RiskReversal Pod Sep 18, 2026 · recorded Sep 17, 2026 54m 36m saved
With Dan Niles, founder of Niles Investment Management, previously a top-ranked computer hardware and semiconductor analyst on the sell side
The average price charged per AI token has halved since the end of May, Dan Niles said, while the number of tokens produced has quadrupled.
The usual reading of that is Jevons paradox doing its work: cheaper inference, more of it, no harm done. Niles's version is less comfortable, because most of the volume growth is accruing to free open-weight models rather than to the two companies the market is paying for.
"You've seen the cost per token being charged on average go down 50% since the end of May till today. The good news is that the number of tokens being produced quadrupled."
Niles was a top-ranked computer hardware and semiconductor analyst before moving to the buy side, and now runs Niles Investment Management. He published a piece on 20 June warning of a speed bump in the AI trade; semiconductors then fell 29% between 22 June and 29 July.
The full episode is covered here so you can skip it. 54 minutes of audio, 18 minutes of reading.
Here are the 15 calls that matter.
Key Takeaways
Token prices are down 50% since late May while volumes quadrupled, with the volume accruing to open-weight models
Texas placing a moratorium on data center connections is what turned him more negative, because it is a supportive red state
The top 1% of AI spenders cut spend 10% from July to August on the Ramp index, the first fall since ChatGPT launched
Inflation has run above target for 65 months, which is why he expects a series of hikes rather than one
Google went cash flow negative for the first time since listing, and is now issuing debt against a 6% federal deficit
He is still in the speed bump camp, with a year or two before a top, and says he is holding the view loosely
The three motives behind the safety statements are concern, a duopoly, and a better IPO valuation — all can be true
He puts Anthropic at $65B of revenue against OpenAI's $40B, on figures he says are not fully comparable
Meta is his dark horse: the number two free app behind ChatGPT and a market multiple against peers in the low 20s
Apple's defense is that it is not spending $200B on capex, and licenses from Google for a couple of billion instead
1. Don't Fight Three Things
Asked to start with macro, Niles reached for the oldest rules in the business.
His first principle
I always say like you don't want to fight the tape. And seasonality works for a reason, right?
Dan Niles
He said he reads S&P 500 statistics back to 1957, the index's modern inception, alongside shorter windows, and takes the pattern seriously rather than arguing with it.
The other two
And I don't want to fight the Fed, right? That phrase exists for a reason. And I definitely don't want to fight the bond market.
Dan Niles
He credited the bond-market line to James Carville, who said he wanted to come back as the bond market because it can bully anybody. The host explained for listeners that Carville ran Bill Clinton's 1992 campaign, and connected the point to a midterm election five or six weeks away where the economy is the issue and a president who wants lower yields is not on the ballot.
And the calendar
September being the only month since 1957 that's down on average and is down more often than it's up.
Dan Niles
All three are pointing the same way, he said, with bond yields at multi-decade highs in some countries.
2. The Data Center Turn
Pressed on what could actually produce a deeper drawdown, given the market has absorbed tariffs, a war and a 5% 10-year while sitting within 2% of its highs, Niles named one thing.
His answer was specific
Well, the big one's data centers, right?
Dan Niles
His reason is concentration: semiconductors are up more than 60% on the year and are somewhere in the low teens as a percentage of the S&P 500, so strip them out and the rest of the market has produced very little.
Semiconductors carried the year
If you look at the semiconductor index for the year, it's up over 60%.
Dan Niles
The event that changed his view was political rather than financial.
Texas moved against connections
The state of Texas, which is a big proponent of data centers, they went ahead and started to put a moratorium on data center expansion
Dan Niles
He called that unusual because Texas is a deeply red state that has been supportive, and the second-largest data center market in the country. The second event was the weekend before the recording.
Three founders said the same thing at once
You saw Sam Altman of OpenAI, Dario of Anthropic, and Elon Musk all talk about the fact that they needed to slow down the pace of innovation.
Dan Niles
3. Tokens Down 50%, Up 4x
The third element is the model economics, which is where his June warning came from. His description of what customers started doing was a metaphor they gave him.
The customer's own framing
We've been using a Ferrari to go to the corner store to get milk and we should be using a Ford, right?
Dan Niles
Which shows up in the price
You've seen the cost per token being charged on average go down 50% since the end of May till today. The good news is that the number of tokens being produced quadrupled.
Dan Niles
The volume, he said, is accruing mainly to open-weight models, mostly Chinese with some American entries including Meta's new release. He also described what he has picked up on profitability: OpenAI cut the price of its leading model substantially and appears to be taking share from Anthropic, while its profitability metrics appear to have worsened. These are private companies, he was explicit, and the data comes from fragments heard when they raise money or speak publicly, so the exact figures are hard to pin down.
He added architectural changes on top, such as separating the cache from the inference step, which may reduce how much compute is needed.
Which leaves one hope
You better hope that Jevons paradox, which everybody keeps quoting, continues.
Dan Niles
His own June call gives the scale of what a speed bump looks like.
The last one cost 29%
It turned out semiconductors went down 29% from June 22nd to July 29th. So it was one heck of a speed bump.
Dan Niles
4. The Ramp 1% Signal
The leading indicator he pointed at is corporate card data.
The heaviest spenders cut back
The top 1% of users actually saw a 10% reduction in spend from July to August if you look at the Ramp index.
Dan Niles
That series, he said, has gone straight up every month since ChatGPT appeared, and the other 99% are still increasing. His read is that the largest spenders are moving to cheaper models, including open source.
What it reduces to
And at the end of the day, it comes down to profits and revenue growth for the AI space.
Dan Niles
The host's aside was admiration for Ramp building the index from anonymized customer data, much of it from cutting-edge technology firms, and he credited Eric Lyman by name.
5. December Is the Hike
On the rate path, Niles and the host agreed there is no meeting in November, which makes the sequence unusual.
October is off the table politically
I doubt the Fed's going to want to raise right in front of the elections because the amount of blowback they're going to get from the White House would be ridiculous, right?
Dan Niles
The October meeting falls about five days before the midterms. He expects the next one, on 10 December, to deliver, and his reasoning is that a single quarter point achieves nothing.
The duration of the overshoot is the argument
Inflation's been running above the 2% target now for 65 months
Dan Niles
He noted Kevin Warsh talking about bringing inflation down in a timely manner and said nothing is close to that.
So one hike is not the forecast
You should expect a series of rate hikes.
Dan Niles
And history says what that does
The market, just based on history, underperforms when the Fed is hiking.
Dan Niles
6. Hyperscalers Borrowing
The link Niles drew between a 5% 10-year and the AI trade is that the companies doing the spending now need the capital markets.
A first for Alphabet
Google went cash flow negative for the first time since they went public.
Dan Niles
They are raising debt and equity, he said, and all of them are.
Against a very large borrower
They're now competing with the federal government that's running a 6% deficit relative to GDP
Dan Niles
He called that the largest deficit outside a major war that the country has run. His broader point was about investor conditioning: anyone who did not lose money through the financial crisis has only ever seen problems solved by throwing money at them.
The reflex he distrusts
Everybody seems to believe it's always buy the dip.
Dan Niles
Central banks and governments are running out of ammunition, he said, which is part of what the bond market is expressing. The alternative side of the trade is the saver, and he put Japan first.
Where the competition for capital comes from
Right now, you have yields the highest since the 1990s that they can invest in.
Dan Niles
A guaranteed 5% over ten years, against valuations he described as among the highest since 2000 on a market-cap-to-GDP basis, is not an inconceivable alternative.
7. The Midterm Drawdown
The host brought two structural risks. The first is the yen carry trade: investors borrowing in yen to buy long-duration American assets, with a rate-hiking cycle now running in Japan, the euro area and the United Kingdom as well as the United States. He said the unanimous 25 basis point vote surprised him, having expected at least one dissent.
The second is a statistic of his own. From 31 July to 9 November, in midterm years since 1990, the average peak-to-trough drawdown in the S&P 500 is about 10%, double the figure in non-midterm years, and the market is nowhere near that.
He also raised the political composition question: seats being picked up by candidates whose platform is, by his reading, anti-large-business and specifically anti-data-center, which would make the period after the midterms worse for the sector rather than better.
8. Speed Bump, Not a Top
Niles took the carry trade seriously and then drew the distinction the episode is named for. On valuation he was careful to say he means the aggregate, market cap to GDP over ten years, not individual names: Nvidia on his read trades at a mid-teens multiple of calendar 2027 earnings, which he called ridiculously cheap if the estimates hold. His hesitation is whether companies are over-earning, with memory as his example.
He is holding the view loosely
I'm in the speed bump category, but I want to keep my mind open to that might be wrong
Dan Niles
Because of where he thinks the cycle is
I think we're just so early in the agentic AI trade, which really started only on January 30th with OpenClaw, that I feel like we still have at least a year or two to run.
Dan Niles
9. Three Motives at Once
Asked whether Musk, Amodei and Altman actually believe what they said, Niles refused the binary.
His framing
Well, remember two things can be true at once.
Dan Niles
Genuine concern about the technology and a desire for a government-sanctioned duopoly can coexist, he said, and Musk has been warning about AI for a decade. The third motive is the calendar.
The listing is the third reason
Both OpenAI and Anthropic are trying to come public, and so if they can show better cash flow because they're spending less, that's going to help them be able to come public.
Dan Niles
Choking off open-weight competition while two companies sit that far ahead would suit them, he said, as would presenting a spending cut as a safety decision.
All three at the same time
So you can have multiple motivations all for the same thing.
Dan Niles
He returned to the same question later and answered it differently: he does not think any of them will actually slow down, because slowing while Google keeps going is intolerable.
The reason it cannot hold
Because if you slow down and Google keeps going, you're spooked.
Dan Niles
The wrinkle he allows is a company slowing for a quarter or two purely to improve the numbers ahead of a listing.
The host's version of the risk was about the rate of change rather than the level.
Capital spending doubled this year
Don't forget capex last year was up like 75%. This year it's like up 100%.
Dan Niles
And the analogy he drew
It's not like the internet stopped growing in 2001, 2002, it doubled in each of those years
Dan Niles
Growth does not have to contract, he said; it only has to slow. He also remembered forecasts made in 2000 that ran out to the end of the decade, and how those ended.
His stated position on television that week
The market can keep pitching, but you don't have to swing
Dan Niles
The internal contradiction he pointed at in the bull case is about constraints.
You cannot have it both ways
If you're telling me power and land is what is dictating how much you can spend, right, then how is bipartisan pushback on data center spend and connecting to the grid not going to matter?
Dan Niles
10. The 2001 Cisco Release
Asked what breaks first if spending hiccups, and whether the enormous reported backlogs are meaningful, Niles reached for a document.
The precedent he keeps
For those viewers, Cisco was Nvidia back then. It had the biggest market cap in the world.
Dan Niles
He pointed listeners at Cisco's May 2001 earnings release, where he recalled the commentary describing orders going from up around 50% year on year to down about 30% within months, while flagging that he may have the exact percentages wrong.
Which is his view of backlog talk
And so this backlog stuff is ridiculous, because I go, it can change very quickly.
Dan Niles
And the mechanism that inflates it
When you can't get what you want anyway, you're going to double and triple order.
Dan Niles
His concession is that the overbuilding is rational for the participants. He quoted Microsoft's chief executive on the point.
The industry's stated logic
The biggest risk is underinvesting, not overinvesting.
Dan Niles
The winners make a fortune and there are wrecks along the way, he said, naming AOL against Google, Webvan and Pets.com against Amazon, whose stock fell 95% before it became dominant.
Where that leaves his call
What I'm trying to say is I still think we have another year or two before we hit the top.
Dan Niles
The host's list of what gets hit first was blunt. On a pullback in spending, the companies that are not Nvidia, meaning AMD, Intel, Oracle, the neoclouds and the memory names, have the most to lose, and the semiconductor index is already 25% off its July highs after doubling from April to June. He used Intel up 10% on a single headline about a possible contract for its Ohio fab, which he noted would not be built out until around 2030, as the example of a market trading every headline.
11. Anthropic Over OpenAI
Asked whether this becomes a winner-take-most market like search, Niles said it coalesces to a few players but not the two people expect.
His enterprise argument
I think Anthropic will do very well in enterprise because that's what they focused on from day one.
Dan Niles
Enterprises pay and consumers do not, he said, which is why OpenAI's consumer-first focus eventually pushed it toward enterprise too.
The revenue figures he has
The last figures we have is Anthropic's at 65 billion in revenues and OpenAI is at 40 billion.
Dan Niles
He said he is not sure exactly how each counts revenue, particularly on infrastructure costs.
His conclusion, hedged
By most accepted measures, Anthropic is now much bigger than OpenAI
Dan Niles
He added that Anthropic reportedly reached adjusted operating profitability in the second quarter with further improvement in the third, and that he dislikes the term. His expectation for the consumer side is Google. On the internet comparison he listed Google in search, Amazon in e-commerce, Netflix in streaming and Meta in social, and asked who the number two in each is and how far behind; even three winners in AI, he said, is more than the internet produced.
12. Meta as the Dark Horse
The name he wrote up a few weeks earlier, and the one he is most animated about, is Meta.
The adoption data point
The Meta Agent is the number two downloaded free app behind ChatGPT.
Dan Niles
He said Meta trades at a market multiple while its peers sit in the low 20s on earnings, and that the historical problem was that capital spending could only be monetized through its own advertising business.
The base it would be monetizing
They had 3.6 billion daily active users.
Dan Niles
If people start doing agentic work inside the Meta app, he said, it becomes a dark horse winner.
Which gives it three labels at once
It's sort of got the holy grail of it can be a value stock, a growth stock, and an AI play all at the same time.
Dan Niles
He was explicit that valuation was not what drew him: the pattern he was buying is Google's from 2025, where the legal overhang cleared, good products arrived and the multiple moved. Meta settled with state attorneys general on youth safety, released an interface for developers, then released a consumer agent, and he said a frontier-class model called Watermelon is due by October at the latest with a company event the following week where a preview might appear.
The pattern he is trading
You settle a legal overhang, and then you're introducing good products, and you can rerate very quickly.
Dan Niles
13. Why He Loves Google
Asked why Google is not being rewarded for the same attributes, Niles said the answer is timing rather than fundamentals.
His view of the business
Oh, I love Google, because if you look at the fundamentals of their business, they saw the most acceleration of any of the big hyperscalers, right?
Dan Niles
He gave the sequential cloud growth figures: Amazon's from 28% to 37%, Microsoft's from 39% to 43%.
And Google's
GCP went from growing 63 to 82%.
Dan Niles
The stock did not respond, he said, because Google reported before the unwind in the AI trade while Microsoft and Amazon reported after, which he described as the market structure deciding how a print is received rather than the numbers.
He expects the frontier leadership to keep changing hands, and noted Meta's own path: it had the leading open-source model, lost ground after DeepSeek, then Mark Zuckerberg acquired Alexandr Wang and his company, and a new model followed.
Google's original claim on the technology
Don't forget these guys invented the transformer technology that OpenAI went out and then commercialized first.
Dan Niles
And the advantage he thinks persists
What makes a model smart — it's the amount of data you have to train it with.
Dan Niles
Free products, he said, mean users hand over training data, and both founders remain closely involved in pushing for the frontier, so he would not count Google out. The host's contribution was the distribution case: Google Cloud, its own accelerator, Chrome's browser share, YouTube and more than two billion daily actives, plus a new application he recommended after reading about it in the Wall Street Journal.
14. Microsoft's Free IP
On circular financing, the host laid out the web: Microsoft the largest investor in OpenAI, OpenAI among Azure's biggest customers, Nvidia an investor in OpenAI and OpenAI among its biggest customers, and a list of contracts including Oracle. Niles's answer was about timing, and he noted the 1990s had the same arrangements and Cisco had at least two drawdowns of more than 30% along the way.
It only becomes a problem later
The circular financing stuff's only going to matter when things start to slow.
Dan Niles
On Microsoft specifically, the stock rose sharply after a print where cloud growth accelerated less than Google's.
The variable was spending, not growth
Well, that's because they didn't raise capex
Dan Niles
He then made the structural argument: Microsoft is not in the frontier model business and does not need to be, because it is entitled to OpenAI's intellectual property.
Which he called the trade
Why should they be spending money when OpenAI is spending a ton of money and they can get the IP for free, right?
Dan Niles
The numbers cited around the table were $13bn invested for 27% of the company, roughly 450 million Office users, and the part he finds most interesting.
The ramp he is watching
They've got 30 million co-pilot users
Dan Niles
His thesis is security. On agents going rogue, he said people in the industry tell him the problem was poor sandboxing rather than autonomous misbehavior, and compared it to a knife that can chop vegetables or do harm depending on the user. If companies conclude they want AI inside something they already trust, he said, the assistant inside Microsoft's office suite is the default, and the 30 million becomes a much larger number.
15. Apple and the Foldable
The host's premise was that the new folding device will not drive much and that Apple needs its assistant working before anyone gets excited about handsets. Niles disagreed flatly.
He rejected the premise
I actually disagree with that premise.
Dan Niles
His precedent is the screen-size change on the iPhone 6.
What the last form-factor change did
Revenue growth went from 7% in fiscal '14 to 28% growth in fiscal '15.
Dan Niles
Samsung has had a folding phone since 2019, he noted, and there are a great many people with money whose eyesight would welcome a larger screen, which is why he expects a large upgrade cycle next year. He picked Apple as a top pick for this year partly for a defensive reason.
The capital-spending argument
They're not spending $200 billion on capex. They're letting Google go spend that, and then they're licensing that for a couple of billion dollars a year.
Dan Niles
And he likes the new leadership for products
I think having Ternus in charge, who's a product guy, is going to be great for product innovation.
Dan Niles
He credited Tim Cook with financial engineering and supply-chain management and said the multiple expanded a great deal under him, while being unsparing about one product.
On the headset
The Vision Pro is a complete bust.
Dan Niles
Near term he expects the stock to be choppy, because the AI features ship as a beta and the cheaper base models are not available until the first half of next year, leaving questions about the December quarter with the consumer already under pressure from prices and oil. Apple's customers, he noted, are the higher-end ones.
The defensive case in one line
The stock is very defensive when you're not spending $200 billion a year on capex.
Dan Niles
Bonus Insights
The two of them spent the last few minutes arguing about the folding phone as consumers rather than as investors, and agreed to bet one. The host's objection was physical: Apple has sold cases and screen protectors at very high margins for 19 years, and a device that is three-quarters glass and folds cannot be protected the same way, which he thought carried the potential for a defect story on the scale of the old antenna problem. Niles's answer was that he is normally an early adopter and has no interest in this one, that he wants to consume less content rather than more, and that he will wait for the version after the beta.
Both landed in the same place on timing. The host said he wants millions of users to find the problems first and expects to buy sometime next year once the product is hardened; Niles said the same about not wanting a beta version of anything. Both flagged that the new chief executive cannot repeat Cook's financial engineering, and that every piece of hardware arriving over the next few years was built under his watch as head of hardware.
This was Niles's thirteenth appearance on the show in nearly six years.
Niles's bottom line is that the AI trade has a year or two left before a top, that the risks worth watching are the political fight over data center connections, the migration of token volume to free open-weight models and a Fed that is hiking into a 65-month inflation overshoot, and that the way to hold the view is to be far more selective rather than out of the market.
Products, Companies & Tools Mentioned
Nvidia (On his numbers a mid-teens multiple of 2027 earnings, and the stock holding up while the semiconductor index fell 25%)
Anthropic (His enterprise winner, at $65bn of revenue on the figures he has, and reportedly at adjusted operating profitability)
OpenAI (At $40bn on the same figures, having cut its leading model's price and, he believes, worsened its own profitability)
Meta (His dark horse: the number two free app behind ChatGPT, a market multiple against peers in the low 20s, and a frontier model due by October)
Google (The cloud business he says accelerated most of any hyperscaler, and the owner of the transformer research OpenAI commercialized first)
Microsoft (Entitled to OpenAI's intellectual property without paying for the training, with 30 million assistant users on a 450 million seat base)
Apple (A top pick for the year on defensive grounds, with a folding phone he expects to drive a large upgrade cycle next year)
Ramp (Its spending index is where he saw the top 1% of AI spenders cut back 10% from July to August)
Cisco (The May 2001 earnings release he tells listeners to look up, as the precedent for how fast a backlog can turn)
Intel and AMD (Named by the host as the most exposed if spending slows; Intel rose 10% intraday on a fab headline)
Oracle (One of the contract counterparties in the circular-financing web, and among the names the host says are exposed)
Samsung (Has sold a folding phone since 2019, which is his argument that the form factor works)
Amazon (Cloud growth accelerating from 28% to 37%, and the survivor of the dot-com wreckage after a 95% fall)
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: