Milk Road AI Sep 17, 2026 38m 29m saved
With Martin, an analyst at Milk Road who runs the firm's macro index and publishes his own portfolio
Neocloud companies borrow to buy GPUs, and CoreWeave alone carries more than $30 billion of debt and asked for another $3 billion the morning this episode was recorded. The Fed had raised rates the day before, with the median projection pointing to one more hike this year.
The obvious conclusion is that the whole category is impaired. Martin, a Milk Road analyst, argues the opposite for one of them, on a basis that does not appear in any earnings line.
"We might have enough GPUs, we might have enough demand, we just do not have enough power."
Martin holds Galaxy as his largest position and was researching whether it had cleared ERCOT, the Texas grid operator, when he found a competitor's approval instead. He scored the company on Milk Road's research framework, got a buy, and is still waiting.
The full episode is covered here so you can skip it. 38 minutes of audio, 9 minutes of reading.
Here are the 10 arguments that matter.
Key Takeaways
Rate hikes have historically produced negative S&P returns at one, two and three months, but higher prices 12 months later in every case except 2022
The Fed's median projection implies one more hike this year, which Martin treats as the base case for any stock decision
CoreWeave carries more than $30 billion of debt and asked for $3 billion more, while simultaneously locking in contracts at $40 million per megawatt
Iron received 2 gigawatts of base-load classification from ERCOT, which Martin says makes energization close to certain
The thesis is that the market prices only Iron's contracted book, missing 5 gigawatts of secured power that appears in no financial statement
The bear case is dilution: the company reportedly needs about $25 billion next year, and Martin models 20%
1. What Hikes Do to Stocks
Martin opened on a chart of S&P 500 returns following the start of previous hiking cycles, naming 1994, 1997, 1999, 2004, 2015 and 2022. The short-horizon returns were negative in nearly all of them. The host described the chart for listeners: 1997 was the standout in which the index only went up, and 2022 wavered before ending lower a year later. Martin's reading is that the horizon decides the answer.
"The message is that in the short term, yes, there, like, rate hikes are, you know, a headwind, but you should probably, like, zoom out and look at 12 months or plus, and then it doesn't really matter, where typically it ends up being higher 12 months later, except for 2022." — Martin
He also noted the market took the hike well, on the reading that the Fed would not have raised if it saw cracks under the surface. Asked whether the move was what Dario's essay had been front-running, he said the two are unrelated.
2. One More Hike This Year
Rates sit in the 3.75% to 4% range, and Martin said the median projection is one hike higher for this year, with 2027 either flat or up again. He treats that as the planning assumption rather than a forecast, and said it would change quickly if the data forced the Fed to support growth.
"So, you know, the message from the Fed is we are probably going to stay higher for longer. So we are not going to get, you know, cheap money anytime soon" — Martin
"obviously hiking rates is not like super constructive for the whole bull market because, basically, the price of the money, the costs, are going to be higher." — Martin
He said hyperscalers are insulated because they hold enough cash not to borrow.
3. Why Neoclouds Are Exposed
Martin defined the business for listeners: neoclouds buy GPUs and rent them out, which means buying the land, the buildings, the interconnections and the liquid cooling before any revenue arrives. That makes them dependent on debt markets at exactly the moment debt gets more expensive.
"CoreWeave is the best example, like, they have over 30 billion in debt, and actually today they announced that they are looking for additional 3 billion today." — Martin
The host recalled an earlier episode of the show that laid out a five-year path to profitability for these companies, running to 2031, and asked whether higher rates break that schedule. Martin's answer was that credit ratings are improving as cash flow arrives, so a company that had to issue at 10% might now fund at 8%.
"It's not all a promise, like their facilities, everything that they do already creates cash. So they're not, they're not, like, horribly unhealthy businesses with zero cash coming in." — Martin
4. GPU Rental Prices Rise
The offsetting news is on the revenue side. Martin said CoreWeave announced contracts at $40 million per megawatt the same morning, and that Nebius has also said it is raising GPU rental prices. He reads the two together as demand outrunning supply.
"So the whole neoclouds market is actually seeing very positive, the developments in terms of demand is simply much higher than the supply, and therefore they can raise the prices of their businesses." — Martin
5. Residual Value of GPUs
The host set up the standard assumption: GPUs have a five-year life and are worthless afterward. Martin's earlier work argued that five years gets a neocloud to break even, which makes everything after that the source of returns. He said current data shows older GPUs still renting.
"If you believe that they will be able to rent out those, let's say, out-of-life GPUs, then it's insanely profitable business, but if not, they are cooked." — Martin
"You might not need, like, the newest chips for, like, everything, right? So you might still be fine having access to all the GPUs because it might just, you know, be enough, good enough for your needs." — Martin
6. ERCOT and 2 Gigawatts
Martin was checking whether Galaxy, his largest holding, had been cleared by ERCOT to connect to the Texas grid when he found that Iron had 2 gigawatts classified as base load. He explained what that classification means and why it changed his view of the sector.
"Now base load means that ERCOT already knows that you want access to that capacity, 2 GW, and if you meet all the requirements, it's guaranteed you will get it." — Martin
"We might have enough GPUs, we might have enough demand, we just do not have enough power." — Martin
The host said the firm's other analyst holds the same thesis, that compute without power cannot be brought online.
7. Three Different Models
Martin's argument is that three companies treated by the market as one category run different businesses. Iron owns the land, the buildings and the interconnections and is now adding GPUs and selling bare metal. CoreWeave owns no property at all and sells a training operations layer, software that keeps long training runs from failing, which is why the large labs use it. Nebius is closer to a general cloud provider, with compliance, regional servers in Frankfurt and APIs onto models such as Llama or DeepSeek, and it intends to become asset light by running other people's data centers.
"But I like the downside here because they are not just, you know, a kind of software company or a service company, because they own the land. They own the buildings. They own the interconnections, and that's what gives that company kind of a floor value" — Martin
"So you can see that they are all, like, a different business, like, they seem like the same one, that they are all selling GPUs, but I would actually argue that their businesses are very different." — Martin
8. The Iron Thesis
Martin ran the company through Milk Road's research framework and got 68 out of 100, which reads as a buy. His one-line version is that valuation captures the contracted book and nothing else, while the power position sits outside the financials entirely. On contracted business alone he said a price target lands near $50 against a share price around $40, which is not much upside; the secured megawatts are the part he thinks gets priced eventually. He put the market capitalization around $17 billion, and said Galaxy is the second-largest holder of secured power at 1.6 gigawatts.
"I think the market prices Iron at what is already contracted on the book, but I think the market is missing that they have secured access to 5 gigawatt of energy" — Martin
He added a second leg: the company is building four buildings for Microsoft, one of which was approved and has been operating since August, generating what he put at $1 billion of revenue, with all four expected online by year end.
9. The Dilution Risk
The bear case Martin volunteered is the funding requirement, which he said the team has put at roughly $25 billion next year.
"Now, the bear case was something that people are afraid of, and I think rightfully so, is it going to survive all the dilution, because they need a lot of money." — Martin
He models 20% dilution as his base case and said a higher number would erode the upside enough to change the conclusion. He also uses $30 million per megawatt in his model rather than the $40 million CoreWeave is contracting at, on the grounds that Iron has no software layer yet, though it has acquired a company to build one.
10. Why He Is Waiting
Asked why he is not buying now, Martin gave two reasons: he does not know whether markets rise from here with another hike likely, and he started analyzing this company this week. He also described a portfolio move made a fortnight ago, prompted by Milk Road's own macro index turning down.
"I sold, you know, 10 or 12% of my portfolio like two weeks ago, and the reason is that I was way too much allocated in the market, and I had only 4% in cash, which made me feel pretty uncomfortable about the markets and what is going on, because I felt like I have no edge." — Martin
"I am not concerned about a bear market. I think that we are still gonna be higher in couple months, but I wanted to increase my cash position." — Martin
The host explained the index for listeners: more than 20 macro metrics combined into one chart and compared against the S&P, designed to flash before the market moves. It has just gone red for the first time since the Iran war scare in February and March.
Bonus Insights
The host's closing question was political: the administration wants a hot economy going into the midterms, so what is left to play if rates are rising instead. Martin said Trump nominated his own Fed chair, that Kevin Warsh appears to be working with him, and that Trump is now posting that rates should be 1% after a hike was delivered. He said he could not reconcile the two, raised the possibility that direct payments matter more to the administration than the market does, and predicted the period will be looked back on with some amusement.
The host closed on the first chart of the episode: if the pattern holds and markets are higher 12 months after a hiking cycle begins, a drawdown over the coming weeks is a buying window rather than a reason to leave.
Martin's bottom line is that the neocloud trade has split into two questions — who can fund the next build, and who already controls the power — and that only the first one is in the prices.
Products, Companies & Tools Mentioned
Iron (The company Martin scored 68 out of 100; he says the market prices its contracted book and misses 5 gigawatts of secured power, four buildings for Microsoft, and roughly $17 billion of market value)
CoreWeave (More than $30 billion of debt and a further $3 billion sought, against contracts at $40 million per megawatt and a training operations software layer the large labs rely on)
Nebius (Raising GPU rental prices; positioned closer to a general cloud provider, with regional servers and model APIs, and moving toward an asset-light model)
Galaxy (Martin's largest position; a data center business with land, buildings and grid access, and second in secured power at 1.6 gigawatts)
ERCOT (The Texas grid operator whose base-load classification Martin treats as near-certain energization)
Microsoft (Iron is building four facilities for it; one has been approved and running since August)
Llama and DeepSeek (Named as examples of the models Nebius can connect customers to through its APIs)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: