It took 25 years in Africa to go from one phone landline per village — to over 75% of the population carrying a mobile device.
In the late 90s, telecommunications in many African nations looked like this:
One landline for an entire village. If you wanted to make a call, you waited in line. If you wanted a phone in your home, you waited years for a copper wire that might never arrive.
How did the transformation take place?
✅ Pre-paid over Contracts: By selling airtime in small increments, telcos made connectivity accessible to the informal economy.
✅ Mobile Money (The Game Changer): When banking infrastructure failed to reach rural areas, platforms like M-Pesa turned the phone into a digital wallet. Millions of “unbanked” people suddenly had access to global finance.
✅ Infrastructure Economics: It’s easier to plant a cell tower on a hill than to bury thousands of miles of copper wire across the Congo Basin or the Savannah.
The Impact:
• Agriculture: Farmers check global market prices via SMS, cutting out predatory middlemen.
• Energy: Pay-as-you-go solar kits are managed and paid for via mobile, bringing light to off-grid homes.
• Health: SMS-based diagnostic tools are saving lives in remote regions.
The Lesson for Leaders:
Innovation doesn’t always have to be linear. Sometimes, the greatest growth happens when you stop trying to fix an old system and “leap” straight to the future.
Africa didn’t just join the digital age—it redefined what the digital age looks like for the emerging world.
Now that said, what are we going to do about AI and data center technology transformations about once a week now overtaking real world construction, infrastructure, and investment velocity?
Over 25 years, such as Africa in telecommunications, you can work it out. But what about transformations in 25 minutes? That poses a problem. A lot of “gambling“ has needed to go on in the attempt to keep up with the AI transformation.
__________
Pejman GolkarBuild-to-spec data centers are dead.
What replaces it? Scaling in Time.
The physical world is being redesigned for AI — Part 2
For decades, data center construction followed one rule: finalize the specs, then build.
Design-build. Fast-track. Modular. Different methods — same assumption.
You’d eventually lock the endpoint.
AI compute just killed that assumption.
There is no endpoint anymore.
The timeline broke:
→ GPU refresh cycle: 12-18 months
→ Data center construction: 24-36 months
→ By the time you’re done, the chip has changed twice
Nvidia’s roadmap proves it:
→ Now: Blackwell GB200
→ Late 2025: Blackwell Ultra B300 — different power, different cooling
→ Late 2026: Rubin — different power, different cooling
→ Late 2027: Rubin Ultra — different power, different cooling
Four generations in three years.
Why it breaks:
You break ground today for GB200.
Month 12: Building 50% complete. Blackwell Ultra ships. Already last-gen.
Month 18: Rubin ships. Different power. Different cooling.
Month 24: Building opens — optimized for hardware two generations behind.
You just spent $800M on a facility designed for an obsolete chip.
The old way worked when chip cycles were 5-7 years.
At 12-18 months, it’s impossible.
So how do you build for a chip that doesn’t exist yet?
Scaling in Time — and Microsoft is already doing it.
They call the unit that makes this possible a “cell.”
I watched Microsoft executives tour Fairwater 2 in Atlanta.
They walked into a cell. Everything finished:
→ Cooling loops installed
→ Power busways in place (800V DC ready)
→ Fiber trunks connected
→ Floors reinforced for 3,000+ lb racks
But no racks. No servers.
Some cells are running. Others are waiting for next-gen chips.
The cell is the buffer — it lets them wait for Rubin Ultra without stalling the build.
Satya Nadella: “You don’t want to build to one fixed spec. You want to be scaling in time, not scale once and get stuck.”
What “Scaling in Time” means:
→ Build the backbone early
→ Leave the last mile open
→ Defer final decisions as long as possible
→ Fit-out at the last responsible moment
You’re not building for a spec anymore. You’re building for optionality.
What this breaks in construction:
→ Procurement can’t lock equipment too early
→ Mechanical installs backbone, pauses terminations
→ Electrical installs busbars, not final connections
→ Schedules must absorb uncertainty, not eliminate it
This isn’t poor planning. It’s adaptive planning.
The contractors who understand this will dominate hyperscale.
The ones who bid like traditional data centers will get burned.
The physical world is being redesigned for AI.
