The current bottleneck is silicon. Every chip that is manufactured gets housed and powered. (It makes sense: the cost of compute is dominated by capex, the power costs are irrelevant, so they're ok paying a premium for power). The space data center hypothesis relies on compute supply growing faster than power supply. (Both are bottlenecked on parts of the supply chain that will take ages to scale.) Even if you believ…

The current bottleneck is not silicon. There is plenty of silicon locked up in previous gen GPUs that are no longer efficient enough to run relative to newer models. The bottleneck is the economics of owning the older GPU models - which is why all the GPU neoclouds are gonna go bust unless they can get customers to continue renting old GPUs.

The economics are vastly different when opex is near zero for these things

Re: Higher usage limits for Claude and a compute deal with SpaceX

#377

Post by kcb »

exabrial wrote: post #375
kcb wrote: post #253

Earlier quoted context omitted.

There's no magic bullet for inference on cheap accelerators. Any accelerator will still require large amounts of high bandwidth memory.

The way to do it _today_ requires enormous amounts of HBM! However, we've never designed inference accelerators, which is actually a quite "trivial" problem, but we've just never had a need. Groq (acqui-hired by NVidia) came up with a different processor architecture: metric shit-tons of SRAM attached to a modest single core deterministic processor. No HBM needed on this card, and 32x faster inference than today's be…

The point is metric shit-tons of SRAM is still large amounts of expensive memory.

Re: Higher usage limits for Claude and a compute deal with SpaceX

#378

Post by dymk »

everfrustrated wrote: post #142
quinncom wrote: post #20

Earlier quoted context omitted.

One of the reasons I refuse to use xAI’s models is because of the outsized negative environmental impacts of the methane gas turbines. Now I have to avoid Claude too.

You realize natural gas is one of the more environmentally friendly methods of generating power. Lots of work went into moving to natural gas generation to improve the environmental impact for electricity generation. This is nothing like burning coal.

More greenhouse gases in the atmosphere, speeding up global climate change. Renewables or don’t do it.

Re: Higher usage limits for Claude and a compute deal with SpaceX

#379

Post by dymk »

everfrustrated wrote: post #162
stuaxo wrote: post #154

Earlier quoted context omitted.

This natural gas burning is suited too close to where people live.

You do understand natural gas burns very clean? People burn it in their houses to cook with it!

Most municipalities ban nat gas in new construction because it’s so unhealthy and unsafe compared to an induction or resistive electric range. No, it doesn’t boil water faster than electric either.
Post reply on HN