Most enterprises are not running frontier models anyway, so Uday Cheruvu says slowing the frontier changes when the compute gets bought rather than whether it gets bought.
The market spent the morning treating a safety debate as a spending cut, marking down Nvidia, the memory names and the neoclouds together. Cheruvu, who owns the names being hit, argued the sell-off confuses a timing question with a scale question.
"So what this means is that slowing down the frontier doesn't actually lead to a slowdown of the total demand curve, but just basically spreads the curve out over a longer period of time."
Cheruvu is a portfolio manager at Harding Loevner and holds the compute, memory and neocloud names at the center of the sell-off, so he was being asked to mark his own book in public.
The full segment is covered here so you can skip it.
Here are the 5 calls that matter.
👤 Guest: Uday Cheruvu, Portfolio Manager at Harding Loevner, who invests in the compute, memory and neocloud names that fell on the pacing news
🎙️ Host: Ed Ludlow, who anchors Bloomberg Tech from San Francisco
📰 Published: 14 September 2026 on Bloomberg Tech
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ length not available
Key Takeaways
A slower frontier stretches the AI demand curve rather than cutting it
Most enterprises already run older or open-source models and save the frontier model for hard reasoning tasks
The labs asking for rules are the labs the rules would protect
Cheruvu says a common rulebook removes the competitive penalty for being the careful one, and slows everyone behind them
Pacing makes the IPO story easier to tell, not harder
An IPO wants the revenue environment and the cost structure settled before the roadshow, not a live argument about the demand curve
The reported 80% margin is unlikely to survive the next generation of infrastructure
Revenue per token is falling while the cost per token and the depreciation behind it are rising
2027 is already bought, so the pacing debate is a 2028 question for Nvidia
Cheruvu says there is significant overordering, and anyone not locked in by now is not getting chips
1. Timing, Not Total Demand
Ludlow opened by asking for Cheruvu's reaction as someone directly invested in the corner of the market being marked down.
Cheruvu's read is that the weekend introduced a second variable to a market that had been trading on only one. He said the events set in play that it is not just about demand and demand growth, but the realistic timeline for it and how implementation of AI is really going to happen
"We're going from the face of super optimistic Bluesky to, hey, hang on, there's a few more hurdles here that we need to get through," he said
Ludlow pressed on the mechanics, asking whether it is linear that if the labs slowed at the frontier they would reduce aggregate spending on compute. Cheruvu's answer was "No, absolutely not."
The reason is that the frontier is not where most of the demand is. He said most frontier models are not being used widely anyway, and that enterprises are looking at how to optimize their costs by using some open-source models and some of the older models, keeping the latest model for extremes and for tasks that require a lot of reasoning
The conclusion he wants investors to take away is that the curve moves rather than shrinks: "So what this means is that slowing down the frontier doesn't actually lead to a slowdown of the total demand curve, but just basically spreads the curve out over a longer period of time."
"To us, that's what investors would need to focus their minds on is this is a timing issue as opposed to a total scale of demand issue," he said
2. Why the Labs Want Rules
Ludlow put David Sacks's response to him — that if the frontier labs think they should slow down they control their own destiny and can simply do it — and the counter-argument that government has to be the one to act.
Cheruvu split the labs' case for regulation into two motives, and only the first is the stated one. The safety argument, he said, is that the technology can be used for dangerous purposes, and that if one frontier lab allows something and another does not, the one that does not is at a competitive disadvantage
So what they want is a set of rules that applies to everyone, "so therefore no one has an incentive or a disincentive to slow down or speed up"
The second motive is incumbency. He said that if regulations are in place, the companies already ahead — he named OpenAI, Anthropic and Gemini — have a strategic advantage, because anyone behind them has to move at a much slower pace than they do
His verdict on the request is blunt: "So from my perspective, there's a bit of self-serving from the Frontier Labs that they do want rules because that is going to help them."
3. The IPO Case for Pacing
Asked whether he worries that an Anthropic or OpenAI listing gets kicked into the long grass, Cheruvu argued the delay is the point.
What a lab wants before an IPO is a settled cost structure, not a live debate. "If you're one of these two companies, you'd want the revenue environment set and the cost structure set before you come to an IPO. What you don't want investors is talking about what the demand curve slowdown might look like, might not look like," he said
On the timing, he said it may be better for them to wait into next year
The unresolved number is profitability. He said the market still does not know these companies' profitability: they say they are going EBITDA positive and increasing profitability, but the pace of that might not be enough to attract investors at the level they want
So the slowdown helps the story rather than hurting it: "So the pacing argument actually helps them get a clearer IPO story."
4. Anthropic's 80% Margin
Ludlow brought in the show's own weekend reading — the Financial Times reporting that Anthropic had told some investors it was on track for a second consecutive quarter of profit at 80% margins — and noted that the figure excludes the revenue share the company owes several parties and the cost of compute.
Cheruvu's answer is that the cost side is moving against that number. He said the cost of compute is going up, because buying the same servers is far more expensive than it was a year and a half ago, so the next infrastructure coming online arrives at a much higher cost and has to earn more revenue
The revenue side is moving the other way. He said the top line is being affected by token prices coming down, so revenue per token is falling while the cost per token and the total depreciation coming through are increasing
His conclusion is an explicit refusal to extrapolate: "So maybe what you saw from 80% is not likely to be the same number going forward. It could be better or it could be worse, but we just don't know."
The resolution comes with the next set of results. He said the clarity that comes through over the next few months, as the new infrastructure comes online and earnings and profitability come through, might be a better story for Anthropic to be listing into
5. Nvidia's 2027 Is Sold
Ludlow's last question was whether any of this touches Nvidia's own guidance: "We know what's going to happen next year. Nvidia told us 70% growth based on supply. Does that change at all or is this a 2028 story?"
Cheruvu took the later date. "No, I think this is a 2028 story because 2027, most of the companies have bought their demand in," he said
He went further and said the order book is inflated by fear of missing out on supply. "In fact, there's a significant amount of overordering because companies know that they need these chips to do the agentic AI that they want to achieve," he said
The scarcity cuts both ways. He said that for 2027, if you have not locked in by now you are not going to get the chips, and if you have them you are not going to let them out of your hand
He extended that to the whole chain, saying 2027 is most likely locked in for Nvidia and across the supply chain, including the neoclouds
Where it does bite is a year out. He said that as we get into this time next year, pacing, AI regulation and these companies' IPOs all come into play — but that over the next 12 months, "I don't think we should worry too much about revenue growth and profitability"
Bonus Insights
Cheruvu allowed that the improvement in model performance itself could change the hardware question by 2028, saying a different type of infrastructure may be required by then
Ludlow flagged that the program would carry OpenAI President Greg Brockman later in the hour saying the open-source work continues and that pacing applies to the absolute cutting edge, which is the framing Cheruvu's answer about older and open-source models runs alongside
The introduction named where the damage was concentrated on the day — Nvidia on the compute side, the memory names, and some of the neoclouds
Cheruvu's bottom line is that a pacing agreement changes the shape of AI spending rather than its size: the compute for 2027 is already committed, the frontier labs asking for rules are the ones the rules would entrench, and the argument that matters for investors starts about a year from now.
Products, Companies & Tools Mentioned
Nvidia (The lens Ludlow used for the whole question; Cheruvu says its 2027 is locked in and any pacing effect lands in 2028)
Anthropic and OpenAI (The two IPO candidates whose listings the pacing debate touches, and whose profitability Cheruvu says is still unknown)
Harding Loevner (Cheruvu's firm, which holds the compute, memory and neocloud names that sold off)
Google Gemini (Named alongside OpenAI and Anthropic as a lab already far enough ahead to benefit from a common rulebook)
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: