Hyperscalers are paying three times an electrician's salary and twice the going rate for electricity to get a data center built, according to UBS Global Wealth Management's chief investment officer for the Americas.
The standard argument is that a higher cost of borrowing will eventually slow the artificial-intelligence build-out. Ulrike Hoffmann-Burchardi said it would take a great deal more than a quarter or a half of a percentage point to make any difference to it, because the companies spending the money believe the prize is the largest in commercial history.
"Yet when you look underneath the surface, they all have one common hidden factor and that is AI."
She spent nearly 25 years at Tudor Investment Corporation, the last 14 of them running the fundamental global equity portfolio and the flagship fund, before taking over investment direction for the Americas and global equities at UBS Global Wealth Management.
I listened to the full interview so you can skip it. 54 minutes of audio, 18 minutes of reading.
Here are the 14 takeaways that matter.
👤 Guest: Ulrike Hoffmann-Burchardi, chief investment officer for the Americas and global head of equities at UBS Global Wealth Management, who spent nearly 25 years at Tudor Investment Corporation
🎙️ Host: Dean Curnutt, founder and chief executive of Macro Risk Advisors, a derivatives strategy and trade execution firm, who has run this podcast since 2018
📰 Published: 1 September 2026 on the Alpha Exchange podcast feed
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 54 min | ✅ Time saved: 36 min
Key Takeaways
A quarter or a half point of rate rises will not slow the data center build-out The companies doing the spending already pay three times an electrician's salary and twice the base electricity rate Housing starts are at a three-and-a-half-year low on the same interest rates
A portfolio that looks diversified by asset class can hold one factor five times over Large-cap technology, utilities, real estate investment trusts, emerging markets and commodities are all exposed to AI capital spending
Circular financing between AI companies is an amplifier, not a trigger The trigger she watches for is demand disappointing, and she does not see that yet
Valuations tell you the size of a fall, not when it starts
Electricity, not chips, is the constraint she expects to bind by 2030 She puts the shortfall at roughly 100 gigawatts, against demand from AI, transport and industrial electrification together
If inference is profitable at the gross-margin line, the whole capex trajectory can be underwritten A frontier lab published its own inference economics, and the arithmetic is public
Cheaper Chinese models mean more AI bought, not fewer chips
Her own first job at Tudor, writing quantitative stock models, no longer exists She said it disappeared over about five months
1. 5 of the '90s Top 7 Are Gone
Hoffmann-Burchardi opened on her master's thesis at the University of St. Gallen, written in the 1990s on the convergence of audio, text and video into what was then called multimedia. The thesis argued that convergence threatened three separate industries at once: computing, media and telecommunications.
The seven largest technology companies of that era were IBM, Intel, Hewlett-Packard, Digital Equipment Corporation, AT&T, Apple and Microsoft, and only two of them are among today's seven largest. Digital Equipment was later bought by Hewlett-Packard
The point she drew from it is that a technology wave creates growth and destroys incumbents in the same move. "The internet digital economy is a large part of the US and world economy now, but is also very disruptive and changes leadership in the market."
Curnutt pushed on the second half of her academic training, political science alongside economics, and noted that when she took the PhD, negative interest rates were not yet a thing anyone had to model
2. Markets Beat the Lab
Her doctoral work was in financial econometrics, which she said also gave her a first look at what was then the frontier of artificial intelligence: machine learning.
The reason she left academia was that the harder laboratory was outside it. "When you actually study financial economics or econometrics, you learn pretty quickly that the most challenging laboratory is not academia, it's the markets themselves."
She joined Tudor in 1999, starting in its offices southwest of London rather than in the City
3. A Loss Hurts Twice As Much
Curnutt framed this section with a story of his own, from a recent Paul Tudor Jones podcast appearance about the Hunt brothers cornering the silver market in the early 1980s.
The host's own anecdote: "And his quote just sort of knocked me off my chair which was one of the Hunt brothers was the richest person in the world by a factor of 3x at least for a period of time when they ramped up silver and then it all came crashing down."
Hoffmann-Burchardi spent almost 25 years at Tudor across quantitative macro, global tactical asset allocation, a quantitative equity business she helped build, and 14 years running the fundamental global equity portfolio and the flagship fund
She described the firm as an unusually academic place for a trading floor, where a colleague kept a book on the seven unsolved millennium problems on his desk and the lunchtime subjects were the Fermi paradox, causality against correlation, and the singularity
The behavioral arithmetic is what makes trading hard, and she cited Daniel Kahneman for it. "Kahneman I always love they're finding that when you experience a loss in trading it's twice as painful as the equivalent gain gives you joy." She added: "So that 2:1 ratio is pretty challenging" Her claim about Paul Tudor Jones is that the culture improved that ratio rather than the arithmetic, and that this is why portfolio-manager tenure at the firm runs longer than the industry average
4. Nothing New Under the Sun
One sentence is what she took from a quarter century there. "I think the biggest lesson for me or the biggest takeaway from my whole 25 years there was this one quote is that there's nothing new under the sun." It comes from "Reminiscences of a Stock Operator", the fictionalized life of the early-1900s speculator Jesse Livermore, which she said Paul Tudor Jones has given every new trader at the firm for more than 25 years
The mechanism underneath the phrase is that the swing between fear and greed is constant even when the events are not. "History might not repeat exactly, but it rhymes."
Three trading lessons came out of it. Ride trends, but pair that with a willingness to sell negative convexity repeatedly, which means taking on positions that lose disproportionately in a large move in exchange for income On sizing: "So you might not need a 60% hit ratio if you can size appropriately at the right time." Where conviction is high, she said, the position has to be large, and options are one way to express that On liquidity: "We have seen over and over again that certain times in the markets we think that certain products are liquid only in times of duress to find out that there's suddenly no bid"
5. Blend Macro and Bottom-Up
Asked which episodes taught her the most between 1999 and 2020, she answered with a method rather than an event.
Crises arrive from company balance sheets as often as from the macro data. She named Bear Stearns, Lehman Brothers and Silicon Valley Bank, and said an asset-liability mismatch, meaning short-term funding against long-term assets, is the marker of vulnerability
The macro lens still has to be there alongside it. She called the Federal Reserve's December 2021 pivot a very important signal for equities, particularly at a macro firm
The organizational fix is having portfolio managers with different focus areas in the same meeting. "The famous computer scientist Alan Kay likes to say that perspective is worth more than 80 IQ points"
Curnutt tied perspective to time on the desk, and made an aside about the market unwind a month earlier: "You put a lot of capital in the hands of someone that clearly, had a tremendously interesting insights on AI, but at 25 years old, it's just hard to gather a lot of perspective along the way." He attributed that view to others rather than stating it as his own
6. Three Lenses, Not Two
At UBS she covers all asset classes, public and private, for the Americas, and global equities. She put assets under management at "7.3 trillion".
Portfolio construction, in her description, is three inputs: expected returns, the risk of each investment, and the correlations between them, balanced against what the client is trying to achieve
Most firms run two lenses, macro and bottom-up. Her team runs a third: structural trends. The three she has picked are AI, electrification, which they label power and resources, and longevity The teams are organized to match: top-down macro analysts, bottom-up stock analysts, and three global teams on the three trends
The same three lenses are used to find risk, not only return. She starts with wherever inflows and excesses have been largest, which gives her two: the AI capital spending trajectory, and the amount of outstanding public debt Her comparison on the second: "The balance sheet of government is certainly much more stretched than the balance sheet of the largest cap companies in the world"
7. Rates Barely Dent AI Capex
Curnutt asked what would tell her the long end of the bond market had become disorderly enough to threaten the AI trade.
She said the Federal Reserve is stuck between two economies running at different speeds. "You think about the housing market and the data this week showed we are at a 3 and 1/2 year low in terms of US housing starts." Against that, AI capital spending looks unconstrained
The size of rate move needed is not the size being discussed. "I do think it will take quite a bit. Not 25 50 basis points." The reason is what the buyers are already absorbing on cost: "And the reason is that right now already these hyperscalers are paying three times the salary of an electrician to build a data center." She added: "They're paying twice the base rate for electricity in order to build data centers"
Her explanation for the insensitivity is competitive, not financial. The companies spending believe being first to artificial general intelligence is worth a great deal, and they have watched incumbents get displaced before, which is the same 1990s story from the first section
Curnutt's summary of the asymmetry: for a homeowner the current rate is restrictive, but for an AI buyer the cost of capital is small next to the size of the outcome being chased. "So for a homeowner, and Warsh has said this, it's restrictive"
A separate risk sits behind all of this, she said: a confidence crisis in the Treasury market over whether the US can repay its debt would arrive independently of anything happening in AI
8. Power Is the Binding Limit
The risks she watches on AI are supply and demand bottlenecks rather than the interest rate. Permitting is first on the list, and she expects it to slow into the midterm elections because political opposition to data centers is already visible in Republican primaries
The shortfall she expects is in electricity: roughly 100 gigawatts by 2030, counting demand from AI alongside transport electrification, building electrification and an electrified industrial base That translates into periods of shortage in turbines, transformers and other grid components
On the demand side, the question is absorption: whether organizations can take up AI capability fast enough, and what happens if monetization disappoints against what the market has priced
The cheaper-Chinese-model argument does not worry her, and she inverts it. "That only means we'll have more AI because the cheaper AI gets, the more we're going to use it"
Her evidence that use cases are still multiplying is her own former job. She said coding has become autonomous enough that a developer watches the code check itself, and the quantitative-model-building role she started in at Tudor is gone: "That job doesn't even exist anymore and that probably happened in the period of 5 months" She also pointed to a run of new mathematical proofs and to drug discovery in biochemistry
The prize she thinks is being competed for is the wage bill. "It addresses probably at a lower boundary, at a minimum the 50 the 50 trillion of knowledge worker salaries"
9. Offense, With Defense Too
The team runs dedicated portfolios in each of the three structural trends, which it calls the trios, for transformational innovation opportunities.
AI is the connector across all three, she said: power and resources are what build and run the data centers, and longevity is a beneficiary through drug discovery and the rest of the healthcare chain
The portfolios are built to attack and to absorb at the same time. "So we have as CIO portfolios that we recommend to our clients to position for these opportunities to play offense but we always like to build in defense when we are playing offense as well"
Inside the AI portfolio, the positioning moves along the value chain — the enablers and chip and networking companies, the intelligence layer of frontier model builders, and the application layer of companies putting AI into advertising or enterprise work Her current view is that parts of semiconductors have already run past what is coming
In the power and resources portfolio, the team moved into the materials sector early in the year, on the argument that the build-out consumes commodities
Longevity is the slowest of the three and she said so plainly. Biochemistry is harder than a math problem, and how the human cell works is still not understood. "We now have the first drug in phase three clinical trials in the US." Many more AI-discovered drugs are in phase 2
The evidence that the balance works is in the daily returns. On days when the semiconductor and AI-enabling trade sells off, she said the longevity portfolio, currently weighted to pharmaceuticals, has produced positive returns
10. One Hidden Factor: AI
Curnutt raised an exercise she had run on her own podcast, Signal Over Noise, and noted the puzzle behind it: macro forces look ubiquitous, yet day-to-day correlations inside the equity market are very low. "It's almost as if the S&P 500 internally diversifies itself."
Her argument is that standard risk models are built on growth, inflation and real rates, and have no term for the structural trends that have become macro forces. AI is the one she says has to be added to a portfolio stress test
The worked example is a portfolio that passes the pie-chart test and fails underneath it. A large-cap technology stock, a utility, a real estate investment trust, emerging markets and commodities look like five different asset classes The hyperscalers earn from cloud spending; the utilities, particularly the independent power producers, from the power demand; commodities go into building the data centers; data center real estate is part of the build-out "And then lastly emerging markets have been dominated by three stocks and those stocks are all semiconductor companies that are very key to the buildout of the data centers"
The conclusion is one line. "So the one variable that they all have in common is AI capital expenditures"
11. Demand Is the Catalyst
Asked what an AI shock factor actually looks like when superimposed on a portfolio, she broke it into three.
Supply bottlenecks are temporary, so the way to play them is to own whatever is scarcest and therefore has pricing power. Supply eventually arrives Her framing of why they occur at all is that AI capability is compounding while the human-built parts of the chain expand in a straight line
Demand is the harder one, because it is where an expectation gap would show up. A gap between the monetization the market has priced and what arrives is what would cause ripple effects
On the circular financing arrangements between AI companies, she made a distinction that separates her from the bears. "It's not a risk in my mind as a catalyst. The catalyst will be if demand disappoints." Those structures mean more players are hit when it happens, so the downside is larger, but they do not start it
The same logic applies to valuation. "Valuations themselves are not a catalyst for a market correction." And: "It's a fundamental disappointment. That's the catalyst and valuations just tell you how much the market is likely going to fall and correct"
She also said the circular deals are not new: developers and landowners have done similar things for more than a hundred years, and Intel has done many of them over past decades. What is new is the scale, and therefore the coverage
She sized the mismatch between price and revenue. The appreciation in public and private markets associated with AI is, on her rough estimate, "about 25 trillion", while revenue generated directly from AI in the application layer is "less than a trillion" — which makes the question how quickly the market has discounted the benefits, rather than the return on investment itself
Where the balance comes from: healthcare, consumer staples and quality businesses, whose earnings visibility does not depend on the build-out, plus short-dated Treasurys as the ballast against every scenario, including a disorderly long end
12. Inference Turned Profitable
Curnutt put the pricing problem directly: markets are pulling forward enormous future benefits, and at some point corporate profitability has to arrive to justify it.
Her answer starts with the economics of running a model rather than training one. "There have been references to the inference workloads now being profitable for the frontier models." She pointed to a public blog post from a Chinese frontier lab last year showing decent gross margins on inference once the stack is optimized, with the arithmetic published on GitHub
The reason this matters is the mix shift. As models approach artificial general intelligence, less of the spend goes on training and more on usage, so inference economics become the economics
If inference clears at the gross and operating margin line, the trajectory is underwritable. "If you can establish that inference is profitable from a gross margin and an operating margin perspective, I think the market will underwrite that and will underwrite the trajectory from here." If that gets questioned, she called it another very big risk factor
Looking to 2030, she said the industry will have to raise another multiple of trillions, in equity or debt — "four potentially 5 trillion in additional capital"
The compute requirement behind that number comes from a UBS CIO paper that estimated the floating-point operations per second needed by 2030 for inference and enterprise and consumer AI alone, before physical AI is counted. It concluded UBS needs "five times the amount of compute that we had installed at the end of last year"
Her base case: "Again, the biggest risk is bottlenecks on the supply side, but the trajectory is only going to be going in one way and that is up"
13. Plan the Scenarios First
Curnutt described a volatile July that the VIX did not reflect. "You wouldn't really know it looking at the VIX. I think it got to 21, but underneath the surface there was just some gigantic factor rotations, momentum rotations, and so the presence of investors in the markets can matter too."
Her answer was process, in three parts: scenario planning first, then robust optimization, which she described as assuming a level of uncertainty and checking the portfolio still sits inside its risk limits even in the bad cases
The part she stressed is that scenarios interact. Rates high enough to make capex financing difficult would stress private markets, raise questions about the build-out, take the equity market down, remove the wealth effect and reach consumer spending, which is where the economy's strength is coming from
"So building in robustness to downside cases is very important." Alongside it, she said a long-term portfolio still needs short-horizon flexibility, which the team gets inside the equity portfolios and through tactical asset allocation
14. Where Humans Still Add
Curnutt asked how AI has changed her own process, given her data science background.
On research, the change is already large. "AI has certainly changed the way you can analyze things in quite profound ways especially of course agentic AI where you can now ask agents to do certain tasks for you and help orchestrate big research reports"
On portfolio construction, the old tools are still the tools. Mean-variance optimization, predictive quantitative models and factor models remain relevant, though she expects generative AI to produce its own set of quantitative models, taking in text and unstructured data as well as numbers, and treats that as a matter of time
Her claim for the human is a specific one, not a general one. "But at this point and you mentioned it earlier perspective and experience do matter in our view and asking the right questions, looking for hidden factors, thinking maybe outside of just pattern recognition"
The other piece is the client relationship: understanding risk appetite and needs, and putting a portfolio together against them. Trust, she said, is the variable that is hardest to replace
Bonus Insights
Curnutt agreed on the last point in his own terms, calling trust a uniquely human-to-human asset and one that is valuable
Asked what she looks for in someone leaving school for markets, she gave one reason to come. "I would encourage them to look into a career in markets. I think there is nothing that's more interesting because you always have to imagine the future." And: "Whether it's on where you position from an investing from a return lens or to think about how you manage risk, it's probably the purest intellectual pursuit outside of academia"
Curnutt raised Treasury Secretary Scott Bessent's activist interventions, including through the yen and Treasury buybacks, as part of the backdrop to the long end of the bond market; the conversation moved on to the AI question without settling it
On the yield curve, Curnutt argued that the pickup for extending from two-year to ten- or thirty-year maturities is not enough compensation given how fast the environment is changing. Hoffmann-Burchardi's position is short duration, which she said is where the ballast is across every risk scenario
The two had met a few months earlier at a conference, which is where the conversation started
Hoffmann-Burchardi's bottom line is that AI capital spending has become a macro factor that most risk models do not contain, that it is close to insensitive to interest rates because the prize is the entire knowledge-work wage bill, and that what would break it is a shortfall in demand rather than a valuation nobody likes.
Products, Companies & Tools Mentioned
UBS (Her employer; she put group assets under management at "7.3 trillion" and described the CIO's job as setting investment direction across public and private assets)
Tudor Investment Corporation (Where she spent almost 25 years, and the source of the trend, sizing and liquidity lessons she carries into portfolio construction)
IBM, Intel, Hewlett-Packard, AT&T, Apple and Microsoft, plus Digital Equipment Corporation (The seven largest technology companies of the 1990s, five of which are no longer in that group; Intel also came up as a company that has done many circular deals over past decades)
Google, Meta and Amazon (The platforms that emerged in the 2000s and took share from those incumbents, which is the displacement pattern she says AI buyers are trying not to repeat)
Bear Stearns, Lehman Brothers and Silicon Valley Bank (Her examples of crises that began in company balance sheets, where an asset-liability mismatch was the marker)
DeepSeek (The Chinese lab whose published inference economics she cites as evidence that running models can be profitable at the gross-margin line)
Books & Resources Mentioned
Reminiscences of a Stock Operator (The fictionalized life of Jesse Livermore, and the source of "there's nothing new under the sun"; she said Paul Tudor Jones gives it to every new trader at the firm)
One More Thing: DeepSeek-V3/R1 Inference System Overview (The public post whose inference gross margins she treats as the evidence the AI capex trajectory can be underwritten; the arithmetic is on GitHub)
Signal Over Noise (Her own UBS podcast, where she first ran the hidden-AI-factor exercise Curnutt asks her to repeat here)
The UBS CIO compute paper (Its estimate of the floating-point operations per second needed by 2030 is where the five-times-current-compute figure comes from)
The Millennium Problems (The book on the seven unsolved problems that a Tudor colleague kept on his desk, in her illustration of the firm's culture)
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: