Citi's Heath Terry says the price of every input into artificial intelligence — inference, memory, power — is still rising, and that is the reason he expects the companies selling that capacity to earn more rather than less.
Rising input costs usually read as a margin problem. Terry says that in this build-out they arrive as revenue at the companies investors already own.
"And so, inference is getting more expensive, memory is more expensive, power is more expensive, everything in that supply chain is getting more expensive, which for the companies that you just referenced means higher revenues, higher margins, faster growth, and ultimately, we believe, higher stock prices that go along with that."
Terry runs technology and communications research at Citi and covered the sector at Goldman Sachs and Credit Suisse before that.
I listened to the full segment so you can skip it.
Here are the 3 calls that matter.
👤 Guest: Heath Terry, Global Head of Technology and Communications Research at Citi
🎙️ Host: Paul Sweeney, who anchors Bloomberg Surveillance and previously ran equity research teams at Credit Suisse
📰 Published: 10 September 2026 on the Bloomberg Surveillance podcast feed
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 6 min
Key Takeaways
Demand is rising into a supply base that cannot move, and the release valve has been price
Terry's read is that higher prices for inference, memory and power show up as revenue and margin at the companies selling capacity
The AI trade wobbled because Chinese open-source models gained ground while US frontier releases were paused
He says releases in the past week or two reversed that, and that is why the trade worked again
Towns are not fighting data centers as such; they are fighting the ones built without them
A Google site in Huntsville, Alabama on an old coal plant drew no objection; another an hour away put the town's leadership under NDA
The newest frontier models cost more per token and less per task, which is the number an enterprise buys on
1. Something Had to Break
Asked how he frames the AI investment theme now that clients already own the chipmakers and the hyperscalers, Terry described a collision between demand that keeps climbing and capacity that cannot be added quickly.
The mismatch is the whole setup. "Look, I think the moment that we're in right now is still best defined by what Sarah Friar at OpenAI called the vertical wall of demand that they're running into. What I would add to that is the far more horizontal line of supply that we have in data center infrastructure and capacity. And what happens when you get into these sort of immovable object, irresistible force kind of moments is something has to break. And that break has been pricing."
Every input is repricing at once, and Terry reads that as good news for the sellers. "And so, inference is getting more expensive, memory is more expensive, power is more expensive, everything in that supply chain is getting more expensive, which for the companies that you just referenced means higher revenues, higher margins, faster growth, and ultimately, we believe, higher stock prices that go along with that."
The trade's recent wobble came from China, not from demand. "We had this break, so to speak, over the last few months where there was concern about that overall AI trade because of the momentum that open source models, particularly those out of China, picked up relative to the U.S. models, which were kind of on a pause as the government restricted a lot of the frontier releases."
He dates the turn to the past week. "That's changed within the last week. You see the most recent frontier model releases. They have now reestablished that Western dominance, which is what you need for that trade to work. And I think it's a big part of why that trade's been working this week."
2. Moratoriums Hit New Permits
The host raised the local opposition to data center construction and the moratoriums some jurisdictions have passed. Terry drew a line between what is already under construction and what is still on the drawing board, then argued the objections are about method rather than principle.
The restrictions bite later, not now. "Look, the moratoriums, the concerns that you have around this, I think are largely going to be more of an issue sort of for future builds. You look at most of the moratoriums that have been issued, those are against new permits versus what's being built now."
The opposition is specific, in his account. "People don't have blanket issues with all data centers. They have issues with data centers that are being built the wrong way, that are being built without consideration for the communities that they're being built in."
He gave two sites in the same state as the contrast. Of a Google project in his home state: "There's a data center that Google's building in Huntsville, Alabama. It's being built on the site of an old coal-powered power plant, which means it's got connectivity directly into the grid. They're using nuclear and solar, so carbon zero on that. They worked with the community ahead of time to build it. No issues."
The other one: "You go an hour down the road, and there's another data center company that's trying to bulldoze 800 acres of farmland, use gas-powered generators on site, and has the entire town leadership under NDA associated with it. Those are the issues. Those are the data centers that people have issues with."
He welcomes the pressure rather than treating it as a risk to the theme. "And so I think what you're going to see in all of this, all of this political pressure, is you're going to see companies forced to do this the right way. And I don't think that's a bad thing."
3. Cost Per Task, Not Token
Asked to explain frontier models against the cheaper open-source competition, much of it Chinese, Terry said open weights were always going to matter — and then made the case that the metric the debate has been using is the wrong one.
Open source was never the anomaly. "Open source was always going to be a big part of AI the same way it was a big part of software. And we're seeing a lot of open source growth, not just from Chinese models, but from all of the Western models as well." He named Poolside, Cohere and Mistral on the Western side and DeepSeek, Qwen and Kimi out of China
Their advantage has been price against a specific kind of workload. "Those are getting a lot of traction because they have been price performant and they have been able to get a lot of usage, a lot of volume usage among a certain type of enterprise and a certain type of workload."
The newest frontier releases are both the most capable and, he says, the cheapest to finish a job with. "These models are now not only the most performant models in terms of absolutely pushing what can be done with AI, but they're also economically efficient when you look at it on a cost per task basis."
The measure everyone has been quoting is cost per token, and he argues it inverts the answer. "And what we're seeing now is that these newest models are actually a lot more efficient. And so even though their cost per token is higher, they burn fewer tokens to answer the question or complete the query or the task that's been assigned to it. So the cost per task is actually lower. And I think that ultimately is what's going to matter at the enterprise."
He stopped short of declaring the threat over. "So I don't want to say that we're completely past open source being a concern, but I do think it's going to be significantly less of a concern going forward."
Bonus Insights
The host introduced Terry as a technology analyst who has worked at Goldman Sachs and Credit Suisse, where the two of them were colleagues, holds an MBA from Columbia and did his undergraduate degree at the University of Alabama — the last of which drew the segment's running joke that Terry "peaked when he graduated from the University of Alabama"
Terry's Huntsville example is in that same home state, which is how the data center question got there
This is Terry's own segment of a longer Bloomberg Surveillance episode that also carried separate interviews with PGIM's Greg Peters on the bond market, Fordham Global Foresight's Tina Fordham on geopolitical risk and Yacktman's Molly Pieroni on value investing, each written up on its own
Terry's bottom line is that the constraint in artificial intelligence has moved from demand to physical capacity, and that as long as pricing is what gives way, the companies supplying inference, memory and power collect the difference — with the open-source discount mattering less now that the newest frontier models finish a task for fewer tokens.
Products, Companies & Tools Mentioned
OpenAI (Terry borrowed its chief financial officer Sarah Friar's phrase, "the vertical wall of demand," as the starting point for the whole argument)
Google (Building a data center in Huntsville, Alabama on the site of an old coal-fired power plant, running on nuclear and solar and, in Terry's telling, community-engaged from the start)
DeepSeek, Qwen and Kimi (The Chinese open-weight models he says have taken volume share on price-sensitive enterprise workloads)
Poolside, Cohere and Mistral (His examples that the open-source push is Western as well as Chinese)
Citi (Terry's employer; the cost-per-task framing is the research view he is putting to clients)
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: