Architect for Detachability: Build IT So the Carve-Out Is Configuration, Not Reconstruction

In December 2015 my status report carried a project codename — Project Carbon — that read, on paper, like infrastructure housekeeping. Should a datacenter in Karlsruhe stay or go. How many internet circuits we actually needed. What hardware the Cisco call-manager required to keep running. Underneath the housekeeping, it was the first datacenter-level engineering to separate the SaaS business that would spin out of Citrix — and the deal team had not formally shown up yet.

Architect for detachability is the discipline of building the IT estate so that any business unit can leave at a clean, known seam — a configuration change rather than a reconstruction. This is the long version of one worked example: a single quarter in 2015, at the keystroke level, that decided how fast a separation could run years later. Not the thesis in the abstract. The actual moves, and the number they produced.

The room nobody was watching

My weekly note from December 9th says it plainly. I was “identifying if the Karlsruhe IPC3 Datacenter should be kept in place to support a smooth spinoff of spinco,” and working Facilities to “reduce the footprint / consolidate internet circuits … to ensure a cost effective way going forward.”

Read at face value, that is a cost line. Underneath, it is a separation decision made before anyone called it one. I was reducing and modularizing the footprint of a business unit so that, when the spin-out arrived, the seam was already cut. The hard part of a carve-out is not the cutting. It is having drawn the boundary cleanly enough, early enough, that there is something clean to cut.

That is the unglamorous core of detachability, and it almost never looks like architecture while you are doing it. It looks like circuit consolidation. It looks like a datacenter keep-or-kill memo. It looks like the kind of work that gets filed under housekeeping and forgotten — right up until the day it is the only reason a deadline is survivable.

What designing-to-detach looked like at the keystroke level

Drop down one more level, into the same stretch of weeks, and the pattern gets concrete.

The boundary you draw is the boundary you can cut. In early December 2015 I filed a change to add an IP address on the Sonus session-border controllers across the datacenters — specifically, in the words of that status note, “to keep Exchange UM configurations and Enterprise SIP trunks separate on the SBC layer.” Exchange voicemail traffic on one address, enterprise SIP trunks on another. On its face it is a minor hygiene decision. It is actually a separation decision in miniature. Two flows that share an address are two flows you will spend a fortnight untangling the day someone needs them apart. Two flows you partitioned on purpose are two flows you can sever in an afternoon. Every clean seam is a future move-out you have pre-paid.

The identity boundary is the one most estates get wrong. The same instinct shows up wherever identity meets jurisdiction. On the global Cisco voice estate I designed logical partitioning and geolocation controls so that India’s Department of Telecommunications toll-bypass rules were enforced inside the architecture, not bolted on after — a regional boundary that held because it was drawn into the design. Carry that forward to the spin-out itself: as the EMEA footprint separated, regional identity governance let access be localized and severed cleanly, and access overlap fell 70 percent. An identity store built as one shared organism cannot be cut along a business-unit line without tenant-wide surgery. An identity store built jurisdiction-aware comes apart where the business actually divides. That is the difference between a divestiture and an excavation.

The decommission is the proof the boundaries held. When the IPC3 datacenter was finally scheduled to sunset — by the end of January 2016 — it was not an emergency. It was a planned move-out: a gameplan to relocate the remaining voice hardware, the circuits already consolidated, the numbers already mapped to where they needed to land. The expensive part had been done in advance, on purpose, while it was cheap and invisible. The sunset was just the bill arriving on a balance someone had quietly paid down for months.

The receipt, six years on

A worked example is only worth the result it produced, so here is the result — one continuous line from that datacenter to the number that closes it.

By 2021 I was directing the corporate-IT estate through a consolidation that took 10-plus locations down to three global hubs — a lean Frankfurt build stood up despite a global supply-chain shortage, an Andover-to-Grand Rapids consolidation, and the remaining Americas footprint. Across that consolidation we pulled services out of Santa Barbara, e-wasted two NetApp SANs and 8-plus racks, and shipped 10-plus pallets of hardware out the door. None of it disrupted the business.

Here is the receipt. The all-in IT datacenter run-rate went from $27K a month — $324K a year — to roughly $13.5K a month, about $162.6K a year. Cut close to half. That is the same instinct that named IPC3’s circuits in 2015, applied at enterprise scale six years on: fewer seams, drawn cleaner, costing less to hold — and, not by accident, far easier to cut when a divestiture needs them cut.

That estate fed a real M&A record: eight deal events, five of them run end to end, roughly $1B-plus in cumulative deal value. The Project Carbon work matured into the 10-country GetGo spin-off that carried 2,500-plus users — a 120-day separation, and the origin story for everything that followed. Two later separations — a divestiture to Genesys, and a 30-plus-country carve-out — each came in under 60 days. The speed of those two is the part people remember. The architecture underneath them, going back to a circuit-consolidation memo in 2015, is why the speed was available to spend.

Why the speed is a lagging indicator

There is a temptation to tell M&A stories as feats of velocity under fire. I have lived those nights and they are real. But the velocity is the trailing number. A separation you can run in 60 days is a result you banked years earlier — partitioned circuits, jurisdiction-aware identity, a footprint always meant to come apart at known seams.

The leaders who get caught flat in a carve-out are rarely slow. They inherited an estate built to be permanent — every system assuming it would never be detached — and then got handed a date. The cost of pulling it apart was set the day the boundaries were drawn as if nothing would ever leave. I had finally put words to this myself by the end of 2021, in my own year-end review: among the things I’d come to see as the through-line of the work was “modularising our footprint across all teams and solutions to allow for easy tuck-ins and spin outs.” It took me the better part of a decade — and a datacenter in Karlsruhe — to be able to name what I had been doing since before I had the language for it.

So the discipline I would argue for is almost boring while nothing is leaving, and the whole game the moment something is. Build as though every business unit might one day need to leave cleanly, even when none of them are going anywhere. Draw seams you can cut. Keep identity partitioned where the business actually divides. Hold circuits and datacenters so that “spin this out” is a planned move, not a salvage operation.

What it asks of you, and the one question that operationalizes it

It asks you to spend a little capital, repeatedly, on optionality the business has not asked for. To partition the session-border controller when fusing it would close the sprint faster. To consolidate circuits before any deal justifies the line item. To treat “easy to detach” as a design requirement sitting quietly beside cost and uptime — knowing most of those bets will never pay off, and the one that does will pay off enormously. That is the honest price: a few percent of design effort, paid on the boundaries, most of which you will never visibly cash in.

If you want one portable test, it is this. Before you fuse two systems to save a sprint, ask: if this business unit had to leave in 60 days, where is the seam? If you cannot point to it, you are not saving a sprint — you are borrowing one, at interest, against a deal nobody has announced. Designing the seam now is the cheapest it will ever be.

Architect for detachability. The carve-out you will be glad you designed for is the one nobody has announced yet.

Common questions

What does “architect for detachability” mean in M&A IT?
It is the discipline of building the IT estate so any business unit can leave at a clean, known seam — a configuration change rather than a reconstruction — by drawing modular boundaries across identity, network circuits, and datacenters in the ordinary months before anyone calls it a separation. Christian Merkel traces the discipline through one worked example: datacenter-level engineering on the Citrix/GetGo spin-off in 2015 (internal codename Project Carbon), which matured into an estate that by 2021 cut its IT datacenter run-rate close to half. The core idea: a carve-out is won or lost in the architecture, years before the deal.

What was Project Carbon?
Project Carbon was the internal codename, on Christian Merkel’s contemporaneous status reports from late 2015 and early 2016, for the datacenter-level engineering that prepared the SaaS business spinning out of Citrix (the GetGo spin-off). It covered deciding whether the Karlsruhe IPC3 datacenter should stay or be retired, consolidating internet circuits, reducing the footprint, partitioning the session-border-controller layer, and mapping the voice hardware needed to support the spun-out entity — separation work begun before the deal team formally engaged.

How much did the 2021 datacenter consolidation save?
The IT datacenter run-rate fell close to half, from $27K per month ($324K per year) to roughly $13.5K per month (about $162.6K per year), per the 2021 IT wins record. The consolidation took 10-plus locations down to three global hubs (a lean Frankfurt build stood up despite a global supply-chain shortage, an Andover-to-Grand Rapids consolidation, and the remaining Americas footprint). Across it, two NetApp SANs and 8-plus racks were e-wasted and 10-plus pallets of hardware shipped out, with no business disruption.

Why does designing IT to be modular make M&A faster?
Because the speed of a separation is a lagging indicator of the architecture. If boundaries across identity, circuits, and datacenters are drawn cleanly in advance, a divestiture becomes a planned move-out at known seams. If the estate is fused with no clean seams, every carve-out becomes a reconstruction under someone else’s deadline. Across eight M&A deal events (five run end to end, roughly $1B-plus in cumulative value), the GetGo spin-off was a 120-day separation and the origin story; two later separations — a Genesys divestiture and a 30-plus-country carve-out — each came in under 60 days, on estates built to come apart cleanly.