I can’t believe I am going to talk about CoreWeave again. It pains me. I don’t care about the stock. I don’t even think it matters that much to the overall market in even the medium term.
But CoreWeave epitomizes the bull vs. bear tug-of-war in the market. So I keep being pulled back in..
Fortunately, nowadays it is much easier to run a few sets of numbers on a stock and get an idea of where it stands. Which is what this post amounts to.
Yesterday I got pulled back into asking questions about CoreWeave with this tweet.
Randolph Duke is basically saying that the IRR of CoreWeave’s GPU purchases is negative if the life is 5 years. Looking at the numbers, they seem reasonable enough. I was like, hmm, this looks pretty bad for CoreWeave.
I asked Claude about the model and whether it was accurate.
The key is the assumptions. The −15% IRR rests on two big assumptions:
A 9% a year GPU failure
Pricing new $40M-per-MW builds at the fleet’s blended revenue per MW
The failure rate does seem high. I think it comes from Meta’s Llama 3 training run that is detailed in a few places (like here and here). If GPUs and their memory fail at Meta’s rate, the annualized failure rate will be about 9%.
But those were job interruptions on a training cluster, not dead GPUs. Which doesn’t mean they failed permanently.
Also, these were failures during training, which is more intensive than inference. I don’t know CoreWeave’s breakdown between training and inference, but it isn’t 100% training.
Bottomline, I think the failure rate is pessimistic. I’ve seen other places that 3% failure is about right.
The second assumption is pricing. This is the big one.
The table is based on revenue coming in at $9.15M/MW. That is the fleet average for CoreWeave.
That’s the whole fleet. It includes the older Hopper-era capacity that costs less per MW, as well as some even older A100 chips that are basically running for free at this point (they are fully depreciated).
CoreWeave’s Q4 guidance implies about $10.9M/MW across the whole fleet. It’s already higher than the $9.15M/MW number. CoreWeave’s exit run-rate guide of $20.5B on 1.85 GW works out to about $11.1M/MW. New capacity has to be above both, because older capacity pulls the average down.
All this underscores the linchpin, which is simply price. Randolph Duke is using this $9.1M/MW number. What if I use $11M/MW as a base case with Claude? With a failure rate of 3% instead of 9%.
I get a somewhat more palatable IRR. At least it’s positive!
But this is still not a great result for CoreWeave. Both the 5 year and 6 year IRR are below CRWV’s cost of capital.
So what’s going on here? Is CoreWeave a time bomb? Can their management not do math?
The trick here, and this is why CoreWeave really sums up the AI trade so well, is that the $11M/MW may be too low.
CRWV said they were signing deals for $14M/MW earlier this year and $20M/MW+ now. This comes from Bank of America:
So what does the IRR look like at those prices? A LOT better.
At the high end, we have a pretty good 51% IRR.
Now is that revenue number legitimate? I think that is entirely dependent on your view. What I will say is that it is not necessarilywrong.
CoreWeave isn’t just making that number up. SpaceX is thought to have signed a deal at $50M/MW with Anthropic for short term contracts.
Just what this number should be in the medium term is the whole debate. Should it be $9/MW? $25/MW? Somewhere in between? Somewhere even higher? Lower?
Within such an extremely large range, CoreWeave could easily be a $0 or a 3-4-bagger from here.
Given where the stock is right now, I would probably lean long over short. I recognize that it is easy enough to justify both positions. The downside is real. CoreWeave has $47 million of debt. There are lots of claims above equity. If those data centers don’t produce a decent IRR, the equity is toast.
But with the range of outcomes so large, the stock feels like a long-dated option. Honestly, the thing this reminds me of the most is the E&P shale-shitcos when the shale revolution was going full-steam ahead. And how many of those companies traded more like options than equities when the sledding got rough. Much blood was spilled explaining why they should go to zero, but some never did.
$40 for an option that could go to, say, $140 (this is Bank of America’s target on the stock) feels like a better risk/reward than going short at $40 and hoping the move from here is a straight line to $0.
A month or two ago I listened to a podcast interview with a credit investor. I have racked my brain and my listening history to try to figure out what I was listening to at the time, and for the life of me I can’t find the exact podcast or the name of the investor. But I do remember exactly what he was talking about. It was the credit market for AI.
What stuck with me from the interview was (clearly not his name but) the analogy he made between AI credit and commercial real estate credit. He used the investing paradigm for valuing commercial real estate credit to poke some holes in AI financing.
On the surface, what he said made sense. Credit is credit and doing credit analysis is doing credit analysis. Judge the risk of default under different scenarios and price the credit based on relative risk.
Nevertheless, I felt like something was off about it. I just couldn’t put my finger on it. It wasn’t until credit and AI started to become a real issue that I did.
The issue has to do with the nature of commercial real estate and the nature of AI.
AI has lots of upside and lots of downside. Things could turn out really good for AI in general, and they could turn out really good for the winners. But they could also go really badly. And even if AI wins, there could still be some big losers.
AI stocks, other than the biggest ones like Amazon or Microsoft or Meta, are largely options on being one of the big winners. That is why they can move around so much. No business’s underlying value fluctuates 10% or 15% in a day. But when you are treated as an option on an outcome, you can see those fluctuations as the probability of that outcome shifts.
The contrast with commercial real estate couldn’t be more different. Commercial real estate can have real downside if it is done carelessly. But if it is done prudently, the downside can be mitigated. The upside is modest. You will collect a spread, and if you are lucky your property will appreciate at a reasonable rate over time.
This makes commercial real estate ideal for credit. As a credit investor, you are all about understanding the downside. The upside is not really that relevant to you, because by owning a bond you don’t participate in it anyway.
But take the same credit framework and apply it to AI. You are focused on protecting the downside. Looking to price risk appropriately to mitigate it. If everything goes well you get coupons and your principal back. If everything doesn’t, you could lose everything.
Given the risks of AI, unless the debtor has pristine credit, you better price it dearly to be compensated for it. You don’t care what the upside might be.
And that is the problem. It is sort of missing the whole point right?
With that in mind, let’s turn to Akamai.
Akamai has just been a mess of a stock. And much of it is connected to them getting into the business of selling compute for AI.
Akamai began as a Content Delivery Network. They evolved into an edge security platform. Those are the core businesses.
Akamai owns servers all over the world. These aren’t big data centers. They are small and co-located with other servers by design. Most of them are located around big cities. The intent is to be located close to traffic, so they can deliver and monitor content close to the source.
If you are a streaming service, you might pay Akamai to house some of your most popular content in local servers around large populations. Then if someone wants to watch one of these shows, instead of you having to send the content from your server farm in timbuktu, Akamai can play it from much closer. It can be cost effective.
On the security side, you just flip it around. You might pay Akamai to monitor the traffic to and from your website, and if you are attacked or if something suspicious happens, Akamai will shut that traffic down.
For Akamai, it is a business they have been in for 25 years.
Because Akamai didn’t own the sort of space you’d use to stand up 1,000s of GPUs and start training models, it wasn’t a natural play in the first few years of the data center buildout. In fact the stock has done nothing for a long time.
I didn’t buy Akamai because I thought they’d get into the business of selling compute, though I did get stuck as a bagholder because of it. I bought the stock after listening the Cloudflare Q4 conference call back in February, and I started to think that Agentic AI may change the game for not only Cloudflare but Akamai as well.
It made sense to me, and still does, the AI agents will generate far more internet traffic over time and that this traffic should drive traffic to the security services and content delivery services that Cloudflare and Akamai offer.
You just think of how agents scour every website to find information, will read 10x or maybe 100x as many pages of information as a human. I think the clincher for me on that Cloudflare call was when CEO Matthew Prince said “if humans looked at 5 sites when they were making a decision, agents might look at 5,000”.
Cloudflare says bots, including AI agents and crawlers, now make about 57% of web requests. They do much of this through APIs, and websites will have to act to protect themselves and guide these agents through those APIs using tools like what Akamai and Cloudflare sell.
As well, if traffic is going to explode I thought that proximity to content may, over time, gain importance. I’m not sure if the CPU based servers that Akamai and Cloudflare use would ever be used directly for inference (both do own some limited GPU servers that do inference), they would be used to fetch web pages, run agent’s tools and process data, tasks which seems likely to grow.
There is a world here of a lot more traffic and more dangerous traffic and what Cloudflare and Akamai do is manage traffic. So while this business has not been a good on in the past, I thought that may change.
Anyway, that is the thesis. It has worked for Cloudflare, because everything seems to work for Cloudflare, but not for Akamai.
Akamai is maybe seeing a slight sign of stabilization in their content delivery revenue which has narrowed to a -3% decline, but there has been no inflection at all in their security business, which has bounced around between 9-11% growth for two years.
I don’t know whether Akamai is just going to miss out on this opportunity, or if I’m just early and we aren’t going to see a lift until AI video becomes a thing. But I was not ready to give up on Akamai, especially after doing pretty well on the stock on my first go round. So I bought the stock in September at $105, thinking really, how much further could it fall?
Shortly after that Akamai did a pivot of sorts. On September 18th, about a week after I bought the stock, Akamai announced a large compute deal with Anthropic. The total commitment had Anthropic buying at least $11.6B of capacity from Akamai over 7 years.
This is not chump-change for Akamai. Akamai’s revenue run rate is a little over $4 billion a year
Sounds great right? Wrong!
Akamai’s stock has done just horribly since that time.
What the heck happened?
Not really knowing what I had just gotten myself into, I set out to find out. Shortly after the deal was announced I ran the numbers. At the time, there were only sparse details on the deal so I did the best I could, which is what I have reproduced below.
My interpretation of the deal with Anthropic was that it was going to be a positive for Akamai earnings, but just not right away. And markets don’t like lags.
For context, at the end of Q2 Akamai was guiding to a range of $6.40 – $7.05 per share of earnings for 2026. If you look at the impact to earnings in 2027, it is within guidance for 2026. So no real growth. It isn’t until 2028 and 2029 that the deal ramps and Akamai sees its earnings rise.
In fact the deal will likely depress earnings for at least a few quarters. You can see that my margin estimate from the Anthropic deal for 2027 is only about 5%. If you parsed out quarterly numbers, Q1 and Q2 of 2027 were likely to be even more negatively impacted by margins as Akamai built out their cloud ahead of any revenue. They would have to begin paying for colocation costs and increased headcount even before Anthropic started paying for capacity.
On top of that, as we saw with the release of Akamai’s Q3 results, big deals like this have delays. Which push out earnings even further.
Markets don’t like it when things get worse, even if the promise is that they will get better after that. There are just too many examples of things getting worse and then just getting even worser.
Markets also don’t like it when things get worse and it costs money to do it.
Akamai would have to pay to install the capacity it is selling to Anthropic upfront. They estimated back in September that they would have to add $5.5 billion of capex to deliver the capacity to Anthropic from the deal. Of that, $1.7 billion would be spent in 2026, before any revenue is generated, $3.1B in 2027, when very little money is being made and the rest in 2028.
The market was being asked to digest the financing risk of capex and the time risk of cash flow all at once. And give Akamai capital to do it.
It was a fairly straightforward exercise to go through Akamai’s cash and debt maturities over the next few years. And from that to conclude that they didn’t quite add up.
Even ignoring the later maturing convertibles, Akamai was short of about $1.4 billion of cash. And that is with full use of their outstanding revolver. If they wanted flexibility, you could raise the shortage to $2.4 billion.
You put all of this together and I think you can see why the stock price tanked.
Well what’s done is done. Akamai reported earnings on November 5th. They gave us more detail about the large Cloud Infrastructure Services (CIS) deals.
The most interesting information they gave was this slide:
The slides validated a lot of my original assumptions, but made me modify a few others.
The revenue from all four of Akamai’s CIS deals is $2.2 billion per year, versus the $2 billion I was using based on the Anthropic-only deal release. The 3 other deals that add up to about ~$0.5B are likely these one’s announced in the first half of the year:
Capex for the 4 deals is $6.8 billion instead of the $5.5 billion that I had been using.
The hardware depreciates over 7 years, not 5 years.
Some minor changes but it really didn’t make a big difference to the earnings per share.
We still have the same problem though. Akamai will not see much in the way of earnings growth in 2027, but they will in 2028 and 2029.
So it is kind of like a microcosm of what I was saying with AI in general. The upside here is pretty good. If earnings do hit $11+ in 2029, and the stock trades around 15x, it is a $160 stock. Maybe it even gets a higher multiple if my original thesis, that traffic growth will grow their legacy businesses and they can shake the label of being a commodity CDN provider, plays out.
What Akamai needed to do to stop the bleeding of the stock was figure out a way of leveraging that potential upside into the capital it needed to fund it.
Four days after earnings Akamai announced and then closed an offering for $2.2 billion of convertible senior notes.
If you read through the details of the notes, they have a couple of features worth pointing out.
First, these notes carry an interest rate of 1%. A very low rate. So just in terms of the cash interest costs of this financing, it is very cheap.
It is cheap because of the convertible feature. When these notes come due Akamai will pay the principal in cash. But they will pay the excess value, which will be determined by how much Akamai’s stock price exceeds the exercise price of the notes at the time of maturity, in stock or cash. The example below is with the 2031 notes:
In plain English, Akamai has structured participation in the upside.
I think it’s the upside that is the important feature of these notes. And it is something that we will see more of.
What’s more, Akamai bought back the upside they sold through a separately made hedge with Bank of America. They paid $250 million for a hedge that offsets dilution up to the bank warrant strikes. The note buyers get the upside, but Akamai paid cash to cancel most of the cost to shareholders. They paid for it of course, but not with dilution.
COREWEAVE STRUCTURING AN UPSIDE
So that’s Akamai and that is mostly the end of this post except that I want to relate it briefly to CoreWeave. Why? Because no one really cares what happens with Akamai but everyone seems to care what happens to CoreWeave. Yet there is an analogy there.
I listened to the CoreWeave call Tuesday night. This was their third quarter call, and because CoreWeave is being treated as a bellwether of AI health, I knew the call would be important.
It was one of those calls where immediately you feel like it’s a turning point of sorts.
The problem with CoreWeave is the problem with Akamai on steroids, and the problem I described at the start of this post. It has been trying to raise capital to fund an AI infrastructure build by offering investors no upside and participation in the downside.
Who wants that? I get a 7% coupon but if things go south in AI I’m hooped? It just doesn’t make sense. Especially with a levered, optionality rich bet like CoreWeave.
The whole reason to invest in a company like CoreWeave at all is because of its optionality. There’s no option with the debt.
CoreWeave can’t get quite the sort of deal that Akamai did. Akamai can price convertible notes at 1% because they have a growing $2.4 billion security business and a $1.5 billion cash flowing delivery business if the AI bet doesn’t work out. CoreWeave doesn’t have that to fall back on.
But just as Akamai found a way around this with its converts, CoreWeave seems to be finding a way around it with its customers.
CoreWeave said on their call that they would not be accessing the unsecured debt markets through at least 2027.
Instead, they are going to get money from those that benefit from their success. Who is ultimately going to benefit from CoreWeaves capacity? Its customers. So CoreWeave asks them to prepay, which many of them can.
You can call this circular financing if you want, but in my opinion it is just a logical outcome of the asset being financed. It is the outcome of trying to finance an asset with lots of upside and lots of downside. You need financiers that get the optionality of the upside.
That could be by offering participation in the share price, or by asking customers, who need the capacity to benefit, to pay upfront. Either way, the counterparty is invested in the upside. Which is the whole point with AI.
This rabbit hole starts from the comments of Elon Musk on the SpaceX conference call last Tuesday night.
Musk said that he expects to deliver 13 GW of data center compute, with 15-20 GW of power and cooling by the end of 2027. And that the compute could be as much as 20 GW.
This is a huge amount. Like way bigger then I would have imagined. Consider the following forecast from Bank of America.
The number Musk is saying includes capacity being added in 2026 of 2.5-3 GW. Which means that SpaceX would sit at about 10 GW, or 28%, of Bank of America’s total adds in 2027.
That is a really big bet for SpaceX. Its raises a bunch of questions such as how profitable it is going to be? What can Musk sell that compute at? Can Musk really raise the money he needs to do this? And so on.
But the biggest question it raised for me was: can we actually use all this supply?
There are three possibilities of what is going on in Musk’s head.
Musk has gone off the deep end ala Howard Hughes. He is in his own world and its simply not seeing things clearly
Musk is exaggerating significantly for effect as he is want to do
Musk sees something that is mostly being ignored by the market
Look, 1 and/or 2 are very possible. Don’t discount them.
But given that we can’t rigorously evaluate 1 or 2, our best approach for affirming or denying them is to evaluate #3. Does Musk see something the market isn’t?
That question led me to investigate just how much inference the expected level of compute would deliver.
We start with an estimate for 2027 capacity adds. I’m going with 25-35 GW.
The reason I am using 25-35GW rather than just the Bank of America bottom-up chip view of 35 GW is because I think the bottom up chip view is the most optimistic. It ignores all the things that could go wrong downstream of the produced chip and there are a lot of them. The most capacity we can add is the amount of chips we can produce. But it could be less.
Next we have to turn that into what additional inference capacity that total capacity translates to.
I assume that 55-60% of the 25GW-35GW goes to inference. The reason inference is only 55%-60% of the total is because compute is also used for training and reinforcement learning, and that will continue to take a portion of the total. To take an example from Musk himself, he said that roughly half of their 13 GW of additions would go towards training and RL for Grok and Cursor.
Now we are at 15-21 GW. But that number assumes that the compute is 100% utilized, which it won’t be.
The capacity has to be sized for peak and runs at maybe 60–70% average utilization
Which gives us ~10–14 GW of average inference load that will be added through to the end of 2027.
The next question is, what sources of demand will use that 10-14 GW of load?
To figure that out, we need to determine how much power it takes to do different tasks.
I decided to focus for now on one type of task. I started asking Claude about Muse, because I had already been wondering how much compute I might expect Muse to use as it scales. I asked how many tokens a user of Muse is likely to go through in a day and what would that be for 2.5M users, which is what Muse roughly has right now.
After some back and forth this is what it came up with.
These numbers are both large and not large.
On the not large side, Google has said that all their products combined use 107T tokens per day across billions of users. That is a lot more than the 2.5T tokens per day of the mid-case here.
On the “large” side, 2.5 trillion does seem like a big number. But what is a token really? It’s a word fragment. Does that big number actually mean anything?
What really matters is how tokens convert into power, and more specifically, into GPUs.
To give the conclusion up front, it turns out that to deliver the mid-case quantity of 2.5 trillion input tokens per day and 75B output tokens per day you would need 8,000-13,000 GB200 GPUs.
You can get to this result by breaking up the input tokens (which are usually called pre-fill) from the output tokens (called decode) and add them together to determine the compute, in either kW or GPUs, that is required.
Getting to that conclusion is worth stepping through because I learned a few things that I didn’t know in the process.
Let’s start by reviewing what input and output tokens are.
Input tokens are the like prep stage – you are loading the model, loading all the prior relevant context, loading the current prompt (or question) that you are asking the model to respond to. And then doing a bunch of math on it.
Output tokens are the answer the model is giving. One token at a time the model writes out (or draws a picture or a video or spreadsheet or whatever response is required) tokens that form the response. Again by doing math.
The key difference between input tokens and output tokens is that output tokens need to be done sequentially while input tokens can be done in parallel.
Every token needs two things done: weights read from memory and math done on them. For the input tokens, the memory read is shared across many tokens. There is one memory call and one big matrix multiplication and then you move onto next layer.
Output tokens read weights for one token, read the 42 GB of model weights (for Muse), do a little math and then move on to the next token. One token at a time.
Per token, the input and output tokens do the same math and take the same energy. But because input tokens can be done in parallel they don’t take nearly the GPU time as output tokens.
The reason is because the speed of the GPU is much faster than the speed of getting something from memory. A GB200 can do 5,000 trillion FLOPs of math a second, but it can only pull 8 TB a second out of memory.
So even though there are 2.5 trillion input tokens in our example, they actually take less GPUs than the much smaller 75 billion output tokens.
Here are the steps to go from tokens to GPUs on the input (prefill) side. It’s basically determined by how fast a GPU can do the math:
We assume that we are using the Muse Spark Open level of model, which is a 816B parameter model but where only 8 of the 256 experts fire per token, so each token touches 42B parameters.
Compute FLOPS are 2 x the parameters so you use 84 GFLOPS per token.
For the case of 2.5T input tokens, it is 2×10²³ FLOPs/day
A GB200 at ~40% utilization delivers ~8×10¹⁹ FLOPs/day – the 40% utilization is about the average for prefill.
The math on that equals 2,700 GPUs.
For the output (decode) side, the tokens are constrained by memory. To create output tokens there is a lot of back and forth from memory. Each time you calculate a new output token, you have to go to memory, grab all the stored context that is relevant to calculate that token, and only then do you get to do the math.
The bigger the context the more back and forth, which is why sometimes Claude will tell you to stop with your current session and start a new one – it is getting too expensive for it to keep reading all that context.
Thus to go from tokens to GPUs on the output side, we have to look at how long it takes to read in input tokens from memory and from that determine how many output tokens you can calculate per GPU per unit of time.
The determining factor here is how many GPUs you need to get the number of tokens through in the time required – in this case 75B tokens per day.
That is about 870K tokens/second. NVDA’s best case benchmarks run several hundred tokens/second for each GPU.
But for real-world long-context stuff it is going to be 2x-4x worse.
So if we assume we can run 150-300 tokens/second/GPU and then put a 1.5x to 2x peak to average buffer, you need 5,000-10,000 GPUs to run the decode step
As you can see, even though there are way, way more input tokens, it’s the output tokens that need more GPUs.
All of this was sort of an “aha moment” for me: that yes, the size of the model matters, but context size (so the size of all that history you want included in determining your output tokens, what is really your answer) may matter even more.
Anyway, between the prefill and decode tokens together you get to the answer of needing between 8,000 to 13,000 GPUs to run Muse for 2.5 million users. That’s the conclusion that matters for this post.
What does that mean in terms of power? A GB200 NVL72 rack draws ~130 kW for 72 GPUs, ~1.8 kW per GPU including CPUs and networking. With a 1.25 multiple of all the other power you need to power the GPU, that is 2.3 kW per GPU.
This translates into the following GPU capacity requirement, in MW, for the Low, Mid and High cases, for 2.5 million users:
Low: ~5 MW
Mid: ~20–30 MW
High: ~100–140 MW
Under these scenarios, it doesn’t seem like a lot of power is needed. An average mid-case user is using about 10W of power consumption. The high-end user is using about 50W.
As (another) aside, it is interesting to guess-estimate what this would cost Meta. We don’t know what it costs Meta to run their Muse model but we can get something in the ballpark by looking at what Meta sells their Muse Spark for through an API.
Meta charges $1.25 per million input tokens and $4.25 per million output tokens. If you apply that to the mid-case you get:
2.5T input tokens/day × $1.25/M = ~$3.1M/day
75B output tokens/day × $4.25/M = ~$0.3M/day
Total ≈ $3.4M/day, or about $1.2B a year at list price
You might notice something about these numbers. Meta charges a little over 3x for output tokens compared to input tokens. But I just went on a fairly long tangent about how much more compute an output token takes.
This is not an error, at least as far as I can tell. it’s how the industry prices. Input tokens are priced well above what they cost to process, and output tokens closer to (or, at long context, below) what they cost to produce.
This appears to be a real thing. Meta, and the industry, makes its margin on input tokens and sells output at below cost. Input heavy workloads like reading big documents makes Meta money. Output heavy, long context workloads, do not. I feel like there is more to this, but I’m not sure what. Yet another thread to pull on.
For now, going back to that number of about $1.2B per year, which includes Meta margin, you could assume the underlying cost to META is 1/5 to 1/3 of this, which would work out to $200M-$400M of cost for those 2.5M users at the mid-case.
You can get to similar numbers by looking at hardware. So take 10,000 GB200 GPUs that cost $60K per GPU all-in, so chip, server, networking, and their share of the infrastructure. That is $600M of capex, or over a useful life of 4-5 years, the depreciation is $120M to $150M. The power is $18M per year for the 25 MW case, and the facility operations and staff are another $60M-$70M. Which gives $200-$240M, in the same ballpark.
Again, the amounts we are talking about aren’t a lot in the grand scheme of things. Muse doesn’t take a lot of compute or cost a lot of money. At least right now.
But its not hard to come up with a much higher number for heavy users.
My “heavy user” example is interacting with Muse a modest 4 times a day and that takes about 50 W of average GPU load. But if you had, for example, a portion of users, maybe many million users, interacting with Muse or another agent many more times a day, you might need more like 1 kW (20x the high end) of average GPU load. If that were the case, it would only take about 10-14M of them to get to my estimated inference addition. If you had 20-28 million users at 10x my high end, that would also get you there.
But 10x and 20x is a lot right?
The thing about agents like Muse, Grok Bot and other assistants, is that getting to 10x or 20x your current usage is not as hard as it sounds because so much of what you do is cumulative ad recurring.
Which brings me to my own Muse usage. Which I actually don’t consider to be that much. In fact I use Claude a lot more than Muse.
Here is where I am with my Muse usage, as explained by Muse:
Maybe my high estimate is not high enough! In fact, Claude agreed and upped the number for the Muse cases to:
Low: ~8-12 MW
Mid: ~25–40 MW
High: ~120–170 MW
But it does all beg the question: why was my own usage so much higher then the generic cases we estimated based on tasks?
Best to ask the source. From Muse (with my underlining):
Almost none of it is your typing. Two drivers dominate. First, context re-processing — this main chat alone accounts for 288M of the 506M input (~57%), averaging ~600K input tokens per message, [which it actually breaks up into 3 200k pieces] because every reply re-reads the full standing context (system prompt, memory files, this long conversation) plus tool results. Second, the standing machinery: 726 distinct agents have run for you in total — The Brief, the idea-a-day, earnings checks, proactive research workers, browser tasks — most of them burning tokens on days you’re not even chatting.
This is an interesting statement. What is driving my compute demand is context and recurrence. The more Muse tries to integrate what it knows about me into its answers, and the more its trying to be helpful by working when I am not asking it to do things, the higher its compute demands are.
What is really interesting about this to me is that its not really about whether or not Muse is a frontier model or whether it is closed versus open-source or all the other arguments that get a lot of attention.
Its actually just that the amount of context you want to give matters a lot.
And there is math behind that. It turns out that while compute increases linearly with model size, it increases with the quadratic with context size. If you want a model to be very personalized, its going to cost you. Given that I think context is probably key for where we want to go with models, this seems like an important thing to recognize.
This is not to say that the model size doesn’t matter. Muse is not a frontier model. The model itself is mid-sized. If I was using a frontier model the power consumption for those tokens would be higher, by something like 2x to 5x.
Back to my own Muse usage. What sort of compute draw do I have? About 150W-250W. A lot more than our High case, but still not super high:
Bottomline: Today, I don’t maintain a 1 kW average use or even a 500 W average use on Muse. Though I might in another year if I keep adding tasks for Muse to do, and if Muse keeps adding features that make it easier to use (like the voice stuff they just put in).
But Muse also isn’t even the primary tool I use. At the moment I just use it like an assistant. I use Claude much more because it seems better for research. Yet in terms of compute, my Claude usage is actually less.
Claude is “cheap” because much of it is just one-off questions and answers. Where it can get expensive is when I have a thread that gets very long, because then its having to recall the entirety of the thread when it answers the next prompt.
I suspect I would be a much bigger than 1kW user with Claude if Claude remembered things in the way Muse does (by remember I mean have the standing context that gets loaded every time) and had standing agents proactively doing stuff. But it doesn’t, Claude remembers only a little from conversation to conversation and you’d have to set up agents to work for you offline. They likely do this because if they did all that Muse does it would cost them a fortune.
Of course, the entirely separate question, which I have no real thoughts on so far other than that Zuckerberg took years to monetize Facebook or Instagram, is that if Muse is very successful and can develop the kind of user base that Zuckerberg is aiming for (he said 1 billion users in a few (couple?) years, Meta is going to need a whole lot of compute and a whole lot of money to pay for that.
Though on the call Zuckerberg also said he wants to drop costs by 10x. So that is going play in the other direction.
But that’s not really what I am here to talk about. Like my other post, Muse is a starting off point of where things may be going, not a destination.
The bigger point here is that going through this exercise gives some context for that 10-14 GW of capacity this year and that 13 GW that Musk is talking about adding. It might still be baloney, much of what Musk says is, but its not inconceivable. 13 million people using agents drawing 1 kW on average per day, or 26 million drawing on 500W per day, doesn’t seem like an impossible ask.
And just thinking for a second beyond the personal assistant, there are actually a number of tasks that would, continuously, use >1kW of draw. Some of these, like the autonomous coding agents, are already well along their way.
From these higher end tasks and all the more mundane, lower compute but higher volume tasks, you can start coming up with a demand picture for the next year that seems reasonable. The table below roughly assumes agentic growth roughly halves from its 2026 pace, and that observed demand is held down by capacity limits.
What backs these numbers up? They are estimates, so it’s fair to ask whether anything we can actually observe backs them up.
Ramp, which sees the AI bills of around 70,000 US businesses, shows business token usage up roughly 30x over the past year, and nearly 3x in the last three months alone.
On OpenRouter, a marketplace developers use to route requests between models, agent traffic is up 34x since February and now outweighs human traffic almost eight to one. They also said “agentic traffic on its model-routing platform reached 132.6 trillion tokens in the week ending October 5th, nearly eight times the 17 trillion tokens attributed to human users.”
Anthropic’s own token demand, which has been the source of so much of the FUD in the space, seems to be accelerating as well.
There is also Notion’s tokens roughly doubling between July and October. Grok Imagine went from about 1.25 billion videos a month in January to over 2 billion by March. And then there is Muse, which has been off to a very fast start. Lots’s of different data points and they seem to all point to acceleration. And all of it is happening while every major provider says it’s short of capacity, so what we’re seeing is likely a floor on demand, not the full picture.
On the other side, we don’t really know how this all gets paid for yet. And we don’t really know just how much in the way of efficiency gains will offset this new demand.
A lot of questions, but that is to be expected when we are dealing with such a new technology where we are forced to project blindly into the future with very little precedent to back us up.
To bring this back to where I started, on Musk and his 13 GW. Can he actually deliver on his 13 GW of compute by year-end 2027. I have my doubts. But on the question of whether that much compute could actually be used? I’m a little more inclined to believe the demand will be there. Not convinced, but leaning a bit more in that direction then when I started down this rabbit hole.
Apart from the shitshow that is Coreweave, Oracle is probably the most maligned of the datacenter providers.
Oracle gets a lot of attention because it is the tweener. It sits in between the hyperscalers (Google, Amazon, and Microsoft) and the Neoclouds (Coreweave, Nebius, all those bitcoin miner conversions, etc).
Unlike the neoclouds, Oracle has businesses outside of the AI datacenter build. But unlike the hyperscalers, these are not insanely profitable businesses throwing off wads of cash that can be shovelled back into the buildout.
This all makes Oracle sort of an interesting case study of what the risk and reward of the AI buildout is.
One of the best parts of AI for me has been its ability to make financial models that I otherwise wouldn’t have been able to make. While making a model is not that hard, the math is easy, digging through all the filings and transcripts to gather all the assumptions and projections from management, analysts, competition and customers is a lot of work and usually requires shortcuts. But with AI, much of this is at your fingertips.
I did it for Oracle this weekend and it really didn’t take that long.
I’m going to step through it, starting with the assumptions. The guts of what I did was focused on one segment, called Oracle Cloud Infrastructure (OCI). I don’t think the rest of ORCL is all that important, certainly not to AI and it certainly isn’t going to determine the big question that I am interested in, which is under what circumstances Oracle is likely to succeed or fail.
The conclusions to watch out for are:
Oracle has a decent amount of debt and there is a reason that S&P cut Oracle to just one notch above junk in July
To complete the buildout of the announced capacity, Oracle will have to add to their debt beginning in about a year, and the amount is not insignificant
Over that same period, if the base case holds out, Oracle is going to go through a period of revenue growth and significant earnings growth. Yet because of the buildout of capacity, they will have negative free cash flow
The outcome of whether this all plays out positively or negatively to the business and the stock hinges mainly on demand for AI compute over the next couple years.
One last conclusion that is worth mentioning. Even though Oracle seems to be constantly in the headlines for their New Mexico datacenter, which is always under some sort of opposition, these very large 1 GW+ datacenters are really not anywhere near the whole story of the buildout. In fact, much of Oracle’s datacenter build is from small, unnamed capacity that we don’t even know where it is. I bring this up because, in general, we hear about the massive datacenters as being so central to the story. And they are to an extent, but many, smaller capacity additions are equally important.
Here’s the model.
Let’s start with the top line numbers and where they come from. I am going with this forecast for total revenue, updated by the $5B increase they expect now for 2027.
I am estimating the following prices for Oracle’s new datacenter capacity. These are flat number through the time frame because really, who knows what pricing will be. They are a little higher than what Oracle said at their Investor Day last year and inline with what you need to get their stated 35% margin on $30M/MW equipment costs.
Revenue per MW per quarter for Blackwell: $2.75/MW/Q or $3.67/GPU-hr
Revenue per MW per quarter for Rubin: $2.75/MW/Q – $4.58/GPU-hr (The shift to Rubin from Blackwell begins in the August 2027 quarter.)
I can tell you, these prices are well below spot prices today. The Google/SpaceX agreement, which is a short-term spot agreement from June, was priced at $11.50/GPU/hr.
I am assuming the following large datacenter buildout ramps:
One of the interesting things I learned is that these five sites don’t really comprise all of the capacity add that Oracle is doing. Oracle must have multiple smaller sites that they are building to get to the capacity adds they expect in 2027 and 2028. I say must, because Oracle, and no one really, details out every datacenter they have, so you just have to assume there are a bunch of small ones.
The model estimates 300 MW of capacity adds of these small, unnamed datacenters every quarter between now and F229. That works out to 2.7 GW of the total 7.7 GW of added capacity between now and then.
Revenue from this additional capacity is determined by prices that are largely fixed and we are assuming they are fixed roughly inline with the above pricing.
Margins on the new capacity being added depend on depreciation, power and any other operating costs. Oracle has pegged gross margins on their new capacity at 30-40%. I am using 35% at steady-state, increasing from 0% at the beginning of the ramp.
Apart from the AI datacenter business, the model assumes non-AI OCI growth of 35%, which is high, but it is what you have to assume to get to ORCL’s own guidance numbers.
Using the above assumptions about when Oracle’s AI datacenter capacity comes online, how much revenue it generates and the margins of that revenue, you can come up with an OCI segment model over the next 2 years that looks like this:
It’s a lot of rows and columns but you can focus on just a few of them. The six rows under “Cumulative MW delivered by site (since Aug-2026) are the new datacenter additions (in MW) coming online each quarter. The “Total new MW since Aug-2026” row is Oracle’s total capacity coming online.
The row labeled “Total OCI revenue ($B)” is how revenue for the segment grows as those datacenters come online.
Finally, the “Gross capex, modeled” is how much capital expenditures Oracle is ploughing in to bring the datacenters online each quarter.
And that’s it for the OCI segment. You can’t really draw conclusions from it on its own so now I have to take those results and integrate them into the rest of Oracle.
Oracle also has non-OCI businesses which are largely cash generating and slow growth. This is their software business (what most people know Oracle for), their hardware business and their services business.
Their software business is on-premises, not a cloud business. It is products like Oracle Database, Java SE, E-Business Suite and other older applications. Most of this is a license business where customers pay annual fees for updates, patches and support.
Oracle’s hardware business comes from a very old acquisition of Sun Microsystems in 2010. This includes racks and servers that run Oracle database, and an older server and storage rack business. This is a funny business that has actually been in decline for years but has started to grow more recently. I haven’t seen an explanation.
Oracle Services business is the fastest growing of the three legacy businesses. This is Oracle Consulting, which has professionals that come in and help you migrate to Oracle Database or use the other products. More recently it seems that some of the OCI integration and education gets uptake here, which may be why its growing a bit faster (~10%).
None of these businesses are forecast to grow all that much. They just generate consistent cash. I have growth from 0% to 8% for these segments. I am not spending nearly as much time figuring out the details here because these aren’t the businesses that are under the microscope.
Below I model out how these non-OCI businesses and the OCI business together impact Oracle’s cash position from quarter to quarter.
The inputs from the OCI model come into play in the “Incremental OCI” row, which is the additional revenue Oracle gets every quarter when a new datacenter comes online, and the “Capex” row, which is the gross capital expenditures Oracle has each quarter.
Oracle has a cash position today of $37B. That cash will pay for the buildout through the rest of their fiscal 2027. But come 2028, assuming they need to fund all gross capital expenditures with cash and want to maintain a balance of $15 billion or so, they will need to begin to take on new debt in fiscal 2028. Oracle will likely get ahead of this, and they have said they plan $20 billion of debt for 2027.
The row “Total New Debt” is showing how much additional debt Oracle has to add to complete the buildout to Q229. This debt accumulates every quarter and reaches $56 billion by F229.
This is a super important point to reflect on. Because $56 billion (second last cell from the bottom right) of more debt by Q229 is a lot! Unless…
The word “gross” in “gross capital expenditures” is important for watching how cash changes. Oracle’s capital expenditures do not necessarily line up with their cash usage.
The reason is that Oracle can finance those expenditures in ways that don’t require them to put up cash.
One way, which is in the above model in the row “Customer Prepayments” is through, well, customer prepayments. Oracle has said that they expect FY27 capex of $90–95 billion, with net cash capex (after customer prepayments) of no more than $70 billion.
Customer prepayments are lumpy but will average a little over $5B per quarter in F2027. I modeled them at $4.5B for the rest of the fiscal year. I don’t know what they will look like in FY28 and FY29 so I assumed they are the same.
The other way that Oracle can fund capex without cash, which is not in the above model, is with vendor financing. This is, quite honestly, the #1 scare tactic of bears right now. They point to Nvidia funding purchases of their own chips by their own customers. We see AMD and Broadcom doing it as well.
Multiple things can be true at once and if you don’t consider all the sides, it is easy to get lost in one narrative. Which is the case with vendor financing.
The nefarious narrative is simply that vendor financing is bad because it is usually hiding something. I mean, it’s true, most business-based revenue scams have some origin in financing revenue growth via your own balance sheet. It reminds me of banks allowing for the payback of unpayable loans with the proceeds from a new loan. Its not sustainable.
The less nefarious narrative is that this is everyone’s best interests. It is an acknowledgement that this is a very capital intensive business and if you want to participate you have to raise a large amount of capital before you can hope to see payback on it.
For a vendor like Nvidia to finance an investment in Nvidia GPUs and agree to participate in the return they believe it will generate instead of cash upfront is not necessarily nefarious. Especially if they are doing it in plain sight, which they are. If they believe that return is likely, which if you listen to Jensen Huang he clearly does, then it probably makes a lot of sense.
It certainly makes sense for Oracle. Without it, they wouldn’t be able to grow their datacenter business as fast and they would likely lose out to hyperscalers that can grow more easily.
Consider this exchange from Oracle’s F127 call in September:
Mark L. Moerdler
Bernstein Institutional Services LLC, Research Division
Ken, we’re going to be sorry to see you go, but I’m sure you have an amazing bucket list ahead of trips and things to do. So I’d like to focus in on the RPO. Can you give us more details on the drivers of the growth of RPO that is not requiring additional CapEx, the prepaid and bring your own chips? Is this AI labs? Is it semiconductor companies? Or is it that sovereign? In fact, we’ve not discussed sovereign cloud in a while, and we haven’t discussed sovereign AI at all. Can you explain how that side of OCI is going and how that will impact CapEx and margins?
Clayton M. Magouyrk
Chief Executive Officer & Director
Sure. So let me make sure we clarify one thing first, and then I’ll dive into the pieces. I didn’t say — and I don’t think myself nor Hilary said that it doesn’t require additional CapEx. We said it doesn’t require additional cash from Oracle, right? So the — I’m not an accountant, but I get to play one on TV sometimes. But the intention of what we’re trying to say is that while there clearly are capital expenditures, it does not require Oracle to go out and find additional cash to do it.
Now the question becomes, well, how do you do that? Well, we have a variety of different models. Sometimes it’s working with our suppliers through different financing arrangements that allows us to pay for the capacity as the customers pay us. That’s one mechanism. Another mechanism is that a customer says, “Hi, I’d like to pay for the hardware, but use your operational ability and your cloud infrastructure, technology assets in your data center to go out and actually turn that into an AI cluster.”
Magouyrk says that additional Capex “does not require additional cash from Oracle”. He then outlines the two ways – vendor financing and customer prepayments.
Now to be clear, he is not necessarily talking about all the capex in the model. Its not really clear at all how much of the F28 and F29 capex is going to come from customers or vendors. We can only guess. But some of it will.
Which leaves us to consider that Oracle’s cash debt needs will be something less than $50 billion. Maybe $30 billion? Maybe $40 billion? Tough to say.
How hard will it be for Oracle to do that?
That’s the big question that has the stock, and the stocks or Coreweave, Nebius and other even more precariously positioned datacenter plays, in volatile straits.
If you listen to the bear voices that were yelling this last week, you would conclude that financing is not going to happen.
But before I just assume that Oracle can’t possibly fund those requirements, I want to see how the income statement actually performs as these commitments begin to generate revenue.
So far, I’ve just focused on the free cash flow. And the easy tack, if you want to just make the bear case and leave it at that, is to stop right there.
But the more correct thing to do is to look at the whole picture. Which shows Oracle massively growing revenue and earnings over that time.
The relevant numbers are:
2027 EPS = $8.20 per share
2027 Revenue Growth = 38%
2028 EPS = $11.69 per share
2028 Revenue Growth = 45%
Yes, Oracle is committing to a large buildout in capacity. On the flipside, they already have revenue commitments for the space from Meta, Nvdia, OpenAI, xAI and even TikTok. And those commitments are going to put Oracle on one heck of a growth path.
The strange confluence that Oracle is going to see over the next few years is that they will transition to accelerating revenue growth, accelerating earnings, and increasingly negative free cash flow.
Here’s my take.
If we get to the end of Oracle’s fiscal 2027 and Oracle puts out the above 2027 numbers and they are forecasting numbers inline with what I have above for 2028, it is hard to imagine that A. Oracle won’t be able to secure the debt they need and B. that the stock will be at $130.
That’s the base case. Now we can look at some scenarios.
I have one big take-away from these assumptions. Nothing really matters all that much except for one thing: demand.
Here again it is worth revisiting the assumptions. Or at least the 10,000 foot view that instructs those assumptions. And the big assumption here is that demand for AI compute grows enough to cover the growth in capacity so that Oracle can bring on its capacity, have it used, and generate (at least) the margin it expects.
We can speculate this and postulate that and what about the CDS spread, and maybe the 10-year yield will rise and well you know the Anthropic IPO may not go off in November, and is OpenAI losing market share to Muse and blah, blah, blah but really, what really matters is demand. If AI demand growth meets the expectations of tech executives a year from now then it is impossible for me to imagine Oracle stock being lower than today, let alone at risk of some sort of debt default.
If demand growth expectations are wrong, then the bears are right and this is an overbuild that will materialize in the next year – and let’s be clear, I mean in the next year, not what is going to happen in 2030 or 2032 or some far off year where no one really knows and the market isn’t going to pretend they do – well if it all falls apart and Oracle is going to be tough sledding.