"Smaller than a pizza box and can fit in any data center" sounds like marketing for "We put it in a standard rackmount form factor"

If we ignore cloud interfacing solutions[1], there is a simple answer. "Enterprise scale" solutions really mean "Enterprise SCALING".

Pre-sales for 90 servers in many cases would be impossible. Pre-sales of a single rack mountable item that can then be expanded from depending on the workload is really attractive, however. Write the software right, and it should be able to cope.

Pre-sales probably starts with a customer renting a single box to try out, possibly with some professional services time. The customer decides they like it, and commission a project to develop a Watson solution. IBM pre-sales then work out how many servers would actually be needed, once the customer is interested. As the solution can scale, when management goes "that cost looks little high" they can safely say "well, if we reduce it by sixteen servers, you'll be slightly slower but that might be ok. Then if the business decides to increase we can simply add more servers[2] at a later date". And, of course, adding more servers at a later date means more professional services time, meaning more money and more quality customer relationship building time.

[1] Which for all kinds of commercial reasons, we have to discount a cloud solution being the default here. Buzzword, yes, but enterprise is slow and many really exciting uses cannot be based in the cloud.

[2] Or, y'know, just turn this key....

I'm assuming they mean 1 server is a pizza box (ie. it's a 1U rack server), so you'd still need ~90U of space.

Either that or IBM has made some breakthroughs in the last 2 years.

You can almost hear the article author completely misunderstanding what he was told.

I'm betting it was something like, "For Jeopardy, Watson was using 15 Terabytes of data stored on 90 servers. And those servers were large back then, now they're your standard pizza box 1U's. And because of advances in storage technology we can store a lot more in the same space."

Isn't there some sort of basic tech literacy class for people who write these articles? It's like reading the sports section and seeing "The Seattle Seahawks are a group of big men and they ran at another group of big men, and they've been doing that a lot, and faster lately. Some of the men in the Seahawks group are smaller and faster, so they are harder to catch by the other group of big men. The Seahawks grouping of men is like a successful business, because they've been winning competitions against other groups."

There is some more prize gibberish in that article:

Saxena said IBM is ready to make a big a bet on Watson, as it did in the 1970s when it invested in the emergence of the mainframe.

Leaving aside the grammatical error (it's probably just an editing artefact), what bed did IBM make in the 1970s on the emergence of the mainframe? Is that a typo for 1950s? Or just unvarnished ignorance of computing history?

It's stuff like this that makes me mistrust a news source: if they get it blazingly wrong about something I know enough about to have a working bullshit detector for, what are the chances that they're not also getting it wrong about everything else?

To be fair, the System/360 (which was the first to run the classic OS/360 operating system) was released in '64[1], and was replaced in the '70s by the S/370.

Certainly mainframes were in use prior to the '70s, but the S/360 was the first "classic" mainframe in that it was designed for both business & scientific use[2], and it wasn't until the 70's that they began to become more popular in areas outside their traditional use.

[1] http://en.wikipedia.org/wiki/IBM_System/360

[2]

16 Terabytes of RAM on P750s.

From http://en.wikipedia.org/wiki/Watson_(computer)

"Watson is made up of a cluster of ninety IBM Power 750 servers (plus additional I/O, network and cluster controller nodes in 10 racks) with a total of 2880 POWER7 processor cores and 16 Terabytes of RAM. Each Power 750 server uses a 3.5 GHz POWER7 eight core processor, with four threads per core. "

The main change from the Jeopardy competitions is they don't need to find an answer in under a second. So you can probably reduce to computitional capability significantly.

I would bet on optimization of the software, lower costs of memory, and database sanitization.

I think prior to that it mentions its cloud based, so the head end is probably only 1U.

A stab in the dark would be chips made from Graphene or IBM's Silicene, if I am to understand what you quoted.