Reminiscences of a Stockblogger

Oct 26

Pulling on the Token Thread

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.

Oct 19

2 Comments

Framing Oracle

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.

That’s the frame.

Oct 11

2 Comments

On AI, Math and Cryptography

Because I haven’t written in this blog in a little over two years, and because AI has been my primary vehicle for investing for the last two years, I have a lot of thoughts about AI that need to be prefaced before I get to anything else.

Even though some of them may sound a little dated by now, well, I have to start (again) somewhere. And while I was going to talk about a couple of stocks at the end of this post, Akamai and Circle, both of which have not done well, and explain why I think that was, but the post got out of control long so I will leave that for next time. So instead, this is just a second, state-of-where-my-mind-is-at post, and I’ll talk about more specific stocks shortly.

I wanted to bring up Akamai and Circle because they have been AI related stocks that really haven’t worked. Which makes them outliers, at least for me. AI has been the only thing that has consistently worked for me for about the last 2 years.

When I have tried to wade into an non AI name, it has had about a 50/50 chance of blowing up in my face.

Just thinking of just the last few months, I have, at various times, gotten long Boston Scientific, Abivax, BuildersFirst Choice, and Amer Sports. All of these stocks have been somewhere between disasters and duds.

My non-AI loser stocks have definitely outnumbered the winners for some time now.

Fortunately, the AI names I have owned have mostly worked.

Yet I am always uneasy about the AI names. Not for the most commonly held reason though. My uneasiness is NOT that I buy into the commonly held bear case of the big AI bears, the guys like Michael Burry and Ed Zitron. I actually suspect they won’t be proven right for quite a while yet (they will eventually of course, because stocks always have a bear market).

Even though there are a lot of different ways of saying the commonly held AI bear case, I believe you can boil it down to a few points and one big unknown that determines their outcome. Those points being:

Spending is so big that it can’t possibly realize the return that it expects
GPUs depreciate too fast to have their investment realized
Creative financing mechanisms are being used to keep the boom going and those will collapse like they always do

Those three points gather their potency if one future development occurs: AI does not deliver on demand.

If AI delivers enough demand, then none of the above worries matter. Money will be made that will paper over shortcomings and the boom with continue.

For now, I think that demand will prevail and that the rally has legs. That is my tentative conclusion (always tentative because it depends on what is happening next and, I admit, before Muse I beginning to waffle that the doomers may be right).

But anyway, this is the common bear-side worry. But not the source of my own unease.

Instead, my uneasy feeling has more to do with how quickly it is getting better. And my inclination that this is just going to accelerate.

For a bit of background on where I am coming from…

In the fall of last year, I sort of decided that I was going to figure out how AI worked.

I had been in and out of different AI stocks through some of 2024 and 2025. But I never really had conviction. I was always torn between the “it’s a bubble” bear camp and the promise of productivity and profits. And so I was always selling out too soon.

I don’t know why it takes me so long to realize obvious things, but it was then that I realized the only way I was going to gain that conviction (or not) was if I figured out how the dumb thing worked.

I got this book called Why Machines Learn, that gave me a good headstart. I followed that up by going through AI papers, a lot of papers, starting with this list from AI researcher Ilya Shutskeyver, who contributed many papers to the subject and was a co-founder of OpenAI, of the 30 most important papers he thinks you need to read to understand AI.

These papers and others are all available for free and are very useful for understand the math and science that goes into AI and how we got to this point. Once you get through them you can tackle pretty much any new idea that comes out and have a good understanding of whether its meaningful or not. I’ve kept at it, reading about what the open labs are doing and, where possible because unfortunately it is all behind closed doors now, what OpenAI, Anthropic, Google and Meta are doing as well.

This learning process was useful for conviction, just as I hoped it would be. What has stuck with me the most is that this stuff is surprisingly simple. It’s all based on math that I learned in undergrad engineering. Today it involves matrices that are enormous, I mean they are hundreds of billions of parameters now, but the actual operations you are doing are not that different than what I learned in first and second year university.

I’m not the only one who has been blown away by the simplicity. Sutskever said himself that one of the reasons he got into the field was because when he read a couple AI papers after his under-grad and couldn’t believe how simple the math was compared to what he was doing with his other math classes and projects. It was a hint to him that this was onto something.

So that’s the first piece of my uneasiness. AI is pretty simple.

Second bit of background. One of the papers that Sutskever lists above is what is known as the Scaling Laws paper.

You hear a lot about scaling laws and what they mean for AI, from bulls and bears alike. What they are usually talking about is this paper called Scaling Laws for Neural Language Models. One of the writers of this paper is, interestingly enough, Dario Amodei, who is the guy in charge of Anthropic now.

This is one of the easier papers to read because it isn’t really exploring any new method or technique. It is just investigating what happens when you make models bigger.

It was written in 2020. It is hardly ancient history.

It didn’t really set out with the expectation to find out what it found. In fact, it was sort of assumed by everyone in AI at the time that eventually as models got bigger things would break down. If you go through the previous 20 or 30 years of AI research, there is this sort of underlying assumption that you can’t just keep making things bigger.

Which sort of goes with the territory. Science is all about simplifying things down through clever solutions. Brute force is frowned upon.

With that expectation, the big discovery of the scaling laws paper was that no, that’s not right. The expected wall doesn’t happen and, in fact, model performance seems to follow a remarkably consistent pattern, called a power law, whereby output gets better regardless of model size.

A power law is just a really simple equation that says that when you make some quantity, say x, bigger, the resulting y scales by a much larger (or, depending on the exponent, smaller) amount. It looks like this:

When they found that models followed a power law is was an aha moment. It meant that A. you could basically extrapolate how well a model would perform as it got bigger. And B. It followed a law where outputs got big much faster then the inputs, which tends to be profitable over time.

The paper identified three key inputs by which model performance scaled: the number of parameters (ie. How big the matrix is), the size of the training dataset, and the amount of compute you use.

They found that these relationships held across more than seven orders of magnitude. Again, no sign of anything breaking down or even hinting that it would break down across a really wide range of sizes.

We are now six years past when that paper was written and so far the power law still hasn’t broken down.

Maybe we shouldn’t expect it to. Or at least, maybe we will be surprised with how big it can get before it does.

Power laws are everywhere in nature. Earthquakes, animal sizes, tree networks, blood vessels, river networks, there is just a lot of nature that follows a relationship of power laws.

Which is to say that when you start thinking about all this you start to wonder if we (being humanity that is) really stumbled onto something here. We’ve tapped into something very simple but that scales. Which certainly makes you go hmmm.

None of this is to say anything is for certain about any particular outcome for AI. Nothing is.

But when you put together the simplicity and the scaling I think we are more likely to be surprised by how quickly AI improves, and as it improves creates more use cases, and with that creates more demand, then we are likely to be surprised that its been a big capital spending black hole that has no use.

The problem with the internet, which everyone likes to compare as the example of a bubble that popped, is that the technology didn’t move fast enough to keep up with the infrastructure being built. If we could have had a Blackberry (the phone, not the pager) in 1998, an iPhone in 2001 and streaming Netflix by 2004, maybe the bubble wouldn’t have burst? We just would have just kept laying fiber and building cell towers.

All of this is by way of background to understand where my uneasiness lies. My uneasy feeling is about how quickly AI is getting better. Which brings me to why I was reminded of that this week.

It started with this sort of cryptic (yes, pun) tweet:

Who is toptickcrypto? It has been a long time and I don’t remember who exactly toptickcrypto is the monikor for anymore but I do remember he was a fund manager, maybe he still is, that he got into crypto in like 2018 or something, and he knows a lot about stocks and crypto.

Him not feeling so good made me want to investigate what he was feeling not good about. Which took me down the rabbit hole of AI models and math. And it all was a little unnerving.

It all starts with this release by OpenAI.

OpenAI published 722 math papers that together solved 372 math problems that had previously been unsolvable. They did it with an unreleased version of their model.

So that’s pretty crazy. And while I see an instance or two of people refuting the solutions, I see much more acceptance admitting what OpenAI did was real.

But beyond that, what Toptickcrypto is pointing out, and what makes him as a bitcoin/crypto guy not feel so good, is the rather ominous omission of those 722 problems: that 0 of them are cryptographic.

Cryptographic means solving a problem that deals with encryption techniques (you know, those things we depend on for literally EVERYTHING on the internet).

Which of course makes one wonder if A. they just happened to decide not to tackle cryptography or B. the cryptographic results are a little scary (ie. good at breaking encryption) and they don’t dare publish them or everyone would freak out.

This was not unnoticed in the crypto world, where there was some freaking out done.

Let me just preface this by saying that crypto, as the name implies, uses cryptography. Also, because of the way crypto is structured, crypto is probably the #1 most vulnerable place to be hacked if cryptography techniques are hacked.

To point to just one of the posts from the community, from Vitalik Buterin, the guy who founded Ethereum and therefore should be taken seriously:

This isn’t the whole tweet. The whole tweet is worth reading because Vitalik has put a lot of thought into this. I have two thoughts on his thought.

First, saying “I don’t recommend anyone scramble to move their funds” is what every bank manager says before the bank run. I’m not saying there is going to be bank run, I’m just saying this isn’t something you want to be saying.

Second, he is admitting that AI is making major steps in math, and that it is unlikely that those steps aren’t happening in the area of cryptography. And based on that, his solution for you, for now at least, is to start moving your crypto to wallets that haven’t had their public key exposed (it turns out that if you have never sent money from a crypto wallet you have never shared your public key and thus it can’t be hacked). Which again, is a bit like saying, take your money from the bank and put it under your floorboard and you’ll be okay.

I’m going to spend just a minute explaining how this all works and why this could eventually be a problem.

First, in crypto, pretty much all crypto, you have a public key and a private key. Both of these are just very long strings of numbers.

The public key is what you send out to the world to transact. The private key is what you would never tell anyone about, and ideally you would keep it locked in a bank vault somewhere. If someone knows your private key, they essentially own all the crypto in your wallet.

The path between the two looks like this:

private key ──(step 1: elliptic-curve math)──▶ public key ──(step 2: hash)──▶ address

The worry is that first “elliptic-curve math” step. The real problem with that step is the “math” part.

I haven’t really bothered to understand exactly what elliptic curve math is, but I can tell you a few things about it.

First, it’s a nifty technique whereby if you go in one direction (calculating the number that is your public key from the number that is your private key) it is relatively easy to do. But if you go the other way, calculating your private key from your public key, it becomes very, very hard. So hard that no one has ever figured out how to do it in a way that wouldn’t take many, many, many years.

However, it is not necessarily impossible. No one has ever proved that going backwards in elliptic-curve math is theoretically impossible. It just has been proven over decades of use that no one has been able to hack it, which means it’s practically impossible.

That is an important distinction, and suddenly quite a relevant one. Because yes, we have something that no human being has figured out a shortcut around. But AI is figuring out all kinds of whacky shortcuts.

Last detour, I promise.

When I was looking into all this I stumbled on this blog post, which talks about the mathocalypse. Anyway, the author of the blog’s wife is a complexity theorist, which means she deals with really hard math. His wife was reviewing one of the AI proofs from OpenAI. And she had this to say about it:

So there are a couple of ways you can take this.

First, you can take it in the mainstream “AI-isn’t-all-it’s-cracked-up-to-be-and-makes-things-up-all-the-time” sort of way. If you did, you would conclude AI is clearly spitting out gobbledygook and we have nothing to fear with this whole math proof stuff.

Or… the other way to take it is a little more seriously. What if AI is simply making connections that are so far from being obvious to us that they sound like gobbledygook (to an expert in the field, keep in mind) but are really fundamental to what you need to know to solve the problem?

Again, you sort of go hmmm. Can’t know for sure but it is alarming.

Of course, all this strikes me as something that could be bad for crypto at some point. And particularly bad for companies like Coinbase or Robinhood. The idea that everyone in crypto is not going to be able to transact without exposing themselves is not a good scene.

But not just yet. None of this matters just yet because fortunately OpenAI has tackled 700+ math problems but is leaving cryptography alone…. riiight. But most of crypto was up on Friday so it is true to say that no one cares about this yet. Citrini, the research substack, had this big piece on AI agents using stablecoins on Thursday, which helped crypto and actually makes a lot of sense and should send the crypto space much higher as long as (or until?) AI doesn’t cut it off at the knees.

The bigger thing for the plain vanilla common stocks I own is why this is a great example of where my uneasiness lies.

What I have been trying to lay out here is why things are progressing very quickly, why things are likely to continue to progress quickly, and from that, why the proper conclusion to draw is not the bear case that its all going to fall apart because of lack of demand, but instead that the real bear case is that at some point AI could break something in bad way.

There is no immediate action to take from that conclusion. In fact, everything I am saying is actually very bullish on AI as long as and until something bad happens. I think token usage will just keep rising, use cases that we haven’t even thought will keep materializing and Ed Zitron, Michael Burry and Paul Kedrosky will keep yelling at the sky.

But I think you also need to stay vigilant here, because we have quite possibly latched onto something that is lightening in a bottle. And I don’t think any of us can say with certainly where it is going to take us.

Sep 30

7 Comments

WHY I THINK META MUSE COULD CREATE THE NEXT LEG FOR THE AI STOCKS

I am back posting. I have been gone for 2 and a half years but have not be gone from investing. Or from writing, I just haven’t been putting it up on a blog. I’ll maybe talk about that at some point, but maybe not. I mean who really cares right? Live in the present.

As for the present, I have been using Meta Muse for 5 days now and I think this is could be the next game changer for AI.

Which is good. Because I have been a little nervous about what will come next with AI. And that if nothing did come next, that the market may have been in for a fall.

Let’s step back in time and review. In January, it became pretty clear that the whole AI-for-coding thing was going to take off. AI was naturally good at coding because it is, after all, a language model, and coding is pretty much the easiest language to learn, at least in terms of definite structure.

And that happened. It was a little delayed is all.

During the first days of the Iran war AI stocks sold off along with everything else. But even then it felt like they were just biding their time, waiting for a pause in the war so that they could go on a run as coding took off. Which is what happened in April.

I remember what I was thinking in March. That period was a tug-a-war for me, trying to answer this question: why is the market not going down more with the Iran war a clear disaster?

I had two competing thoughts at the time.

One was the recognition of the old maxim (well, at least my old maxim) that goes something like this: “the market doesn’t care about bad things until it can’t ignore them any longer”.

This is something I have noticed over time. It is basically saying that markets don’t actually like to go down. So they don’t go down unless they have to.

However, there is a flip side to that maxim. Just because nothing really bad has happened to the market yet, that doesn’t mean it won’t. The market may still go off the cliff if things get just a little bit worse.

To that very worrisome possibility, I weighed the much more positive possibility that AI coding agents were just that big of a deal.

It was possible that the market was strong in the face of war and high oil prices simply because it saw that coding agents were that big of a deal. War wasn’t going to stop them. Inflation wasn’t going to stop them. And that you couldn’t push stocks down with them coming on the horizon.

As it turned out, it was this second possibility that was correct. Coding agents were a big deal. The market was right not to allow itself to be pushed down.

When the Iran war eased, even just at the margins, the result was a steep bull market that lasted through the spring and into the summer before it hit a wall.

What happened at that point, and caused the pause that we are still in today, is a recognition that coding can only take AI so far.

You can replace programmers with AI. You can generate more code faster. And for a subset of the population (ie programmers and people that work with them) this can be very meaningful.

It is very meaningful to the programmers that lose their job.

It is meaningful to the companies that gain in productivity from being able to generate more code faster.

It is even meaningful to software companies, as we see now that the doomer-ism that software was about to be obsolete was wrong. That, in fact, some software was growing faster then ever, and that most software companies were more profitable then ever, because they no longer had to couple growth with more programmers.

But AI-coding agents weren’t meaningful to most people. AI had disrupted a segment. But what would happen when that segment, which also just happened to be the fastest adopters of AI (programmers love technology!), became saturated?

That question is what got me thinking about what comes next? It is something I have thought about all through the last month.

It is a big question, much bigger then it seems at a glance. To understand why “something coming next” is really, really important, consider the AI build out that is upon us…

WHY WHAT COMES NEXT, MATTERS MOST

So… the thing about the AI buildout that makes it matter really amounts to this: It is TRULY MASSIVE.

AI capacity is usually measured in GW of power consumption. Power is directly related to how many GPUs and CPUs you can run. Measuring power is the same as measuring compute capacity.

There are different ways of defining what AI capacity is. I am going to describe it as usable AI capacity, which would be what you call “frontier-grade”. Frontier grade means you could use that capacity to train or serve a big frontier model on it, just so we can put a consistent set of numbers on it.

AI capacity is expected to be about 15 GW by the end of 2026.

It is expected to rise to 45 to 55 GW in 2027.

Planned datacenter capacity is even higher. We are up to 210 GW (!!) of planned capacity that will come out in maybe the next 5 years or so.

So the thing is, these numbers are essentially insane. Consider that Alberta’s total electrical generation capacity is like 12 GW.

We are increasing AI datacenter capacity by two Albertas next year. And the pipeline is for like 12-15 more Alberta’s.

Apart from the rather unreal size of what we are talking about, what made me even more uncomfortable in the summer was that it seemed like wealready had enough capacity for what we were doing with AI right now.

I use Claude regularly. I code with it. I use it in Excel. I ask it a bazillion questions, mostly about stocks.

I’ve never had Claude say, “Woah, slow down there buddy, I can’t handle all this right now, I’m a little overloaded”.

I’m simplifying this a little too much of course. There are lots of strings Anthropic can pull to manage usage without it becoming apparent to a user like me. But still, if we were really, really short of compute, I would sort of expect Claude “brownouts” on occasion, just like when the electrical grid gets overwhelmed.

That isn’t happening. Even at the margins.

Where I’m going with this is that if this isn’t happening now, and we are adding two Alberta’s of compute next year, well you start to wonder if we are going to tip over into a glut of AI capacity at some point.

Which wouldn’t be good for stocks.

AI IS THE MARKET?

There are two things that the casual observer of the stock market is not aware of right now:

Just how many stocks have participated in the AI binge
Just how badly stocks that are not participating in the AI binge are doing

On point #1, do you realize that companies like Enbridge, TC Pipelines and AltaGas trade at price to earnings multiples today that are about 50% higher then they were 4 years ago?

It’s true. I spent a bunch of time on this a few weeks back. These stocks used to trade at 4-6x EV/EBITDA. Now its like 8-10x.

While it is always impossible to separate out what specific reason a stock is trading higher, and you can argue that with these pipeline stocks maybe it’s the Carney government and their pro-development stance, or maybe a decline in hatred of oil and gas, and yeah, okay I concede some of that might be it.

But if you look at the charts of these stocks, they all really started going up with the AI boom.

The reason, I think, was simply the recognition that more AI means more power (ie. That 10 Albertas in the next few years) and more power means growth in moving natural gas (and a lesser extent oil) and so these companies get higher valuations because they aren’t just stagnant dividend paying pipeline stocks, they are sort of growth stocks now. If a pipeline stock is actually allowed to be called that?

That’s just one example. But it holds broadly. There are a lot of stocks and sectors that have been buoyed by AI, well beyond a bunch of chip stocks.

On to Point 2. The market is pretty close to all-time highs. But you’d never know it by looking at the average non-AI related stock. In fact, 52-week lows on the NYSE are skyrocketing and 52-week highs plummeting.

There are a lot of stocks not doing very well at all. Don’t get this wrong, I’m not making the argument that the market is a disaster. I’m just saying that, if it weren’t for AI, it probably would be doing pretty blah.

The overarching point I am making here is that if it is true that AI is being overbuilt in the short term, the market could be in some trouble.

Which brings me to Meta Muse.

MUSE DOES EVERYTHING YOU COULD ALREADY DO WITH AI, BUT BETTER (MAYBE MUCH BETTER)

Look, I’m not going to pretend that I am an AI expert with my finger on the pulse. But I do keep pretty close tabs on what’s out there, I use it every day, and so I think I have a (slightly) better than average grasp on the state of the industry.

And if I’m right, then Meta Muse is something new and different.

This isn’t immediately obvious. Muse doesn’t necessarily do anything new. I haven’t found something with Muse where I’m like wow, I couldn’t do that before.

But what Muse does, is it does things much, MUCH more easily then you could have done them before.

The product is billed as an assistant. And that is truly what it is. It is like having someone to do and keep track of all the things that are a hassle for you to do. And to do the things you never did do because you didn’t have the time to do them.

You couldn’t really say that about Claude or ChatGPT. Yes, you can get Claude to do anything, but half the time you are stuck with a python file or installing a C++ compiler. And most people don’t want to go there.

Here are some of the things with Muse I have done in the past 48 hours:

Had it build a process whereby every Friday it scans all the local flyers for certain groceries that we buy on a regular basis. It then finds the best deals on each of them and sends an email with the list to my wife, mom and myself.
Recreate (I already had something like this in Claude) a Daily brief newsletter in HTML format that comes out every day and summarizes all the AI content on my AI X.com list that I have curated
Had it remind me of a bunch of tasks I keep forgetting about such as replacing a light, watching a particular video, doing a warranty call, replacing the car key battery
Summarize a livestream on stocks that occurs every Friday afternoon. I also had it schedule in summarizing future livestreams for me, which it will do every Friday and send me an HTML of the summary
Helped me haggle with SiriusXM to reduce our subscription price by about 65%. Also add a reminder of when this deal ends in a year so I don’t start paying full price.
Build a Tetris game
Create a short newsletter-style PDF of the Akamai/Anthropic deal based on my notes, model and discussion I had with Claude on the deal
Add weekly reminders to take out recycling/green bins/waste at a particular time
Check Calgary to Toronto flights every week for each airline to look for a deal and send an email to us with what they find each week and whether they think there is a deal worth acting on.

There are a dozen more I am mulling around in my head. All of them could have been with Claude. But Muse made them so easy to do that, unlike Claude, I actually did them.

It was literally like a few sentence description for the grocery flyer check and it was done in a few minutes. It built the Tetris game, which plays just like Tetris, in 3 minutes. The AI Daily newsletter was something I did with Claude but I had to run .bat files every day to actually generate and so I only got around to it once or twice a week, whereas with Muse it just builds and creates the newsletter every day and I have it there every morning with my coffee and I don’t have to do anything to get it.

The things that Muse has that make it more useful are, I think, twofold.

One, with Muse you get your own virtual computer with memory and compute. It doesn’t need to save things on your computer. It isn’t always asking to access your computer or install something on it. And it is set up to remember stuff for you.

That last point is really important I think. Muse can be your memory, just like an assistant would be. And it can be long-term memory. Muse is going to remind me about the SiriusXM renewal in 12 months.

Two, Muse is a chat, so you have it up on your computer and phone and its just like you are talking to your assistant.

Now of course, the worry here, which is already all over the news, is privacy.

Do you want to give Muse your data?

The thing is, none of the tasks I listed above actually had me give Muse any of my data. In fact, the only task where it would have been helpful was in haggling with SiriusXM, because if I was willing to give Muse my name, phone number, address, it would have just haggled with the online Sirius assistant, who I think it just a bot itself, for me. Because I wasn’t, instead I just copied the prompts from Sirius into Muse and asked Muse how I should respond. It worked.

But other then that, Muse doesn’t really have to access to anything. In fact, because you get this little virtual computer in the cloud, it doesn’t even need to access your computer at all (whereas Claude or GPT did for some tasks).

Of course as I have it do more things, it will learn more about me. It already knows some things about the groceries we buy or the flight we take most often. Just like any assistant would.

But these stories I’m reading about it wanting your bank info or your email access, well… you can certainly do that, and maybe at some point everyone will do that, but it is far from necessary to get value out it.

WHY ANY OF THIS MATTERS

What all of this does is make me think that maybe we aren’t about to run into an AI glut.

I am starting to warm up to the idea that Muse ushers in the next leg of AI.

I am certainly not the only one that sees what Muse can do and, more importantly, how easy it can do it. While AI was basically something that was super useful for programmers, researchers and people that liked to just ask lots and lots of questions, now its useful for, as Chris Camillo calls it, “the normies”.

Normies would define people who don’t really care about AI one way or another but if you can tell them where they buy coffee cheaply or alert them to a deal on flights and do it without them having to build and run a python script, and do it for free, they are happy to take advantage of it.

Muse isn’t perfect. It fails in some tasks and there are certainly some people out there that are gleefully pointing out those failures.

But I think that misses the point. Muse is just going to get better. If AI’s trajectory so far is any indication, it is going to get better quickly. And if I’m right about how this is going to open up a whole new, and larger, addressable market of people, there are going to be other Muses that come around. There will be an Anthropic Muse and an OpenAI Muse and a Google Muse in short order.

With coding agents, you had quick adoption of what is really a pretty small group of people – programmers.

With Muse, and whatever other coding agents are on the horizon, I am sure adoption will be slower. Most people won’t use it. But the pool of potential adopters is now essentially anyone that uses Facebook. So its really, really big.

Muse is free. Which gives you this massive pool of people that can afford it, even if only a small fraction of them are willing to be early adopters.

The other consideration, insofar as what this means for AI demand, is that about half the tasks I’m getting Muse to do are recurring.

With Claude or ChatGPT, it is all one-offs. Ask this, ask that. With Muse, much of what it is really useful doing is reminding, checking or doing something over and over again – every day or every week or every second Thursday.

That is real recurring usage. Which is additive.

And for now, the trend is your friend.

Finally, Meta announced just today that they have Muse for small businesses. Which makes a lot of sense. And given some of the things I am having it do for me, it will clearly be helpful to small businesses, especially if they can’t afford that assistant employee.

You add all this together and you start to see a picture of how this could play out. If Muse and its soon-to-come descendants gain traction, usage is going to increase, maybe a lot. Agents use A LOT more tokens then people.

And that, is where I was going with all this.

Maybe we aren’t about to hit an AI supply wall. Maybe things are going to be okay. And maybe the AI trade has yet another leg higher in the near future?