By asahifile 
If you want to see what an AI boom actually looks like, skip the keynote and stand in a coatings lab in Pittsburgh. At PPG, a $24 billion company, engineers fire torches at steel panels to test whether a coating can stop the steel underneath from softening. Paul Fenton, a chemist and engineer there, showed the New York Times a panel whose paint had swelled from a millimetre thick into a black crust nearly ten millimetres tall. “Just like a toasted marshmallow likes to expand,” he said — the trapped air keeps everything cold. The product, Steelguard 951, has become one of PPG’s flagship lines for data centers, and PPG expects its data-center paint sales to triple this year from last year.
A trillion-dollar boom shows up in a paint can
The logic is unglamorous and enormous. Tech companies coat gray concrete and gray steel with more gray: intumescent fire protection, rust inhibitors, and conductive and anti-static finishes. Data centers have become a fast-growing market for paint makers as a result. PPG’s overall sales rose 7% to $4.5 billion in its most recent quarter; the data-center paint line, by the company’s own forecast, is growing several times faster. Industry estimates cited by Techmeme put the specialised coatings on a single data center at roughly $3 million to $8 million — small next to a billion-dollar campus, and paid on every one of them.
The long tail of unlikely winners
Paint is only one entry on a longer list that also includes gravel, wiring, switchgear, transformers and cooling equipment — what The Verge summarised as a “long tail of unlikely winners”. The same buildout is also exposing limits that no amount of chip supply can fix. Microsoft pulled out of a deal with the British company Nscale to build an AI data center after grid officials said the site would not be supplied with power until the 2030s. Google, meanwhile, is funding one of the largest “virtual power plants” in the United States, paying homes and businesses to trim their electricity use when the grid is strained — a data-center company buying demand reduction because it cannot buy enough generation quickly.
Analysis: follow the physical, not the narrative
Two lessons are worth keeping. The first is that in a gold rush the durable businesses are often the ones selling picks. Coatings, cabling, gravel and switchgear carry no model risk, no token-price deflation and no platform war: they sell into whatever gets built, whoever wins. That is precisely why a paint forecast is a better indicator of real construction than any capex slide. Coatings also behave like a subscription: they are consumables that get reapplied on a maintenance cycle, so a fleet of data centers keeps buying paint long after the racks are bolted in. The second is that electrons, not GPUs, are now the binding constraint. Compute can be ordered in quarters; transmission capacity, interconnection queues and substations cannot. The Microsoft-Nscale episode and Google’s demand-response funding show the same constraint from opposite ends — one company walked away because the power was not coming, the other is paying people not to use it. For investors watching the AI trade wobble over financing, this is the counterweight: the physical buildout has its own momentum, and its beneficiaries are not the companies anyone is arguing about.
What to watch
- Whether industrial suppliers — coatings, cable, switchgear, transformers — keep raising data-center revenue guidance into 2027.
- Grid connection queues, and whether more announced projects slip to the 2030s like the Nscale site.
- Whether hyperscalers start buying generation or demand-response capacity outright rather than contracting for it.
- Whether data-center construction starts slow in response to the AI financing squeeze.
Sources: The New York Times, The Verge, Digg/Techmeme citing JLL.