Anthropic is in talks to pay $6 billion for a startup that nearly doubles chip output. That's a hedge against the $50 billion in data centers it already committed to build
Anthropic is reportedly negotiating to buy compute-efficiency startup Decart for roughly $6 billion — its largest deal ever — to squeeze more work out of the same AI chips. The company making that bet is the same one that committed $50 billion and 2,400 construction jobs to new data centers in Texas and New York less than a year ago.
Anthropic is reportedly negotiating to acquire Decart, an Israeli startup whose software squeezes far more output from the same AI chips, for roughly $6 billion — what would be its largest deal to date. The company weighing that purchase is the same one that committed $50 billion to new data center construction less than a year ago. That contradiction, not the deal itself, is what a GC or developer pricing data-center work should sit with.
What does Decart actually do?
Decart's product, the Decart Optimization Stack, rewrites how an AI model talks to a chip — custom kernels, a proprietary compiler, and hardware-aware model design layered on top of Nvidia GPUs, Amazon's Trainium accelerators, and Google TPUs. The company claims it pushes Model FLOPS Utilization — the share of a chip's raw computing capacity actually doing useful work — above 80%, against an industry norm closer to 40-50%. In plain terms: roughly double the output from the same rack of hardware, without adding a single new chip. Decart raised $300 million in a Series B in May at a near-$4 billion valuation; a $6 billion acquisition would be a substantial premium over that, three months later.
Why would a company building $50 billion in data centers pay billions to need fewer of them?
Because capacity, not cash, is what's rationing Anthropic's growth right now. In November 2025, Anthropic committed $50 billion to build new data centers in Texas and New York with partner Fluidstack — a buildout projected to generate about 2,400 construction jobs and 800 permanent ones as sites come online through 2026. That commitment hasn't changed. What's new is Anthropic apparently deciding it's also worth several billion dollars to extract more usable compute from the chips it already has and the ones still coming. One analyst framing of the deal: it's a bet that software efficiency gains arrive sooner than new data centers do — a reasonable wager when capacity, not demand, is the constraint holding back revenue.
Does this mean the data-center construction pipeline is about to shrink?
Not necessarily, and there's a real case it goes the other way. This is the same dynamic construction has watched from the demand side before with the Texas grid-approval freeze that put a chunk of the AI data-center pipeline on hold — capacity constraints move fast and don't always resolve in the direction anyone expects. Here's the split:
| If efficiency gains dominate | If demand dominates (Jevons paradox) |
|---|---|
| Same AI output needs fewer chips → fewer racks, less power draw, smaller halls | Cheaper effective compute unlocks workloads that weren't affordable before → more total demand |
| Announced campuses get resized or phased more slowly | Announced campuses proceed as planned, new ones get added faster |
| MEP and power-infrastructure subs see smaller data-center footprints per dollar of AI spend | MEP and power-infrastructure subs see the same crunch, just running longer |
Nothing in the reporting on this deal says which way Anthropic itself expects it to break — and the fact that its own $50 billion buildout isn't being pulled back suggests Anthropic isn't betting the efficiency gain replaces the construction, at least not yet. It's buying optionality on cost per unit of compute, not canceling concrete pours.
What should a GC or developer actually do with this?
Nothing today. The deal is still in talks and could fall apart, and even if it closes, one acquisition doesn't reset how much data-center capacity the industry is chasing. The useful discipline is tracking the leading indicator this points toward: whether hyperscalers start citing compute-efficiency software — not power availability, not chip supply — as a reason to delay, downsize, or re-scope an announced campus over the next couple of quarters. That would be the first real evidence efficiency gains are substituting for square footage instead of just padding margin on top of a buildout that goes ahead regardless.
If your firm has data-center or MEP work in backlog, ask the developer's team directly whether their compute forecast has moved since this story broke — it's a five-minute call that tells you more than any industry projection will.
Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.
- What is Decart, and why is Anthropic reportedly paying $6 billion for it?
- Decart is an Israeli AI infrastructure startup whose software, the Decart Optimization Stack, raises Model FLOPS Utilization on Nvidia GPUs, Amazon Trainium chips, and Google TPUs above 80%, against an industry norm of roughly 40-50%. Anthropic is reportedly in talks, as of August 13, 2026, to acquire Decart for about $6 billion — its largest known acquisition to date. The deal has not closed and could still fall apart.
- Does this mean Anthropic will build fewer data centers?
- No confirmation of that yet. Anthropic committed $50 billion in November 2025 to new data center construction in Texas and New York with partner Fluidstack, work expected to generate roughly 2,400 construction jobs and 800 permanent ones, with sites coming online through 2026. Nothing reported about the Decart talks says that plan is being resized.
- What is Model FLOPS Utilization and why does it matter for how many data centers get built?
- It's the share of a chip's theoretical computing capacity that's actually doing useful work on a given job — the rest is lost to inefficiency. Running training or inference at 80% MFU instead of the 40-50% industry norm means getting close to double the output from the same chips, and by extension the same power and cooling infrastructure a data center provides.
- Could an efficiency gain like this actually increase data center construction instead of reducing it?
- That's the standing counterargument, sometimes called the Jevons paradox: making a resource cheaper to use has historically increased total consumption of it, because the lower cost unlocks demand that wasn't economical before. A deal like this could just as easily accelerate the case for more AI compute — and the construction pipeline that supports it — as slow it down.
- Should a GC or developer bidding data center work change anything because of this?
- Not based on this deal alone. It isn't signed, and Anthropic's own $50 billion buildout is proceeding regardless. The signal worth tracking is whether hyperscalers start citing compute-efficiency gains as a reason to delay or resize announced campuses over the next two to three quarters — that would be the leading indicator, not one acquisition.