I don't think you understand the size of the US capital market. We are talking probably ~150 trillion.
It's easy as fuck for Google to raise this money because they are a money printing business. They are the most profitable company in the world, so for anyone this is basically the same as buying US debt.
Yeah, but Google has the money for this. They are quite literally the most profitable company in the world. They are only raising because they don't want to harm there other businesses buy eating up their capital for this.
"If all of this was done to better humanity, AI development would be done in public, data would be legally obtained, models would be released for free, access wouldn't be gatekept behind ever increasing subscription costs."
The vast majority of AI development is public. There are papers literally every single day to read. In fact everything you need to build Claude and GPT models is public. Thanks to Google, DeepSeek, and all the other research labs. There are more research labs than there are closed shops. In fact there really is only one Anthropic, and lately maybe OpenAI. Google still releases papers all the time on AI.
There are more open source models than closed source models and all of them are accessible without a subscription. Yeah you still need to pay for them, but hey as we build out infrastructure and more time is put into efficient models today will easily run on person compute of the future.
What do you mean with "open-source"? Of course, the inference code for all the open weight models is publically available - see llama.cpp or hf transformers.
There are, however, very few models where also the full training pipeline is available. Olmo by AI2 comes to mind.
Harnesses aren't really going to change much of the performance on models like Opus, and GPT.
You literally can just give the model a bash tool and it will do just fine in fact it will most likely do better than majority of harnesses due to how well models are at bash.
The model do all the lifting. It really doesn't matter which harness you use.
People need to stop thinking that LLMs actually know what they are. They don't. They don't know their Qwen, They don't know they are Opus. They don't even know they are an LLM.
This is why literally every single model's system prompt starts with something like:
"You're Claude Opus a large language model from Anthropic"
They can easily add this "individuality" in RLHF. The base model won't know, correct, but the final user-facing model very much can if that's what they so desire.
Crazy they bring up honest, when Claude models are literally known for straight up lying about things it has done and tries to act like it did what you asked.
I have tested almost every language and Go is pretty good but for some reason LLMs get paranoid over races and just start spamming locks next thing you know every fucking struct has mutexes and every function has locks lol.
The best language I have seen an LLM use was Kotlin. It actually surprised me how well it wrote the language. I wrote a project in it and I think I didn't have to correct it once. Like I was seriously impressed. I just wish Kotlin had better tooling so I didn't have to use gradle or maven lol.
I have a very, very hard time believing that. Surely the acquisition left his wealth largely in the form of Anthropic stock, so his personal definition of success is "rep Anthropic so my stock goes up" and at that point he has succeeded.
Me, I still have to be competent to succeed. I don't just get to declare that because I used AI the effort was a success, and I have 0 desire to work with those kinds of people.
"in May 2026, we entered into Cloud Services Agreements with Anthropic PBC (“Anthropic”), an AI research and development public benefit corporation, with respect to access to compute capacity across COLOSSUS and COLOSSUS II. Pursuant to these agreements, the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee"
Anthropic is paying them 1.25 billion per month to serve Claude in their data centers. That's more revenue than Starlink. In fact that's their largest revenue stream lol.
> "Thinking this will prevent war, the US government gives an impenetrable supercomputer total control over launching nuclear missiles"
That's not even close to what the Golden Dome is.
> The Golden Dome is a planned multi-layer missile defense system for the United States, intended to detect and destroy ballistic, hypersonic, and cruise missiles before they launch or during their flight.
Napkin math on 5 year depreciation is 5.5 billion per year for 28 billion. However the 28 billion is cash upfront, spacex is probably paying 10-20% interest on the 28 billion for another 2-6 billion per year.
So net you are looking at finance expenses of 7-11 billion per year. The electricity costs will be significant on top of that, but harder to get a solid read on.
Net of everything, spacex may be getting a 14-28 percent yield before paying for electricity. After electricity/insurance/data/taxes/other expenses - I’d guess it’s anywhere between 0% and 7% yield.
Odds are good that Anthropic abandons the deal before the depreciation schedule completes. Who is going to rent the GPUs then?
5.5 year depreciation is only on the chips. Power, networking, cabling, the actual construction of the building is probably closer to 60% of that number. Also, they are only renting out colosus 1 ($10B), not colosus II ($18B).
So, it's 10B, with $4b of that being attributed to a 5 year depreciation. The rest of the facility probably has a depreciation of around 20 years, and you can easily swap out GPU's, TPU's, Trititum, Tesla's own GPU's, as they start failing, so the normal depreciation curve only "kinda applies here".
There is no interest, as he was venture funded not debt funded.
Electricity is coming from Nat Gas Turbines, so again even though you have a some depreciation on the equipment there, you are getting it for far below meter prices.
So, from my math, he gets ROI on the chips in 3 months, and ROI on the entire facility in 9 months? That's literally the best investment of all time.
> There is no interest, as he was venture funded not debt funded.
Who is "he"? SpaceX has $20bn of debt and $9bn in "other financing" corresponding to "obligations related to certain AI infrastructure assets recorded as failed sale-leaseback transactions."
Edit: Page 122 of the S-1. They are paying SOFR + 0.75%. So around 5.5% on the $20B.
I'll keep the below for integrity sake:
Well, i'm sure SpaceX bought Xai using some kind of prefered share/debt financing, but that's not to say that XAI had the original debt financing.
We can never know what the exact details, and the exact financing is on this debt. Maybe it's tied to Elon's Tesla Shares, Maybe it's tied to a convertible, maybe it is actual "loans" from a bank. Even at $9b in debt, and you naivly assume they are paying 10% (Def not 20% as OP claimed), you are paying $900m a year, for the entirety of xAi. Including that in the calculations to rent out the entire compute is folly. Not only is 900M not directly attributed to c1, cause it's split between c1, c2 and all the training runs, but you can never verify the interest. And even then, one month of this deal pays for the whole year of interest expense.
So go ahead and lower my estimate by 10%... doesn't make a difference.
If only there was some SEC filing available disclosing additional information about the 6-months $20bn bridge loan which was on the news four weeks ago…
oh, nice! thanks for this. This confirms spaceX has $20B at a SOFR + 0.75%. So around 5%-5.5%. Found on page 122 of the S-1. I assume they will retire all debt with the $50B raised from the IPO, plus have $30B to... dominate with starship. If starship is around $100m per launch right now, they can launch 300 ships with the IPO. or 30,000T in orbit.
> If starship is around $100m per launch right now, they can launch 300 ships with the IPO. or 30,000T in orbit.
Uh, starship is still a development program. There's 1 launch pad right now able to launch V3. No starship has flown with an actual live payload. The starlinks going out the PEZ dispenser are probably the only thing launching on it anytime soon.
Basically, Starship launching thousands of tons to orbit isn't constrained by money but by time.
But you are right to call out the launch infrastructure as the true bottleneck. They have 3 pads currently under development. So in 6-9 months they'll have 4 operational pads.
Also, how do they heat tiles hold up? How fast can they catch, refurbish and relaunch is what remains.
I'm confident, and will be putting my money where my mouth is (By investing in the IPO) that they will have useful orbital payloads this year.
Before Starlink we only send up like 1k ton of payload.
Starlink is the only reason why it jumped to 3k tons.
So SpaceX builds all of this to send its own stuff up which is basically only Starlink and in the future its own competition (amazon and leo). For something which is only consumed by 10 million customers right now and they increased the price for starlink which makes it even less competitive.
And Spacex has to send up Starlink every 5 years which keeps revenue low and Starlink is hard llinear growth as one Starlink Satelit can't handle that much traffic.
If his IPO makes all of that money, he will entertain us with funneling billions into a system which will then deliver a handful people onto mars if even.
A person on mars doesn't make money, it costs money.
uh.. orbital data centers? Are we on the same page? 1.25B a month for half a gigawatt, with it rising up to 1g is the deal they just made with anthropic (total deal size 2-2.5B a month). Starship cost is probably something like (y = (400m - 20m*(0.85)^t + 20m, where t is times starship is launched. At 30 launches, they are close to their target of $20m. Falcon 9 has launched 649 times, they've resused the same booster 34 times.
a single nvidia NVL72 rack's peak performance is around 135w so 1gw (1,000,000 kW /135 kW per rack‚ is 7,407. each rack weighs around 1.47 metric tons. so you have 7407/1.47t = 10,888.29t+ 15,243.606 (plus 15,243.606 is for an additional 140% for foldable radiator and solar panels... so 26,131.9t to orbit for 1gw of compute. each starship can do 100T, so 26,131.9t/100 gives us 261.32 starships. given the cost curve earlier:
So $7.5B for the required tonnage to space. 3 million to $3.5 for each rack is 7407 * 3.5m = 25,924.5b. + 7,558.19 is 33b. if we can rent 1gw for $2-3b a month we get buyback in 13ish months? Literally best business model ever. if they last 5 years, each gw is worth $160-180B for the cost of $33B. once block 4 comes out with 200t... AGI ;)
Starship has launched today and hasn't finished everything it needs to even proof that it works reliable. For a Datacenter you need to put it higher than low orbit.
After that, you will need to start to even develoop server racks full of GPUs for space. You literlay need to engineere from scratch a cooling system which will be a few hundred square feed big. You need to be able to transfer massive amount of energy reliable from a small GPU rack to all of this area.
Than you need to send a few thousand of these constalations up there. Every single micrometeroid, sun storm, broken component means loosing a whole rack immediadlty.
Then you need to actually verify that you can send up all of that infrastructure in space and keep it alive there.
Then you have to assume a certain amount of lifetime which will be a lot shorter than on earth and you can't sell it of. These resources are gone.
To all of this cost, you will have to add latency, data syncronisation between racks has to be complelty engineered from scratch too. We talk about 100 of gigabits between these small constelations.
In the same timespan with a lot less money, you could already build a normal datacenter somewhere on the aquator or just as south in texas as possible. You can service it, you can upgrade it, no issues.
all of these are already solved with starlink besides the high orbit. The real problem is power.
>you could already build a normal datacenter somewhere
Can you? from where i'm sitting it's pretty illegal to build a datacenter in the united states. There has literally been hundreds of projects to build data centers around the US that have gotten canceled due to NIMBY's. Collosus itself is plagued by lawsuits.
All those things you listed aren't deal breakers, and maybe some earth based compute will be used but we simply don't have enough power, political will or the legal system to be able to push these data centers through.
So despite the fact that it is theoretically "Cheaper" to build a ground based system, and it will take 261 starship launches just for a single GW... The math is still mathin'. $33B to make $180B.
Generating your own electricity really isn't that much far below grid prices - you'd typically be looking at a central estimate of 15 years to get a return on investment for a large Combined Cooling and Power cogeneration setup.
But, crucially, there's a huge level of uncertainty. You're making bets on the relative cost of nat gas and grid power, both of which have historically shown extreme volatility over that sort of timescale.
The level of risk means that very few proposed schemes go ahead in full unless there's some other factor involved (lack of sufficient grid connectivity, availability of subsidies).
yes, it says they will be paying for "additional capacity" at a reduced rate as it becomes availble in May and June.
Essentially, they are using most of C2 for Grok5. That training run is coming to an End, and they will be leasing more capacity. So taht 1.25b per month will go up to around $1.5B-$2B per month as they finish grok 5.
So the 1.25B is for c1, and the revenue will scale into c2. No idea how much scale, but c2 is almost double the compute? So potentially $3b a month, but probably closer to $2.5B since they get a discount.
At the time of the announcement IIRC the deal was only for Colossus 1. Is Anthropic also leasing Colossus 2 new?
At the time the consensus narrative was that SpaceX no longer needed Colossus 1 for Grok and that was why it could be leased to Anthropic while Colossus 2 would handle Grok training and inference. Does Anthropic also leasing Colossus 2 change this?
Right. This compute still being powered by an illegal amount of gas turbines in a residential neighborhood?
Claude is eating so much compute, the threat of that power being tuned down by lawsuit (rightfully) is worth the risk to Anthropic in the short-term. Instead of declaring "bubble", I'm just going to say that's so crazy.
Yes and both are getting sued? I wouldn't necessarily classify the NAACP as environmental activists, but they are concerned with the wellbeing of the people they represent.
"I suspect your definition of a lousy place to build infrastructure in might overlap with my definition of a relatively good place to live."
Which is quite close to the standard NIMBY attitude: "I want good grid, cheap electricity and other infrastructure, but not in my backyard. Either someone else's, or somewhere where no one lives at all."
Can't you see the fundamental problem with that attitude? We mostly live in densely populated regions. We do need roads, rail and power plants to sustain our way of living, you too.
Of course we need infrastructure. It’s annoying when a project that seems like a net good is held up due to environmental concerns, but this planet is very beautiful and I’d like to be able to continue to enjoy it.
In practice, everything is tradeoff. Your house, and mine, and everyone else's here stands in some place which was once a beautiful natural spot. But few people would support tearing their own homes down in order to restore that beauty.
A technical civilization of 8 billion people cannot exist without doing at least some damage to the environment, but most of the time, outright bans on further development are demanded instead of some reasonable mitigation.
Maybe you really deeply care about the environment. Most of the NIMBYs I met don't. When it comes to infrastructure, housing etc., the environmental concerns are quite often just a legal tool, and the real motivation of the people who wield them is more along the "I have mine, I don't care about yours, just sod off".
has anyone done the math on:
1. cost to build out and run the data centers
2. cost of compute (hardware and energy)
3. depreciation of legacy GPU and thus value at the end of 3 years.
And then compare the $45B revenue from Anthropic to see if it's mostly break even or if one of Anthropic/SpaceX came out ahead on the contract.
Maybe it is a win/win. Anthropic gets desperately needed compute at a fair price. SpaceXAI sells compute at a fair price and gets desperately needed revenues.
XAi purchased Teslas allocation of GPUs, in exchange for Tesla purchasing XAis allocation of GPUs at a later date. This they claim was done because Tesla didnt have their datacentre ready to receive those gpus at that time. I dont see how or who was robbed here.
SpaceX is already indicating their strategy on this, because they’re renting their last-gen data center to Anthropic and keeping the current-gen data center for themselves. Rinse and repeat.
Ed Zitron https://www.wheresyoured.at/ has done the math, and it's pretty bleak. His somewhat voluminous rantings contain raw figures on investments, data centre builds, energy availability and depreciation.
He believes Oracle has already signed it's own death warrant, and that Meta is close behind. MS, Amazon and Google have massive revenue streams to sustain them, but looking at the numbers, each has to earn from AI the equivalent of their existing real revenue. I can't see that happening.
And he believes from multiple perspectives of the data that Nvidea are either massively overstating their GPU sales, or that there are warehouses full of unused GPUs. There just isn't the energy capacity to run them all, let alone data centres to put them in.
He is complaining that there are no 1GW+ data centers, with evidence like this:
> For example, CNBC’s MacKenzie Sigalos reported in October 2025 that Amazon’s Indiana-based (allegedly) 2.2GW Project Rainier data center was “operational,” but only seven out of a planned 30 buildings were actually operational, and her comment of “with two more campuses [of indeterminate capacity] underway.” This comment was buried two videos and 600 words into a piece that declared the data center was “now operational,” with the express intent of making you think the whole thing was operational.
But if you read the report that "buried" comment is far from buried - the whole thing is about how it is still under construction!
Of course 1GW data centers don't all come online at once! You get them online in the parts you can as soon as you can!
From a previous comment of mine – the quotes are all from a single article:
He comes across as just a ludicrously unpleasant, spite-filled person.
> I'm fucking tired of having to write this sentence.
> I am so very bored of having this conversation
> I don't care about this number!
> Shut the fuck up!
> This isn't the early days of shit.
> Didn't we just talk about this? Fine, fine.
> $3.25 billion a quarter is absolutely pathetic.
> This isn’t real business! Sorry!
> He said in one of his stupid and boring blogs that
> This man is full of shit! Hey, tech media people reading this — your readers hate this shit! Stop printing it! Stop it!
> It's here where I'm going to choose to scream.
> Dario Amodei — much like Sam Altman — is a liar, a crook, a carnival barker and a charlatan, and the things he promises
are equal parts ridiculous and offensive.
> Why are we humoring these oafs?
> Despite Newton's fawning praise
> Nobody talks like this! This isn’t how human beings sound! I don’t like reading it!
> Ewww.
> I'm sorry, I know I sound like a hater, and perhaps I am, but this shit doesn't impress me even a little.
> I know, I know, I'm a hater, I'm a pessimist, a cynic, but I need you to fucking listen to me: everything I am describing is unfathomably dangerous
> expensive, stupid, irksome, quasi-useless new product
> I know this has been a rant-filled newsletter, but I'm so tired of being told to be excited about this warmed-up dogshit.
> I refuse to sit here and pretend that any of this matters.
> I'm tired of the delusion. I'm tired of being forced to take these men seriously.
When I read this kind of thing, it’s very apparent that this is being driven entirely by spite not insight. He’s just so angry about everything. There are 57 exclamation marks in this article!
In the 90s we had people talking like this about The Internet. They're all over on FB now, with a detour in between to say stuff like "my isp can track me!?"
If this doesn't completely destroy any benefit of doubt that people have in Dario, I don't know what to say. The guy has been speaking out of both sides of his mouth since the foundation of Anth. Over and over he repeated that Anth's investment in compute would not be reckless; it would represent sanity and be proportional to growth. Over and over Anth told users that the models weren't nerfed overnight, we were just prompting wrong. As expected, Anth simply failed to make the early compute deals when everyone else did, and are now forced to be on the wrong end of price gouging just to keep pace. Actions speak louder than words.
It's true that Anthropic didn't buy as much compute as OpenAI. But OpenAI's compute purchases are one of the largest investments in human history.
It's also true that they are now scrambling for compute, and might be paying more than OpenAI paid. But now they have the revenue to justify it!
To me it is the opposite of "speaking out of both sides of his mouth" - he's been consistent in his "we won't be reckless in buying compute too far ahead of demand" message.
I don't think this is right. Anthropic's growth in the last 6 months went hockey stick in quite an unexpected way (eclipsing OpenAI), so they've done what is sensible - they've increased their compute spend. I don't know if what they're buying from SpaceX is good value, I think there's plenty of reasons to think they got a fine deal. X AI failed. Everyone left. So SpaceX is sitting with a bunch of empty server farms. Yes, Anthropic are desparate for compute, but SpaceX are desparate to IPO a company with double digit billion dollar revenue for $2T so I think there's good reason that this deal represents reasonable value.
skeptical of someone who talks about anthropic enough that they think 'anth' is a reasonable abbreviation lol. maybe youre not wrong but youre definitely biased and not coming to conclusions in a rational way
So you're complaining that his actions matched his words?
He didn't make compute deals until he saw the growth necessary to justify them. As a result, they're paying over-the-odds compared to if they'd have make deals earlier. Maybe that was a poor business decision, but I'm not sure how it represents speaking out of "both sides of his mouth"? Sounds like he was honest.
Even if we consider what they pay for colossus high (I've no idea, I haven't looked at the numbers), wouldn't this a bit be different from "investing in infrastructure"? They're not building the DC themselves, they're just renting, they can scale down to 0 anytime and not have pressure to recover costs, they don't have debt on the HW, etc.
At least from how it's described in the filing, it sounds like they're committed through 2029 on the deal with SpaceX. So they can't just pull out without a massive lawsuit from SpaceX.
"Pursuant to these agreements, the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee. The agreements may be terminated by either party upon 90 days’ notice."
What would be interesting to know how much did it cost xAI to build it ? Ai says between $18-$40 billion to just build, without running cost, but no idea how close to reality this is.
Yes and no. They paid what everyone else pays for those gpus. NVIDIA make the profit and leaves crumbs for the rest. For other components they paid less, but since the gpus are the majority of the cost…
Anthropic is getting capacity from Colossus 1 not Colossus 2 it sounded like. The initial colossus capex was under $5B, making that an even more astounding payoff.
Edit: S1 states both are being leased so the 20-25B initial investment probably more relevant
I thought I saw a report from someone at Google saying that they were still running 7+ year old hardware because of demand. Even if it is not state of the art, if it generates more than the electricity costs, keep it running until it dies.
The article says:
In December, an internal Nvidia memo seen by CNBC said, “Elon prioritizing X H100 GPU cluster deployment at X versus Tesla by redirecting 12k of shipped H100 GPUs originally slated for Tesla to X instead. In exchange, original X orders of 12k H100 slated for Jan and June to be redirected to Tesla.”
There is a lot of money coming in from industry, actually an exponentially increasing amount of money - money that would otherwise go to employees is now going to AI companies. There are more startups than you can shake a stick at that are basically creating virtual workers for xyz task and selling them to companies. All that money is funneling back to AI companies. It's a gold rush and employee salaries are the gold.
You can quibble about the exact numbers, but I think it’s fairly clear at this point that inference is profitable with decent margins. Like you say, unit economics are more interesting than the profitability of the company as a whole.
This thing is going to explode. From this I have to imagine OpenAIs numbers are also going to be much worse than people imagine / what has been shared.
If my napkin math is right, Anthropic will make SpaceX/Xai whole on the initial investment for Colossus 2 + operating costs + a small profit by the end of the contract. No massive profits, although they may be able to fudge accounting to make it seem that way.
Sure, either side could cancel. But Anthropic needs compute, and they found it in SpaceXAI. Why would they cancel the deal unless they don't need more compute or if they could get compute for less elsewhere (but where would that be realistically)?
Anthropic have deals in place to get another 5GW from Amazon, and 5GW from Google, just that it is not available yet. For comparison the total of Colossus #1 and #2 is 1GW
They certainly have some big shoes to fill. But I'm glad they didn't die with their boots on, and got their foot in the door at this new opportunity. It certainly didn't help that they were running on a shoestring budget.
Whoa. I've said before, but I think Dario severely underestimated the coming demand and ensuing need for compute, and would need to pay through the nose when the crunch hit. I suspect that Google deal also worked out better for Google. This data point supports that view.
While Altman got laughed out of the room as a "podcasting bro" asking for trillions in investment in compute, Dario was going on about how difficult it is to forecast capacity on the Dwarkesh podcast. Seems like a major unforced error on Dario's part. What I cannot understand is how they both came to such different perspectives; my best guess is that ChatGPT has so much more traffic that OpenAI could gauge the trends much better.
This won't hurt Anthropic long-term of course, but this won't look great on that balance sheet, that too right around the time they plan to IPO.
They have different personalities. I can only imagine Altman wants to stay on top of the chaos, and believes he will come out ahead whatever happens, while Dario is trying to stay realistic and mitigate worst-case scenarios.
Oh, for sure it could be much worse -- they could have been in xAI's place! ;-)
But while this is a "good problem to have" it would have been an even better problem to avoid in the first place, because it seemed avoidable.
Now, I'm totally armchair billionaire-CEO-ing here, but anybody with any compute has been so obviously capacity constrained for so many quarters all the while scrambling like mad and spending obscene amounts of money to acquire even more compute. With lead times of 2 - 3 years, something Dario explicitly called out on Dwarkesh, it seemed prudent to acquire first, ask questions later. Worst case, they could have rented any extra capacity out, like Elon is doing!
Outsiders are reasonably questioning this mania but Dario, as one of the biggest believers in AI and even AGI, showing hesitancy seems uncharacteristic. I wonder if this is one of those rare cases where it would have been better to drink his own Kool Aid!
Anthropic got somewhat lucky that Elon wanted to stick it to Altman, but boy, even then he drove a hard bargain.
That's fair, I'm not a finance person, just meant to say that this is costing them way more than they would have liked at a critical stage in their corporate evolution. I'm curious though what your take is.
They literally own their own data centers. That's whats surprising about this. They are lying to their customers when they say they operate their own data center because obviously they don't if everyone's apps are down with GCP blocking their account.
It's easy as fuck for Google to raise this money because they are a money printing business. They are the most profitable company in the world, so for anyone this is basically the same as buying US debt.
reply