A deep dive into the Anthropic & xAI agreement
MMaximilian Schwarzmüller
컴퓨터/소프트웨어창업/스타트업경제 뉴스우주/천문AI/미래기술
Transcript
00:00:00Yesterday, Anthropic announced two interesting news which have an impact on developers and
00:00:07where there is a bit more behind those news than you might at first think.
00:00:14Because they announced that there would be higher usage limits for using Claude, which
00:00:19especially matters of course if you're using Claude code, because chances are that like
00:00:24many others you frequently hit those usage limits and you felt them to be a bit too restrictive
00:00:31especially for example when compared to the Codex usage limits.
00:00:35And the other big thing they announced yesterday which actually is kind of related to the first
00:00:40announcement is that they entered a compute deal with SpaceX and that they would rent the
00:00:48entire Colossus 1 data center which is that data center Elon Musk built in record time
00:00:55in 2024 that they would rent the entire capacity which is around 300 megawatts so they would
00:01:03use that for training or inference and therefore have more capacity to work with.
00:01:11That of course has an impact on the usage limits they can provide though there is a bit more
00:01:15to it but it is important because Anthropic was heavily compute constrained and they still
00:01:22are, that is fair to assess, all these big AI companies are because we're all in this
00:01:29situation or they are all in this situation where essentially demand for tokens does or
00:01:36did not scale linearly but instead essentially exponentially because of course we nowadays
00:01:45have all these models that allow for longer running agentic sessions, agentic coding sessions
00:01:53or whatever you're doing with them it's not just about coding these modern models and their
00:01:59agentic harnesses are pretty good at other tasks like data science and so on as well because
00:02:04they also generate code there but that's a different story but because these models and
00:02:10the tools in which they run became so much more capable and they can now run for longer
00:02:16you and me and all the users are burning through way more tokens and that is simply not necessarily
00:02:23what these companies, what Anthropic planned for and definitely not what they have the capacity
00:02:29for hence that deal where they actually enter this compute partnership with SpaceX.
00:02:38Now what's interesting is that in this post here they share that they would as I mentioned
00:02:43immediately add those 300 megawatts of new capacity which is over 200,000 of NVIDIA GPUs,
00:02:51again keep in mind that is from 2024 but still that's a lot of compute which they can add
00:02:57right now and that joins their significant other compute announcements which have even
00:03:02more eye watering numbers of up to 5 gigawatts agreed both individually with Amazon and Google
00:03:10but here it's worth noting that this is all about capacity that's mostly planned and not
00:03:16necessarily online yet.
00:03:17The 5 gigawatts they agreed on with Amazon includes nearly 1 gigawatt of new capacity
00:03:25by the end of 2026 so yeah we're not even close to those 5 gigawatts yet and for the
00:03:33Google agreement it's also 5 gigawatts which will begin coming online in 2027 so that's
00:03:39also not there yet so we're talking about 10 gigawatts here which are not even close to
00:03:44being available right now.
00:03:47But the 300 megawatts from the XAI data center are.
00:03:52The question or one of the questions that of course comes up is what does that mean for
00:03:57XAI?
00:03:58Don't they need that capacity?
00:04:00Is that the end of Grok?
00:04:03And the answer is no maybe well not the end of Grok obviously but I think it's fair to
00:04:11say if Grok had some demand that would be somewhat similar or even close to what Anthropic is
00:04:21seeing they would not be renting out that extra capacity instead of course what is the case
00:04:27is clearly they don't need that capacity for Grok you might wonder why?
00:04:31And the answer to that is that this Colossus data center here which was built as I mentioned
00:04:37in record time 122 days in 2004 this was just the first data center they built essentially
00:04:46but XAI now has an even more powerful Colossus 2 cluster available for themselves so they
00:04:54can use that for training and inference and that's not the only data center they also acquired
00:04:59another data center the macro hardware data center and XAI is on a roadmap toward nearly
00:05:062 gigawatts of power with operations ramping up in 2026 so it's not like XAI gave away all
00:05:14of their compute power but of course a part of it and clearly they wouldn't do that if
00:05:18there were more demand for Grok but as it is that of course gives them a nice way of utilizing
00:05:24this otherwise unutilized capacity and earn some extra revenue off that and for Anthropic
00:05:29it of course means that they now got more capacity available that then leads to those higher usage
00:05:35limits and I think this is all a bit more nuanced than you might think so clearly Anthropic was
00:05:43heavily compute constrained and I mean all these AI companies are struggling to get enough
00:05:50compute I talked about that in another video an episode already I won't bore you with that
00:05:56here but of course compute is a big problem because of that exponential growth because
00:06:02of these longer running agentic sessions and all the tokens that get burned by those sessions
00:06:08which need to be inferred where we need the inference for so yeah that is the point here
00:06:14but now of course Anthropic can increase those usage limits because they now got a bit more
00:06:21capacity available that's great but in addition I would make the argument that they're also
00:06:27under pressure to stay competitive there and that's also something I already argued for
00:06:32in another episode because as I mentioned in the past we had of course GitHub Copilot move
00:06:39to usage based billing now they have their own reasons they also had a pricing model or
00:06:47they basically gave you a certain number of requests for your payment for your subscription
00:06:54fee and of course charging requests instead of tokens is not ideal or counting requests
00:07:01instead of tokens is not ideal because since we have these longer running sessions it may
00:07:06be one request but way more tokens than in the past but of course they could have also
00:07:10used to a token-based measurement system and they instead moved right away from subscriptions
00:07:16to usage based pricing in the end so of course they did that Anthropic had those relatively
00:07:24strict usage limits still a subscription still token based but relatively strict when compared
00:07:31to Codex which of course kind of is their big competitor OpenAI with their Codex models
00:07:37or the GPT 5.5 model now and the Codex tool and their subscription offerings and you got
00:07:43way more generous amounts of usage out of Codex you still do get lots of usage out of Codex
00:07:50and GPT 5.5 I will say is a pretty good model so naturally I think Anthropic of course has
00:07:58a big incentive to stay competitive to grow its market share and they are seeing amazing
00:08:03revenue growth by the way if you want to learn how to get the most out of Codex and Claude
00:08:09code how to use the advanced features as well and not just the basic stuff you already know
00:08:15then you may want to check out my courses or even get a membership which gives you access
00:08:20to all my courses for a flat monthly or annual fee and you find links below but as mentioned
00:08:26of course Anthropic is seeing strong growth they probably want to continue on that growth
00:08:32trajectory also of course since they have all these capacity or since they have all that
00:08:39a capacity here coming up in the next year and so so they want to utilize that they want
00:08:43to grow even more based on that and they need the revenue the money in the end to pay for
00:08:50that capacity so clearly they want to keep on growing aggressively obviously without burning
00:08:56lots of money but they want to keep on growing they want to grow their market share so that
00:09:00eventually at some point they can raise the prices that will certainly happen at some point
00:09:05and that's how we end up at those higher usage limits obviously as I mentioned combined with
00:09:11the fact that they're now no longer as compute constrained as they were before now that they
00:09:16got that SpaceX deal but there is more there is more to that deal because I already mentioned
00:09:23yeah, XAI which is of course part of SpaceX in case you didn't know they wouldn't have
00:09:31leased that colossus one data center to anthropic if they needed it themselves clearly they don't
00:09:38but they can get another advantage out of that because in that anthropic announcement there
00:09:43is this interesting sentence here as part of this agreement we have also expressed interest
00:09:49in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity huh
00:09:56that's interesting because SpaceX in case you missed it is planning to launch orbital data
00:10:02centers which kind of makes sense I mean it is their thing shooting stuff into space it
00:10:09is the the argument that was being made in the end for acquiring XAI which is a totally
00:10:15interesting deal on its own by the way Elon Musk buying Twitter on accident then founding
00:10:22XAI then having XAI buy X Twitter and then having SpaceX buy XAI all with help of our
00:10:29investors that that's a smart way of getting richer to put it like this but one of the arguments
00:10:34is of course that they want to launch those orbital data centers that was the reasoning
00:10:40for the deal of acquiring XAI and now of course they have an anthropic on board too or do they
00:10:46well maybe not necessarily they have expressed interest so this could very well just be some
00:10:52sentence they agreed on to put in there so that SpaceX has some some nice story for when
00:10:58it's trying to have its public offering its IPO later this year that they can say it's
00:11:04not just for their own data centers but they also got an anthropic on board and so that
00:11:09is a nice deal. Anthropic gets the Colossus 1 data center may also use orbital data centers
00:11:14in the future but doesn't commit to it right now because there are some problems with orbital
00:11:19data centers it is yet to be proven that this is really a good investment to put it like
00:11:26this and that they really beat data centers on earth and deliver significant advantages
00:11:32but they may of course join and again for SpaceX it's nice to be able to point at anthropic
00:11:37as a potential future customer for their IPO. And I think this is what's worth knowing about
00:11:46this deal. The good thing is higher usage limits also of course in the future lots of capacity
00:11:53coming online for anthropic but as I showed you hopefully there is a bit more to it. If
00:11:59that is good or bad for us as developers or humanity as a whole is a whole different story
00:12:05you can have different opinions on AI and I certainly do as you know if you heard other
00:12:11episodes or watch other videos from me but that's what's happening here for anthropic
00:12:16and what it means near term for us developers.