I don't think that we'll see that kind of compute power on the desktop simply because most users have little to no use for the floating point compute power they have now.

Consider this: The Core 2 Duo chipset in this Macbook Pro I'm using now has about, oh, 30 times the floating point capacity of a Cray X-MP. I feel pretty comfortable saying that for general floating point crunch that it would be on par or beat a Cray 2. It isn't until you get up to the Cray "T" series of CPU's that you pull away from the floating point compute power that you likely have on your desk today, unless we're talking specialized vector problems (eg, FFT's) and then the specialized vector architecture on the Crays might eek out an edge.

Most PC's have great gobs of floating point processing capacity now - you have both the single/double precision crunch in your Intel or AMD CPU's, and then you have smokin' fast vector single-precision power in the video chipsets. Your average 64-bit game console can render images so much faster than the early Cray machines that it isn't remotely funny.

What are people doing with the incredible computing power they have already? Playing games. Rendering porn. Sending 10K "Cheap Viagra!" spams per minute. A minority of computing users are doing actual/useful work with their cycles on things like spreadsheets and various stats/modeling packages. What am I doing with this 30X speed of a Cray XMP? Typing this in one window and listening to a tape of a Dead show in another window. The CPU usage graph thingie shows that I'm using about 8% of one CPU and maybe about 12% of the other. More cycles ain't gonna make typing go any faster or make Jerry sound any better.

Once we got to a point where we achieved full motion video on the desktop (say, 30+ frames per second), the push for "more, more, more" started abating - and going after even 10's of teraflops, much less petaflops, just seems like spec one-upmanship. I'm now at a point where I can take on computing problems that when I started playing with computers in the late 70's on a PDP-11, I thought I'd never, ever be able to take on. Geeks like me dreamed about the power of machines like the Crays, CDC 2xx's, IBM's big vector iron, but that's all we could do, because there was no way we geeks could buy a Cray, much less feed the monster with the 480/3-ph power, cooling plant, disks, IO processor, etc.  All through the 80's and 90's, there was no machine that could keep up with the workload we were shoveling onto it. At cisco, we had 8-way and 16-way SMP servers at our disposal, and I had a Sparc 20 with two CPU's and gobs of memory on my desktop besides that - and still, I was waiting 45+ minutes for a build of the system. For 20+ years, I wanted systems that were 10's of times as fast, and all I could get was 20 to 100%  as fast in any upgrade, no matter how much money we threw at the problem.

Now, short of playing with these climate modeling programs to verify and test various apocalyptic claims, I don't have any problems that require even the floating point throughput I have now, much less massively parallel systems like this latest one you reference.  My compiles are done in a blink now, even on a laptop. When the latest round of Intel/AMD CPU's came out, a threshold was finally crossed - I can honestly say that even when I'm developing big s/w builds, playing with big Lisp workspaces or working on SPICE simulations, I no longer wish that the system would be 10's of times faster. Maybe twice as fast, but not 10's of times as fast. And if I put up the money, I could get a Mac Pro or something similar with 8-CPU SMP, 16 GB of core, RAID disks, the whole enchilada - for under $10K --  a trivial amount of money to gain access to absurd amounts of capacity by comparison to the supercomputers of only 10 years ago. With this laptop (which is constrained compared to a desktop machine due to heat dissapation and memory configurations), I can emulate any machine I ever used in my career faster than the native iron ran.

This, BTW, is a problem for the computing industry. Users have reached a point where many no longer feel a need to upgrade. Once most users had upgraded to WinXP and a sufficiently large machine to run it well... there is little perceived "need" to upgrade. Everything runs as well as users want. Most users will never go to gigabit Ethernet - 100Mbit/sec Ethernet is more than enough for them. Once they get 3 to 5 Mbit/sec broadband, there is nothing that is crying out for a bigger pipe. Bigger disks for users than 1Tbyte? I'll have to wonder why. Unless someone is a photographer and is laying down a lot of raw images, how are most people going to fill a 500GB disk, much less a Tbyte? Sure, Intel/AMD will keep pushing out faster and faster CPU's - but I think now that the emphasis will be on "operations per dissapated watt" rather than on all-out computing power.