GARPify · Full Picture
Data Centre Boom (THEME)
Price is in a green uptrend while the multiple expands — the multiple is doing much of the lifting. Over the past year price +78% decomposes into earnings +49% amplified by a multiple +20%. The engine factors (ROIC, FCF, margins) are near their own highs; the one yellow flag is earnings yield.
| Factor | Direction | 1-year | 5-year (per yr) |
|---|---|---|---|
| Earnings (EPS) | rising | +49% | +29%/yr |
| Multiple (P/E) | cheapening | +20% | -4%/yr |
| = Basket | uptrend | +78% | +23%/yr |
How to read it. Every factor reports two things — where it sits in its own history (level) and which way it's moving (direction: up / down / range-bound), sign-unified so up = constructive. Price = earnings × multiple, so the top two rows multiply exactly to the price line. Context rows confirm the engine but don't enter the price math. Sources. YCharts (price, P/E, EPS, ROIC, margins, revenue). Scope. Describes condition across factors — not a score, rating, or buy/sell call.
The Big Story
Watch the money change hands
There is a simple problem with a big number. When a giant company announces it will spend billions, you have no real way of knowing whether it means it. Announcements are free. Plans get quietly cut. So the trick is not to listen to what the giants say they'll spend. The trick is to watch what actually leaves their hands — and where it lands.
Start with what changed. For thirty years a data centre was just a warehouse of servers that grew along with the internet. Then it turned out that the same chip that draws a video-game frame is the chip that trains an artificial intelligence. And the four companies that essentially own the internet — Amazon, Microsoft, Google, Meta — began buying that hardware at a scale nobody had ever seen.
How much? Roughly $410 billion in 2025. And instead of easing off, they told their own investors they'd spend somewhere around $630 to $725 billion in 2026 — up about 77% in a single year. That $700-billion-odd is a hard number to feel, so here is a yardstick: it is nearly double what the entire Canadian federal government spends, on everything, in a year.
Those are still just announcements. Here is the part you can actually check: a large slice of that money has already changed hands. It sits as signed backlog — contracts, not hopes — on the books of the companies that sell the picks and shovels. That is the whole difference between a story and a stock. And it is why this basket's earnings keep compounding, and why, over time, its price has followed those earnings up.
Why does any of this matter to your money? Because of an old idea worth holding onto: over the years, a company's share price tends to follow its earnings. Month to month, the market's mood swings the stock around — but across years, it is the earnings that decide where it lands. So a group of companies whose earnings keep climbing is a group whose share prices have, over time, tended to climb with them. That is the bet this theme is really about.
What the money builds
The thing itself — and its scale
It helps to picture where the money actually goes. Somewhere outside a mid-sized town — Texas, Virginia, Arizona, it hardly matters which — a windowless building is going up on land that was a field eighteen months ago. From the road it looks like an ordinary warehouse, except for the electrical substation beside it, which is about the size the whole town's substation used to be. It will cost more than a billion dollars before it opens its doors. And here's the strange part: the most expensive machines that will fill it are already scheduled to be obsolete before the parking lot is repaved for the first time.
That last detail is the whole business in a sentence. These buildings burn through their most expensive contents every few years — which is precisely why the companies that supply those contents have such a long runway. The building doesn't get equipped once; it gets re-equipped, over and over.
And they are going up at a pace with no precedent. By one industry count, roughly 770 of these facilities are in the construction pipeline right now — enough to double the world's data-centre capacity in about three years, on top of the ~1,360 already running. Not over a generation. In three years.
What the spenders are saying
Ask the miners, not the man selling shovels
If you want to know whether a gold rush is real, don't ask the man selling shovels. He'll always tell you it's real. Ask the miners — the people actually spending the money. In this boom the miners are the four giants who run the world's cloud — the ones the industry calls hyperscalers — and in their own earnings calls they are not hedging. They are racing each other.
| Buyer | 2025 spend | 2026 plan |
|---|---|---|
| Amazon | ~$125B | ~$200B |
| Alphabet / Google | ~$91B | ~$185B |
| Meta | ~$72B | ~$125B |
| Microsoft | ~$90B | ~$120B |
| The big four | ~$410B | ~$630B+ |
The line that keeps recurring on their calls is the opposite of caution: they can't build capacity fast enough. When four of the most disciplined spenders in business all decide to nearly double their outlay in the same year, and say out loud they still can't keep up, that is demand confirmed by the people paying for it.
Sources: company FY2025 results & 2026 capex guidance (Microsoft, Amazon, Alphabet, Meta). Full sources in the theme fact sheet.
Where the money lands
The part you can verify
Now follow the money one step further, to where it actually arrives. This is what turns a headline into a theme: you can already see the spend showing up as signed orders on the sellers' books.
Broadcom is carrying a $73 billion backlog of committed AI orders — not a pipeline, actual purchase orders — and its chief executive told investors he has "line of sight" to more than $100 billion of AI revenue in 2027. Eaton is sitting on a record $19.8 billion order book. Vertiv, about $15 billion. Add just those three and roughly $108 billion of the giants' spending is already contracted. The other two operators don't even quote a number: Nvidia's data-centre demand runs ahead of what it can supply, and TSMC's leading-edge factories are effectively sold out.
Put the two halves together and the story closes on itself. The buyers are nearly doubling a ~$700B-a-year spend and saying they can't build fast enough. About $108B of it is already signed onto these five companies' books. And the earnings that fall out the far end are what, over time, their share prices have followed. The rest of this report is about the five operators catching that money — and how well each one is doing it.
Sources: Broadcom Q2 FY2026 results & earnings call ($73B AI backlog, ">$100B in 2027"); Eaton & Vertiv FY2025 filings; NVIDIA / TSMC FY2026 disclosures. Every figure here is a reported order book or disclosed demand signal, not a forecast. A description of the record, not a recommendation.
Operators
The five operators catching the money
The story and the money are settled; what's left is execution. A spend only turns into earnings if the businesses standing in front of it are well run — the right spot on the curve, and the right people running it. Below is the one-line read on each of the five: what they do, who leads it, and a link straight into the full analysis in the company's own report.
Five companies, five very different jobs. Here's the short read on each — the full story lives in its own report.
Huang made a strange bet years early: he handed his software free to university labs long before anyone was training AI on it. So when the moment came, the whole field already spoke NVIDIA's language. Today he ships a new generation every single year, and the industry builds to his calendar. The steepest curve in the basket, run by the man who drew it. See how —
NVIDIA Leadership →Everyone designs the chips. Almost nobody can actually make them at the leading edge. TSMC can, and Wei's whole job is to stay one generation ahead and build enough capacity to meet demand he can already see coming. Right now that capacity is effectively sold out. The one link in the chain no one can go around. See how —
TSMC Leadership →Tan built Broadcom one acquisition at a time, and pointed the whole thing at AI. Its custom accelerators and networking silicon are the second pillar of the data centre, right beside NVIDIA. The proof isn't a pitch — it's the order book: a $73 billion AI backlog, and a stated line of sight to $100 billion of AI revenue by 2027. See how —
Broadcom Leadership →Strip an AI data centre down to physics and the problem is heat and power — feeding the chips and stopping them from cooking. Vertiv sells exactly that. Albertazzi took a sleepy business and turned it into a margin turnaround riding the build-out, with about $15 billion of backlog to show for it. See how —
Vertiv Leadership →Between the utility line and the machine sits a lot of electrical equipment — and a hyperscale data centre needs an enormous amount of it. That's Eaton's aisle. Ruiz's job is disciplined portfolio management pointed at the electrification wave, and it's carrying a record $19.8 billion order backlog into the build. See how —
Eaton Leadership →Each capsule is the short read. The full story on each operator — the curve, the track record, how they spend their money, and how their promises compare to what they delivered — is in the linked report. As always, GARPify lays out the facts and leaves the call to you.
What's Changed
The short read since last update
As of 21 August 2026
The story is durable. The tape moves. Here's what actually shifted since last time — and the one thing that would make us tear the story up.
The return, valuation and price figures here are as of 21 August 2026; the guidance and backlog moves trace to each operator's latest filing. This is a status update, not a recommendation.
The Numbers
Three questions decide whether a compounding thesis is worth owning — the basket answers them in order.
Earnings Growth
Over time, price follows earnings. For a theme the test is whether the basket’s growth is real — more customers, more volume, healthy pricing — or flattered by buybacks and one-offs — and whether something durable is paying for it. This tab separates the two and asks whether the engine can keep running for years.
Read each chart like a price chart: the actual metric (green when rising / red when falling) with its trend line (grey dashed). On mean-reverting measures like margin and the share count a tinted band marks the level versus the basket's own median; on cumulative-growth measures (revenue, EPS) it carries no signal and is omitted. These are composite, basket-level figures.
Combined revenue across the five operators has reached $440B, compounding roughly 22% a year over the past eight years. This is the top line - real customer demand - so its slope is the evidence that the basket's earnings growth is backed by sales the businesses actually made.
The equal-weight average of the five operators' earnings per share now stands at $7.25, compounding roughly 17% a year over the past eight years. The actual line rides above its rising trend: the bottom line of the whole theme is doing the work, not just the story around it.
Taken as one business, the basket keeps about 54 cents of operating profit on each dollar of sales - up roughly +22 points over five years. Margin this high means each extra dollar of revenue drops heavily to profit, so the expansion multiplies revenue growth on its way down to earnings.
The combined share count of the five operators is about 34.9B - it has held roughly flat (-1% over five years). Because the float has not been shrunk, none of the per-share growth is manufactured through buybacks: what shows up in EPS comes from the businesses, not from a smaller denominator.
How it adds up - and what it has done to the price
Putting it together. The four lines describe an earnings engine running on real fuel: combined revenue compounding about 30% a year, the basket's aggregate operating margin still expanding at ~54%, and a share count that has held roughly flat - so equal-weight earnings per share has grown roughly 26% a year without help from buybacks.
Price is simply earnings multiplied by the valuation the market puts on them. Over the past 5 years the basket returned about +54% a year (8.6× in total) — earnings supplied roughly +26% a year (3.1×) while the multiple expanded about 22% a year (2.7×, basket P/E 17 to 46). The gains here have been driven by both rising earnings and a richer multiple.
For a single company the earnings question is 'real growth or financial engineering?' For a basket it widens to whether the whole group's earnings power is growing - and here it plainly is, across revenue, margin and the share count alike. Whether that continues is a question of demand, which the next tab takes up.
Where the Earnings Land
The basket grew earnings at roughly 32 percent annualized over the window and is currently accelerating. For most of that period the price tracked the earnings closely; in the most recent stretch the price ran somewhat ahead. The thesis rests on earnings continuing to grow at rates that justify both the current price and the recent multiple expansion. The growth driver is identifiable and externally visible - hyperscaler capital spending - which is the tab's reassurance and its central risk at once.
Whether the current price is a reasonable entry point — measured against what these businesses have normally cost — is the Valuation tab's question. What this tab establishes is that the earnings underneath the basket are real, large, and still growing. The Risk tab examines what would break the thesis at the operator level.
Future Earnings
Every durable theme rides an adoption S-curve: a slow base, a steep middle as the shift compounds, then maturity. T. Rowe Price built a firm on catching the wave early. This tab places the basket on that curve — is the runway still long, or is the easy growth behind it? — because where you sit on the curve shapes every other reading.
Where this theme sits on the technology growth cycles that drive it
Imagine a child who wants to build the world's biggest lemonade stand next to a highway. She has the money and the customers. But the town approves only so many permits a month, so she cannot build any faster than the permits allow. The permit is the binding constraint - not demand, not capital.
That child is every large technology company today. Their permits are data centres: the buildings have to be permitted, powered, cooled, and constructed before any chip can be sold into them. Right now the binding constraint on the whole theme is physical - power and construction, not demand. This is the curve that matters most, and we return to it below.
What an S-curve is
An S-curve - or sigmoid - is the shape almost every lasting technology shift follows as it spreads through the world, and it has three phases. A slow base, while early adopters experiment and most people ignore it. A steep middle, where the shift compounds and adoption accelerates - this is where the bulk of the growth, and of the investment returns, is made. And a maturity plateau, as the market saturates and growth slows. The investor's whole job with a theme is to judge where on this curve it sits: the same business can be a wonderful holding in the steep middle and a poor one at the plateau. The hard part - as the next section shows - is that the top is rarely visible until you are past it.
The shape of an S-curve. Adoption is slow at first, steep in the middle, then levels off - the returns are made in the middle, the mistakes are made buying near the top.
Why S-Curves Matter for Investors
In December 1999, Cisco Systems briefly became the most valuable company in the world, selling the equipment that connected the internet. Within thirty months it had lost about ninety percent of its value. The internet was real; it just was not yet generating enough revenue to pay for all the equipment being bought in advance of it.
That is what S-curves help investors avoid. An S-curve tracks how a technology spreads - slow at first, then steeply, then leveling off. The expensive mistake is buying near the top while mistaking a maturing curve for an accelerating one. The trick is that you can rarely see the top until you are past it. GARPify's approach is to track several curves at once, with a clear hierarchy of which drives which.
The buildout — where the earnings come from
Of the basket's operators, the future earnings that are easiest to see sit in the physical layer - the power and cooling equipment that turns data-centre plans into operating capacity. It is the slowest part of the theme to move - you cannot energize a substation as fast as you can sign a contract - which is exactly why it has the longest runway: years of committed work still to deliver. Two operators in the basket earn here, and the numbers below describe a physical buildout compounding off a low base.
The buildout earns through the companies that move the power and cool the racks. Combined revenue at Vertiv and Eaton has grown from about $20B to $39B - Vertiv alone has more than doubled as data-centre work crowded its order book, while Eaton's electrical business compounds more steadily. This curve moves at the pace physical capacity allows, not the pace contracts are signed - which is exactly why its runway is long: years of committed work still to deliver.
The picture is consistent: a physical layer compounding fast off a small base, with the quickest growth in the cooling and mechanical work most directly tied to data centres. The part that matters for a forward read is that much of this revenue is not a forecast at all - it is already contracted, sitting in backlog, which is where the tab turns next.
The backlog — earnings already on the books
There is one forward read more concrete than any curve: the backlog. A backlog is revenue a company has already signed but not yet delivered - contracted, not forecast - and in this basket it is most visible in the physical layer. Vertiv builds to order, and its reported backlog has climbed to record levels alongside the data-centre work, with a rising share tied to large technology facilities; Eaton's electrical order book and backlog have expanded on the same demand. A backlog at record highs means a meaningful slice of the next year or two of these operators' revenue is already booked - the clearest evidence that the buildout curve has runway the income statement has not yet caught up to. It cuts both ways: backlog is also the early-warning gauge. When book-to-bill slips below one and the backlog stops growing, it shows up here before it shows up in revenue - which is exactly why we read it on the way up and on the way down. (Backlog figures are reported in each operator's filings; the company reports behind this theme track them quarter by quarter.)
How Today Compares to Past Cycles
Place the Data Centre Boom alongside three earlier hardware cycles - PC chips in the 1990s, internet infrastructure in the late 1990s, and smartphone chips from 2008 to 2020. At roughly three years past inflection, which is about where this theme sits, all of them looked similar. They diverged in what came next.
PC chips kept compounding for another five years and ended through maturation, not collapse. Smartphone chips peaked around year four and plateaued. Internet infrastructure accelerated for three more years and then collapsed inside twenty-four months. All three outcomes are possible from here, and the factor that most distinguishes them is the same gap between spending and revenue: the PC cycle closed it favorably, the internet cycle never did.
Both lines are rebased to 100 at the start of the five-year window. For most of the window the two move together: when earnings rose, the index rose. From late 2024 onward the price line pulls clearly above the earnings line - the price index compounded to roughly eight times its starting level while equal-weighted earnings grew to a little under four times theirs. That gap is the multiple expansion that the Valuation tab examines.
The honest read: for most of the window, the return came from compounding earnings. In the most recent stretch, part of the return came from earnings that were still growing fast, and part came from investors paying a higher multiple for those earnings. If earnings keep compounding, the multiple eventually grows into the price; if earnings decelerate, the price corrects toward the slower earnings line.
Valuation
A wonderful theme and a wonderful entry price are not the same thing. Buffett’s line holds for a basket too: price is what you pay, value is what you get. This tab sizes up what the basket costs today — against its own history and against the growth it would have to deliver — so you can judge whether today is a fair entry.
What you are being asked to pay today
Howard Marks's most practical insight is that there are no good ideas, only good ideas at the right price. Reading price discipline for this basket takes two readings: how it is priced against its own history, and how it is priced against the growth ahead. This tab does both, briefly.
Charts read like a price chart: the actual basket multiple (green rising / red falling) with its 1-year average (grey dashed); the amber dashed line is the yardstick — the basket's own 5-year average, or the 1.0 PEG fair line. This is about what the market pays for the group's earnings and growth, not market timing.
The basket trades at 41.8× trailing earnings — below its own five-year average of 46.5× (the dashed line). So against its own history the theme is priced a little under normal rather than stretched: when the line sits below the dashed one, the group is cheaper than it has typically been. What the single line cannot show you is how far apart the individual names sit underneath it.
Divide that multiple by the 47% the group is growing earnings and the PEG is about 0.9 — right around the 1.0 fair line. Taken as a group, you're paying roughly a dollar per point of growth: priced about where the growth justifies, neither a bargain nor a stretch.
A History of the Basket's Sentiment Periods
Equal-weight basket P/E mapped to the sentiment cycle, 2015–2026.
Sentiment is easiest to read in hindsight. The basket's equal-weight multiple traces the crowd's mood toward the data-centre operators over the past decade — from indifference before the AI era, through euphoria, to a present that reads as optimistic rather than frothy.
The basket's sentiment has travelled from indifference, through euphoria, to a present that is optimistic rather than frothy. The pendulum is no longer at greed; it is on the optimistic side of center, which is historically where the more interesting entry points have appeared.
How it adds up — the basket's price against its own history
Against its own five-year history the basket sits a little cheap: 41.8× today versus a 46.5× average, about 10% under normal. Divide that by the roughly 47% the group is growing earnings and the PEG lands near 0.9 — about fairly priced for what it is compounding. Quality at a quality price.
The average is doing a great deal of hiding, though, and the spread underneath it is the more useful read. NVIDIA trades 53% below its own five-year norm — the share price rose, and the earnings rose faster. Eaton sits 32% above its norm and TSMC 24% above; in both, the multiple has moved further than the business. Broadcom is 3% above, which is to say priced almost exactly like itself. The cheapest name in the group is 53% under its own history and the dearest is 32% over it. That is one basket and no two identical price decisions.
A basket changes the risk more than the math. The exposure here is spread across five businesses rather than riding on one bet, so a single disappointment doesn't sink the thesis — but neither does a single blow-out rescue it. What it does not do is make the entry price equally attractive in every name, which is why the last column of the comps table is worth more of your time than the basket average.
The risk is the one every growth story carries: the multiple holds only while the growth does. If the group's earnings decelerate, the PEG rises above 1 even if prices never fall — and the names already carrying a premium to their own history have the furthest to fall back. For now the basket is priced fairly against what it is delivering — whether that delivery continues is the question the company reports behind this theme are built to answer.
Ways to Play
There are two ways in: buy the individual operators, or buy a fund that holds the basket. This tab lays out both — the operators side by side on quality, growth, price and the job each does, then the funds that give you the theme in one purchase. GARPify presents the options and leaves the choice to you.
A four-factor view of how the five operators differ within the basket thesis
The basket view treats the five as one. The four-factor comparison below shows how they differ — on quality, growth, price, and the job each does in the stack — so you can weigh which one fits. That is the question every subscriber has to answer for themselves.
The basket is the analytical unit; the operators are the investment units. Applying the four-factor framework - Quality, S-Curve exposure, Sentiment, and Price discipline - at the operator level surfaces a pattern the basket-level read cannot. The five names sort naturally into two groups: a silicon trio (NVIDIA, TSMC, Broadcom) and an infrastructure pair (Vertiv, Eaton) - Vertiv the maker of the power and cooling systems inside the building, and Eaton the electrical distribution that carries power from the grid to the rack.
The five operators, by the numbers
The same four factors as the rest of the report - quality, growth, sentiment, price discipline - shown as the figures behind them, so every reading traces to the data rather than to a colour.
| Quality | Growth | Sentiment | Price discipline | |||
|---|---|---|---|---|---|---|
| Operator | ROIC | Rev YoY | EPS YoY | P/E | Own 5-yr avg | vs own 5-yr |
| NVDA NVIDIA Compute platform | 93% | +65% | +150% | 32.9× | 69.9× | -53% |
| TSM TSMC Foundry | 30% | +36% | +17% | 30.2× | 24.4× | +24% |
| AVGO Broadcom Networking silicon | 16% | +24% | +22% | 61.3× | 59.5× | +3% |
| VRT Vertiv Power & cooling | 21% | +28% | +78% | – | – | – |
| ETN Eaton Power & electrical | 14% | +10% | +19% | 42.7× | 32.3× | +32% |
| Basket equal-weight | 21% | +32% | +47% | 41.8× | 46.5× | -10% |
Each column is the data behind a GARPify factor: return on invested capital (quality), revenue and earnings growth (where it sits on the S-curve), the P/E the crowd is paying today (sentiment), and that P/E set against what the same operator has averaged over the past five years (price discipline). Vertiv's current multiple was not available for this update, so its price columns are left blank and it sits outside the basket average. Source: YCharts, as of the report date.
What the numbers say
Quality is high but uneven. Return on invested capital runs from about 14% at Eaton to roughly 93% at NVIDIA, and every operator is growing both revenue (10-65%) and earnings (17-150%) double digits. This is a basket of genuinely expanding, profitable businesses rather than a single bet on one name.
On price, read each name against its own history. The last two columns are the ones that pay for the reading: what each operator's multiple has averaged over five years, and where it sits against that today. NVIDIA trades 53% below its own average - the share price went up, and the earnings went up faster. Broadcom sits 3% above, which is another way of saying normal. Eaton is 32% above its own average and TSMC 24% above; in those two the multiple, rather than the business, is carrying more of the load.
So the comparison surfaces a high-quality, fast-growing basket priced a little below its own history on average - and the average is the least useful number on the page. The spread runs from 53% below normal to 32% above. That is exactly what a comps view is for: not to rank the operators, but to show where each one sits, and where the price rather than the business is doing the work.
Each operator here is a starting point — the full analysis, the management, and the risks live in its own company report.
In a fund
Own the theme in a fund — and the one thing to know first
Say the thesis convinces you and you're wondering how someone would get exposure through a fund rather than a single name. There's a catch worth knowing first: no single fund holds all five operators. The basket lives in two different corners of the market — the silicon (the chips) and the power and electrical gear that feeds and cools them — and the fund world splits along the same line. So there are two shelves to look at. What follows is a best-fit map across the major fund families, fees shown, ordered by how cleanly each one actually holds the operators — "fit" describes overlap with the basket, nothing more, and is not a ranking of quality or a preference.
Best-fit fund map as of mid-2026, across VanEck, iShares/BlackRock, Invesco, SPDR/State Street, First Trust and Global X. MER and holdings/fit are from each provider's fact sheet; AUM and 5-year return are market figures, updated with the rest of the market data in this report. Listing is US unless marked CAD (Canadian, TSX). GARPify has no commercial relationship with any fund named here — inclusion is editorial. This is a description of investable vehicles, not investment advice.
Risk
Charlie Munger solved problems backwards: tell me where the theme breaks, and I’ll know what to watch. This tab inverts the thesis — what would have to go wrong for the whole basket, how likely each threat is, and how survivable. The goal isn’t a theme with no risks; it’s knowing exactly what you’re accepting.
What could break the thesis
Every thesis has a counterweight. Charlie Munger insisted the only way to understand a business is to invert - to study not what would make it succeed but what would make it fail. The Data Centre Boom basket has five identifiable risks, ordered here from most to least observable.
Step 1 - The Buildout's Funding Source
The single largest risk is also the most observable. The basket's earnings depend on the capital spending of a handful of cloud providers. If that spending slows or reverses, the earnings of all five operators slow or reverse with it. Everything else in this tab is secondary to this one dependency.
The spending has surged for several years. The interpretive question is not whether it is large - it plainly is - but whether the growth rate is sustained or about to digest. The first sign of moderation will appear in each hyperscaler's forward-spending commentary before it appears in the actual capital-expenditure numbers, which is why earnings-call language is the cleanest early signal.
Step 2 - The Five Risk Categories
Risk One - Capital-spending digestion. After a multi-year surge, the buildout could pause to absorb capacity already installed. Digestion need not be a collapse; even a flat year would compress the growth that current multiples assume. The signal to watch is the year-over-year growth rate in combined hyperscaler capital spending.
Risk Two - Hyperscaler concentration. A small number of customers fund most of the demand. That concentration means a strategic change at even one large buyer - an efficiency push, a budget reset, a shift to internally-designed silicon - is felt across the whole basket. The signal to watch is customer commentary about alternative suppliers being qualified.
Risk Three - Taiwan geopolitical risk. The manufacturing layer is physically concentrated in Taiwan. This risk is low-probability and high-severity, and it cannot be hedged inside the basket because TSMC has no leading-edge substitute. The signal to watch is the broad geopolitical posture rather than any single quarterly data point.
Risk Four - Build-out timing and supply chains. The power and cooling layer faces a different constraint than the silicon layer: grid interconnection queues, transformer and switchgear lead times, permitting, and - critically for the equipment makers - supply chains for power and thermal components and skilled installation labour. These can throttle how fast new capacity actually comes online. The constraint cuts both ways for Vertiv and Eaton - the bottleneck supports pricing but can also delay revenue, and for Vertiv specifically, large fixed-scope projects carry execution risk if lead times or input costs are mis-estimated. The signal to watch is order backlog, book-to-bill, and lead-time commentary.
Risk Five - Competing inference architectures. The thesis assumes AI workloads keep demanding large numbers of NVIDIA-class accelerators. If inference economics shift toward designs that need fewer or cheaper chips, the compute layer feels it first. The signal to watch is the performance-per-dollar trajectory of each new chip generation against the last.
Step 3 - Which Risks Are Currently Flashing
Some of these risks have data suggesting they are activating; others are dormant.
Capital-spending digestion is currently dormant. Reported spending growth has not yet shown deceleration, though by its nature this is the risk most likely to turn first and the one to watch most closely.
Hyperscaler concentration is structural and permanent. It is not flashing so much as always present. Every customer is also a potential designer of its own silicon, which keeps this risk live regardless of the spending trend.
Taiwan risk is dormant but unhedgeable. Nothing in the current data changes the picture; the exposure is constant and binary rather than something that builds gradually.
Grid and power timing bears watching. Lead times across electrical equipment remain extended. This currently supports the infrastructure pair's pricing, but it is the place where revenue timing could disappoint.
Competing architectures is dormant. No data yet suggests the next chip generation will underdeliver against the last. The risk becomes more readable as new-generation benchmarks publish.
Step 4 - What Would Resolve Each Risk
Each risk has specific events that would clearly move it in one direction. A reader monitoring the thesis can watch these rather than the news cycle.
Capital spending: the year-over-year growth rate in each hyperscaler's reported spending. Sustained growth confirms the cycle; a decisive deceleration would be the clearest single warning.
Concentration: major-customer announcements and any meaningful shift of volume toward internally-designed accelerators.
Taiwan: the broad geopolitical posture; there is no clean quarterly metric, which is exactly what makes it hard to underwrite.
Build-out timing and labour: order backlog, book-to-bill, lead-time commentary, and supply-chain availability from Vertiv and Eaton.
Architectures: new-generation chip benchmarks. If performance-per-dollar keeps improving at the historical pace, the replacement cycle stays intact.
Step 5 - The Integrated Bear Case
Each risk above was analyzed individually, which makes them easier to monitor but harder to weigh. The harder question is what happens if more than one materializes at once. The following is not a forecast - it is the path that would have to unfold for the bear case to become real, written so a subscriber can recognize it if it begins.
It would likely start with language. On some earnings call, a large cloud provider describes its spending plans with a word it has not used before - measured, disciplined, optimizing - and the market reads it as the first sign of digestion. In the months that follow, new-generation chip benchmarks come in merely good rather than transformational, softening the replacement-cycle logic, while electrical-equipment backlogs stretch far enough that the infrastructure pair's revenue slips to the right.
At the destination, perhaps eighteen months out, combined spending growth has decelerated sharply, the silicon names still earn well but no longer grow into their multiples, and the infrastructure pair is delivering on backlog built at the peak. The companies remain good businesses with intact moats and strong cash flows. What has changed is that the cycle which justified the multiples has visibly slowed.
What this scenario does not require. It does not require Taiwan disruption, and it does not require any operator to lose its competitive position. It only requires the growth rate to normalize and the multiples to follow. That is what makes it the base bear case rather than the tail risk.
How to weigh this. The bear case is not the base case. Reported spending growth remains intact, new-generation benchmarks have not disappointed, and backlogs are full. The point of writing it down is to define, in advance, the signals that would tell you it had started.
Where the Risks Land
The thesis behind the Data Centre Boom basket has identifiable, observable risks. The capital-spending and architecture risks are currently dormant; the concentration and Taiwan risks are structural and constant; the grid-timing risk is live but currently supportive. The integrated bear case describes what materialization would look like - the linguistic shift in guidance that precedes the numerical deceleration - so that it can be recognized early rather than in hindsight.
Think
This is where it comes together. You have read the businesses, judged the growth, weighed the price, and named the risks. No one can take this last step for you — Buffett’s filter is to act only within your circle of competence, and only when the answer is a clear yes. So, knowing what you now know: all in, is this a theme you’d own?
A personal record of how you assessed this report. Print for committee discussions or client reviews.
Disclaimers
Important information about this research and how to use it
Not Personalized Investment Advice
The content of this report is impersonal analytical research. It is general in nature and does not take into account the specific investment objectives, financial situation, risk tolerance, or particular needs of any specific person who may read it.
Nothing in this report constitutes an offer to buy or sell, or a solicitation of an offer to buy or sell, any security. Nothing in this report constitutes a recommendation that any particular security, portfolio of securities, transaction, or investment strategy is suitable for any specific person. The descriptions, analyses, and observations in this report are presented to help readers form their own judgments about whether the securities discussed are appropriate for their own circumstances.
Readers should consult their own registered investment advisor, portfolio manager, or other qualified financial professional before making any investment decision based on information contained in this report.
Registration Status
GARPify is not registered as an investment dealer, portfolio manager, exempt market dealer, or investment advisor with any provincial or territorial securities regulator in Canada, with the United States Securities and Exchange Commission, or with any other securities regulator in any jurisdiction.
GARPify operates as an impersonal financial publisher providing analytical research and commentary. GARPify does not provide personalized investment advice, manage client accounts, or hold client funds or securities.
Personal Holdings Disclosure
The author of GARPify research may from time to time hold positions in securities covered by GARPify reports. Specific holdings as of the publication date of this report:
As of the publication date, neither GARPify nor the author of this report holds any position in the equities or exchange-traded funds discussed in it.
The author may buy or sell securities covered by GARPify reports at any time, subject to the trading restrictions described in the GARPify Code of Conduct, which prohibits trading in covered securities for a specified blackout period before and after publication of new or updated research on those securities.
Source of Data and Information
Quantitative data in this report is sourced from YCharts and from primary regulatory filings retrieved from SEC EDGAR and SEDAR+. Specific data points are cited where they appear in the report. All data is current as of the publication date shown on the Cover tab.
These reports are updated quarterly. Because prices, valuation multiples, earnings, and fund data change continuously, the figures in this report become dated between updates. Readers should check current information against a major data provider, such as YCharts, before relying on any specific figure.
GARPify makes commercially reasonable efforts to ensure the accuracy of the data presented but does not warrant or guarantee the accuracy, completeness, or timeliness of any data, content, or opinions in this report. Readers should verify any specific data point against primary sources before relying on it for any investment decision.
Forward-Looking Statements
This report contains forward-looking statements about market conditions, business performance, valuation multiples, and potential investment outcomes. Forward-looking statements involve known and unknown risks, uncertainties, and other factors that may cause actual results to differ materially from any future results expressed or implied by such statements.
The forecast scenarios in this report represent conditional outcomes under specific stated assumptions about earnings growth and valuation multiples. They are not predictions. Actual outcomes may differ materially from the scenarios presented.
Past performance, including the historical returns of the Data Centre Boom basket presented in this report, is not indicative of future performance. Securities mentioned in this report can decline in value, and readers can lose money on investments in those securities.
Conflicts of Interest
GARPify is funded entirely by subscription revenue from its readers. GARPify accepts no advertising, sponsored content, paid placements, marketing support, or compensation of any kind from any company covered in its research, from any exchange-traded fund issuer, from any mutual fund company, from any wholesale distribution team representing any fund or issuer, or from any other intermediary that may distribute GARPify research to its own clients or advisor networks.
GARPify does not engage in investment banking, corporate finance advisory, brokerage, or any other business activity that would create a financial interest in the performance of securities or funds covered in its research, other than the author's personal holdings disclosed above.
GARPify research is selected, written, edited, and published solely by the GARPify author. Editorial decisions about which companies to cover, which exchange-traded funds to catalogue in the Exposure section of each theme report, what analytical framework to apply, and what conclusions to present are made on objective methodology that is documented and defensible. The inclusion criterion for any exchange-traded fund in the Exposure section is the concentration of that fund's holdings in the basket operators of the relevant theme. This criterion is applied uniformly. No exchange-traded fund issuer, wholesale team, or other intermediary has any role in determining which funds are catalogued or how they are described.
GARPify may, from time to time, enter into agreements that permit exchange-traded fund wholesalers or other intermediaries to distribute GARPify research to advisors in their professional networks under a non-redistribution license. Such agreements involve no monetary or other consideration flowing in either direction. The existence of such distribution arrangements does not influence GARPify's editorial decisions, fund selection methodology, or analytical conclusions.
Trademarks and Attribution
The analytical frameworks referenced in this report are attributed to the investors who developed them: Warren Buffett (quality and long-term ownership), Peter Lynch (earnings as price driver), Howard Marks (cycles and price discipline), Charlie Munger (inversion-based risk analysis). GARPify is not affiliated with, endorsed by, or sponsored by any of these individuals or their respective firms.
Company names, tickers, and product names mentioned in this report are trademarks or registered trademarks of their respective owners and are used for identification purposes only.
Distribution and Reproduction
This report is intended for the use of subscribers to GARPify. Subscribers may print or save individual copies for personal use. Republication, redistribution, or commercial use of this report or any portion of its content without the prior written permission of GARPify is prohibited.
Privacy
GARPify collects and uses personal information of subscribers in accordance with the GARPify Privacy Policy, which is compliant with Canada's Personal Information Protection and Electronic Documents Act (PIPEDA). The full Privacy Policy is available at the GARPify website.
No Warranty
This report is provided on an "as is" basis. GARPify disclaims all warranties of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, and non-infringement. GARPify shall not be liable for any direct, indirect, incidental, special, consequential, or exemplary damages arising from the use of or inability to use this report or any information contained in it.
Governing Law
This report is published in Ontario, Canada. The interpretation of these disclaimers and any disputes arising from the use of this report are governed by the laws of the Province of Ontario and the federal laws of Canada applicable therein.
Contact
Questions about the content of this report or GARPify's research methodology should be directed to: bob.simpson@garpify.com.
For specific concerns about regulatory status, personal holdings disclosure, or any other compliance matter: bob.simpson@garpify.com.
This disclaimer text is a working draft prepared by GARPify for review by qualified Canadian securities counsel before publication. It is not legal advice and has not been reviewed by a lawyer admitted to practice in any jurisdiction. Subscribers and readers should not rely on this text as a substitute for advice from qualified legal counsel.