Oracle's return on assets has fallen from 39% to 27% to 17%, and Altimetry Research forecasts 11% for next year.
Every other large cloud business is spending as heavily and holding its returns near the mid-20s. That gap is why Joel Litman and Rob Spivey read the same negative free cash flow as a reason to buy Microsoft, Alphabet and Amazon and a reason to sell Oracle.
"If you look over time, sometimes the best signal for buying a stock is when free cash flow goes negative."
Litman is a forensic accountant and the chief investment officer of Altimetry Research and its institutional arm, Valens Research; Spivey is Altimetry's director of research. They restate company accounts under their own framework rather than using the reported figures, on the argument that standard accounting misstates what capital-heavy businesses actually earn.
I listened to the full episode so you can skip it. 37 minutes of audio, 17 minutes of reading.
Here are the 9 takeaways that matter.
👤 Guests: Joel Litman, a forensic accountant and the chief investment officer of Altimetry Research and its institutional arm, Valens Research; and Rob Spivey, Altimetry's director of research
🎙️ Host: Bridget Bennett, MarketBeat's digital media producer, who hosts the firm's video interviews
📰 Published: 1 September 2026 on YouTube (MarketBeat)
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 37 min | ✅ Time saved: 20 min
Key Takeaways
Negative free cash flow at a high-return company is a buying signal, not a warning Amazon's deepest cash burn, when it was building AWS, was one of the best entry points the stock ever offered
A 5% policy rate is a rising rate, not a high one The 25 years it is being compared against were the lowest rates on record, so the comparison flatters nothing
Companies are borrowing at 5% because the projects they are funding return 20% to 50%
ASML is lifting output of its most advanced machines from 85 a year to 110 It raised prices for the first time in a while and says it would still be fully booked if it built more
GE Vernova is going from 12 GW of turbine capacity to about 30 GW within five years One machine delivers 400 to 650 MW, which Litman compared to the draw of inner-city Boston
Comfort Systems earns a 50% return on assets by building in its own factories rather than on site
Oracle is the one large cloud borrower whose spending is destroying its returns Its return on assets has run 39%, then 27%, then 17%, with 11% forecast for next year
Rocket Companies is the position that loses if corporate borrowing keeps long-dated yields up
1. Rates Rise, But Aren't High
Bridget Bennett opened on Fed chair Kevin Warsh's Jackson Hole speech the previous Friday, and on the fear it left in the headlines. The answer from Altimetry was that the market has read the speech backwards.
Altimetry's reading is that Warsh signaled resolve on inflation without believing inflation is a live problem. The firm's own analysis of what Warsh says points the same way, and a rate rise does not follow from it "But in reality, interest rates in terms of what the Fed's going to do, it's unlikely to have to hike because Warsh is really telling you when you get down into the roots, inflation isn't really an issue in the US."
Altimetry's account is that Warsh has moved back and forth since taking the job — hard on inflation at his first press conference, softer at the last one — and that this speech was a correction back
The 25-year headline flatters the comparison, because the 25 years being compared against were the cheapest money in history. "In the 25 years of the lowest interest rates in recorded history of humanity, you've never had such low interest rates." "We had negative interest rates for many many of the years of like negative real interest rates."
The distinction they drew is between a high rate and a rising one. "But at 5%, that is not a high rate, it's a rising rate." "And so, of course, we want to pay attention, but it's only high rates that cause a problem for the economy."
Bennett made the same point from the other end, saying that anyone who ran a business in the 1980s would treat a 5% rate as a treat
The borrowers can carry it, they said, because the companies taking the money are earning returns far above 5% on what they build with it
2. Who's Borrowing, and Why
Altimetry separated two explanations for rising yields that look identical on a chart. One is investors refusing to hold US government debt. The other is companies competing for money because they have somewhere profitable to put it.
Their answer is the second, and it changes what an investor should do about it. "If the real reason is there are companies that are borrowing massive amounts at 5% because you know that they can get a 20 30 40 50% return on that money and they're going to put that much more at work."
The spread between the borrowing cost and the return on the project is the whole test. "I would stay away from the companies and that's what we'll talk about that are also borrowing at 5% and maybe aren't going to get a 5% rate of return and those are the companies you should stay away from."
They rejected the macro framing outright: "But either way, the battlefield is not, oh my goodness, the US economy is in trouble. Let's just get out."
For two months, Spivey said, the debate has been about bond vigilantes, and about the largest cloud operators turning free-cash-flow negative. He named Oracle, Amazon and Alphabet as negative on free cash flow now
3. Negative Cash Flow Is Good
The bear case Spivey was answering is that a company with negative free cash flow — one spending more on new assets than its operations bring in — is proof the AI build-out has stopped paying for itself. He said the reverse is usually true.
His rule is that the cash burn is a buying signal when the business already earns a high return on what it owns. "If you look over time, sometimes the best signal for buying a stock is when free cash flow goes negative."
The screen he applies is return on assets, the profit a company earns against the money tied up in it. "Alphabet has a 25% return on asset business. That's 2x corporate average." He put Microsoft at 25% to 30% on the same measure
Spivey said Satya Nadella repeats the same line on every Microsoft earnings call, that "we would have sold more Azure revenue if we had more Azure capacity", and that demand is outstripping supply
The precedent he leaned on is Amazon's own history. "When Amazon went free cash flow negative, really deeply free cash flow negative to invest in AWS in the late 2000s, early 2010s, that was one of the best buying opportunities for Amazon."
The reverse reads as a warning. "On the flip side, when companies like Home Depot or Starbucks in the early 2000s went positive free cash flow, that wasn't a buying signal. That was a sell signal cuz they had stopped growth."
What separates the good case from the bad is where the cash is going, not whether the number is negative. "So, if you have negative free cash flow because the company's losing money in operating cash, that is a bad company." "Well, when investment exceeds operating cash flow, you have negative free cash flow."
4. ASML Picks Who Builds Chips
The first of three buys. ASML makes the lithography machines that print circuit patterns onto silicon, and Altimetry's case is that nothing in the AI build-out gets made without them.
The company sits at the single point every chipmaker has to pass through. "ASML is arguably the lynchpin of the semiconductor chip manufacturing companies"
The demand comes from two directions at once. Altimetry put the largest cloud operators' capital spending on a path from $350 billion to $800 billion to $1.2 trillion, and said the memory makers have committed a comparable sum "Well, all the memory companies, Micron, Samsung, SK Hynix, they just announced that they're going to spend a trillion dollar in capex to make new chips." That figure does not include Taiwan Semiconductor, Intel or anyone else, they said
Pricing has already moved. "They just raised prices for their first time in a while."
The number Altimetry cares about is the return on assets under its own restated accounts. "And that's leading return on assets, uniform ROA for this business, to double from 20%, which is already 2x market average, to 40% in the coming years."
"And the market is totally sleeping on just how big of a move this is for ASML."
The teardown that failed
The story Altimetry told to explain why the monopoly holds is a teardown that went wrong. A Chinese chipmaker bought an ASML machine and took it apart intending to copy it.
"In the words of the ASML engineers, they took it apart, and they couldn't put it back together again and work." The buyer then called ASML to report a fault, and the engineers who came out found what had happened
The point is that the barrier is not the design. The maintenance, the calibration and the assembly are what cannot be copied "And simply having the pieces and the blueprints isn't enough."
"So this ASML has a monopoly that's going to stay for a long time no matter how many people try to copy it."
From 85 machines to 110
Altimetry runs audio analysis over earnings calls, looking for where a management team sounds confident and where it sounds like it is holding something back. Spivey said ASML's most recent call produced a clear signal on capacity.
"Management was specifically highly confident when talking about their capacity expansion."
The scale of the ramp is the news. "They've historically produced around 85 of these a year. They're ramping that to 110." "So that is basically 25% plus ramp in terms of capacity that they're planning on ramping." Spivey said the company is also confident it would still be fully booked even if it built well beyond that
These are single pieces of equipment "the size of an elephant", which have to be shipped and installed on site
A company raising prices, adding capacity and still booked out for years is what tells Altimetry the growth has time to run. That is why Spivey said he is still bullish after a large move in the stock
5. GE Vernova Sells It All
The second buy, and the one Litman knows from the inside. GE Vernova builds the large gas turbines that power stations use, and Litman has spent the past year working on a gas-fired power project built for AI demand.
Litman's argument is that the constraint on AI is electricity, and gas turbines are the only way to add a lot of it quickly. Nuclear takes five or ten years; a turbine does not
"One machine can get you 400 to 650 megawatts of power. That's like inner city Boston." A few of them together produce a gigawatt
The demand is far larger than the current fleet. "The United States needs at least a thousand plus gigawatts of power."
The order book is already longer than the factory. Litman said the backlog runs more than five years out, and that the company is adding capacity against it "Whatever capacity they build, they will sell."
The machine is only the first sale. Maintenance and recalibration follow it, in what Litman called a razor-and-blades pattern "Once they get installed, they keep making money. It's not like they just sell it once."
Altimetry's restated return is four times what a screening site shows. "The real return on investment, by the way, is something like 20%." Litman said a site such as Yahoo Finance shows about 5% "Because GAAP accounting totally messes up companies like this in terms of what's really on the balance sheet and what really they're generating for operating cash flow."
Project Matador's 17 GW
Litman said he has been involved with Project Matador in Amarillo, Texas, which he described as the largest gas-powered electrical plant built for AI in the world, at 17 gigawatts "That's like 17 Bostons' worth of electricity, all geared towards AI power."
The turbines are about the size of a small house, with blade tips that break the sound barrier, and Litman said one draws in as much air as a blimp holds in 10 to 15 seconds
The siting logic is the gas, not the land. "So that's one of the reasons Project Matador, the one that I've been working with for the last year, happens to be in Texas is because that's the right place to build these things." Litman said the gas is effectively unlimited in parts of Texas, and cheap when the pipeline run to the site is short
He was careful not to set it against nuclear: "Gas is one that you can get up and running right now and it's amazing."
Bennett made the timing argument herself, saying nuclear plants and even small modular reactors take so long to build that gas is the only thing that meets demand now
The electron election
The stock is volatile because it is the market's proxy for whether the AI build-out continues. Spivey said the pullback since June was not about the company, which reported well, but about that doubt
He described investment cycles as an S-curve — slow start, acceleration, then deceleration — and said the argument is over where this one sits
The second worry is permitting, which Spivey calls the electron election. He said states from Maine to Virginia to Texas are deciding whether power plants and data centers get built, and whether consumer power prices rise if they do
The capacity plan is what he says settles it. "This is a company that's going from 12 GW of capacity for turbines to probably 30 gigawatts of capacity for turbines in the next 5 years." "If they can triple their capacity, they can probably more than triple their earnings"
"The volatility is a buying opportunity. This is one where you can really buy the dips."
6. The $60B Contractor
The third buy is Comfort Systems USA, ticker FIX, a $60 billion engineering and construction company that does heating, ventilation, air conditioning and electrical work. Altimetry's line on it was that almost nobody watching would have heard of it.
The backlog is the demand evidence. It has gone from about $8 billion last year to as much as $14 billion this year, on data centers, power plants, and infrastructure for power and for fiber
The competitive advantage is where the work happens. Most contractors put people on a site and build there, which runs into a shortage of skilled labor. Comfort Systems prefabricates as much as it can in its own specialty facilities and cuts the time crews spend on site "And the thing is they can do this because of their scale because they're bigger than everybody else." Running those facilities across the country is what a smaller competitor cannot copy
The reported numbers understate what the business earns. "It's not a 15% return on asset business. It is a 50% return on asset business." That is four to five times the corporate average
7. Third or Fourth Inning
Bennett asked the question that hangs over all three names: how long the demand lasts. Spivey's answer moved the subject off AI.
The management teams he talks to say demand for data-center capacity is not slowing, at Microsoft, Alphabet, Amazon and at Elon Musk's companies
The cycle is bigger than data centers, and started before the tariffs. "It is about a reindustrialization of the United States of America." "That was happening, by the way, long before we started getting tariffs and everything in 2024, in 2025, I should say." Semiconductor fabrication plants, broader factories and returning supply chains are all part of the same spending, he said
Altimetry tracks how depreciated US corporate assets are as a measure of where the cycle stands, and Spivey said even three years of heavy investment has not carried it far
His summary of the position was a baseball metaphor with a number in it. "This is the third or fourth inning of this baseball game for AI and re-industrialization spend. It is not the ninth. It is not the back end."
8. Why Oracle Is the Sell
The first of two to avoid. Altimetry's objection is not to Oracle's software business but to the part of the AI supply chain it has moved into.
Oracle is the one heavy borrower among the large technology companies whose credit rating shows strain. Most are rated double-A or single-A; Oracle is triple-B-minus, which Altimetry described as half a grade above junk, and said even the ratings agencies are signaling concern "So Oracle is issuing debt. They're not getting the phenomenal interest rate everyone else is getting."
Litman was careful about the company's record first: "What Larry Ellison has done with Oracle is absolutely fantastic over the years." He told the old joke on the way through — "What's the difference between Larry Ellison and God? God doesn't think he's Larry Ellison."
The problem is which link in the chain it chose. Data centers themselves are the commoditized part, bought from suppliers with long backlogs and sold into a market where new capacity keeps arriving "It's the least profitable." "It's the one where they may be issuing debt at 5 and a half or 6% and not getting a five and a half or 6% return on the investments they're putting into."
The difference from the other cloud operators is what gets sold on top of the building. "When Oracle is building these data centers for OpenAI in terms of Stargate and everything else, they're building a dumb shell" Microsoft, Alphabet and Amazon sell storage, security and infrastructure as a service on the same asset, which is why their returns hold near the mid-20s while they spend
The returns are moving in one direction. Oracle's return on assets has run 39%, then 27%, then 17%, with a forecast of 11% for next year
"And so Oracle is one of our big stocks that so many people own and we'd say stay away from it. Sell it if you own it."
9. Rocket Is on the Wrong Side
The second to avoid, and the only name in the episode with no connection to AI. "And this company has absolutely nothing to do with AI at all."
The mechanism is the one the episode opened with. Corporations borrowing heavily to invest keep long-dated yields up, so mortgage rates do not fall, and anything geared to the American homebuyer struggles
Rocket's exposure to that is close to total. "Rocket companies are Rocket Mortgage, which touches, I think, something like one in six, if not more of every single mortgage in the United States today." "Also, Mr. Cooper, which is the servicing of mortgages, and Redfin, which is, you know, basically along with Zillow, the way that you find what house you want to buy."
The valuation needs an outcome Altimetry does not expect. The market needs the company's return on assets to triple for the current price to be fair, they said "And we just don't see a setup for that to happen." "It's just not going to happen with interest rates where they are."
The counterexample is the same company five years ago. In the very low rates of 2020 to 2022, mortgage firms were adding staff because they could not keep up with refinancing demand The point was how fast that reversed, and Bennett agreed that the story can change quickly
Bonus Insights
GE Vernova's largest turbines have a nickname. Litman said the company's engineers and customers "often call them Jersey boys", after the state where that part of the business was based
Spivey set the bond-vigilante debate aside deliberately: the two months of argument about who will buy US government debt is, on his reading, the wrong explanation for why yields rose
Bennett noted that other analysts on the show have taken the opposite side on Oracle, and asked viewers to say in the comments where they land
Litman's defense of gas was that it is not a bet against nuclear but an answer to a timing problem, and that the energy build needs more than one source
Altimetry's bottom line is that a 5% rate is only dangerous to a company that cannot out-earn it, which is why the pair would own the businesses borrowing hardest to build AI infrastructure and sell the two whose returns are falling as they borrow.
Products, Companies & Tools Mentioned
ASML (The lithography-equipment maker Spivey calls the lynchpin of chip manufacturing; raising prices, doubling its restated return on assets from 20% to 40%, and going from 85 of its most advanced machines a year to 110)
GE Vernova (Builder of the large gas turbines Litman says are the only fast way to add power for AI; one machine delivers 400 to 650 MW, and capacity goes from 12 GW to about 30 GW in five years)
Comfort Systems USA (The $60 billion contractor whose backlog went from about $8 billion to as much as $14 billion, and which Altimetry restates as a 50% return-on-assets business)
Oracle (The sell. Triple-B-minus rated, building what Spivey calls a dumb shell for others, with return on assets running 39%, 27%, 17% and a forecast 11%)
Rocket Companies and Rocket Mortgage (The other sell — a lender, servicer and listings business whose value depends on mortgage rates that Altimetry says are not coming down)
Redfin and Zillow (Named as how buyers find a house, and as part of Rocket's exposure to a frozen housing market)
Microsoft and Azure (The 25% to 30% return-on-assets business Spivey says is capacity-constrained, not demand-constrained)
Alphabet and Amazon Web Services (Alphabet at a 25% return on assets, twice the corporate average; Amazon's deepest cash burn to build AWS was, Spivey says, one of the best times to have bought it)
Micron, SK Hynix and Samsung (The memory makers whose announced trillion dollars of capital spending lands on ASML's order book)
Nvidia, AMD, Broadcom, Taiwan Semiconductor and Intel (The chip designers and foundries powering the build-out, all of whom need ASML's equipment before anything gets printed)
Home Depot and Starbucks (Spivey's counterexamples: both turned free-cash-flow positive in the early 2000s, which he says was a sell signal rather than a buy)
Project Matador (The 17 GW gas-fired AI power campus near Amarillo, Texas, that Litman says he has worked on for the past year)
SpaceX, xAI and OpenAI (Named as the demand side: Musk's companies as builders that say demand is not slowing, and OpenAI as the tenant Oracle is building Stargate for)
Altimetry Research (Litman and Spivey's own firm, which restates company accounts and runs audio analysis over earnings calls to grade management confidence)
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: