I live in Claude. I think Anthropic has built an amazing product. I also read the S-1 this morning, and a few things jumped out at me.
As of June 30, Anthropic disclosed approximately $265.6 billion of expected future fixed or minimum guaranteed commitments, largely for compute and infrastructure, with terms running from one to 20 years. Separately, it had approximately $321.4 billion of undiscounted obligations under leases that had not yet commenced. Let the scale of that sit for a minute.
It has committed to spend at least $114.5 billion with Google through July 2033 and at least $110 billion with Amazon through April 2036, in both cases "regardless of actual usage." Where spending falls short of those minimums, Anthropic pays the difference in cash.
Against that, Anthropic generated $16.3 billion of revenue in the first half of 2026 and held approximately $132 billion of cash and short term investments.
The company identifies the tension itself: "our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years."
The part I find most interesting is the connection to the neoclouds. Some of the capacity Anthropic has contracted for has not been built yet. Its agreements run to approximately $44.6 billion with Nscale and approximately $41.3 billion with Lambda, and both firms still have to finance the chips and build the sites.
Nscale's own S-1 says that as of September it "had not obtained binding commitments for any of the financings required to fund performance under the Anthropic Services Agreements."
Both are also on a clock. Nscale must obtain that financing within a specified period, and Lambda's largest tranche, about $34.9 billion, can be terminated if it does not timely secure its facility lease. So these companies are incentivized to come to market whenever a window opens rather than when pricing is 'good'. Financing conditions for leveraged AI infrastructure have become materially tougher over the last couple of months.
I am not an IPO expert. What interests me is what this says about credit, and about the sheer scale of AI financing still to come. Demand has to grow into the compute Anthropic has committed to. The infrastructure providers have to finance and deliver the capacity. The first is an Anthropic risk and AI adoption risk. The second is not entirely in Anthropic's control, and a lot of paper will have to get placed to make it happen.
As of June 30, Anthropic disclosed approximately $265.6 billion of expected future fixed or minimum guaranteed commitments, largely for compute and infrastructure, with terms running from one to 20 years. Separately, it had approximately $321.4 billion of undiscounted obligations under leases that had not yet commenced. Let the scale of that sit for a minute.
It has committed to spend at least $114.5 billion with Google through July 2033 and at least $110 billion with Amazon through April 2036, in both cases "regardless of actual usage." Where spending falls short of those minimums, Anthropic pays the difference in cash.
Against that, Anthropic generated $16.3 billion of revenue in the first half of 2026 and held approximately $132 billion of cash and short term investments.
The company identifies the tension itself: "our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years."
The part I find most interesting is the connection to the neoclouds. Some of the capacity Anthropic has contracted for has not been built yet. Its agreements run to approximately $44.6 billion with Nscale and approximately $41.3 billion with Lambda, and both firms still have to finance the chips and build the sites.
Nscale's own S-1 says that as of September it "had not obtained binding commitments for any of the financings required to fund performance under the Anthropic Services Agreements."
Both are also on a clock. Nscale must obtain that financing within a specified period, and Lambda's largest tranche, about $34.9 billion, can be terminated if it does not timely secure its facility lease. So these companies are incentivized to come to market whenever a window opens rather than when pricing is 'good'. Financing conditions for leveraged AI infrastructure have become materially tougher over the last couple of months.
I am not an IPO expert. What interests me is what this says about credit, and about the sheer scale of AI financing still to come. Demand has to grow into the compute Anthropic has committed to. The infrastructure providers have to finance and deliver the capacity. The first is an Anthropic risk and AI adoption risk. The second is not entirely in Anthropic's control, and a lot of paper will have to get placed to make it happen.
0
0
1
58
The AI we already have and the next trillion dollars or more of AI capex are not the same bet. And yet they keep getting defended with the same argument.
The money that has already been spent has bought remarkably capable models. For the vast majority of knowledge work, I am not sure model intelligence is even the constraint anymore. The constraint is training people on how to use it and helping them adopt it in their workflows.
There are definitely reasons to keep pushing the frontier like national security, scientific research and capabilities we haven't discovered yet. But that can be financed at a very different scale.
Maybe the argument is that enough additional compute eventually unlocks something fundamentally different, like self improving systems and harnesses that can suddenly do domain specific work far better than they can today. We aren't there yet.
All of this leads me to a question I haven't seen addressed. What does the efficient frontier of compute spend versus useful AI intelligence look like? And where on that curve is good enough for the vast majority of us?
The money that has already been spent has bought remarkably capable models. For the vast majority of knowledge work, I am not sure model intelligence is even the constraint anymore. The constraint is training people on how to use it and helping them adopt it in their workflows.
There are definitely reasons to keep pushing the frontier like national security, scientific research and capabilities we haven't discovered yet. But that can be financed at a very different scale.
Maybe the argument is that enough additional compute eventually unlocks something fundamentally different, like self improving systems and harnesses that can suddenly do domain specific work far better than they can today. We aren't there yet.
All of this leads me to a question I haven't seen addressed. What does the efficient frontier of compute spend versus useful AI intelligence look like? And where on that curve is good enough for the vast majority of us?
0
0
0
24
If you're trying to work out what AI is actually worth to you, I'd break it into two categories.
The first is productivity, and it's the easy one. An experienced investment professional's time has an hourly value and you know roughly what it is. So when analysis that used to take three days takes an afternoon, you count the hours, multiply, and put a dollar figure on it. That number is real. It is also the smaller half.
The second is harder, and I think it matters more. It is how much better an investor you become because you can cover more ground.
The names you can now triage far enough to know they aren't worth your time. The situations you can screen that you never used to get to. The second-order exposure in the book nobody had the hours to check. None of that was slow before. It simply didn't happen.
So there are no hours to save on it, and it never shows up in the productivity math. You just arrive at the decision having looked at far more of the universe than you used to.
That is not time saved. It is being augmented to do more than you could do on your own, and walking into the decision more confident the work was actually done. That is much harder to measure, but I think is worth much more.
The first is productivity, and it's the easy one. An experienced investment professional's time has an hourly value and you know roughly what it is. So when analysis that used to take three days takes an afternoon, you count the hours, multiply, and put a dollar figure on it. That number is real. It is also the smaller half.
The second is harder, and I think it matters more. It is how much better an investor you become because you can cover more ground.
The names you can now triage far enough to know they aren't worth your time. The situations you can screen that you never used to get to. The second-order exposure in the book nobody had the hours to check. None of that was slow before. It simply didn't happen.
So there are no hours to save on it, and it never shows up in the productivity math. You just arrive at the decision having looked at far more of the universe than you used to.
That is not time saved. It is being augmented to do more than you could do on your own, and walking into the decision more confident the work was actually done. That is much harder to measure, but I think is worth much more.
0
0
0
15
Ran the first cohort of AI for Credit Professionals on Friday with a good mix of buy-side and sell-side professionals. It was two hours, part presentation and part hands-on training.
This isn't generic AI training. It is specifically for investment professionals in credit.
Running a second cohort in the next few weeks given the demand. If you are interested, reach out to me at prashant@eigenstrategy.com
This isn't generic AI training. It is specifically for investment professionals in credit.
Running a second cohort in the next few weeks given the demand. If you are interested, reach out to me at prashant@eigenstrategy.com
1
0
0
24
Ran the first cohort of AI for Credit Professionals on Friday with a good mix of buy-side and sell-side professionals. It was two hours, part presentation and part hands-on training.
This isn't generic AI training. It is specifically for investment professionals in credit.
Running a second cohort in the next few weeks given the demand. If you are interested, reach out to me at prashant@eigenstrategy.com
This isn't generic AI training. It is specifically for investment professionals in credit.
Running a second cohort in the next few weeks given the demand. If you are interested, reach out to me at prashant@eigenstrategy.com
0
0
0
20
Check out my latest article: Where the Risk actually sits in the AI Buildout
linkedin.com/pulse/where-ri… via @LinkedIn
linkedin.com/pulse/where-ri… via @LinkedIn
0
0
0
24
I miss being in an investing seat right now. The volatility makes the opportunity set in credit markets genuinely interesting.
The real edge today is how you use AI across credit markets, corporate or structured. It can triage a pipeline of ideas across primary and secondary. It can run more and better screens. It can synthesize information faster and more effectively, which matters when the first bid or offer is what lets you trade ahead of the rest of the market. And it can help separate signal from noise in a space like AI infrastructure right now.
Most of the press around AI infrastructure credit is the concern around the neoclouds. The interesting question is how the financing of this sector evolves; whether it stays a mix of public and private near term, or whether, like in the past, issuers are forced to go to the private markets and pay a premium for certainty of execution. I expect more structuring creativity to show up; warrants, equity kickers, and components that bridge a wider cost of capital gap. Also, I would expect more granularity from issuers around contracts with tenants and more differentiation and dispersion based on whether they are take or pay, cancellable etc.
And that is where AI becomes particularly valuable in credit: not because it tells you what to buy, but because it lets you evaluate more situations, more variables and more documents fast enough with your lens, to make better investment decisions.
The real edge today is how you use AI across credit markets, corporate or structured. It can triage a pipeline of ideas across primary and secondary. It can run more and better screens. It can synthesize information faster and more effectively, which matters when the first bid or offer is what lets you trade ahead of the rest of the market. And it can help separate signal from noise in a space like AI infrastructure right now.
Most of the press around AI infrastructure credit is the concern around the neoclouds. The interesting question is how the financing of this sector evolves; whether it stays a mix of public and private near term, or whether, like in the past, issuers are forced to go to the private markets and pay a premium for certainty of execution. I expect more structuring creativity to show up; warrants, equity kickers, and components that bridge a wider cost of capital gap. Also, I would expect more granularity from issuers around contracts with tenants and more differentiation and dispersion based on whether they are take or pay, cancellable etc.
And that is where AI becomes particularly valuable in credit: not because it tells you what to buy, but because it lets you evaluate more situations, more variables and more documents fast enough with your lens, to make better investment decisions.
0
0
0
33
The quiet, unsexy gap in AI for finance isn't orchestration. It's data cleaning, and the work that goes with it.
Foundation model labs and AI wrapper companies may be underestimating how much labeling and cleaning credit data actually requires before a model can use it well, especially once you start dealing with non-public securities.
If you are working on any solutions related to this or are focused on the credit data space and perhaps more unique data sets, I would love to connect and hear about it.
Foundation model labs and AI wrapper companies may be underestimating how much labeling and cleaning credit data actually requires before a model can use it well, especially once you start dealing with non-public securities.
If you are working on any solutions related to this or are focused on the credit data space and perhaps more unique data sets, I would love to connect and hear about it.
1
0
2
50
There's understandably a lot of consternation about AI related debt out there.
I read a NYT interview with Peter Orszag, tied to a paper he co-wrote on the new supply side era, where he compares the scale of AI capex debt to the fiber buildout of the 1990s.
The scale of capex investment is enormous, but what is going to matter the most is the speed of adoption of AI, particularly at 'traditional' firms that are not VC or tech related. Whether real adoption happens over the next few years or the next decade, changes the return on capital completely.
For the debt financing this capex, there is another question: who actually bears the adoption risk?
If the contracts are long-term, take-or-pay and backed by strong counterparties, the pace of end-user adoption may matter much less to the lender. If the contracts leave utilization, pricing or renewals exposed to the market, that is a very different credit.
Compute pricing is showing a similar pattern. Rent it now and it's expensive, because near term demand is the one thing everyone can see clearly. Go out 6 months, a year or two years and the compute pricing curve inverts, meaningfully, because nobody has real visibility into usage, adoption or which GPUs will be standard by then. Thanks @stanleylee21 for sharing this info.
So if you're pricing any AI related financing, the pace of adoption is a key underwriting input. And just as important is determining who has contractually agreed to bear the risk if adoption takes longer than expected.
I read a NYT interview with Peter Orszag, tied to a paper he co-wrote on the new supply side era, where he compares the scale of AI capex debt to the fiber buildout of the 1990s.
The scale of capex investment is enormous, but what is going to matter the most is the speed of adoption of AI, particularly at 'traditional' firms that are not VC or tech related. Whether real adoption happens over the next few years or the next decade, changes the return on capital completely.
For the debt financing this capex, there is another question: who actually bears the adoption risk?
If the contracts are long-term, take-or-pay and backed by strong counterparties, the pace of end-user adoption may matter much less to the lender. If the contracts leave utilization, pricing or renewals exposed to the market, that is a very different credit.
Compute pricing is showing a similar pattern. Rent it now and it's expensive, because near term demand is the one thing everyone can see clearly. Go out 6 months, a year or two years and the compute pricing curve inverts, meaningfully, because nobody has real visibility into usage, adoption or which GPUs will be standard by then. Thanks @stanleylee21 for sharing this info.
So if you're pricing any AI related financing, the pace of adoption is a key underwriting input. And just as important is determining who has contractually agreed to bear the risk if adoption takes longer than expected.
0
0
3
1,109
A number of people I know have been approached recently by consultants offering to buy their anonymized data including their workflows. Meaning, how they actually do their work, anonymized, then sold to companies training AI models on it.
And it makes a lot of sense, if you're Anthropic, OpenAI etc. and you are potentially concerned about model capability flatlining based on your current training data sets.
For most companies, the real moat is the data, the relationships and how the work actually gets done. That is the part that models can't easily copy because the training data typically doesn't include this information. But if you choose to monetize this way, that can change quickly.
So now I'm wondering, if enough of that gets handed over, do these models eventually get good enough to replace the experienced professional who built it? I don't know, but I am genuinely curious and a bit concerned.
And it makes a lot of sense, if you're Anthropic, OpenAI etc. and you are potentially concerned about model capability flatlining based on your current training data sets.
For most companies, the real moat is the data, the relationships and how the work actually gets done. That is the part that models can't easily copy because the training data typically doesn't include this information. But if you choose to monetize this way, that can change quickly.
So now I'm wondering, if enough of that gets handed over, do these models eventually get good enough to replace the experienced professional who built it? I don't know, but I am genuinely curious and a bit concerned.
0
1
1
57
A few thoughts on the recent vol in credit markets
1) I think it's too soon to call the CCC weakness purely idiosyncratic. Distress usually starts with the weakest links. Right now it's happening with CCCs, weaker BSLs and pockets of private credit, and that's before we've seen any meaningful remarks from SaaS portfolio companies.
2) People are underestimating what rates do to leveraged credit. Most sponsor owned businesses don't hedge much interest rate risk. That's not a theory, that's what I saw play out in 2022. A real move in rates hits floating rate debt and any fixed rate debt coming up for refinancing feels it just as fast.
3) On the AI side, if the 25-45% ROI from the hyperscalers on AI capex holds up, these issuers can absorb another 50-100bps (or more) of financing cost and still want the deal done. Doesn't make it a bad credit, just less rate sensitive.
4) The issuance we are seeing from the hyperscalers further out the curve will inevitably have an impact too, since that duration has to get hedged somewhere, and that flows into treasury yields.
1) I think it's too soon to call the CCC weakness purely idiosyncratic. Distress usually starts with the weakest links. Right now it's happening with CCCs, weaker BSLs and pockets of private credit, and that's before we've seen any meaningful remarks from SaaS portfolio companies.
2) People are underestimating what rates do to leveraged credit. Most sponsor owned businesses don't hedge much interest rate risk. That's not a theory, that's what I saw play out in 2022. A real move in rates hits floating rate debt and any fixed rate debt coming up for refinancing feels it just as fast.
3) On the AI side, if the 25-45% ROI from the hyperscalers on AI capex holds up, these issuers can absorb another 50-100bps (or more) of financing cost and still want the deal done. Doesn't make it a bad credit, just less rate sensitive.
4) The issuance we are seeing from the hyperscalers further out the curve will inevitably have an impact too, since that duration has to get hedged somewhere, and that flows into treasury yields.
0
0
1
54
This is well written and exactly why domain expertise matters so much. Very applicable in finance.
Yannis and I led forward deployed AI engineering at Palantir.
Much of today’s boom is not true FDE work.
Companies are filling a product gap with engineering headcount. What comes after that? x.com/i/article/2104…
0
0
1
47
P_Rad retweeted
Yannis and I led forward deployed AI engineering at Palantir.
Much of today’s boom is not true FDE work.
Companies are filling a product gap with engineering headcount. What comes after that? x.com/i/article/2104…
Much of today’s boom is not true FDE work.
Companies are filling a product gap with engineering headcount. What comes after that? x.com/i/article/2104…
63
91
1,033
464,949
How are you managing the pipeline of opportunities in credit right now, especially given the volatility? It isn't just a volume problem, it's also a selection problem and volatility makes it worse.
Take Project Jupiter, for example: $18bb of loans behind Oracle's New Mexico data center, quoted around 90 cents. Oracle sent a force majeure notice on September 24. People underwrote this at new issue. If you passed then, is it on your list today, or did it leave your universe when you said no?
For most desks, that's settled by whether someone called and told you or you read it on Bloomberg, Octus or 9fin. You didn't choose that filter.
That's the case for pointing AI at the top of the funnel and triaging your pipeline of primary and secondary ideas. Not that it removes bias. Your filter is bias. But it's one you deliberately wrote down versus one you inherited. And it also allows you to look at ideas that net out the effect of crowding, which is a big issue in credit investing.
The scarce input was never information. It's what you chose to look at.
Take Project Jupiter, for example: $18bb of loans behind Oracle's New Mexico data center, quoted around 90 cents. Oracle sent a force majeure notice on September 24. People underwrote this at new issue. If you passed then, is it on your list today, or did it leave your universe when you said no?
For most desks, that's settled by whether someone called and told you or you read it on Bloomberg, Octus or 9fin. You didn't choose that filter.
That's the case for pointing AI at the top of the funnel and triaging your pipeline of primary and secondary ideas. Not that it removes bias. Your filter is bias. But it's one you deliberately wrote down versus one you inherited. And it also allows you to look at ideas that net out the effect of crowding, which is a big issue in credit investing.
The scarce input was never information. It's what you chose to look at.
0
0
0
49
I don't think 'Does AI capex have ROI' is one question. I think its five.
And I am skeptical of anyone who claims to already have a single number that answers it.
1) Start with backlog - Hyperscalers report enormous committed compute demand. And a lot of commentary treats that as proof the spend will pay off. The real question is how much of the backlog is take or pay versus cancellable or usage based. And I haven't seen enough disclosure around that split.
2) Cost of compute - This is getting a lot of attention right now for a reason. GPU rental pricing is elevated because memory and packaging are genuinely scarce, booked out years in advance. And the scarcity is a driver of what's happening in semiconductor pricing and equities well beyond the hyperscalers themselves. Some share of today's implied return is a scarcity premium, not a durable structural margin. At least it's not clear to me yet.
3) Adoption - This is a key piece at the center of my company EigenStrategy, and we need a more precise measure than 'companies say they're using AI'. In my opinion, usage inside AI native startups spending venture money is a weak signal, because the spend doesn't have to justify itself the way it does everywhere else. The real test is whether traditional industries, running on their own PnL discipline with no subsidy, are getting enough value to keep paying for it without a mandate from the top.
4) Closed versus Open source - If open weight models keep closing the capability gap, pricing power moves away from the foundation model labs and down towards whoever owns the compute or the application layer. That changes who actually captures the return on all this spending, independent of whether the spending pays off at all in aggregate.
5) Security- Incidents like the ones making news recently are a reminder that none of the return math above prices in a trust shock. It's the factor more likely to show up as a sudden break than a slow curve. And it's the easiest to ignore until it isn't.
A hyperscaler funding capex off its own balance sheet, a neocloud financed against leased GPUs and a data center financed against a single customer's credit are exposed to each of these five factors in genuinely different ways. And until the reporting gets granular enough to separate them, a single ROI figure deserves real skepticism.
And I am skeptical of anyone who claims to already have a single number that answers it.
1) Start with backlog - Hyperscalers report enormous committed compute demand. And a lot of commentary treats that as proof the spend will pay off. The real question is how much of the backlog is take or pay versus cancellable or usage based. And I haven't seen enough disclosure around that split.
2) Cost of compute - This is getting a lot of attention right now for a reason. GPU rental pricing is elevated because memory and packaging are genuinely scarce, booked out years in advance. And the scarcity is a driver of what's happening in semiconductor pricing and equities well beyond the hyperscalers themselves. Some share of today's implied return is a scarcity premium, not a durable structural margin. At least it's not clear to me yet.
3) Adoption - This is a key piece at the center of my company EigenStrategy, and we need a more precise measure than 'companies say they're using AI'. In my opinion, usage inside AI native startups spending venture money is a weak signal, because the spend doesn't have to justify itself the way it does everywhere else. The real test is whether traditional industries, running on their own PnL discipline with no subsidy, are getting enough value to keep paying for it without a mandate from the top.
4) Closed versus Open source - If open weight models keep closing the capability gap, pricing power moves away from the foundation model labs and down towards whoever owns the compute or the application layer. That changes who actually captures the return on all this spending, independent of whether the spending pays off at all in aggregate.
5) Security- Incidents like the ones making news recently are a reminder that none of the return math above prices in a trust shock. It's the factor more likely to show up as a sudden break than a slow curve. And it's the easiest to ignore until it isn't.
A hyperscaler funding capex off its own balance sheet, a neocloud financed against leased GPUs and a data center financed against a single customer's credit are exposed to each of these five factors in genuinely different ways. And until the reporting gets granular enough to separate them, a single ROI figure deserves real skepticism.
2
0
4
117
A few people have asked what the AI training for credit professionals will look like.
I am going to keep it to about 2 hours- start with the basics, what the models are and how they work. Then the downsides, where they fail and the guardrails you need around them. Then the credit-specific part, where AI actually helps.
I am going to keep it to a small group and its intended to be interactive. If you are already using Claude, ChatGPT etc we can work through real questions together.
If you work in credit and want in, email me at prashant@eigenstrategy.com
I am going to keep it to about 2 hours- start with the basics, what the models are and how they work. Then the downsides, where they fail and the guardrails you need around them. Then the credit-specific part, where AI actually helps.
I am going to keep it to a small group and its intended to be interactive. If you are already using Claude, ChatGPT etc we can work through real questions together.
If you work in credit and want in, email me at prashant@eigenstrategy.com
0
0
1
34
There's been a lot of discussion around Harvey AI's margins, which went from 50% to -50% in 6 months. Some of it is the result of foundation model token costs.
The interesting part is that apparently margins improved after the company built its own model using open-weight models.
What's surprising is not many people in the investing world are talking about post-training small or medium sized open-weight models. Yet. And Anthropic still doesn't have one.
I ran an experiment for credit underwriting in June using a Gemma 4 26B model that I post-trained and I tested it on two ABL deals it had never seen. It scored 55/60, while unaided Claude scored 44/60.
Just another example of the contrast in performance between a model carrying your judgment and one running unaided.
The interesting part is that apparently margins improved after the company built its own model using open-weight models.
What's surprising is not many people in the investing world are talking about post-training small or medium sized open-weight models. Yet. And Anthropic still doesn't have one.
I ran an experiment for credit underwriting in June using a Gemma 4 26B model that I post-trained and I tested it on two ABL deals it had never seen. It scored 55/60, while unaided Claude scored 44/60.
Just another example of the contrast in performance between a model carrying your judgment and one running unaided.
0
0
0
56
I have been thinking incessantly about how much a judgement core harness can improve the results of an LLM for credit investing. Turns out, a LOT!
linkedin.com/pulse/edge-hol…
linkedin.com/pulse/edge-hol…
0
0
0
30
I spend most of my time between two groups that see AI infrastructure differently. The tech community and the finance community are at odds on what the buildout will actually earn, and there is very little hard data on the question. Based on commentary from the hyperscalers and VCs, the tech community is far more bullish.
So I pointed my own AI tooling at the primary filings rather than the coverage. What struck me is how much the word backlog is being asked to carry.
SpaceX's IPO filing discloses a cloud services agreement giving Anthropic access to compute capacity that includes roughly 325,000 GPUs, at $1.25 billion a month through May 2029. That is about $45 billion at the headline run rate. But after an initial three month period, either party can terminate on ninety days' notice, and no termination payment is disclosed. The earliest exit is closer to six months than three years.
Nebius is a different version of the same problem. Its Meta arrangement has a contract value of up to $27 billion. About $12 billion of that is dedicated capacity for Meta's own use. The rest, up to $15 billion, is capacity Nebius says it intends to sell to third parties, with Meta purchasing whatever remains. That second piece is a backstop. It is worth having and it helps make the buildout financeable. But it should not be read as $15 billion of compute Meta has said it needs.
Headline contract value is being read as a demand signal. Some of it is financing architecture, and the mix matters a great deal to anyone underwriting this.
If you are closer to these contracts and are willing to share color, let me know if/where I am wrong.
So I pointed my own AI tooling at the primary filings rather than the coverage. What struck me is how much the word backlog is being asked to carry.
SpaceX's IPO filing discloses a cloud services agreement giving Anthropic access to compute capacity that includes roughly 325,000 GPUs, at $1.25 billion a month through May 2029. That is about $45 billion at the headline run rate. But after an initial three month period, either party can terminate on ninety days' notice, and no termination payment is disclosed. The earliest exit is closer to six months than three years.
Nebius is a different version of the same problem. Its Meta arrangement has a contract value of up to $27 billion. About $12 billion of that is dedicated capacity for Meta's own use. The rest, up to $15 billion, is capacity Nebius says it intends to sell to third parties, with Meta purchasing whatever remains. That second piece is a backstop. It is worth having and it helps make the buildout financeable. But it should not be read as $15 billion of compute Meta has said it needs.
Headline contract value is being read as a demand signal. Some of it is financing architecture, and the mix matters a great deal to anyone underwriting this.
If you are closer to these contracts and are willing to share color, let me know if/where I am wrong.
0
0
0
42
