Buying tin, selling clouds

In the nineties and early naughties I spent a fair part of my life buying computers, servers and printers, as well as working on some big procurement and framework agreements.

This gave me a hell of an insight into how the tier two build on demand manufacturers worked. The big boys, the Compaqs, the HP's and the rest in the main built standard configurations that more or less met their client's requirements, and stored them in warehouses. This of course cost them money, but meant that if VeryBig plc wanted a thousand desktops delivered to Warrington next Tuesday, they could have them.

The tier two guys were different. Some were better than others, but I've walked round enough of their assembly facilities to know they all worked more or less the same.

As build on demand manufacturers they of course had no inventory. They built it, put it on a truck and delivered it. They also had low component stock. If you wanted a big order - more than twenty or thirty units you needed to talk to them, as they would effectively book the production line for you for a day and order in all the components they needed for that run.

Most of them tried to push standard configurations at you, but you could always arrange something different - eg a different network card, hard drive unit. Everything was negotiable, and in the main they were fairly flexible.

Most of them were very conservative on component choice - they bought good quality bits from good manufacturers in the Far East - while we joked about no name units they weren't, they were just name we didn't recognise.

The real difference was quality control - remember you could go and buy the bits build a pc yourself, and some people did, but these guys did it better by testing things, letting pc's burn in. Remember, dealing with warranty claims was expensive as it often meant sending people out to swap things, so their aim was to do everything that reduced cost.

The result was unglamourous companies working out of tin sheds in disticntly low rent locations, building predictable product.

These guys then started making servers. At first they were not good, but they started getting better with better motherboards, disk controllers and more testing to the point where they could sell you a generic server that performed as well as a brand name unit but at half the price.

By the time the virtualisation and blade server revolution rolled around I wasn't buying tin any more but looking at how Dell's blade offerings developed over the years I would not be surprised if they followed much the same trajectory with some attrition - those that couldn't scale up dropping out of the race.

Now we're seeing something of the same thing happening at both ends of the spectrum - the no name Chinese Android tablet makers are producing good and reliable product and gradually starting to eat into the brand name vendors market share. Aldi and others have offered noname tablets in the past and more relevantly Tesco in the UK is launching a no name tablet.

This is a very important move - a supermarket chain - and Tesco is massive - is risking their brand name on a tablet. Tesco want customer satisfaction and near zero warranty claims as a result of this move as they don't want to damage their reputation or pay out money to handle warranty claims.

At the top end something equally interesting is happening. The big commercial cloud service providers are increasingly able to demonstrate that they can provide disk and compute than less than in house. To do this they need to buy racks worth of servers. And they are buying them from the no name server makers. The cloud providers have enough so that they can swap dead units in and out and are of such a scale they can afford to have generic units sitting around as spares, meaning that they don't need the fancy fast response warranties that the top end manufacturers provide - and as an IT house you pay for one way or another.

And warranty and maintainance costs are a major part of running infrastructure. Run lots the same generic kit, and providing the kit's reliable enough you can get by with reduced warranty terms providing you're big enough to afford to hold spares and employ engineers who can fix things.

It's what we found when I bought tin - the tier two manufacturers were happy for you to (hold (their) spares if you had enough of something and do first fix as it saved them money. They'd even help train your in house people to do their fault diagnosis.

It's much the same with servers. You don't want servers to go down. Unfortunately entropy being what it is means that you can never get a 100% reliability. So you need to add redundancy. That costs. And you need fast response, as when you've failed over to your backup nodes you're vulnerable to another failure.

The big cloud providers don't have this tension. They have enough in the way of kit and capacity to cope with routine failure in a way you never can as an individual private cloud provider. They can run with a lower warranty config - I'm sure a 48h return to base warranty would work fine for them.

It's significant that most of the recent spectacular cloud outages are due to configuration problems, not smoke, fire and dead spindles ...

Written with StackEdit.