Mr. Moore in the Datacenter

Yahoo is celebrating its tenth anniversary, which led me to reflect on the fact that a decade is a very long time in Silicon Valley, particularly when viewed through the lens of Moore’s Law. As Ray Kurzweil and others have observed, when humans contemplate exponential progress, we tend to overestimate what can be accomplished in the short term (where the curve is relatively flat), but we tend to underestimate progress in the long term (when the curve gets very steep, goes up and to the right, does a hockey stick, etc.). People tend to think more easily in powers of ten (as opposed to the powers of two prevalent in the technology industry), so a decade is a good duration to look back and consider what Mr. Moore has done for us lately, after we’ve had six or seven doublings of memory density, computing speed and bandwidth. The datacenter is a great place to look to see these trends converge.

Given my background, it is perhaps not surprising that I’m a big fan of consumer-oriented web services such as Google and Yahoo, as well as recent Mobius VC investments Technorati and NewsGator. I’m equally enamored with enterprise-focused software-as-a-service businesses such as RightNow, and Mobius VC portfolio companies Postini, Quova and Rally Software Development. Another cool software-as-a-service startup offers a hosted application wiki and is called JotSpot, which was founded by Joe Kraus and Graham Spencer, two of the guys with whom I co-founded Excite back in 1993. Though these enterprise and consumer oriented companies have different revenue models, their delivery model and back-end architectures for serving their customers are fundamentally similar.

Each of the companies I have mentioned above have benefited greatly from the drastic increase in the amount of storage, computing power, bandwidth and datacenter rack space that a dollar buys in 2005 versus what a dollar bought for the same thing in 1995, back when Yahoo and Excite launched their sites. I spent some time poking around the web trying to find 1995 prices for CPUs, RAM, storage, bandwidth and colo space, but it turns out that this kind of pricing archaeology is difficult to practice online, as all searches for these things turned up advertising and commerce sites focused on selling me these commodities today, not ten years ago.

I related my problem to my colleague Jocelyn Ding, who is the SVP Business and Technical Operations at Postini, and she was able to track down some old price lists from a variety of vendors from 1995, 1998 and 2000 for bandwidth, cage rental, one and four CPU servers and storage systems. Thanks go to Jocelyn for helping me put some real numbers behind my somewhat obvious assertion that a dollar goes much further in today’s datacenter than it did a decade ago. I also dug up a good whitepaper detailing costs of enterprise storage since 1992 which includes projections to 2010, which can be found here. For some items, the prices didn’t go back to 1995, so in those cases, I have extrapolated the price trend to estimate 1995 costs based on 1998 or 2000 costs relative to today’s costs:

Bandwidth: $1100/megabit/month in 1995 vs. $128/megabit/month in 2005

Cage Space: $175/sqft/month in 1995 vs. $25/sqft/month in 2005

Disk Storage: $1,300,000/TB in 1995 vs. $3,300/TB in 2005 (SCSI RAID)

1-CPU Server: $25,000 in 1995 vs. $1,000 in 2005 (web server class machine)

4-CPU Server: $360,000 in 1995 vs. $38,000 in 2005 (with 16GB RAM)

In addition to the price reductions, we also have to look at the compute performance of a web server class machine in 1995 vs. today. Given five or six performance doublings since 1995 courtesy of improvements in clock speed, bus speed, architecture changes from 32 to 64 bit, additional cache memory and faster RAM, a conservative estimate would be that today’s single CPU 1-U “pizza box” web server is roughly fifty times faster than last decade’s model. Couple that with the 25x price difference for this pizza box, and your 2005 dollar buys you more than one-thousand times as much compute power as it did in 1995. Bandwidth is at least ten times cheaper than it was in 1995, floor space in the data center is seven times cheaper and enterprise-class storage is at least four hundred times cheaper than it was only a decade ago. With some smart software and network engineering, the cost per gigabyte of storage can be brought down an order of magnitude further still using a distributed filesystem based on low-end IDE drives. Finally, with the rise of Linux, Apache, MySQL and open source in general, software license costs can also vanish from the equation when running a large-scale web service.

What does this mean for a web-services company? The cost to deliver an application to an end-user has dropped dramatically for these companies and the cost to operate their data centers therefore has much less of an impact on their costs of operations and capex budget than it used to, which means their gross margins for delivering their product have improved significantly since 1995. For companies like Yahoo, Google and more recently, Technorati, this means the cost to deliver a page view or search results page has gone down dramatically, while the average size of a search-results page is perhaps only marginally larger since 1995. Even considering the size of a search index (Google’s 8B pages today vs. Excite’s 10M in 1995) has grown nearly one thousand-fold, the costs of computing power and storage have accommodated this expansion while bandwidth costs and rack space have fallen nearly tenfold.

For enterprise-focused companies like, Postini, Quova and Rally, the story is similar. Add in a subscription-based recurring revenue stream and you have a business model that has all the benefits of a dependable revenue stream and profit margins that can approach those of a traditional software company. Thanks to the low cost and high performance of today’s hardware coupled with an elegant service architecture, Postini is able to process several hundred million email messages per day for its customers with an extraordinarily light hardware footprint and does so quite profitably as a result.

The fun thing about doing a retrospective like this is to realize that when I write about this again in 2015, the increases in CPU speed, memory density and bandwidth will make today’s costs and capabilities look as quaint as 1995’s do today. Thus the environment will continue to become more hospitable to the software-as-a-service model, more entrepreneurs will create meaningful businesses based on this model, VCs will continue to invest in these ideas (myself included), and we’ll all be able to enjoy some mind-blowing applications a decade from now that are simply not possible today.