Schneider Electric's Fastest Growth in Five Years Is a Megawatt Story, Not an AI Story

Schneider Electric reported its fastest sales growth in five years. The attribution was AI data centers. The disclosure was a headline: no organic growth rate, no backlog figure, no megawatt count, no split between power distribution, cooling, and software.

I have audited token vesting schedules with more granular disclosure than that.

The number is probably real. The mechanism is almost certainly real. The evidence, as presented, is not. That distinction matters more than usual right now, because the same physical queue that produced that headline is repricing every bitcoin miner on earth — and doing it with almost no public measurement.

Hype evaporates; receipts remain. This piece is about which receipts actually exist.

The context nobody bothered to supply

Schneider Electric is a French listed industrial, EPA: SU, a CAC 40 constituent, roughly a century and a half old. Core business: energy management and industrial automation. Data centers are one terminal market among several, alongside buildings, industry, and grid infrastructure. By most industry estimates the data center vertical contributes somewhere north of twenty percent of group revenue and is the fastest-growing slice of it.

The coverage that generated the headline came from a crypto-native outlet, not a wire service. That detail matters. A crypto publication covering a French electrical conglomerate is not a media outlet expanding its beat. It is the crypto beat discovering that its own bottleneck has moved.

The arithmetic underneath is straightforward. A conventional data hall rack draws 5 to 15 kilowatts. An AI training rack draws 40 to 130. NVIDIA's GB200 NVL72 configuration lands near 120 kilowatts per rack. Multiply by row, by hall, by campus. The power train that was adequate in 2019 — low-voltage distribution, air cooling, a conventional double-conversion UPS — is not adequate now. Busway replaces conduit. Liquid replaces air. Medium-voltage switchgear moves closer to the silicon.

Schneider sells all of it. That is the thesis, and it is a sound one.

In 2021 I published a four-thousand-word teardown of a major marketplace's royalty mechanism. The lesson that outlasted the article was narrower than the finding: cross-vertical coverage degrades. A desk that covers tokens full-time, assigned to an industrial earnings note, will relay the press release and add no denominator. That is not malice. It is the absence of a filing on the desk. The reader pays for the gap.

What follows is a teardown of what the headline establishes, what it does not, and why a bitcoin miner in West Texas should read a French industrial's order book the way an auditor reads a custody statement.

The four numbers that were not published

A sales growth figure is an output. Auditors do not grade outputs. They grade the inputs that produced them.

Four inputs were absent.

Organic growth. Schneider has been an active acquirer for a decade. Revenue that arrives through consolidation is not demand; it is arithmetic performed by an accountant. Without an organic figure, "fastest in five years" cannot be decomposed into volume, price, currency, or acquisition. A European industrial reporting in euros with dollar-denominated North American revenue carries an FX tailwind or headwind every quarter capable of moving reported growth by several points on its own.

Backlog delta. Data center equipment is ordered twelve to twenty-four months before it is energized. Sales recognized this quarter reflect orders placed in 2023 and 2024. The backlog figure — absolute size and year-over-year change — is the only line item that speaks to current demand. A record sales quarter with a flat backlog is a company shipping its past. A record sales quarter with a growing backlog is a company selling its future. The headline cannot distinguish the two. Ledger balances do not lie; they only wait.

Segment split. Power distribution, cooling, UPS, and software carry wildly different margins and wildly different competitive structures. A group-level growth number averages a high-margin software line against a low-margin switchgear line and produces a figure that describes neither.

Megawatts. The unit that matters is not euros. It is megawatts delivered and megawatts in queue. Every infrastructure analyst I respect converts equipment orders to MW. The article did not.

In 2025, auditing exchange compliance under MiCA in Stockholm, the standard I applied was narrow and non-negotiable: a reserve attestation is meaningful only if it is cryptographically verifiable and independently reproducible. A signed PDF from an accounting firm is a claim. A zero-knowledge proof against a published commitment is evidence. Of the three platforms I examined, one met that bar.

The same distinction applies here. A press summary of a quarterly result is the signed PDF. The filing — with organic growth, backlog, and segment detail — is the commitment. Until you read it, you are holding a claim, not a balance.

I am not being pedantic about disclosure. In 2020 I traced a yield aggregator's outflow to a hidden backdoor by aligning three data points the team never expected anyone to reconcile: withdrawal timestamps, contract deployment blocks, and beneficiary addresses. The public narrative said one thing. The ledger said another. The ledger won, because the ledger was the only artifact nobody had edited.

A press summary is an edited artifact. Treat it accordingly.

Where the growth physically comes from

Strip the brand. What does Schneider actually sell into an AI hall?

Medium-voltage switchgear at the campus boundary. Transformers stepping down to distribution voltage. Low-voltage switchboards. Busway running overhead instead of cable in tray. Rack PDUs. Uninterruptible power, increasingly in high-voltage DC topologies rather than legacy AC. And cooling — chilled water, coolant distribution units, cold plates, immersion tanks in the densest halls.

Two structural shifts are doing the work.

Density is the obvious one. At 120 kilowatts per rack, air stops being a cooling medium and becomes a mixing medium. You cannot push enough cubic feet per minute through a rack envelope to reject that heat at a sane delta-T. Liquid has to reach the die. That is a decade-scale architecture switch, and architecture switches are when value migrates between suppliers.

The other is topological, on the power side. Legacy double-conversion AC UPS loses three to six percent in conversion. At 100 MW of critical load that is real money. At 500 MW it is a power purchase agreement. The industry is migrating toward HVDC distribution and, further out, solid-state transformation. That migration redistributes value among the companies that make converters, the companies that make switchgear, and the companies that make the orchestration software.

Schneider's position is strongest in integration and weakest in the one component growing fastest. It has spent the last two years buying liquid cooling capability rather than building it. That is a rational move and simultaneously an admission. The article did not mention it.

Here is the part that should interest anyone holding tokens rather than shares.

Value per megawatt is not constant and is not published. Nobody in this sector discloses euros of equipment content per megawatt of IT load. Estimates circulate privately. They differ by a factor of two depending on whether upstream utility interconnection is counted, whether the cooling loop is included, and whether software is amortized in. Without that denominator, an order book is a numerator without a scale.

The queue is the ledger

The genuinely important sentence in this story is the one that never gets written: the bottleneck has moved.

Through 2023 and 2024 the binding constraint on AI capacity was GPU allocation. That constraint has partially cleared. What has not cleared is the interconnection queue, the transformer lead time, and the licensed electrical labor to install both.

Transformer lead times that ran 12 to 20 weeks before 2020 have, in the large-power segment, run to 80, 120, and in some cases past 150 weeks. Interconnection queues in several US grid regions stretch beyond four years for new large-load service. These are not financial variables. They cannot be hedged with a derivative. They resolve only when copper is wound and a line is energized.

This is where the crypto transmission becomes explicit rather than rhetorical.

A bitcoin miner holds, in physical form, something a hyperscaler cannot manufacture quickly: an energized interconnect at known capacity, with existing switchgear, existing transformers, existing permits. That asset was priced as a stranded cost during the 2022 drawdown. It is now priced as an option.

Every announced miner-to-HPC conversion — site leases, hosting agreements, colocation contracts with AI tenants — is a repricing of the same underlying object Schneider is selling into. The miner is not selling compute. The miner is selling a grid position.

The scarce asset in AI infrastructure is not silicon and not capital. It is an interconnection agreement with a signed energization date. Everything downstream — racks, cold plates, switchgear, software — is manufacturable. The queue is not.

Which makes the Schneider headline a lagging indicator dressed as a leading one. Orders book when a project is financed. Projects are financed when the interconnect is granted. Interconnects were granted two to four years ago. What is being observed is the physical maturation of decisions already made, not the discovery of new demand.

That is not bearish. It is a timing correction. Anyone reading "fastest growth in five years" as evidence of accelerating AI demand this quarter is reading a receipt for a purchase made in 2023.

There is a second-order effect worth naming. Grid operators facing large-load requests increasingly require load flexibility — the ability to curtail on signal. A bitcoin miner is structurally the most flexible large load on any grid. It can drop from nameplate to zero in seconds without damaging product, because its product is not physical. A semiconductor fab cannot do that. A hospital cannot do that. This is why several system operators have folded mining load into demand response programs, and why some miners earn more from curtailment credits than from hashing during peak events.

That flexibility has a price, and the price is now visible. As AI load crowds the queue, the option value of a curtailable block of capacity rises. Schneider sells the equipment that makes a load controllable. The miner owns the load. Both are selling the same thing to the same grid: predictability.

What this does to hashprice and hashrate

Bitcoin mining and AI hosting compete for the same three inputs — power, land, capital — and compete on none of the same margin structures. The divergence is now visible in filings.

Mining revenue per unit of energy is set by hashprice, which is set by network difficulty and the BTC price. Neither is controllable by the operator. AI hosting revenue per unit of energy is set by a contract, typically with an investment-grade counterparty, typically with escalators, typically running five to fifteen years.

The consequence is a duration trade. Mining is a spot exposure dressed as an infrastructure business. Hosting is an infrastructure business with a credit counterparty. When the cost of capital is high and the grid is tight, duration wins.

I have watched this mechanism before, from the audit side. In 2022 I published a comparative analysis of algorithmic stablecoin designs before Terra collapsed. The failure was legible from the incentive table alone: a pegged asset whose stability depended on a subsidized yield that had to be paid out of the asset's own issuance. The price chart was not required. The payout schedule was sufficient.

The mining-to-hosting migration has the same legibility. A miner that converts capacity to long-duration hosting has changed its risk class. It is no longer levered to a hashprice it cannot influence. It is levered to a counterparty's credit and a contract's term. That is a different instrument, and the market has been slow to reclassify it — which is precisely why the reclassification has been profitable for the operators who moved first.

The Schneider number confirms that reclassification. It does not cause it.

DePIN and the tokenization of the queue

Now the part where the crypto claim gets tested rather than repeated.

There is a class of token projects — DePIN energy networks, tokenized power purchase agreements, on-chain energy credit registries — whose pitch is that physical energy infrastructure can be coordinated and settled on-chain. The pitch is not wrong in principle. Metering data, curtailment events, and settlement can be represented as signed attestations. I have verified more oracle designs than I would care to admit.

But the pitch misidentifies the bottleneck.

A DePIN energy network optimizes for the coordination layer. The binding constraint in 2026 is the physical layer: transformer availability, interconnect queue position, licensed electrical labor. No token shortens a 120-week transformer lead time. No settlement mechanism accelerates a utility's study process. A protocol that perfectly coordinates a resource that cannot be delivered is a well-engineered null.

Volatility is not risk; opacity is. The DePIN sector's real exposure is not price. It is that most of these networks cannot demonstrate delivery against claimed physical capacity. Ask a DePIN energy project for its interconnect agreement. Ask for the nameplate, the energization date, the utility. The ones that answer are infrastructure. The ones that pivot to tokenomics are marketing.

Contrast that with what Schneider's order book is: a signed, dated, physically anchored claim on future electrical capacity, aggregated across hundreds of projects, disclosed under French securities law. It is a boring artifact. It is also, functionally, the most reliable public index of AI power demand that exists outside utility filings.

The irony is structural. Crypto spent a decade building settlement layers for energy markets, then watched the actual energy bottleneck get resolved by purchase orders executed in euros and disclosed in a quarterly filing.

The competitor cross-check that was omitted

A single company's record quarter is not an industry signal until a second company confirms it.

Schneider's direct comparables in data center power and cooling are ABB, Eaton, Vertiv, Siemens, Legrand, and Delta Electronics. Their positioning differs materially. Vertiv is the purest exposure — data center is the business, not a segment. Eaton carries heavy electrical Americas content with meaningful data center weighting. ABB is broader and more automation-leaning. Schneider's distinguishing feature is end-to-end integration from campus medium voltage down to rack-level monitoring, plus a software layer that creates genuine switching friction.

That integration is a real moat. It is also a dilution mechanism. Schneider's revenue spreads across buildings, industry, and infrastructure. A doubling of data center orders moves the group number by a modest amount. Vertiv's does not.

So the honest read of "fastest growth in five years" is this: the direction is corroborated by every other company in the sector, and the magnitude is unverifiable from the article alone. That is not skepticism for its own sake. It is the difference between a fact and a frame.

There is a second omission. Chinese vendors — Huawei Digital Power most prominently, plus domestic switchgear and cooling suppliers — are compressing into the same market at lower price points. In a market with 100-week lead times, price competition is temporarily irrelevant because nobody can deliver. When lead times normalize, it will not be. Every capacity expansion decision made in 2025 and 2026 on the assumption of permanent scarcity is a bet against that normalization.

I have seen this shape before. In 2021 I dissected an NFT marketplace's royalty enforcement and found it bypassable with a wallet switch. The mechanism was marketed as protocol-level protection. It was client-side courtesy. The gap between the marketing claim and the technical reality was the entire story, and the story only became visible when someone read the contract instead of the announcement.

The AI power trade has the same gap, one layer down. The announcement is demand. The contract is a lead time. They are not the same document.

What the market is actually mispricing

The mispricing, if there is one, is not in Schneider. It is in the second derivative.

Public markets have repriced electrical equipment manufacturers and miners with interconnects. They have repriced copper and, partially, grain-oriented electrical steel. They have not repriced the third layer: the licensed electrical contractors, the switchgear integrators, the commissioning engineers, the transformer repair shops. Those are unglamorous, unlisted, and capacity-constrained in exactly the way the bottleneck demands. Capital flows to the visible layer. Value accrues to the scarce one.

That is the same pattern the 2021 liquidity mining cycle produced. The tokens were repriced within weeks. The market makers, the auditors, the RPC providers, and the block space were repriced over years, and most of them were never tokenized at all. The subsidized layer gets the attention. The constrained layer gets the margin.

What the bulls got right

The prevailing skeptical position is that AI capital expenditure is a bubble. The supporting evidence is usually a chart of hyperscaler capex against revenue, or a comparison to the 1999 fiber buildout.

The comparison fails on one dimension. Dark fiber could be lit later. A transformer cannot be wound faster than copper and grain-oriented electrical steel allow. A grid interconnect cannot be studied faster than a utility's engineering queue allows. These are physical serial constraints, not financing decisions. They convert a speculative narrative into a physical backlog with a date stamp.

That is why the AI power complex has held up better than AI software multiples. The bears are shorting a story. The physical layer is not a story. It is a purchase order.

The crypto-native version of the same error is the reflexive dismissal of miners as obsolete. That dismissal was correct in 2022, when hashprice compressed and the marginal operator was levered to a spot commodity with no counterparty. It is wrong now. A miner with 300 MW energized and a signed interconnect in a region where the queue runs four years is not a mining company. It is a power company with a legacy business attached, and the legacy business is the hedge.

The bulls also got one thing right that nobody credits them for: they stopped arguing about algorithms. The 2026 conversation is about electrons. Algorithms were never the scarce input.

The next Schneider filing will contain four numbers: organic growth, backlog, data center revenue share, and — if the company is being honest — megawatts. Demand those. Everything else is commentary.

The same standard applies one layer down. Ask every miner-turned-host for its energization date. Ask every DePIN energy network for its interconnect. Ask every AI infrastructure token for the nameplate capacity behind the claim.

The queue is the ledger. It has already been written. We are only waiting for the balances to be published.