An oilfield services company has agreed to pay $4.1 billion for a German heat exchanger manufacturer. SLB N.V. is acquiring Kelvion Holding GmbH from funds managed by Apollo Global Management and Triton Fund IV, with $3.4 billion in cash and $0.7 billion of assumed debt. That price establishes what installed thermal capacity is worth to an acquirer with no prior position in the segment.
The Capacity Number
The relevant capacity metric in an AI data center is not wafer starts. It is megawatts of heat rejection, and the physics is unforgiving: essentially all electrical power delivered to a rack exits as heat. A 300 MW AI campus therefore requires 300 MW of continuous thermal removal, sized for the worst ambient day of the year.
That one-to-one relationship is why accelerator shipments and heat exchanger shipments are the same demand curve, offset by a construction cycle. Every incremental gigawatt of AI compute booked by a hyperscaler creates an equivalent gigawatt of thermal equipment order flow downstream.
The deal structure reinforces the point. Assumed debt is roughly 17% of total consideration, leaving the transaction overwhelmingly cash-funded. Apollo Funds described Kelvion as a global leader in thermal management for data centers and diversified industrials, and the sellers are exiting for cash rather than rolling equity.
The Allocation Decision
Thermal capacity is allocated the same way HBM capacity is allocated: by contract position, not by spot order. Large custom heat exchangers, dry coolers, and coolant distribution units are engineered-to-order products built on fixed fabrication lines. A hyperscaler that has not reserved factory slots two build cycles ahead does not get equipment on schedule, regardless of what it paid for its GPUs.
This is where the acquisition changes the competitive map. SLB is buying an allocation position, and it is buying it in a segment where the customer base is shifting from industrial process plants to data center developers. Industrial thermal work carries stable but modest pricing. Data center thermal work carries schedule-driven pricing, because a delayed cooling loop strands an entire accelerator deployment.
Of the $4.1 billion in total consideration, roughly 83% is being delivered in cash — an unusually clean structure for an industrial carve-out of this size.
The workload mix makes the allocation question sharper. Google has been courting Hollywood studios to adopt its AI tools, per the Los Angeles Times. Video generation is among the most thermally intensive inference workloads in production, because sustained token generation over long sequences keeps accelerators near peak draw for extended windows. Sustained load, not burst load, determines cooling plant sizing.
The Margin Impact
Cooling economics resolve into power usage effectiveness, and PUE converts directly into salable capacity. On a 100 MW IT load, moving PUE from 1.4 to 1.2 releases 20 MW of facility power. That 20 MW can then be sold as compute rather than consumed by chillers and fans.
For an operator constrained by a fixed grid interconnection, that reclaimed headroom is the highest-return capex in the building. It expands revenue capacity without a new substation, a new permit, or a new utility queue position. Liquid cooling economics improve further as rack densities rise, because air-side handling costs scale badly against high-density heat flux.
The margin structure for the equipment vendor follows the same logic. Fabrication lines running near full utilization spread fixed tooling and labor cost across more units, and engineered-to-order pricing holds when lead times are long. Utilization is the whole margin story in heat exchanger manufacturing.
Where I Was Wrong Before
My two calls on GE Vernova this summer bracketed the stock at $420 and $450, and neither level resolved anything. The error was analytical, not arithmetic: I framed power equipment as a tariff cost pass-through story when the binding variable was delivery schedule. Component cost inflation is visible and easy to model. Fabrication slot scarcity is neither, and it is the constraint that actually moves revenue recognition.
The Inflection
Consensus will name the pure-play cooling vendors as the beneficiaries of this repricing, and that read is largely embedded already. The less obvious shift is upstream, in coil fabrication, brazed plate capacity, and CDU assembly — the sub-tier that supplies every thermal integrator including the one SLB just bought. That layer has no branded equity exposure and no ability to raise capacity quickly.
Note also what SLB is doing with its cash. The company is redeploying balance sheet away from a core market where the global energy import bill has swelled by as much as $330 billion over six months, per reporting on the Iran conflict's price effects. Diversifying into a demand curve tied to compute rather than crude is a deliberate cash-flow duration trade.
The falsifiable version of this view is the deal itself. Terms should hold as struck at $4.1 billion, with the $3.4 billion cash component intact and no superior competing bid inside 30 days. If a rival bid clears $4.1 billion, thermal assets are being valued more aggressively than this analysis assumes and the entry multiple argument fails. If SLB substitutes equity for cash, the acquisition is a diversification hedge rather than a conviction position in data center thermal capacity, and that would change the read materially.
At $4.1 billion, SLB now owns the thermal node of the AI build-out. The next repricing happens one tier up, in the fabrication capacity that every thermal integrator draws on.