Equinix's AI Niche: Selling the Switchboard, Not the Power

The AI data center boom has a famous money story: cloud giants and their wholesale landlords convert tens of billions of dollars in capex into megawatt leases. Equinix (Nasdaq: EQIX) is up about 33% year to date, and it is tempting to file it in that same drawer. But Equinix does not really sell the thing the boom is famous for — bulk raw power. It sells the switchboard that connects the tenants who own the power. That distinction, not the AI label, is what shareholders are actually paying for, and it sets the price of admission.

The boom's money doesn't all flow where the headlines do

Start with the phrase riding the ticker: "AI data center." It sounds like one market, but it is two very different businesses with two very different economics.

Training a frontier model is a brute-force problem. It wants thousands of GPUs packed into a single power-dense building with a huge utility connection, and it is essentially a wholesale real estate deal: build megawatts, lease to Microsoft, Amazon, Google, or a fast-moving private platform like CoreWeave. That layer — the one that grabbed most of the public's attention and the megatonnage of capital — favors the wholesale model of a Digital Realty, where the value is deliverable power at scale.

Inference is the opposite problem. When a trained model answers a question in real time, that answer cannot tolerate the round trip of data hurled across the internet to a distant mega-campus. It needs to run close to where the data and the user live, inside dense, low-latency metros where networks, clouds, and enterprises physically meet. That is Equinix's layer, and it is the reason the company keeps attaching itself to the AI story without being a megawatt landlord. The company and NVIDIA now market a turnkey "AI factory" built for that distributed, enterprise inference work — high-density racks running 40 to 130-plus kVA each, placed across its global footprint so processing can happen where the data lives.

The switchboard is the scarce, sticky asset

Being in the right layer only matters if someone pays you for it. Here is where Equinix's model diverges from a power landlord. It sells proximity to an ecosystem: more than 499,000 customer-to-customer, metro-to-metro interconnections that let tenants privately exchange traffic with clouds, networks, and each other. Roughly 10,000 customers sit on that fabric, including about 60% of the Fortune 500.

The economics of that fabric are better than the economics of renting floor space. Interconnection revenue carries higher margins than space-and-power rent, and every new connection raises the cost of leaving — the network effect is the moat, not the concrete. In its second quarter Equinix added a record 9,700 net interconnections. That is the rent capture mechanism in one number: order growth converting into a stickier, higher-margin revenue stream rather than a lease with a termination date.

The clever part: keep the heavy build off the balance sheet

This is the move that rewards looking one layer farther. A high-margin interconnection business and a capital-hungry hyperscale build make awkward bedfellows on the same income statement — the low-return wholesale machinery would drag down the returns of the high-return core. Equinix's answer is xScale, a parallel program that serves hyperscale, wholesale AI workloads next to its existing campuses but runs through joint ventures, so the heavy capital and thinner margins sit off balance sheet. In October 2025 it expanded xScale through a joint venture with GIC and CPP Investments that raised more than $15 billion, and the program has leased roughly 430 megawatts across twenty-plus facilities.

In plain terms: when the megawatt boom comes to Equinix, the company can say yes without letting the low-margin version of its business pollute the economics of the switchboard it actually wants to be paid for.

What the numbers say, and what they hide

The second-quarter report, out in late July, reads like a boom. Revenue rose to $2.625 billion, up 16% year over year; adjusted funds from operations (AFFO) per share rose 19%. Backlog hit a record as annualized gross bookings grew 23%. Management raised full-year 2026 guidance to 11%–12% revenue growth and 10%–12% AFFO-per-share growth — a raise the company described as the largest in its history — and lifted its long-term 2027–2029 targets to 10%–13% annual revenue growth.

Now read the footnotes with the same care. A meaningful slice of that revenue and profit increase came from one-time xScale fees — a helping hand, not the recurring engine. Strip those out and the engine is monthly recurring revenue (MRR), which grew a solid but unflashy 11% year over year. That is the number to anchor expectations to, because the one-off fees will not repeat.

And the build is devouring cash. Trailing operating cash flow is about $3.9 billion against capital expenditure of about $6.5 billion, leaving free cash flow deeply negative — roughly negative $2.6 billion. Management expects the bill to keep climbing, guiding annual capex to $5 billion–$7 billion. The niche is doing exactly what the boom makes it do: spending today's cash to buy the capacity that should pay back for years. It is a feature of the model, up to a point — and the point is reached when growth has to stay funded by debt and equity for longer than investors expect.

The price of the switchboard

There is nothing wrong with a business like this; the question is the clock attached to it. Equinix trades at roughly 27 times trailing EV/EBITDA and about 66 times trailing earnings, with a dividend yield near 2%. Those are rich multiples for a business whose core recurring revenue grows 11%, not 30%. The market is effectively pricing the AI boom as if it flows through Equinix in full and stays there.

Part of the valuation is deserved — durable cash flows and compounding interconnection usually earn a premium over a flickering AI trade. But the premium also assumes the favorable parts of this story are the permanent ones: that enterprise inference demand keeps landing in metro data centers rather than drifting to clouds or to wholesalers who also add interconnection; that the one-time xScale fees become recurring economics instead of capital being returned to joint-venture partners; and that the negative cash can keep funding 11% growth without the multiple taking a hit.

The two numbers to watch

For a holder or a watcher, keep two signals in view. The confirmation metric is whether MRR keeps compounding at double digits in quarters that do not include one-off xScale fees — that is recurring demand the market is paying for, and if it accelerates, the premium earns itself. The normalization signal is the failing one: watch for interconnection additions to slow, or for enterprise AI inference work to be captured by the cloud giants and wholesale landlords rather than by the switchboard. That is the moment the hidden winner becomes ordinary — not because the AI theme faded, but because the scarcity that justified the price moved to another layer.

The famous boom got its headline winners. Equinix's niche is real, it is profitable, and it is priced that way. Your job is not to decide whether AI helps Equinix — it obviously does. Your job is to decide whether the market has already charged you for the best case, and to hold out for a price that lets the switchboard's returns arrive before the multiple pays out.