The cage in Johannesburg has been ours since last year. Ten Ryzen 9950X nodes sit in it, racked, cabled, burned in and currently carrying nothing but our own monitoring traffic. What is not finished is transit, and until that is finished the site takes pre-orders rather than orders.
We are announcing it anyway because people ask about African capacity roughly once a week, and the honest answer is more useful than silence.
What is in the building
Ryzen only, to begin with. The EPYC line needs memory density and a fatter uplink to be worth selling, and neither is justified until we know what demand looks like. Storage nodes need bulk drives that we have not shipped there. GPU instances need power per rack that the site cannot give us this year.
| Item | At launch |
|---|---|
| Families | Ryzen NVMe only, four plan sizes |
| Uplink | 100 Gbit/s to the site |
| Filtering | 4 Tbit/s of scrubbing at the regional edge |
| IPv6 | Routed /64 on every instance from the first boot |
| Custom ISO | Supported, same as everywhere else |
| Out-of-band console | Included, same as everywhere else |
Why transit takes months here
Bandwidth into and out of southern Africa is priced per megabit in a way most of our other sites stopped being a decade ago, and the capacity that makes it cheaper arrives on submarine cable, on a schedule nobody local controls. We are negotiating with multiple Tier-1 transit providers and one regional carrier group. Two of those conversations are about price. One is about whether a route that currently hairpins through Europe can be made to stop doing that.
That hairpin is the whole reason latency in the region is worse than the map suggests. Our own probes from the racked nodes currently measure the following, and these numbers are the reason we have not opened ordering:
| From | Round trip today |
|---|---|
| Amsterdam | 158 ms |
| New York | 232 ms |
| Singapore | 214 ms |
| São Paulo | 338 ms |
The Amsterdam figure is respectable. São Paulo is embarrassing for two continents that face each other across an ocean, and it is the direct product of traffic between them travelling north first. We expect to improve two of these four numbers before opening. All four would be a promise nobody could keep.
What a pre-order actually means
The first fifty instances go at launch pricing, which is the current Ryzen list price held for twelve months regardless of what the site ends up costing us. A pre-order reserves a slot in the build order. Nothing is charged until the instance boots, no deposit is taken, and cancelling requires an email containing the word cancel and nothing else.
If the site slips past the end of this year, pre-orders convert to a slot at any other location at the same held price, and we will write that offer down rather than leave it to a support conversation.
Our record on this
We have slipped before. Santiago was announced in early 2023 and opened five months later than we said, because a power commitment we treated as settled was not settled. Sofia arrived on time. Dubai arrived three weeks early, which is not a skill, it is a delivery lorry being early.
The pattern is that sites in mature markets land when we say and sites in developing markets land when the physical layer says. Johannesburg is the second kind. We would rather tell you that than pick a date, miss it, and write a different sort of post.
Why bother at all
A quarter of the tickets we get about latency in Africa are from people currently serving African users from Europe, paying for a 158 ms round trip in each direction and telling themselves it is fine. It is not fine for anything interactive. The rest are from people who want the jurisdiction, which is a legitimate reason and one we do not argue with.
There is a third group who want the site because it is unusual. That is also fine, and it is roughly a third of what makes any of our smaller locations viable.
Progress goes on the status page as it happens, and the site page in the location index shows JNB-01 as pre-order until the day it does not.