Where the AI Capital Cycle Stands, and What It Means for Your Portfolio
- 8 hours ago
- 7 min read
Where the AI Capital Cycle Stands, and What It Means for Your Portfolio
You did not choose to own the AI build-out. The market handed it to you. If you hold a global equity portfolio today, a meaningful share of your wealth now sits in a handful of companies spending hundreds of billions of dollars a year on chips, power, and data centres. The position was built by index weighting, not by a decision you made at the kitchen table. That is the part worth thinking about before the cycle turns.
On The Artisan Podcast, Alpine Macro's Noah Ramos argued the AI build-out sits closer to the fifth inning than the ninth. His case: usable compute is still scarce rather than oversupplied, and the bottlenecks worth tracking are rotating from silicon toward power, batteries, and networking. Below is his framework, and the signals he says would change his mind.
Is AI a bubble, and why does it matter for a portfolio you did not choose?
The bubble question dominated the conversation, and Ramos took the unpopular side. His view, and Alpine Macro's house view, is that the cycle has not yet entered over-capacity territory.
His argument separates two things most commentators blur together: magnitude and duration. On magnitude, the intensity of capital spending is comparable to the late-1990s telecom build-out. On duration, he argues we are much earlier. To use his baseball frame, somewhere in the fourth or fifth inning, not the eighth.
The distinction matters because the telecom collapse was not caused by spending alone. It was caused by spending on infrastructure nobody used. Ramos points to the inverse today. Utilization rates on AI server racks are running above 80%, he says, and the constraint on going higher is not weak demand. It is networking, data-transfer speed, and latency.
Here is the figure that anchors his “still early” call.
According to Ramos, of the roughly 12,000 data centres operating globally, about 80% never touch AI. Of the AI-capable remainder, he estimates only about 5% can house the latest generation of chips. On his math, that leaves one to two percent of total capacity ready for frontier work, which is why he describes compute as extremely constrained rather than oversupplied.
He layers demand visibility on top of the supply picture. Frontier labs, he notes, each operate on less than two gigawatts of compute today and have stated ambitions roughly thirty times larger within four years. Cloud revenue backlogs at the large platforms are growing fast. None of that is a guarantee. But it is the difference, in his framing, between dark fibre that sat unlit for a decade and server racks that are full the day they switch on.
What would tell you the cycle has turned?
This is the practitioner question, and it is the one worth keeping. A thesis you cannot falsify is not a thesis. Ramos named four signals he watches.
The first is return on invested capital for the hyperscalers. Take a large platform's capital outlay, divide its realized profit by that outlay, and you currently get an expected return he places in the 18% to 20% range. As long as that number holds while spending climbs, the build-out is still earning its keep. Compression in that figure would be the early warning.
The second is capital-expenditure guidance. If one of the major technology leaders stands up on an earnings call and guides spending downward, that is the operators themselves signaling the top.
The third and fourth are about the technology, not the financials. Does AI keep improving, and can the labs serve inference demand while still training new models. Ramos described this as a “compute allocation dilemma,” a dual mandate not unlike a central bank's. A lab must reserve compute to train the next model and reserve compute to answer the queries that actually generate revenue. When the two compete, something gives.
You see, this is the link most market commentary skips. The bubble does not pop because a magazine runs a cover. It pops when the cash stops compounding, and the place that shows up first is usually credit. Risky credit has historically cracked three to six months before the major equity tops. That is a signal a family office can monitor without forecasting anything.
Why are the chips improving faster than Moore's Law ever delivered?
Ramos called Moore's Law “completely dead” as a description of what is now happening, and he meant it as an upgrade, not an epitaph.
The old rule doubled transistors roughly every two years. The gains he describes are steeper. From Nvidia's Hopper generation to the current cutting edge, he cited an improvement of roughly a thousandfold in compute over about four to six years, driven less by shrinking transistors alone and more by advances in packaging, in how chips are stacked and connected, and in lithography moving down the nanometre scale.
The investment consequence is the part worth holding onto. Each new chip generation does not only deliver more compute. It delivers radical improvements in energy efficiency per calculation and in the cost of producing a unit of intelligence, which for an AI operator is the cost per token. That is why he remains constructive on next-generation silicon despite the premium price. You pay up front to drive down the operating cost as usage scales.
There is a counterintuitive corollary. Faster innovation is extending, not shortening, the useful life of older chips. Ramos noted that cooling and software improvements are keeping three-to-seven-year-old chips running at high utilization, because the demand for AI is diversifying into tasks that older silicon handles well. The stack is not being thrown out. It is being sorted by job.
Where is the value moving as the cycle shifts from training to inference?
Training builds a model. Inference runs it. Training is an investment in future revenue. Inference is the return on capital already spent, earned closer to real time. Ramos argued we are crossing the threshold where inference monetization becomes visible, and that the shift changes which hardware matters.
The economics he laid out are stark. He cited the cost of producing a million tokens falling from roughly $412 on an older Hopper-class system, to roughly 12 cents on the current Blackwell generation, to a projected 1.2 cents on the generation after. At the same time, the amount a model can process at once has gone from a couple of thousand tokens to more than a million. Cheaper to produce, and more produced per transaction. Both sides of the margin move the right way.
His point for investors: the silicon that is best at inference is not the same silicon that is best at training. Inference leans more on custom chips and specialised accelerators and less on the top-end GPUs. As an anecdote, he noted that one leading lab's recent enterprise growth has run on relatively little Nvidia hardware, using chips from other large platforms instead. High-level compute, he argued, no longer has to mean GPUs alone.
Which bottlenecks come next?
Ramos described the cycle as a game of whack-a-mole, where capital floods to whatever layer is currently the binding constraint. First compute, then power, then memory, then networking. His map of the next constraints is the most portable takeaway from the conversation.
In the near term, two to three years, he points to advanced chip production and to grid interconnection. On his figures, the United States has roughly twice as much generation capacity waiting to connect as it currently has deployed, which means the short-term power bottleneck is interconnection, not generation.
Further out, he is constructive on power generation broadly, with nuclear life-extensions, natural gas, and the return of uranium enrichment to the West as themes, and on grid-scale battery storage. His battery argument is worth stating plainly: the immense innovation of the past decade was optimized for electric vehicles, where weight and space are constraints. Grid storage has neither constraint, which allows cheaper materials and larger cells. He also flagged photonic networking, tied to the space economy, and chip power-delivery as longer-dated niches.
A reasonable reader will notice that several of these stocks have already run a long way. Ramos did not dispute that. His framing was about where the structural demand sits, not about entry points, and the distinction between a good company and a good price is exactly the work that comes after a thesis like this.
What does this mean for jobs, and for the next generation?
Joseph asked the question every parent in the audience is carrying: what should a six-year-old aim for. Ramos's answer was not the headline version.
Yes, the work most exposed to AI is white-collar knowledge work: coding, writing, summarization, financial analysis, basic legal tasks. But he resisted the mass-displacement narrative in favour of a force-multiplier one. His radiologist example made the point. The job everyone expected to vanish has not, because reading images is the task, while helping patients is the job, and better image-reading has increased the demand for the human on top of it.
The durable ground, in his view, is work that requires physical embodiment, a human trust layer, or judgment that cannot yet be replicated. He noted a revealing asymmetry in the survey data: roughly 70% of Americans worry about AI putting people out of work, while fewer than a third worry about losing their own job. People may fear the idea more than the event.
How Northland thinks about a cycle like this
We do not earn a fee for predicting which inning the AI cycle is in. We earn it by making sure a family's exposure is the result of a decision rather than an accident. When a single theme drives index returns, broad-market portfolios quietly concentrate, and the families who feel the most pain in a correction are usually the ones who never knew how much they owned.
That is the conversation we have. What do you actually hold, across all your entities. How much of it rests on one cycle continuing. What would you want to have done before a turn rather than after. Mr. Ramos gave a framework for watching the cycle. Our role is to connect a framework like that to your specific balance sheet, your time horizon, and your tolerance for the drawdown that every cycle eventually delivers.
Author
Joseph Abramson, CFA, MBA is Co-Chief Investment Officer at Northland Wealth Management. Based in Montreal, he leads the firm's investment strategy and alternatives research and writes Northland's quarterly investment commentary. He hosts The Artisan Podcast, where he speaks with strategists and managers about the forces shaping global markets.
