By Robert Swift

Every technology mania eventually meets a loading dock. For global AI infrastructure stocks, that moment has arrived.

The first phase is always the exciting one. Models, chips, demonstrations, conference keynotes delivered in a leather jacket. The second phase is duller and considerably more physical. Someone has to pour the concrete, secure the grid connection, order the switchgear with a lead time measured in quarters, and persuade a utility that this particular substation should be built now rather than in 2031.

Australian investors have just been handed a very clear illustration. Goodman Group reported its FY26 result on 20 August 2026 with a development book of A$19.7 billion, of which 78% is now data centres, and a global power bank of 6.4 gigawatts across 16 cities. Greg Goodman described the constraint plainly: scarcity of power and land is what limits AI and cloud growth, not enthusiasm.

That is the whole thesis, stated by a landlord in Sydney rather than a chief executive in Santa Clara. The money has been committed. The bottleneck is physical. And the listed companies that sit inside the bottleneck are, for the most part, not the ones retail investors have been buying.

The size of the cheque

It is worth pausing on the numbers, because they have moved from large to faintly absurd.

UBS estimates, reported in late August 2026, put total hyperscaler capital expenditure at roughly US$1.009 trillion in 2026, rising to US$1.447 trillion in 2027 and US$1.619 trillion in 2028. Cumulatively that is around US$4.1 trillion across three years, against approximately US$1.292 trillion spent over the previous six years combined. The spending base is not cycling. It is resetting.

The more revealing statistic is the ratio. On the same estimates, Amazon, Alphabet and Microsoft will collectively spend about 102% of their cloud revenue on capital expenditure in 2026, easing to roughly 99% in 2027 and 94% in 2028. These are enormously profitable companies with diversified earnings, so this is not a solvency observation. It is a behavioural one. Every dollar the cloud business earns is being fed straight back into the machine, and then some.

Investors have started to notice. The market’s reaction function has quietly inverted. Raising capital expenditure guidance used to be received as evidence of demand. It is now increasingly received as evidence of cost. Meta learned this the hard way earlier in the year when it lifted full year guidance and was rewarded with a sharp fall.

Which raises the question this article exists to answer. If the spenders are being punished for spending, who is being paid?

Why the bottleneck is where the return lives

The useful distinction is between companies that finance the buildout and companies that monetise it.

Financing it means writing the cheque and hoping the utilisation shows up. A data centre does not generate an attractive return merely because it contains expensive silicon. Capacity has to stay full, customers have to keep paying, and the revenue has to cover depreciation, electricity, financing and operations. That is a genuine risk, and it sits squarely with the hyperscalers.

Monetising it means selling something the buildout cannot proceed without, ideally under a long contract, ideally where supply is constrained by something other than your competitors’ willingness to invest.

That second category is smaller than the marketing suggests. It requires physical assets, long lead times, regulatory permissions or a customer relationship measured in decades. It tends to be unglamorous. It also tends to get paid whether or not the model layer ever produces a profit, which is a rather important property to own in a boom.

Five listed businesses sit inside that category. None of them is cheap. That is addressed at the end, as it should be.

Vertiv Holdings (NYSE: VRT): the equipment

Vertiv sells the unglamorous middle of a data centre. Power trains, uninterruptible supplies, switchgear, thermal management, and the higher voltage direct current architectures now being validated for the next generation of accelerator racks.

The second quarter of 2026 delivered net sales of US$3.274 billion, up 24% on the prior year, adjusted earnings per share of US$1.52, and adjusted operating margin of 22.6%, an improvement of 410 basis points. Management raised full year guidance to approximately US$14 billion of sales at the midpoint with adjusted EPS of US$6.65 to US$6.75.

The margin expansion is the part worth dwelling on. Revenue growth in a boom proves demand exists. Margin expansion alongside it suggests pricing power, which is a rather better indicator of where the shortage actually sits.

The complications are two. European organic growth was only 2% in the quarter, so the story is considerably more American than the headline implies. And the shares closed at US$261.95 on 21 August 2026 for a market capitalisation near US$100.85 billion, having risen over 100% in twelve months and roughly 62% in the calendar year to date, before falling about 13% in the past month. A forward multiple in the low forties and a beta above two is not a quiet holding.

Schneider Electric (EPA: SU): the European incumbent

Schneider spent a century selling circuit breakers and industrial components. It now supplies a substantial share of the electrical backbone of the world’s data centres, from power distribution to cooling to rack systems.

The first half of 2026 produced record revenues of EUR 21.2 billion, up 14% organically, with the second quarter accelerating to 16.5%. Adjusted EBITA reached EUR 4.1 billion at a margin of 19.3%. Management raised full year guidance for organic revenue growth to a range of 10% to 13%, up from 7% to 10%, and lifted the adjusted EBITA growth target to a range of 14% to 19%. Data centres and semiconductors both showed triple digit demand growth in the second quarter.

What makes Schneider interesting is not that it benefits from data centres. Everything on this list does. It is that Schneider is not solely dependent on them. Energy management grew 18% in the quarter and industrial automation grew 11%, with strength across North America, China and India. Grid modernisation is a multi decade requirement that would exist even if artificial intelligence had never been invented, and Schneider sells into it.

The company flagged a currency headwind of EUR 400 million to EUR 500 million for the full year and warned of supply chain disruption from the Middle East affecting the second half. Investors buying European industrials from Australia should also be clear that they are taking a currency position alongside an equity one.

Equinix (NASDAQ: EQIX): the meeting place

If Vertiv and Schneider sell the equipment, Equinix owns the room where the networks meet. Its business is interconnection: the neutral exchange points where clouds, carriers, enterprises and model providers physically connect to each other.

In the second quarter of 2026 Equinix added 9,700 net interconnections, its largest quarterly addition on record, with annualised gross bookings of US$424 million, up 23%. Revenue grew 16% with an adjusted EBITDA margin of 53%. Full year revenue growth guidance was raised to a range of 11% to 12%, and the outlook for 2027 to 2029 now calls for annual revenue growth of 10% to 13% with adjusted EBITDA margin at or above 53% by 2029. Management noted that eight of the top ten model providers and eight of the top ten neocloud operators already run key networking workloads on its platform.

Interconnection is a genuine moat. Its value rises with the number of participants already present, which is a difficult thing for a new entrant to buy.

The complication is capital. Equinix has guided to US$5 billion to US$6 billion of capital expenditure in 2026 and US$5 billion to US$7 billion annually through 2029. That pressures free cash flow, and the shares closed at US$1,065.39 with a market capitalisation near US$105.12 billion, up over 41% in the calendar year, on a forward multiple near 61 and a dividend yield of about 1.8%. A wonderful business at an unforgiving price is still an unforgiving price.

Digital Realty Trust (NYSE: DLR): the shell

Digital Realty is the wholesale landlord. It owns the megawatts that hyperscalers lease years before they intend to switch anything on.

The second quarter produced a record backlog of US$1.9 billion at 100% share and a development pipeline that expanded to 1.4 gigawatts under construction at a total cost of around US$20 billion. Renewals were signed with cash releasing spreads above 25%, which tells you something about the balance of negotiating power in this market. Shortly after quarter end, two further hyperscale leases in the United States added US$410 million of annualised rent at 100% share.

A backlog of that size, at those spreads, is about as close to visibility as this sector offers. The shares closed at US$190.62 with a market capitalisation of US$71.81 billion, an indicated annual dividend of US$4.88 and a yield near 2.5%, having risen close to 25% in the calendar year to date.

The risks are the familiar ones for a real estate investment trust operating at scale. Interest rate sensitivity is real, funding requires regular equity issuance, and reported earnings are noisy because of development timing and asset gains. Investors should read the backlog and the leasing spreads rather than the headline earnings comparison.

Constellation Energy (NASDAQ: CEG): the electrons

The last name is the one that has not worked, which is why it is worth including.

Constellation is the largest owner of nuclear generation in the United States. In the second quarter of 2026 it reported adjusted operating earnings of US$2.55 per share, up from US$1.91 a year earlier, and raised full year guidance to a range of US$11.50 to US$12.50 per share. It signed an additional 920 megawatts of long term power purchase agreements with investment grade customers, running 15 to 20 years and commencing between 2029 and 2032, including a 176 megawatt agreement with Walmart. It cleared regulatory hurdles at the Crane Clean Energy Center, targeting a restart in 2027, and agreed to sell the 606 megawatt Brazos Valley gas plant to LS Power for US$860 million as a condition of its earlier Calpine acquisition.

That is a company signing twenty year contracts to supply the scarcest input in the entire chain. Revenue of US$7.5 billion missed expectations near US$7.94 billion, and the shares have fallen roughly 29% in the calendar year to date, trading around US$261 to US$265.

There is a lesson in that divergence. Owning the bottleneck does not exempt a company from being repriced. Power generators carry regulatory risk, outage risk, integration risk from a very large acquisition, and a valuation premium to utility peers that the market periodically decides it no longer wishes to pay. Investors who assume that a good structural position guarantees a good share price have not been paying attention to Constellation this year.

The common thread across these global AI infrastructure stocks

These five businesses share a set of characteristics that has nothing to do with artificial intelligence and everything to do with why they might endure.

They sell into a constraint rather than a trend. They own assets that take years to replicate: substations, interconnection ecosystems, reactor fleets, manufacturing capacity for high voltage equipment. They contract long, with counterparties that are investment grade. And crucially, they are paid for capacity and connection rather than for the eventual profitability of anyone’s language model.

That last point is the entire argument. The great uncertainty in this cycle is whether artificial intelligence generates enough revenue to justify US$4.1 trillion of spending. That uncertainty falls most heavily on the companies doing the spending. It falls considerably more lightly on the companies collecting a contracted payment for a megawatt, a cross connect or a switchboard.

Risks, and the part where valuations are addressed honestly

They are not cheap. Forward multiples in the forties and sixties, on companies whose growth depends on a capital cycle, is a demanding combination. Buying a good structural position at a poor price has historically been an effective way to earn a mediocre return for several years.

Momentum in this sector is violent in both directions. Vertiv fell roughly 13% in a month while its fundamentals improved. Constellation is down about 29% in the calendar year while signing twenty year contracts. Whatever else these are, they are not defensive holdings.

The capital assumption may not hold. The UBS projections assume spending keeps rising even as the growth rate moderates. That is a forecast, not a fact. If hyperscalers decide the returns are not arriving, order books shorten, and companies built for a US$1.6 trillion annual market will be operating in a smaller one.

The financing layer is getting crowded. Hyperscalers are increasingly funding this buildout with debt, which competes for the same capital as everything else and puts upward pressure on the cost of capital across the market. That is not a neutral development for long duration assets.

Customer concentration is the same in every name. Diversifying across five suppliers to the same handful of buyers is not diversification. It is one position expressed five ways.

Currency and timing. Australian investors take a currency position in every name here. And it is worth noting that Nvidia reports its quarterly result on 26 August 2026, United States time. A single result does not change a structural argument, but it will move sentiment across every stock discussed above, in whichever direction it chooses.

TAMIM Takeaway

The artificial intelligence trade has been priced as a semiconductor story. The physical evidence increasingly says it is a power, land and equipment story, and Goodman’s FY26 result made that point from Sydney with a 6.4 gigawatt power bank and a warning that scarcity of power and land is the binding constraint.

What the market may be underestimating is the durability of demand at the bottleneck compared with the volatility of demand at the application layer. Vertiv, Schneider Electric, Equinix, Digital Realty and Constellation Energy all sell into constraints that take years to relieve, under contracts that are paid regardless of whether the models make money.

What the market is very obviously not underestimating is the price. Every one of these businesses has already re-rated substantially, and one of them has de-rated sharply despite good news, which should be a useful reminder that structural strength and share price direction are separate questions over any period shorter than a decade.

For long term investors, the practical approach is not to chase the theme. It is to build a watch list of businesses that own genuine bottlenecks, understand what each would be worth on more sober assumptions about capital expenditure, and be prepared to act when sentiment inevitably wobbles. The buildout will take a decade. There is no requirement to own all of it this week.

______________________________________________________________________________________

Disclaimer: Vertiv Holdings (NYSE: VRT), Schneider Electric (EPA: SU), Equinix (NASDAQ: EQIX), Digital Realty Trust (NYSE: DLR), and Constellation Energy (NASDAQ: CEG) are held in TAMIM portfolios as at the date of article publication. Holdings can change substantially at any time.