You Can’t Expedite a Person
We forecast megawatts to the quarter. We don’t forecast the people who build them — and they’re the longest lead time on the sheet.
A few weekends ago someone asked me why we need so many data centers.
Great question. And the conversation that follows is always the same conversation. Power. Water. Memory. GPUs. Land. Transformers. Grid interconnection queues. We have gotten good at that conversation — there are dashboards for it, analyst reports, board slides, escalation paths.
There’s one more line on that list, and it isn’t on anybody’s constraint model.
People who know how to build this.
Not people in general. Core infrastructure people. The ones who can lay out a facility, design the fabric, turn it up, and then stand in front of it at 2 a.m. when something breaks in a way the runbook didn’t anticipate.
The data center workforce shortage is a hard constraint. It has the longest lead time of anything on your sheet. And it’s the only one where we are actively shrinking supply while we solve for everything else.
The demand is not in dispute
As of September 10, 2026, Aterio tracks 7,027 data centers in the active U.S. pipeline — 2,089 operational, 818 under construction, 4,120 announced. Texas leads on facility count at 1,360, Virginia 1,047, then Georgia, Pennsylvania, Ohio. Roughly 521 gigawatts of estimated capacity sits behind that pipeline.
Lawrence Berkeley National Laboratory’s June 2026 update put 2024 data center consumption at 192 TWh, 4.7% of all U.S. electricity, and projects 649 TWh by 2030 — close to 12% of national load, requiring 148 GW of new interconnection.
Link: Drinking the Future Water
I’ve written before about the water side of this buildout — same pattern, different resource.
And the supply chain is already saying it can’t keep up. CIO reported earlier this month that lead times which used to run 30 to 45 days are now 9, 12, sometimes 18 months.
Every one of those constraints gets tracked, modeled, escalated to someone with a budget.
Now run the same exercise on the people.
The supply curve nobody drew
Uptime Institute’s 16th Annual Global Data Center Survey, released this July, found more than half of operators now report difficulty finding qualified candidates. Turnover stays persistent, with staff routinely lured away by other data center companies.
Read that second part as a market signal. We aren’t growing the pool. We’re trading the same people back and forth at rising prices and calling it a labor market. That’s what a market does when supply is fixed and demand isn’t.
On the network side the trend line is steeper. Enterprise Management Associates has asked the same question every cycle since 2022. Organizations struggling to hire and retain network engineers: 26% in 2022. 41% in 2024. 52% in 2026.
The three hardest skills to source, per EMA: network security, networking for AI workloads, and network automation. Sit with that list for a second. We can’t staff the people who secure the networks. We can’t staff the people who build the networks the entire AI buildout depends on. And we can’t staff the people who would automate our way out of the first two.
Underneath all of it, a retirement wave already in motion. A 2023 OpenGear survey found 86% of CIOs expected a quarter of their network engineers to retire within five years. Three of those five years are behind us.
The shortage has a shape
This is the part that changed how I think about it. The shortage isn’t general. It’s located.
EMA’s research includes a network architect at a multinational bank describing what he actually sees. They’re trying to hire a senior engineer and an intermediate engineer — people with 10 to 15 years of experience. Every resumé is either a 25-year veteran with multiple CCIEs, or a junior who has never done an implementation. Qualified for site support and not much else.
The top is retiring out. The bottom is arriving unfinished. The middle is gone.
A shortage with that shape isn’t a recruiting problem. It’s a manufacturing problem. Something that used to produce mid-career infrastructure engineers stopped running, and we’re now drawing down inventory.
So it’s worth being precise about what that something was, because we can restart it.
How one of these people actually gets made
Paul taught me how to make an Ethernet cable.
How to hold the crimpers. How to seat the pairs before you put the end on. Then router configs — not just the commands but what they meant, why that line was there, what it was doing. And once I could do it the long way, he showed me the shortcut codes.
Then he left me at a customer site by myself.
We had walked the whole thing together first. Where the cable runs go. Where the equipment goes. The configuration. Then he left, and it was mine. My job to own it now, not follow it.
I ran the cable. Built the ends. Brought the router and the switch up.
No lights.
I had mixed up the two ends and built a crossover cable. Two different pinouts, one on each end, and I’d swapped the layout. I stood there with a problem that belonged entirely to me.
And the question I had to answer was: which end do I fix?
That’s the moment. Not the class, not the cert, not the documentation. Somebody hands you the failure and walks away, and you reason from what you know to what’s wrong. That’s where an engineer starts.
Paul showed me the way — not the commands, the ownership.
We were at Darwin Networks — a VC-funded wireless ISP out of Louisville, selling across all 50 states. That’s where the crimpers were, and the customer site, and the crossover cable.
Paul Schultz went from that building to eight years as a network architect at Google, covering backbone and edge for their global network. Then to Gaikai at Sony, running network engineering behind the PlayStation Now streaming service, where he launched twelve data centers across three continents. Then back to Google as a senior staff engineer, tech lead for their global internet edge and CDN. Today he’s a principal network engineer at SpaceX, working on Starlink ground networking.
A startup ISP to space lasers. He taught me how to hold a pair of crimpers.
I’m not the least bit surprised where he ended up. He had a passion for this work that was contagious, and it’s the reason I went and learned more on my own instead of waiting for someone to put it on my calendar.
From there I climbed too. Programming switches. Then compute. Then storage. Then all three together, which is when it clicked, because you can’t design anything real until you understand how the pieces fail each other. Then data center builds — physical layout, hot aisle and cold aisle, top-of-rack design, power and cooling and cable management as one problem instead of three. Then virtual layouts on top of the physical. Then the routing that ties it together: BGP at scale, dual routing protocols, OSPF redistributed into BGP, and the judgment to know why you’d do it that way and what breaks if you don’t.
Every rung was load-bearing for the next one. And that ladder is the entire reason I have a career: Cisco, Nexum, HP, F5, Oracle, Microsoft, six startups, NetApp today. None of it was a plan. All of it traces back to Paul and a handful of others who took a chance on somebody before there was much evidence the chance was a good one.
That’s the process. That’s what stopped running.
We are removing the first rung on purpose
A lot of infrastructure work got redefined as portal work over the last decade. Provision from the console, open a ticket, run the wizard. Click-ops.
Link: Expert Mode: No Guardrails but With Control
This is the same tension I’ve written about in platform governance — guardrails that protect without removing the ability to understand what you’re operating.
Abstraction isn’t the villain. Abstraction is how we scale, and engineers who pair real fundamentals with automation are the most valuable people in this industry right now. But operating an abstraction doesn’t teach you what’s underneath it, and this discipline is defined by its failure modes. Nobody needs a senior engineer when the system is healthy. Then the second path in a redistribution loop starts blackholing traffic at 2 a.m., and you find out exactly what your team knows.
The deeper problem is what’s happening to the entry level.
EMA names it directly: the NOC was the apprenticeship. That’s where engineers got made — a junior in a chair, triaging events, escalating, getting told check the interface counters first, absorbing a craft over a few years from people who had already made every available mistake.
That chair is being automated. One architect at a Fortune 500 retailer told EMA the explicit long-term goal was eliminating junior-level roles. Another described leadership wanting to run with 10 people what used to take 25.
EMA also surveyed 152 IT and security decision-makers: 72% were at least somewhat concerned that AI replacing entry-level roles would damage their ability to develop engineering talent internally.
So the industry knows. It’s proceeding anyway, because the headcount savings land this fiscal year and the consequence lands in 2032.
Here’s the diagnostic. Run it on your own organization this week: if you eliminated every entry-level role on your infrastructure team, where does your next senior engineer come from?
If the answer is “we’ll hire one,” ask who you’re hiring them away from, and ask what that operator does next.
That’s not a workforce strategy. It’s a shell game with a ten-year fuse.
What depth costs, and who pays it
Link: No One Owns — AI
Judgment is also what ownership rests on, which is a problem I keep running into from the other direction.
At some point I decided I needed to go deeper than the job was going to take me, and nobody was going to hand it over.
I paid my own way to a two-week CCNP boot camp in the Pocono Mountains. My money. Two weeks away from home, away from my wife, hands deep in protocols and design. It was also the first time I ever shared a condo with someone I didn’t know.
I got sick while I was there. My instructor took care of me — went out and picked up medicine. First time I ever ate ramen noodles. My wife mailed us Graeter’s ice cream pints and I handed them out to the whole class.
Nobody treated it like a competition. A room full of people trying to get good at the same hard thing, taking care of each other.
I came back with a CCNP — a certificate that said I could build dual routing protocols, redistribute OSPF into BGP at scale, and explain why I did it that way and what breaks if I don’t. But what made those two weeks work was a guy buying cough medicine for a student he’d known four days.
Depth costs evenings, weekends, unreimbursed money, labs that won’t converge at midnight, time away from the people you love. And it has never once been delivered to me by a platform. It came through people who decided to invest in me. Paul at a customer site. An instructor in the mountains with a bag from the pharmacy.
AI doesn’t change that. AI is a real multiplier for an engineer who already has judgment. It’s a very confident liar to one who doesn’t.
What to actually do about it
If you lead infrastructure people:
- Protect the first rung. This is the whole thing. Every junior role you automate away is a senior engineer you won’t have in 2032. If AI goes into the NOC, point it at making juniors dangerous, not redundant.
- Let people own something and be wrong on it. Paul walked out of that building on purpose. Supervised failure is the most efficient teaching tool this industry has ever had, and we’ve engineered most of it out of the job.
- Forecast labor like you forecast megawatts. You model racks, power, and lead times to the quarter. Give humans the same rigor. Labor is a fixed cost of your timeline, not a line in the risk register.
- Fund the depth. The boot camp, the cert, the lab time, the two weeks away. I paid for mine and I’d do it again, but the fact that I had to says something about how we funded development.
- Use the pipelines that exist. Microsoft’s Datacenter Academy, Google’s STAR program, AWS Workforce Accelerator, Oracle Pathways, Equinix’s Pathways to Tech, community college partnerships like NOVA’s. And look hard at the trades — IBEW and UA apprenticeships have quietly run the best talent development model in this industry for a century. Copy it.
If you’re an engineer:
- Learn the layer beneath the one you work at. If you write Terraform, learn what the network does when you apply it. If you run the portal, learn the protocol.
- Break things on purpose in a lab. Fundamentals aren’t a certification, they’re a habit.
- Find someone who built things before the abstraction existed and ask them to teach you. Most of them would love to. Nobody asks anymore.
- If you’re senior, be somebody’s Paul. Take a chance on someone before the evidence is in. That’s how every one of us got here.
The longest lead time on the sheet
Put them side by side.
Silicon: months. A new fab: two years. A journeyman electrician: four to five. Someone who can stand in front of a flapping BGP session at 2 a.m. and know where to look first: a decade — and for a lot of them, we never started the clock.
You can expedite a transformer. You can’t expedite a person.
Every megawatt we announce is a promise that somebody will be standing there to build it, turn it up, and keep it running. We are making those promises considerably faster than we are making those people. That is a resource constraint, and it deserves a slide in the same deck as the power study.
The good news is that it’s the one constraint on the list you can personally do something about this week. You can’t manufacture a transformer. You can hand a junior engineer something real, walk out of the building, and let them figure out which end to fix.
I didn’t earn my way onto that ladder. People put me on it and let me climb. That’s why I spend so much of my time now trying to give somebody else the same push.
Be Human First.
Sources: Aterio U.S. data center tracking (Sept. 10, 2026); Lawrence Berkeley National Laboratory, U.S. Data Center Energy Usage Report: 2025 Update (June 2026); Uptime Institute 16th Annual Global Data Center Survey (July 2026); EMA Network Management Megatrends 2026; OpenGear CIO survey (2023); CIO / Network World, “IT infrastructure shortages are real and lasting” (Sept. 7, 2026).