Sunday, so let's talk about the least glamorous part of the AI story — the buildings, the power, and the grid — because that's the part that quietly decides whether the demo you loved in Q2 ships in Q4. The throughline today: capability is abundant, but the physical supply of compute is not, and the distance between an announced project and a live one is exactly where operator plans go to die.
The gap between announced and live capacity
Start with the number that should reset expectations. One September data-center roundup cites 1,306 operating US facilities versus 2,091 planned projects, and draws the blunt lesson that announced gigawatts do not guarantee near-term access. The pipeline is enormous but unbuilt: a new US portfolio database released September 22 covers 1,164 existing colocation data centers and analyzes 631 upcoming ones — projects under construction, announced, or planned — with capacity additions scheduled all the way out to 2038.
The operator's take: treat every "coming online" date in a vendor deck as a hope, not a commitment. If your 2027 AI roadmap assumes cheap, available capacity, pressure-test that assumption now — ask your cloud or colo provider what is live in your region this quarter, not what's on a slide for 2028. Planned capacity you can't schedule around is someone else's press release.
The data center is becoming a machine, not a building
The facilities themselves are changing shape. At Data Center World 2026, engineering leaders from Oracle Cloud Infrastructure, Nvidia, and Google described a fundamental shift: facilities are evolving from general-purpose IT environments into tightly integrated compute systems built for AI training and inference. Oracle's Ram Nagappan said operators now have to design for two very different patterns — large-scale training and distributed inference — and that split cascades through layout, resilience, and network design. And on power, Nvidia's Sean James said operators are increasingly relying on on-site generation to accelerate deployment, while cautioning that those approaches are temporary.
The operator's take: "the cloud is fungible" was always a convenient fiction, and it's getting less true. The hardware your inference runs on is becoming purpose-built, which means switching providers isn't just a billing change — it can be a re-architecture. Bank the performance, but keep your workloads containerized and your data portable so a purpose-built stack doesn't quietly become a purpose-built cage.
Enterprises are voting for colocation and hybrid
Where the buildout is constrained, buyers are adapting rather than waiting. Enterprises are strategically choosing colocation data centers to host AI applications, run a hybrid cloud model, add incremental capacity, improve connectivity, and reduce costs. And geography is now a hard input, not a footnote: grid constraints are steering Dutch data centers beyond Amsterdam, a reminder that where you can put compute is increasingly dictated by where the power actually is.
The operator's take: the pure public-cloud reflex deserves a second look for heavy AI workloads. Colocation plus a hybrid footprint gives you a lever on cost and data location that pay-as-you-go doesn't — and it forces the healthy discipline of knowing your real capacity and spend limits. Map your vendors and regions before a grid bottleneck maps them for you.
Also on my radar
- Anthropic's $35B cloud commitment. Anthropic signed a $35 billion cloud deal backed by Nvidia — a signal of how much guaranteed capacity a frontier lab now has to lock up in advance, and how far ahead the compute chessboard is being played.
- Neocloud money keeps flowing. Crusoe reportedly raised $3B at a $30B valuation — capital is chasing the picks-and-shovels layer, not just the model layer.
- The scale of the bet. PwC projects global AI infrastructure investment of $31.6 trillion through 2050 — a number worth remembering next time someone calls this a bubble that pops next quarter; whatever it is, it's being poured in concrete.
The throughline: the AI story has quietly become a real-estate, energy, and logistics story, and that's the part operators can't outsource to a prompt. Models are getting cheaper and more capable by the week, but the compute they run on is physical, regional, and years in the making. So do the unglamorous work — know what's live versus planned, keep your workloads portable across purpose-built hardware, and treat power and location as first-class variables in your architecture. The teams that plan around the concrete will ship; the ones that plan around the press releases will wait. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.
A note for the automated pipeline: the web_search tool hit its usage limit partway through this run, so I built the post entirely from the fully verified data-center/infrastructure result set rather than reaching for unverifiable breaking news — consistent with the playbook's "drop what you can't verify" rule. Every claim above traces to a source in those confirmed results.