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 talk big about safety, but when the rubber hits the road, profits matter more than safety.
Illegal is a strong term here. While the wiki link you included indicates there might be some permitting nuances, I've seen nothing claiming the power is "illegal."
xAI removed its illegal gas turbines and obtained permits for the others only after being sued by the Southern Environmental Law Center. They then built another unpermitted site (Colossus 2) across the state line in Mississippi, and they are being sued again. [0]
"The company began operations at its first site, Colossus 1, in June of 2024 and used as many as 35 unpermitted gas turbines to power the facility. Despite receiving intense public pushback over the use of illegal turbines and the lack of public input and transparency around Colossus 1, xAI officials said it planned on “copying and pasting” its unlawful turbine strategy to power Colossus 2."
"xAI removed its unpermitted turbines at the Colossus 1 data center after SELC, on behalf of the NAACP, sent a notice of intent to sue under the Clean Air Act. The company obtained permits for its remaining 15 turbines."
They did not require permits at the time as they were portable
Think transport trailer sized.
If you use portable power for under 365 days a year, an epa permit was not required.
They changed the rules on permitting after and xAI complied
Yes, I believe it's xAI's position that they were technically in compliance at the time. I don't know that a judge would agree. The new EPA rule is more of a clarification; they do not concede that point.
Proper utility scale gas generators come with proper utility scale pollution controls to make sure nasties like fine particulate and NO is filtered or properly reduced into some much less harmful to human health.
CO2 is bad for us long term. But there are plenty of other nasty combustion products that are extremely bad for humans in the short term. Which is why we have pollution and air quality regulations.
Portable generators don’t meet any of the stronger requirements that utility scale systems have to meet, because it’s assumed they’re only operated in small numbers for short periods of time. They’re not designed to safe to operate in large numbers over long periods of time in the same place. For that you need proper pollution controls
If you are burning that much fuel it needs to have emissions regulations. How would you feel if 20 miles upwind of you somebody fired up a few hundred random gas generators and kept them running 24/7 with no emission controls on them, rather than using utility power which is far cleaner and more efficient?
Public power utilities get permits for their operations. xAI tried to get around permitting regulations and environmental laws by claiming the generators were temporary, got sued [0], and even the Trump administration's EPA ruled against them [1]. They are also now trying to do it again in another state with Colossus 2 [2].
I don't quite understand the business logic behind "blocking" openclaw (you can still use it at API rates) but I never saw how this was unethical. Anthropic has no ethical obligation to support other people's software
Blocking openclaw made everyone realise that what anthropic giveth, anthropic can take away.
It is similar to the xAI gas turbines in that it tarnished their image - at least amongst those naive people who saw them as a plucky startup rather than a profit seeking corporation who don't like competition.
I agree with you that the ethics are very different.
I don't get it. On the one hand we had Steve Jobs saying "No App Store!" and everyone getting up in arms, then here we have "no obligation to have Anthropic support other people's software," and that being OK. So which is it? Or does the answer change daily depending on what makes us feel good?
I find the ethics of power generation, resource use, and pollution in a world struggling with climate change to be more of a challenge than whether a few people can run some software. And that’s coming from a Claude user that’s getting tired of their shenanigans.
from perplexity deep research:
"Colossus‑related gas‑turbine power plants have been run in ways alleged to violate the Clean Air Act, in already over‑polluted Black and low‑income communities near Memphis, and Anthropic has now become the main user of that infrastructure."
> "The xAI facility has already deployed *nearly 20 gas turbines, including four large units with a combined capacity of 100MW*, to power its AI system Grok... There are plans to add *15 more gas turbines between June 2025 and June 2030*, and the turbine application projects *annual emissions of around 11.51 tons of hazardous air pollutants*."
> "it is currently *running gas turbines without the necessary permits from the Shelby County Health Department*"
> "findings from the Southern Environmental Law Center indicate that the facility has 'installed' gas turbines. This suggests that new industrial systems are in place and that *xAI is obligated to comply with the new NSPS* [New Source Performance Standards] *to avoid violating the Clean Air Act*"
> "NSPS are authorized under *Section 211 of the Clean Air Act*... All new sources must comply with the *Best System of Emission Reduction (BSER)*, which mandates the use of state-of-the-art technology to minimize air pollutants."
> "there is a history of Elon Musk's companies, such as *SpaceX and the Boring Company, being fined thousands of dollars for violating environmental law* to circumvent regulation"
So they haven't gotten permits, but why? Why where the permits denied?
Just the other day we had news that some Californian environment protection agency denied permits for SpaceX for political reasons as opposed to following objective rules, as ruled by a judge. So the fact that some permits were not issued doesn't tell me anything.
To my understanding: the permits weren't denied, they were never applied for.
Edit: I re-read https://www.tba.org/?pg=Hastings2025AIX and yes, it seems that xAI never applied for permits related to the gas turbines as they're making the argument that the permits aren't required.
I wonder what the pollution from these gas turbines is like. SO2 from trace sulfur compounds? Is it much worse than a traditional gas-fired power plant for some reason? I can't imagine it would be but I have to plead ignorance and beg for hints here.
Well, from a quick bit of searching it looks like it's all NOxes. There are supposed to be known ways of mitigating NOx formation [1][2] but there enough dependencies that I'm not going to do any more digging.
The independent study, conducted by EmPower Analytics Group and commissioned
by the Southern Environmental Law Center, was led by a Harvard-trained
environmental health scientist Dr. Michael Cork and shows that operation
of xAI’s proposed permanent gas turbines would measurably increase
health risks for families throughout the area—even in places as far away
as Germantown and North Memphis.
I live in Memphis, none of this is true. What is true is that there is a concerted effort to smear anything related to xAI‘s presence in Memphis for some reason.
For some facts, the colossus data center is next-door to a steel mill and city sewage treatment plant, a vacated gigawatt scale coal power plant complete with nasty Coal Ash Ponds, and a brand new combined cycle gas power plant. The area is at the far edge of Memphis city limits up against the river, in a heavy industrial area. There’s even a major Valero oil refinery right there too.
Memphis has trillions and trillions of gallons of water, both in a gigantic underground aquifers and the Mississippi River itself. xAI has agreed to shed load in case of impending brownouts. The fear mongering is out of control.
They had a ton of portable turbines that were under operating under a temporary permit, and that was the disputed part. However, the blame should rest with TVA and or Memphis light gas and water for not being able to run an appropriate high voltage connection less than 1 mile from the plant to the data center in a timely manner. However… What difference does it make if the natural gas is burned at TVA plant or very similar gas turbines on site in the same neighborhood. Environmental groups and the county health department tried suing, was struck down, xAI works closely with the State, but the whining continues. xAI is paying gargantuan taxes to the city, no tax breaks.
These environmental groups do not care about the nasty unregulated cars burning oil, that I have to breathe every day. We terminated our motor vehicle inspection requirements due to the “burden” it places on the low income population. So they can burn their oil in my face, but then they sue to stop a SOTA turbine in an industrial area? There are junkyards in these same areas that burn their piles of waste tires every year or so “on accident”. No lawsuits there either.
Agreed there is a huge effort to smear as much as possible. Between parent comment being very highly voted and Wikipedia page being militantly updated and seeing these tired, wrong talking points everywhere, it's pretty obvious
> These environmental groups do not care about the nasty unregulated cars burning oil, that I have to breathe every day
That's a weird thing to say and makes me doubt everything else in your comment.
Just in case you were talking about some specific group I looked up who was doing the suing and they have lots of stuff about promoting trains and EVs etc.
They are literally suing the Trump administration (that Elon helped elect) on this topic.
> There are junkyards in these same areas that burn their piles of waste tires every year or so “on accident”. No lawsuits there either.
I wonder what drew xAI to this poorly regulated hellscape?
Also, one of the organisations suing xAI is also suing a) cement factories that burn tires for energy and b) tire companies that use the additive that kill salmon
We have similar issues here in Wisconsin. Especially when it comes to solar and battery storage facilities. I absolutely think there needs to be more regulations carved out for data centers, just as there is for any other industrial building, but yeah the great mongering is incredible to see. Especially when the argument of "save our beautiful farmlands" is brought up. Do you even know how nasty agricultural runoff is?
Not every allegation that appears in print is true. One should be very skeptical about these kinds of allegations, especially when there are deep-pocketed corporations involved who can be sued or pressured to settle in the face of sufficiently "plausible and persistent" (to borrow Hazlitt's term) claims of harm done by their operations.
So I was just Googling this, and apparently most datacenters don't pay any state tax on revenue generated by said datacenter? Huge loophole if true, no wonder capital investment in datacenters is so high. [0]
> In general, data centers only pay corporate income tax if they generate revenue. Not all data centers do this because many don’t sell goods or services; they simply house servers. By qualifying as business expenses rather than revenue generators, they reduce the tax liability of their parent companies.
> Thus, when it comes to income tax, at least, many data centers – especially hyperscale data centers owned by large companies – don’t generate tax revenue because they don’t generate direct operating income.
"The datacenter" doesn't generate billions. The computations performed inside might, but almost never is the datacenter owner the same person running the servers inside. The owner is just leasing space; selling electricity and cooling to a third-party company. Their margins (ie: taxable income) are thin because competition is high. A landlord's taxable income is not determined by his tenant's income.
You are being pedantic to pretend you don't understand. It's tiring and unproductive.
At the end of the day, people are paying money to utilize the servers within the datacenter. That money is revenue. That revenue ought to be taxed by the state.
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
"The independent study, conducted by EmPower Analytics Group and commissioned by the Southern Environmental Law Center, was led by a Harvard-trained environmental health scientist Dr. Michael Cork and shows that operation of xAI’s proposed permanent gas turbines would measurably increase health risks for families throughout the area—even in places as far away as Germantown and North Memphis." - https://www.memphiscap.org/
> Data centers can inadvertently pollute water through chemical runoff from evaporative cooling systems, including biocides, corrosion inhibitors, and heavy metals that accumulate at scale when facilities discharge up to 5 million gallons daily.
The plan was to develop a recycled wastewater facility, which will pull arsenic from contaminated shallow acquifers, and pump that into the drinking water supply's acquifers.
Now that you have stopped using Claude Code, what have you replaced it with? Would love to know your setup. I am experimenting with local models too, but nothing comes close to Claude (Code), at least for me - not just for coding, mind you.
FWIW, if you want frontier-level performance, as it was a few months back, Deepseek v4 and K2.6 are there. Almost zero chance you can run them locally, but you do have choice in terms of providers.
Qwen-coder-next is considered SOTA for things you could actually run locally.
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-balanced onto the SpaceX datacenter.
Did SpaceX just spend $10B to rent out its own datacenter, juicing their recurring revenue metrics with their own AI services investment?
With Anthropic's help. And when it's time for Anthropic to hype their IPO maybe SpaceX will return the favour and offer some deal that looks great to retail investors.
I don't think it's the conspiracy theory that you're making it out to be.
It is publicly known that the vast majority of deals in the AI space are circular in nature without the need for explicitly encoding any of it in a legal contract or even tacit agreements.
e.g. Nvidia has invested significantly in many AI companies including both Anthropic and OpenAI which rely heavily on Nvidia's hardware and will undoubtedly use some of said investment towards that end.
Nvidia and Oracle are already public companies, they're just aiming for their next quarterly statements.
SpaceX is getting dressed for their debutante ball and is putting on the makeup to make a grand entrance on the auction floor.
Is there a difference? I legitimately have no idea. You are right that we can add another entry to the list of interconnected circular dealmakings. All this ain't gonna end well next time the music stops playing.
Your argument is that since it is common in a bubble to make circular deals, there is no conspiracy. But you seem to suggest that people committing tens of billions of dollars aren’t looking any further down the pipeline than the name on the receiving bank account? Have you ever been anywhere near a large deal?
That's a lot to imply from my simple comment. My viewpoint is actually the exact opposite of what you claim: it all feels like a house of cards that is set to collapse at any moment. I can also tell you're quite passionate about this and I wonder if that emotion is clouding your interpretation of what was meant to be an innocuous comment.
My point was that there is a lot of this happening, it is not a unique statement nor is it surprising to see at this point.
I made no attempt to dismiss or justify any of it.
Sure, if "pretty smart" means overinvest in capital spending on an dirty datacenter powered by unpermitted gas generators that you don't even need anymore because of lack of demand for your product, so you lease it to a competitor (presumably at a huge loss). I am not sure that "major source of revenue" as a datacenter provider is the kind of growth opportunity that IPO investors are looking for.
Anthropic doesn't has that much pressure to pay while Musk has an IPO coming up and he wants to cleanup his numbers.
Its also not a good sign because he should be able to leverage Grok, his billion dollar investment, instead of renting it out to Anthropic. But hey what does it matter to investor? if the IPO explodes, it is clear that people either can't read, don't care or don't understand.
Says who? Oracle spends a lot of money to get ready for AI customers like OpenAI. They aren't there yet. They can't lose money serving what they don't have.
Its not even that. Its better to be involved in the game with a leader/help out a competitor who is competing against someone you don't like and don't want them to win, than to sit it out.
Are you worried about Google too? They're selling compute. Same with Microsoft, and Amazon. As far as I know Anthropic is really the only one that's compute-bound.
> As far as I know Anthropic is really the only one that's compute-bound.
I use gemini models daily. Jetbrains tells me when they are overloaded and switches to alternative (usually to openai which turns everything to shit). I'd say happens about fortnightly.
It's a good litmus and forecaster for AI demand and I wish we had more visibility.
Amazon is a compute specialist, their competitive advantage is in the compute business. And conversely they're not really trying to play in the AI business, so it's not at all suspicious that they don't want to use all their compute themselves.
Amazon is a bookseller and Google is just a web indexer. GCP didn't even open it's preview until 2008. Not sure why you think a business model is in any way a static thing.
It was pretty obvious to me that the merger was a way of quietly shutting xAI down in a way that keeps investors happy. With it also being used as a vehicle to offload the Twitter debt to the public, he certainly has good accountants.
Yep - and in the meantime it's an asset of SpaceX to boost their IPO price, as long as this is done before people realize that xAI is apparently becoming a datacenter company not an AI one.
Then you've got SpaceX buying 1200 cybertrucks from Tesla, so it's serving as failure laundering vehicle for all his endeavors.
> it's serving as failure laundering vehicle for all his endeavors.
Which would be fine to me if Tesla wasn't a publicly traded company and SpaceX wasn't about to IPO. Whereas juicing companies in a way that affects the open stock market feels very inappropriate.
I didn't say he's failing at everything - SpaceX certainly seems a huge success. Telsa had been doing well, although sales are now declining fast, and the Cybertruck has been a failure. He massively overpaid for Twitter, ruined the site, then got X.ai to bail him out. X.ai seems like a failure - evidentially not enough demand to utilize the data center he built for it, and when have you seen anyone say they use Grok for anything ?
And now SpaceX investors are going to be left as the bag holders for X.ai/Twitter.
It is always so odd seeing how many internet people consider any new attempt that doesn't go immediately viral with success as a bad mark on someone's character.
If you're not able to see a whole slew of "bad marks" on Musk's character, then you haven't been paying much attention. It's not either/or - you can be be successful in some areas while being a childish twit and moron in others.
I think I've just seen many more fake/exaggerated "bad marks" than real ones so I've become bitter. He's definitely not been perfect, but the areas I see flaws (he can be extremely rude, some drug abuse, doesn't treat close/loved ones well, seems to lash out when getting too close to someone, can be very Ego driven and doesn't admit it until it's too late, constantly needs to be at war with someone) are always just passed over for "He's a LiTErAL NAzI" and "He is HOARDING all the MonEY because he's SO GreedY"- which are just so demonstrably false.
Overall though, to classify the work he's done and the impact on the world as unsuccessful is just insane. It's almost always from someone who hasn't even managed to lead a team of 10 through one project too.
Yeah his luck is annoying. But he stoped having any character and ethics. He literaly was with his Tesla in front of the White House and bought himself a seat next to a Clown.
But he also plays in areas were market disruption can't be done by many people at all.
But look Tesla: He did the cybertruck debakel. He tanked Tesla as a brand, he is burning money on xAI and Twitter, he destroyed a beloved brand Twitter. He did the boring company garbage.
The only thing this shows is some kind of masterclass between manipulation, public ignorance, luck, economy of high invest high risk and risk adverse industries.
Starlink doesn't scale very well which is a low margin business, especially when Amazon and the others are joining the club.
xAI is just a loss.
Twitter probably still a loss.
Tesla made a lot of money with co2 certificates. And a market were people were quite ignorant for a long.
Space-X he wants to push that to the death, without a real endplan. He now talks about Mars and Datacenter in space like there is any real business up their.
Why would they spend 10B and potentially 60B in cursor if they were to shut xAI down? And I'm pretty sure Elon wants to have a model of his own, even if weaker, so it's "not woke".
Yeah it's corporate subprime. Bundle a load of overpriced "assets" with made up valuations into something that's actually valuable, then shove it on the public markets so everyone has to buy it in their index trackers.
Plot twist but makes perfect sense for both companies.
Anthropic gets the compute they so desperately need to keep growing. Elon rents out compute that xAI couldn't make use of due to little demand for Grok. SpaceX gets revenue on the books for IPO.
PS. I want to translate this part:
We’re very intentional about where we’ll add capacity—partnering with democratic countries whose legal and regulatory frameworks support investments of this scale
To real speak:
We're putting profits above anything else. Yes, Elon is a far right guy who supported Trump, a president who isn't very democratic, but we're just really desperate for more money. We're also trying to make you forget that xAI is funded by Middle East non-democratic governments. Heck, we'll even buy compute from China if we can sell Anthropic models there.
>we'll even buy compute from China if we can sell Anthropic models there.
Considering that Anthropic mass-bans Chinese users accounts based on using VPN (used to circumvent the Chinese firewall) and then demands an ID or a residence permit of a country where Claude officially works to ensure that the user doesn't live in China, seems unlikely.
If the Chinese government tells Anthropic they can freely sell Claude in China, Dario is suddenly going to be kissing China's ass instead of saying how we can't let China win the AGI race for democracy and western values.
They told the US government no on using Claude for approving lethal military strikes.
China can get plenty of value from Claude without needing to use it for anything similar.
They very specifically avoided a trap where the next time the US blows up a school full of children they were very obviously going to blame Claude for it.
While I agree with the sentiment, $200 Million is really not a big contract for Anthropic when they're on $44 Billion annual revenue. It's less than half a percent.
> funded by Middle East non-democratic governments
What's the problem here exactly? Are you insinuating any non-democratic government is bad and evil and only democratic governments are the correct and right way to govern? sort of like: "there is only one true prophet, and it's the one I follow, and all the others are false!"
Anthropic brought up the "democratic" justification, not GP. GP was just pointing out that Anthropic doesn't actually care. If it can get a sweetheart deal from an autocrat, it'll take it.
But assuming there are people that care, if a government doesn't derive its right to govern from the will of the people it governs, under what definitions can it be considered legitimate? Divine right of kings?
Yes, especially in the context of supporting US imperialism and capitalists interests (perpetual war + extraction machine) over what would actually benefit Americans: peace + cooperation initiatives. Something also tells me that American civilians would rather cooperate with peaceful governments than those that feed the blood machine.
America could do so much to compel the world to work in from a human rights perspective rather than petrodollars. I can't imagine any serious person would say the average American benefits from US imperialism. All US politicians did was traded away were secure middle class lifestyle for cheaper widgets, hardly anything worth caring about.
Who benefits from American petrodollar policies? Not Americans, all the wealth gets extracted to the elites while civilians suffer from the imperial blowback/boomerang.
Look at what the new deal coalition brought in and they nearly burnt out enough to allow neoliberalism to flourish during their fall. What do we have in return? No universal healthcare, no universal childcare, a broken welfare system, increasing income inequality, losing the ability to make a better life.
IDK what world you're living in, but in the real world Americans are the richest people on earth and richer than ever before in real terms. And yes international seigniorage is part of that.
Urban/industrial and refinery complex; not at all rural. Located about 8 miles southwest of downtown Memphis (3231 Paul Lowery Rd) on a bend of the river.
Which is now their official position I guess as this whole “AI space datacenter” stuff is a significant part of the whole SpaceX IPO.
I assume privately they may not share that opinion, but it’s not in Anthropic’s interest to talk about this (very little to gain, and may ruffle a lot of feathers if they say the wrong thing).
Anthropic needs any compute they can get. So if Elon wants to build orbital data centers Anthropic would be happy to run models on it. There isn't really any doubt Elon can build orbital data centers the question is if they are economical compared to earth based.
Local models are always going to be useless unless compute get significantly cheaper, and it's not. TSMC might literally run out of capacity to build any consumer compute product.
Once computer constraints ease up, you will see much larger models. The reason LLM seems to have stalled a bit is because there just not enough compute.
You have more people using AI which requires more compute, and you want to build larger models which requires more compute and you have limited compute. What do you do?
Help me understand why not? I know solar power generation in space, and "beaming" the power back, was a naive idea. But this would actually use the power up there, mostly for training, but also for inference.
That claim seems reasonable. I have zero knowledge of the economics of launching and maintaining satellites though.
Which is apparently manageable. Scott Manley isn’t an industry veteran, but he does know a lot about space engineering and science. Here’s his breakdown of the feasibility, and heat management is not really a major issue:
These satellites will be in orbits where they are always illuminated. That means constant temperatures, which means no thermal cycling and no reliability concerns.
When people say 'running it hot is bad for reliability', they mean 'running it hot and then brining it back to room temp from time to time will eventually kill it'.
It's in space which requires liquid cooling. No rocket is big enough so it has to be assembled on orbit. No liquid cooling terrestrial system is 100% leak free.
The existence of starlink proves that this is false. Look at most current pitches, they don’t talk about GW-class monsters anymore. There’s absolutely nothing stopping a 20-30kW satellite bus the size of starlink (or I guess up to 100kW? once starship is available - it’s all about payload fairing diameter) from hosting ~1 rack of compute and antennas. The economics may or may not make sense, we’ll have to see.
There’s very little research work needed to make this happen; it’s all about engineering some satellite buses and having them fly in close formation to get a “data center”. And this group of satellites in sun-synchronous orbit would relay to a comms constellation e.g. starlink itself) and operate as a global scale data center. The heat management and orbital mechanics are all straight forward really.
I've heard this before. A datacenter and a starlink sattelite are not in the same ballpark of power usage and heat dissipation needs. The are orders of magnitude off from each other.
The point is that you don’t need to put a whole datacenter into a single satellite. You can put a single rack per satellite and have different racks communicate via antennas, laser links, or perhaps even wires since they’ll be launched in groups of 10-50 anyway. You could also dock them to each other, but that’s not necessarily needed.
I don't understand what makes these "datacenters" if they're distributed across satellites with WAN-esque interconnect.
Are we overloading the term "datacenter"? Or is it not overloaded but somehow able to achieve datacenter-like speeds / (tail) latency even when distributed across satellites?
It's worth noting that GPUs have a much higher failure rate than traditional CPUs. Over 10x the failure rate due thermal stress. The amount of heat generated is very different. You can't really replace a GPU in a satellite (at least today?) which would place most of these satellites as space debris in a ~5 year horizon.
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 believe that's the case, the point at which orbital data centers start making sense is incredibly sensitive to the exact growth rates.
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
H100 rental prices are still as high as when the cards were brand new. The prices vastly exceed the power costs.
In a world where power or DC permits are the current bottleneck those H100s would be getting retired in favor of Blackwells. But they aren't. They are instead being locked in for years long contracts.
If silicon were relatively abundant and power/DC space scarce, you'd get an order of magnitude more bang for the Watt by replacing the H100s with newer GPUs.
But nobody is doing that. Blackwells are being installed as additional capacity, not Hopper replacements.
So it is pretty clear that silicon is the primary bottleneck.
Because you'd need to trash the old GPUs in order to make room for new GPUs. Right now new GPUs get online mostly in new DCs. TSMC fab capacity is much more limiting than DC building and it will likely keep being the case. It's much easier to build a DC than a fab.
"Space datacenter" -> overpriced starlink with some shitty edge compute -> "look guys, we built a space datacenter; earnings results to follow" -> number go up.
Pretty much everything has been "very dumb because of economics, logistics, serviceability and more". What kind of hacker are you to be on this site lol
SpaceX have presented on this and it's fairly straightforward and they already do it with starlink satellites, just at a larger scale. Sound like you are the uniformed one (or an EDS victim)
Starlink satellites don't generate the sort of heat a datacenter full of GPUs does. The ISS has enormous radiators, and it's only in space because it's a space station. Putting datacenters there is just goofy given the amount of available space on the ground.
All of that has been repeatedly addressed in anything that discusses it, if you care to try to understand. It has ~nothing to do with available space, the US grid can’t handle the current rate of expansion. It’s bad enough that apparently Span, the smart electrical panel company, is pitching a box full of Blackwells that’ll sit outside new construction homes and use all the headroom on residential 200A circuits. Space is starting to look reasonable.
Related, US readers should call their reps and ask them to support a successor to EPRA, the Energy Permitting Reform Act, the vast majority of the generation that’s waiting for approval is from clean energy sources. It nearly got over the line before the last Congress ended, and it’s one of the most impactful things we can do to combat climate change, combined with electrifying various carbon intensive activities.
Not quite, I'm rooting for the solar/battery microgrids down here, one of the startups I've invested in is working on those, but you don't really even need batteries for panels in a dawn-dusk sun synchronous orbit, which is a pretty huge advantage. Also, there aren't weeks where you have 1/4 the output because it's just cloudy all week, and your output isn't crushed during winter.
And the hardest part of my home solar install, by far, was the counterparties (inspectors, power company, and subcontractors). My understanding is that it's much worse when you're trying to get a grid scale install online, the interconnection queue is currently years long. This avoids most counterparties except the ones they're already routinely dealing with.
I've heard this before and these are not comparable at all. Starlink is missing a few digits in it's power usage and heat dissipation needs compared to a datacenter.
Scott Manley, I’d say one of the top pop space youtubers say otherwise. If anything it’s easier in space. On earth most complexity in datacenter is cooling. In space you just radiate it away.
And SpaceX already proven they can launch sort of datacenters 10k times by launching Starlink (up to 20KW of solar each IIRC).
FWIW Musk should support Bernie Sanders more. Putting moratoriums on datacenters would make space based ones far more economical.
He just mentions and walks through idea of having some amount of compute up there and what the heat rejection calculations roughly look like. He doesn't actually explore the economics of doing such a thing or discuss if it's actually worth doing.
It's not that you can't put a server in space, but the costs to do it almost assuredly don't make any sense. Because, if you can do it in space you can do it easier on the ground and save yourself millions in launch cost and extra complexity. Your cooling challenges are way cheaper and simpler in an atmosphere.
There's nothing much being in space really gets you, other than it makes it harder for a government to take your computers away. Not impossible, just harder.
Especially with everyone clamoring to have datacenters built in their backyards. There's absolutely no way there can be an advantage to figuring out compute outside Earth's magnetosphere, especially since none of the engineers as SpaceX would ever think of any long-term benefits of that.
The economics don't work unless Starship is doing flights in quantity, and it has met or exceeded its cost targets.
Roughly, a single rack plus solar to power it in the $15m+ range just to launch. (This assumes power dissipation is handled via some means that does not require launch to orbit. Also does not include batteries.) Choose your own hardware for the rack, but call it < $5m.
SpaceX earning $15m every time someone launches a $5m rack would be a great business for SpaceX.
Use your own calculator/LLM, but mine is suggesting that the ~$7B Colossus 1 data center in TFA would be around $50B if launched on Falcon 9 (still ignoring cooling and batteries).
(There are obviously a lot of other asterisks. I'm ignoring power storage and heat dissipation. Maintenance probably doesn't matter given 75% of cost is in the launch. Network bandwidth could be a problem considering how DCs are used. Competition - if Company A spends $100B for $25B of actual AI infra, how competitive will they be against Company B who gets $100B for their $100B by spending it in Canada or Mexico, which they can do right now? Etc.)
None of this works without Starship, which has not set a date for its first LEO insertion test yet. Yet the whole point of orbital DCs is nothing on the ground can move fast enough, hence the rush to orbit...which can't really move at all right now.
In space you get bit flips fairly quickly when using very small transistors. You would have to run stuff on fairly old hardware, which probably makes the whole thing economically inefficient for serious "computation in space".
I don't think space compute is going to work out, but I would certainly say "yes happy to buy space compute from you in the future if you offer it at a good price"
It makes no sense. We're being presented with a forced choice -- put them in space, or put them in the middle of downtown Seattle.
This is stupid. I don't understand what's happening... specifically, what mental virus is spreading that lowers everybody's IQ by 10-20 points, evidently including my own. Put the data centers in the ocean, powered by solar and networked with Starlink or LEO. Put them in the desert. Put them 20 miles south of Nowhere, Idaho.
Because the US has levied high tariffs on solar cells, can't build their own solar cells economically enough, and has such a torrid permitting system that it can't build transmission lines. Natural gas is the only form of generation that's easy to permit outside cities (due to pipeline agreements and this admin fast-tracking natural gas generation approval) but few cities will allow one. DCs need to be built within low latency interconnect of urban areas or else they become uncompetitive.
Elon claims (which I take with a huge grain of salt because he's made endless broken promises in investor calls and interviews) that he disagrees with the administration's stance on solar and would use it to power his DCs if he could, but contends that permitting is a huge problem.
The US needs to figure out how to build again.
> This is stupid. I don't understand what's happening... specifically, what mental virus
"Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes"
What does that have to do with my point? Space-based data centers need solar cells too. They are just like terrestrial data centers, only more expensive. For every dollar you save on the PV array, you'll spend two more on radiators.
And you don't need permits in international waters, any more than you need them in orbit. Lease space on container ships.
The argument is that it's too hard to gain the necessary approvals on Earth such that space is faster and easier. Not sure I buy it fully (I do see it somewhat), but that's the argument.
Sure but acquiring enough land to build solar profitably near a DC that can hook up to the big US interconnects is very expensive and will often be blocked by residents. If you want to build solar where there are few people and cheap input costs, then you need to transmit the power to the DC.
I keep hearing about brand new data centers they want to create. Seems reasonable to go to sunny, enormous, business friendly Texas and surround the data center site with acres of solar panels, batteries, emergency gas, and whatever sized grid connection you can get approved immediately.
If the DC is for training or text inference, latency seems irrelevant, so go where you can quickly plop down power.
Texas is actually absorbing a lot of the US's new generation capacity (though the grid there remains dirty)
It's fraught to make a DC for a single purpose because it reduces the value of the DC. A DC that serves multiple purposes can handle other workloads. Moreover even if inference is slow, latency does still matter, and it costs quite a bit to light up net capacity (you still have to run fiber to an interconnect and depending on how far you are, this can get expensive fast.)
I guess, but the amount of money getting thrown around is just stupid. Having to spend a few million to light up some more fiber is a drop in the bucket.
Supposedly some of the behind the meter gas turbines that have been getting installed are rated for a ten year service life. The DCs are burning them out in 10 months from rapid cycling. If they are willing to treat $10-100 million generators as disposable, cost seems irrelevant.
Ehh, I think they are just "kissing the ring". This was part of the agreement for the terrestrial datacenter access, pretend like the space orbital compute is more than the boondoggle that it clearly is.
I want to be clear, I do think that one day something like that will exist, I just don't think it's anywhere close to being a reality, much like FSD.
Also it costs them, almost [0], nothing to say it and then later come up with some reason why they are no longer interested.
Most big tech CEOs are people that only "succeeded" due to have an unregulated monopoly or picking the right lotto ticket and not due to any innate above average intelligence. Go look at the 100s of billions in wasted capital and tell me who benefitted from such waste while workers + children suffer from lack of medical care.
You honestly expect this trajectory to continue unabated?
> You honestly expect this trajectory to continue unabated?
Knowing humanity's history, yes. Not sure we're ever going to see a second French Revolution. People are pacified and are not rioting. And they really should. Most of us are kind of privileged. I know people out there who are barely holding on and the recent fuel + food price increases might push them over the edge to actual poverty.
The whole armchair engineer debate online about this is hilarious
I'm just a software engineer, all I need to know is SpaceX is aggressively pursuing this - that's enough for me to believe it's viable
SpaceX operates literally orders of magnitudes more satellites than anyone else. If anybody understands the physics and engineering of space compute, it's SpaceX. Lay people debating this online is just showing their ignorance as far as I'm concerned, and it mostly comes from an emotional place of wanting Musk enterprises to fail
Thank you for a reasonable comment. I know internet people love to comment on how "dumb" things are, but we're seeing a growing group of funded, motivated, and intelligent people working towards a common goal. It's at least something to be curious about, I wish the comments were more oriented towards in-depth discussions on the actual current blockers.
They're in every discussion even remotely related to anything Elon Musk.
Everything Elon does is somehow stupid or evil. Actually, that reminds me of Thunderf00t YouTube streams where he was (or still is?) betting Starship would fail miserably every test flight, and he'd talk about how evil and stupid Elon is for 3 hours with chatters, watch the flight then say something like "it's still bullshit."
I think it's a mixture of cope and a little bit psyop from adversaries like Russia who are being crippled in Ukraine because of Starlink.
Doubling the five-hour rate limits is merely a marketing stunt if the weekly rates are not also doubled. It simply means that you can reach the weekly limits in three days instead of five.
Same. I hit limits after 45 minutes. I'm on a measly Pro plan. I'm usually building small, open source projects, often from scratch. I only work on these projects in a 2-hour window in the morning. This is my "free time" development. I hope this change helps, because I was days away from switching back to Codex, though I like Claude Code a bit better these days.
I also hope that the fact I had OpenClaw in my sandbox once is not why I hit these limits so damn fast. I don't use it anymore and I've tried to rid my sandbox of anything "openclaw" but it is in my git history in various places on various projects. Claude doesn't seem to be transparent about this limitation.
Along with how many 5-hour windows they use in a day.
If you're using it 24/7 then yes, I'm sure the weekly limit is more of a concern.
If you're just using it during working hours - ie. you only use two 5-hour windows per day - then you probably, like me, struggle to hit the weekly limit even if you do max out some 5-hour windows.
last week with claude i saturated a team premium seat at day 6 of its cycle, and a max 20x seat at day 4, plus ~$150 extra usage spend, with a 60hr work week where i am not even primarily an IC, as well as a codex 20x plan at day 3 with a personal project
What does your usage look like day to day? Are you using a low level amount all day long? I'm with the others here, I've never hit the weekly limit ever, only the hourly, and I consider myself a heavy user.
I dedicate a significant amount of time to defining the precise actions that agents should perform (PRD/ADR). I break down the feature sets into Milestones and slices (tasks). These tasks are small, well-defined, and scoped. I have a prompt template that the “architect” agent prepares whenever I want to initiate a new feature. This ensures that the prompt structure remains consistent and standardized over time. The generated prompt is then pasted to the “orchestrator,” which performs context discovery (using Repoprompt) and finalizes the plan then proceeds to launch subagents to do the work.
Based on the size and complexity of the task, as well as any inter-task dependencies, the orchestrator deploys one or more subagents (sometimes 5 or 6 subagents) to work on these mini tasks. Once all tasks are completed, the orchestrator initiates verification and launches a review workflow. This workflow uses the original prompt, acceptance criteria, repository internal guidelines, and relevant skills to conduct a thorough review of the agents’ work.
Typically, there are one or two review iterations, during which the review agent identifies any issues. Sometimes, I may also notice issues and have to "steer" the orchestrator. The time required for a slice to complete ranges from 30 minutes to 4 or 5 hours, depending on its size, complexity, and the number of subtasks it contains.
Only if I run about 3 such orchestration in parallel I can reach hourly limit.
I have found that it uses a lot more tokens if I give it a very detailed todo and loop over every task 1 by 1. I now keep it to phases with detailed tasks underneath and use /loop over the phases and it uses a lot less. I also manage the context windows and tend to clear it often to keep it under around 200k (or less depending on project size)
Yeah, I do that too. Essentially, the system I described begins working on a task that is small enough and clearly defined. Each “slice” in a milestone usually have 5-10 subtasks (for instance, Slice E1 has P1...P6 subtasks). The orchestrator then receives the prompt to implement E1-P1.
That’s because the week ends before you can use them because you’re waiting for your hourly resets. Now the week essentially got longer with the same limit
No, I'm just using it a lot. It's productive enough that I've found it worthwhile tacking on subs for GLM 5.1 and Kimi as well (GLM is fantastic, Kimi is good when it works but temperamental)
For me personally, I have the basic Claude Code subscription that I use to rewind on some evenings or on weekend, to code a bit for 1-2 hours. I have like 3-5 session with it every week.
The 5h windows are frustrating because I can go through them quickly if I have a more complex task. I haven't yet met the weekly limit. I'd say there are many cases similar to mine.
I disagree. I routinely hit the 5 hour limit on Pro with Opus 4.7 just trying to have it do one design task or comprehensive code review on a large PR, and the worst part is, the overhead and bringing all that context back into another 5 hour window blows through 30%+ of my 5 hour usage limit.
I've found with opus 4.6 which im still stubbornly using i can burn about 10% of the weekly within a 5 hour window with my workflow.
Mentally i think about the weekly usage in terms of usage per day so about 14% per day which results in me not using that much early in the week so i can kinda "burn freely" later on. which leads me to a spot where usually on the final two days im sorta thinking about how can i expend that usage ive "saved".
the 5 hour windows make this harder, sometimes the final day of the week im trying to get that 10% in every 5 hour window of my waking hours and i HATE that, i wanna work when i am most productive, not around some ridiculous window of time, i dont wanna think "I am gonna be utilizing claude the most around 11am so i should send a dumb message to haiku to get my 5 hour window started at 7:30am so i can have it roll over at 12:30."
So im happy about this change sure. But it is 100% them creating a problem and pretending having some relief from that problem is them doing their users a favor. I understand they are doing it to lower peak hours usage and all that, I still despise it.
People are waisting tokens by using Opus for everything.
Using Advisor [1], you can use Sonnet most of time; Sonnet can handoff work it can't handle to Opus. When Opus is done, you automatically go back to Sonnet.
I think the main reason that workflow has not worked for me is because im using an ide version of claude code, which means my main agent isn't a crafted agent and is "stock" sonnet or "stock" opus. I'll likely swap to the cli version soon enough and see if that remedies it (this isn't laziness on my part, i instead learned opencode workflows first because it applies more broadly, the only limitation is usage of a claude subscription within it).
So with the stock sonnet i get the chatty confidently wrong sonnet instead of a strict crafted agent. Stock Opus is a lot more reasonable, and hands off simple tasks to crafted sonnet agents with the chatty and more strict workflows, so i guess im literally doing the opposite(closer to what that old article describes).
I rarely use Opus for planning (in the Pro plan). Spec a feature in Sonnet, hand it to Haiku, come back for review. That’s a 5-hour window gone, sometimes 2.
I hit my weekly limit around day 4, with 2 maxed out windows per day (and sometimes a bit of usage at night).
I completely understand why people would use Opus for everything, it’s much more thorough and effective. Sonnet as well, but on Pro it’s gonna be Haiku all the time.
my workflow allows for about 10 windows being maxed out each week(if this threads claim is true that is now 5 windows), i always use Opus for planning and just have strict rules for delegation when its actually crafting the code.
I have a pretty nailed down .claude/ where the goal is single sources of truth, so agent md files all reference the relevant files for what domain they are working within with that domain's conventions and structure etc, i think keeping this stuff up to date is massive compounding context savings, as well as just better for performance because it keeps all agents context windows free of noise by helping them only load in what is actually needed.
I've never really messed with haiku for anything besides absolute low end repetitive tasks, its usually an agent i have crafted when i want to ask it to generate a bunch of seed data or generic questions for tests or something similar. My assumption is that it would just be terrible and even though its super cheap, it is still inevitably bringing the final results back to the better models and if thats not valuable tokens then im wasting the haiku tokens and the passoff to the better models with work that will be repeated anyway.
On the Pro plan you can max 10 windows per week using Opus for planning? I’m impressed. Even with Serena and really tight context management, I use Sonnet for most planning and Haiku for implementation. That gets me through the week doing 1 or 2 features and 10-ish windows of bug fixing.
> Looks like Elon's finally giving up on XAI and just selling the compute
I don't think that's certain yet, but I do think that the open-source models like Gemma and Qwen are getting so good so fast that even Anthropic has real risk around the long-term value of their models and tooling.
Basically, if I'm Anthropic or xAI, I try to get revenue whenever and wherever possible and see what sticks. There's no value in playing for monopolistic control when everything is so volatile.
I don't know if it relates to the same data centers, but this also comes hours after several still recent Grok models were deprecated at short notice. Grok 4.1 Fast is the cheapest way to do research on X (cheaper than the X API!) and it's gone on May 15: https://docs.x.ai/developers/models - freeing up compute to sell?
The details are secret. It very well could be wasted GPU time but Anthropic could have made a killer offering as well.
I'm just speculating, but a particularly killer offering Elon wouldnt be able to refuse would be if Anthropic agreed to give them some training data / technology.
Giving Musk the benefit of the doubt, here's a thought experiment: It doesn't seem like any of the big labs in the US can keep a lead for more than 3 months. The Chinese models are closing in. Even if xAI comes up with the best model, so what?
On the other hand, power and compute are limited. Ridiculous as orbital compute sounds, land/power on earth is not easily scalable. There are too many limiting factors, chief among which in the US is regulation. But in space, if you make one satellite work, you just get more resources and launch more. This also leads naturally to Tesla's plan for a chip fab.
I don't think this is giving up. He's getting inside information on how Claude works, and a huge stream of Claude usage data. This will all inform future grok development, IMO.
> 300 megawatts of new capacity (over 220,000 NVIDIA GPUs)
The scale is just mindboggling here. Are there any blog posts or anything discussing what kind of infrastructure is used for even just the inference side (nevermind the training) for SotA models like Opus? I would have thought it might be secret, but given that you can actually run the models yourself on AWS Bedrock doesn't that give an indication?
I know you're probably talking about the compute infrastructure, but I think the electricity infrastructure side is interesting too, data centers are doing things in dumb ways because the need for operational expansion speed is greater than the need dollars:
> It’s regulation with the utilities. There are ramp rates, there are all of these things that you’re supposed to do to not screw up the grid. Data centers have been in gross violation of that. When you think about what’s wrong with data centers, they have load volatility, which we just talked about, then they decide to power it with behind-the-meter natural gas generators. These natural gas generators, their shaft is supposed to last for seven years. It’s lasting 10 months because of all the cycling.
I haven't bothered to look but I'd guess Mellanox GPU-to-GPU networks, and massive custom code for splitting tensors across GPUs, and for shuttling activations across GPU nodes.
> but given that you can actually run the models yourself on AWS Bedrock
That's not exactly how it works. Anthropic are hosting their models in AWS Bedrock as a managed service. Customers call those LLMs just like calling any other API. There's no visibility into what kind of AWS infrastructure is serving that API request.
All evidence is that the final training runs across thousands to low tens of thousands of GPU, and that a single instance of the resulting model runs (or could run) well within a rack (ie NVL72).
The massive scale is all massively parallel: test-time compute for users, test time compute for RL rollouts (and probably increasingly environments for those rollouts), other synthetic data generation, research experiments, …
Limits were the last straw that made me cancel my subscription and make my workflow completely model agnostic with pi.
While this is good news, I'm not coming back. Anthropic just lost me with too many wrongs in too short of a time period.
Opus has been replaced with GPT 5.5, DeepSeek, Kimi, Qwen and they all allow me to use my own, single harness and switch models easily if any of them start treating me the same.
I wouldn't make any grand stand declarations like this honestly. The models themselves are all hot swappable with minimum effort. The AI labs american or chinese don't really have a moat. Today anthropic is bad and openai is good. Last month it was the other way around. Next month it may be google.
The only certainty is that you can swap models quickly and painlessly.
Same, though I'm reconsidering, in light of the recent bugs (which can happen to any provider) and the increased limits. I guess that's at least 3x more Opus for my usecase.
Say what you like about Sam Altman, but given how Anthropic is scrambling to sign capacity deals for compute we can sure say he was right about the capcity build out needed.
Scrambling? Seems to me xAI built too much capacity (for what they can use in 2026). Does that mean OpenAI built the right amount? I don't see how this proves that just because we see one AI company willing to sell compute. We don't even know the terms/pricing.
> Anthropic CEO Dario Amodei said his company tried to plan for 10-fold growth. But revenue and usage increased 80-fold in the first quarter on an annualized basis, which he says explains why it’s been so hard to keep up with demand.
> “That is the reason we have had difficulties with compute,” Amodei said Wednesday at his company’s developer conference in San Francisco. Amodei added that the company is “working as quickly as possible to provide more” capacity and will “pass that compute on to you as soon as we can.
I think "scrambling" is a fair characterization of the CEO saying "we have had difficulties with compute" and "working as quickly as possible to provide more"
They've also signed new compute deals with Google and AWS recently.
Or the bubble he was pumping hasn't popped yet. We won't be able to say how much of this capacity was actually "needed" until 10 years in the future, if ever.
People are getting real stuff done with the internet. But there were also a whole lot of overhyped companies that rightly crashed back in '99.
Once we've gone through the AI equivalent of the dot.com crash, will Anthropic still be scrambling for more capacity, or will they have more than they can profitably use, like the dark fiber we were left with last time?
Depends when it happens. I'm sure they'll take up whatever capacity is available.
At the moment computer providers are charging more for outdated H100 capacity now than when the H100s were new. That capacity is going to the smaller labs, not the frontier labs.
That hardware has already been depreciated financially so even if all those small labs disappeared it's not sending computer providers bankrupt - they can just cut prices and so long as they can charge more than electricity and maintenance they'll just keep them running.
> First, we’re doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans.
The fine-print-omission appears to be that weekly limits are not doubled. The progressive 5-hour rate limit shrinking was indeed an efficiency blocker that finally convinced me to cancel, but being only able to get 4 full sessions a week as opposed to 8 doesn't compell me to resubscribe.
For my hobbyist purposes Deepseek v4 Flash has replaced Claude Code because I was also sick of hitting 5 hour limits with Claude. Right now, the only thing I miss from Claude is multi-modal image support. I can work around no image support since I can use v4 Flash all day and spend around $1. I am aware Deepseek is currently discounting their API at 75% off so I may try out another provider once the discount is gone at the end of the month.
At this point if feels like if you properly scope your work open weight LLMs are adequate.
Good news, I was wrong. The discount only applies to Deepseek v4 PRO. Deepseek v4 Flash is not currently on discount which means the dirt cheap price will stay the same.
If you can make up an inconsequential arbitrary rationalization to not use a service then I’m sure you can do the opposite to convince yourself to use it.
That’s what virtue signaling is I guess - the action you’re taking is pointless, the only point is to tell everyone you’re taking it therefore feed the narrative forward?
The entire economy runs off gas turbines though this is the thing you boycott?
Obviously I’m virtue signaling, and I hope instilling a feeling of shame in people who support businesses that contribute to climate change.
But more than that, the emissions generated by the Colossus data centers are far worse than typical combined-cycle gas plants or data centers that buy renewable: these turbines emit NOx, fine particulates, carbon monoxide, and formaldehyde into a population-dense area.
Deciding not to spend money with a company you don't like is not pointless. The point is that you're not participating in something that you judge to be wrong.
The world is full of things I feel are wrong yet have near zero power to stop. That does not mean I should willingly support those things.
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.
While the burning of methane is cleaner, the extraction of methane is a massive source of uncontrolled pollution emissions which is made worse by the fact that methane is 20x worse for greenhouse effect than CO2. Clean methane is another green washing myth to encourage people to keep consuming at much as possible
as other have mentioned tons of portable generators are no where near as safe as power plants and these are built right next to people's homes with no oversight or regulation.
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.
Nazi feels like a close enough shorthand for “person who posts white nationalist and antisemitic views, supports authoritarian regimes, and seig heils on stage.”
Nah but his wife did; well I guess that is not true, it was on a stage with pyrotechnics. Not even cold and his wife is making NYE look like a proper funeral, poor lad.
It may be more productive to ask what is right with burning fossil fuels for electricity right in the middle of marginalized communities that have to bear the cost of this pollution for AI slop.
I think so, but that's also really great because I frequently run into the five hour caps, but very rarely use my entire weekly allotment. There are lots of situations where I do things like write the plan for all the work that has to get done, and then set a reminder to execute the plan after I get home, when I'm done making dinner (because e.g. my five hour cap ends at 6pm). Higher caps for the five hour period is a lot more convenient.
The purpose is to control the total amount of requests they need to handle in a given timeframe. If everyone could use up their whole weekly limit in 5 hours, many would do so, thus pushing the GPU/TPU clusters to or above their capacity limits.
That is one take, but here is how I interpret it. They spent a lot of money training a model which isn't doing enough inference to justify continuing to use those GPUs, and now they are buying even more GPUs to build an even bigger model that also won't be very popular.
Oh. Just as I'm in the process of migrating to Pi+Qwen (local). This was probably going to be my last month on the Pro sub as I'm seriously fed up with the limits and degradation that started weeks after I signed up. Let's see how this shakes out.
How does Pi+Qwen (local) compare to Anthropic's offerings? Surely you're not getting the same breadth and quality of output using Qwen? How is the performance?
So far I've only really set things up and done some benchmarking (a set of capability prompts created and evaluated by Claude, HumanEval and MBPP; haven't completed the latter 2) on several local models (Qwen 1.7b, 4b, 9b & 35b a3b; 1.7b got 6/8 correct at ~14.7 tok/s on the capability set, to 35b for 8/8 at ~4.5 tok/s; can share full results if interested), and setup llama-swap so I can dynamically select them. I'll need to decide which of my projects I'll be really testing them on, with the awareness that I'll have to be even more involved.
this feels more like something designed to bolster the spacex ipo than anything else
300MW is peanuts compared to their multiple 50GW+ deals to the point you start to wonder why just 300MW is making the difference in their capacity that they can increase limits this much... also, why couldn't their many existing multi billion dollar deals not allow them to expand capacity?
when you take this into account, then you read their statement about orbital compute it starts to smell quite fishy
What's the current status of the 'biggest computer wins' vs. specialized proprietary research/data in the AI arms race? People had such high hopes for xAI because of the monster machine Elon built. Or has xAI just turned over too much staff too quickly?
I want to believe. A couple of weeks ago I fell into this "trap", they offered a similar thing. I subscribed to the Pro Plan. Had fun for a couple of weeks and then I entered frustration phase. I love the product, but I hate those up and downs. My rant made it to HN front page - which I am not happy of. I want the stuff I build to be seen on the front page.
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
SRAM and HBM are two completely different things though... SRAM is what your L1,L2,L3 caches are made of (most of the time, asterisks exist). This is something we've been doing for years and is a proven technology thats unbelievably cheap. It's all part of the processor.
Strictly speaking there is that one startup that compiles entire models into huge ASIC. With trade off that entire hardware becomes outdated when new model version is released in 2-3 months.
Interesting that the 5h limits are raised, but if I understand announcement correctly, the weekly limit is not. So all this means is that you can burn through your weekly limit faster and be locked out entirely, or having to buy tokens
If Anthropic and SpaceX and OpenAI are all going public this year then this is a clever move to stick it to OpenAI. However, I'm kinda sus of my Claude subscription now
Reading the comments here again surprises me how in an anti-Elon bubble most folks are. They are renting out spare Colossus 1 capacity. Colossus 2 is still coming online. Orbital data centers are really the plan in the next few years. XAi is still behind, but not a disaster considering how late they entered (and Elon’s unfortunate fixation on anime characters).
SpaceX is extremely uniquely positioned to crush the rest of the world combined in order to orbital data centers.
> SpaceX is extremely uniquely positioned to crush the rest of the world combined in order to orbital data centers
Sure, as long as your data center is 3x4m - size of a Starlink satellite (think Spinal Tap Stone Henge) . Anything bigger than that (i.e. actual data center sized) is going to require some assembly.
I've heard TeslaBot is good at folding shirts, and serving drinks (at least while teleoperated) - perhaps it can help?
You're not going to fit a data center in a Starship either, unless you are talking a Tiny Corp Exabox "data center in a shipping container" sized one. Even something that small (1MW) would still need 4x the solar capacity of the ISS, and therefore likely some assembly required. Then you've got latency from satellite to satellite ...
In any case, it appears that Musk can't even generate enough AI demand to utilize his own ground based data center. Maybe he can add "data centers in space" to part of his Mars colonization plan. Maybe have Tesla Bots driving around in Cybertrucks too ?
AFAIK "orbital data centers" are a bunch of nonsense.
1. GPUs create heat. There's no efficient way to get rid of the heat in space (vacuum is an insulator).
2. Die-shrink makes modern processors and memory more and more susceptible to radiation; shielding is possible, but adds cost + mass (which adds cost)
I struggle to understand how orbital data centers can make sense. Is it mainly for continuous solar energy? Surely this can't be enough to offset the costs of launching?
Continuous 5x solar power (relative to on earth), no earthbound construction red tape or protestors, and yeah it can only possibly pencil out with Starship launching routinely. No other rocket system could even come close to making it work.
Hopefully they will work on response time. I've been noticing it taking 5+ minutes for each turn, for not complicated requests. Seems to vary based on time of day too.
For context, xAI GPU utilization is at 11% and they're also expanding.[0] Renting one datacenter to Anthropic doesn't mean that they would be shutting xAI / Grok down.
I shared a couple of days ago why they were not doing like Google and offering oss models, but damn, offering Anthropic models after all the badmouthing. Next news: OpenAI models live on Colossus 2
I could have used this news 2 days ago. I've been trying out Claude Code for a few days and kept running into the limit, so I wanted to upgrade to Max. In the upgrade-flow they hit me with an identity verification through Persona. No problem, I thought, I'll just cancel the upgrade. Nope, all access to Claude Code on the old plan was now also blocked and can't be unblocked without completing Identity Verification, which I'll never do. What a bad experience.
On the plus-side, it told me how much cheaper Deepseek is and that it's on parity for reverse engineering work.
> First, we’re doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans.
Ok I guess, this was a bit of a hassle, but you're not increasing my weekly allowance, you're just not annoying me as often.
> Second, we’re removing the peak hours limit reduction on Claude Code for Pro and Max accounts.
It wasn't a limit reduction (as in, I didn't have a lower 5-hour limit), it was "tokens are more expensive" and it ate my weekly limits faster. This should never have been instituted to begin with.
> Third, we’re raising our API rate limits considerably for Claude Opus models, as shown in the table below:
Meh.
This is why I don't care for all the "it's a subscription, you're free to not use it!" arguments here. It's not an all-you-can-eat subscription with some generous fair use limits, it's a "X tokens per month for $Y", and they keep lowering the X unilaterally and in secret.
Is that what you do when you prepay for a year to get a discount and the supplier just says "oh I'll just give you half of what you paid for"? You "just move to pay again for the rest"?
How many tokens exactly did they guarantee you when you signed up? I don't recall ever seeing a hard number. It was always pretty clear it was flex pricing.
They guaranteed as many as they were offering when I signed up. I tried it for a month, it worked for me, I signed up for a year. Then they reduced the limits.
If you think that's fine, I have access to an all-you-can eat buffet to sell you for only $2000 a year, it's a steal.
What's going to be the hit on our atmosphere when the data centers re enter? I guess it won't matter as the AI will replace the humans by then for the GDP and tax base.
I mean, as someone who has the Max 20x plan and uses it only outside work (so I could not hit anywhere close to the weekly limit at all), I'll gladly take the 5-hour limit doubling.
My first impression to this post is "what the hell are they thinking?", but actually it seems like a decent move by them.
They basically made it so that normal users can better utilize their plan while not benefitting the backgroundagentmaxxers and stealth openclaw abusers in the ranks of their subscription audience. Making their plan more attractive to the people they actually want to sell to.
Hopefully this leads to a loosening of harness restrictions later.
> We’re very intentional about where we’ll add capacity—partnering with democratic countries whose legal and regulatory frameworks support investments of this scale, and where the supply chain on which our compute depends—hardware, networking, and facilities—will be secure.
*Buys compute from actual fascist Elon Musk in a failing democracy during the death throes of late state capitalism.
Oh what is that, the most "ethical" AI company on the planet making deals with literal democracy undermining fascists?
I'm starting to think the problem with "ethical" AI was always that no company could ever act ethically in the long term. They are and always will be a cancer to society and AI will only serve to amplify this further.
I’m sure that’ll comfort all the minorities affected by the rampant amplification of extremists on Twitter. I don’t disagree those are big achievements but also they’re irrelevant to those who feel the impact of Musk’s own extremism, and their lives would be unchanged if none of the Musk companies existed. If you’re unaffected by racism then it’s going to feel easy to only look at the positives of Musk.
I doubt it'll ever happen because heat dissipation will be a big problem, but this is likely in response to the proliferation of data centers. I would rather have data centers in space than convert countryside to concrete and metal jungles.
I was motivated to post this because I was just reading a thread where many users were praising Valve and GabeN for how their company is run, but I'm curious to read more about A & B.
FTA: "SpaceX has done a lot of engineering work to make its Starlink satellites fainter. They are still too bright for research astronomy, but thanks to new coatings, their brightness has not increased dramatically even as SpaceX has launched larger and larger satellites."
I acknowledge there's an issue here, but I don't think it makes sense to label it "pollution". When something is polluted it generally means using it can lead to some form of harm, directly or indirectly. I fail to see how confusing satellites for stars stars causes harm, per se (though of course it would suck to be an astronomer).
Starlink constellations will lead to a world where there is absolutely nowhere you can go where you cant see man-made junk. No truly pristine wilderness anywhere without being able to see formations of glowing dots helping "off-grid" idiots stream Netflix. It's spiritually harmful if nothing else.
Also who said pollution has to be harmful? Light pollution is a thing, and this is the same class of problem.
Why dont they dip the satellites in vantablack to make them truly invisible?
There are those who would disagree on the existence of a "spirit", and so immediately invalidate that argument.
Light pollution is borderline, but actually acceptable is it does cause harm. It disrupts sleep quality and sleeping patterns, also generally affecting plants and animals negatively.
You argument seems to hinge on Starlink not being a massive improvement to how non-broadband connected folk get internet. Your crusade against "offgrid" idiots is intentionally dense as it ignores the millions of people who will be able to access the internet.
Of course it’s a serious argument. Anyone using telescopes or doing Astro photography now sees Starlink satellites leaving trails all over the place. And that’s with a small number compared to the 1 million satellites they are proposing. It’s a public resource that a private company is stealing from all of us.
Models are a commodity, let's say Elon actually figures out building datacenters in space, or maybe he continues to be the leader of building earth based datacenters. Probably better business to not have yourself as your only customer. Dogfood, and open it to all.
The politics and economics of Musk throwing some support towards Anthropic is interesting (samma is probably pissed).
But, if you will pardon a little rant: I hate the idea of subscription inference plans and also 'dumping' by subsidizing non-profitable products. Inferencing should be pay as you go and dumping illegal.
So you can put Anthropic on your list of companies that like to talk big about safety, but when the rubber hits the road, profits matter more than safety.
reply