This article is from the August 15, 2001, issue of Red Herring magazine.

Intira's rambling 100,000-square-foot facility some 40 miles east of San Francisco was originally designed to house scientists designing the infamous Star Wars missile defense system. It has bulletproof windows, reinforced walls (considered more secure than Fort Knox), and a foundation that can handle the angriest rumbling of the San Andreas Fault.

Behind the bulletproof Plexiglas windows, a visitor can see row upon row of EMC (NYSE: EMC) storage systems, Sun Microsystems (Nasdaq: SUNW) refrigerator-size servers, powerful switches made by Foundry Networks (Nasdaq: FDRY), and Cisco Systems (Nasdaq: CSCO) routers. This is a snapshot of today's modern data center. The trouble is, it doesn't have what it takes to pass the stringent requirements of today's large corporate IT departments. "Fundamentally I believe that computing is moving into the network, and complex enterprise applications are moving into the data center," says Mark Shull, CEO of Digex (Nasdaq: DIGX), a hosting company soon to become part of WorldCom (Nasdaq: WCOM).

It all comes down to the equipment and designing it to handle the huge demands of enterprise-level data, a real possibility when several corporations are sharing the same infrastructure. Except for some slight modifications to these hardware systems over the past few years, they aren't ready for the challenge. They use equipment designed to handle the modest but fail-proof requirements of a few thousand users in a company. For corporations to outsource their IT infrastructures to data centers, what is needed is carrier-class equipment.

In the past, when data processing demand exceeded capacity, the typical response was to add more hardware. But experts say this piecemeal approach is shortsighted and expensive. "Unpacking more Sun boxes and Cisco routers isn't the answer, as there is a scaling problem," says Steve Dow, general partner at Sevin Rosen Funds, a venture capital firm.

The dilemma today is that these data-center operators can't add or remove customers without taking them offline. To do so online requires an entirely different architecture. "Most boxes in the enterprise aren't designed that way," says Promod Haque, a general partner at Norwest Venture Partners. "Those are the kinds of changes that are getting made into equipment," he adds. "It's what I call data-center carrier class."

FUTURES AND OPTIONS That's why these data centers need an overhaul to prepare them for the coming data onslaught. But the needed improvements are expensive and require business to support the capital improvements. And that's not likely, given the falloff in customers, like last year's dot-coms, which were big business for these centers (see "Data centers run for cover," June 1).

Today the inside of a data center is filled with disparate hardware, which is nearly impossible to manage. Investors like Mr. Haque believe the answer is virtualization, a way for multiple customers to share common processing and technology resources. "What you don't want is 20 different boxes for 20 different customers, with each one having its own user interface," he says.

Instead you need hundreds of blades chock-full of processors that can handle hundreds of customers. "By partitioning the processing power, you can scale up and down to match the customers' processing needs," Mr. Haque adds. In other words, if a few hundred processors are lying idle, and Victoria's Secret needs more processing power and bandwidth to meet the demand generated by its annual Webcast, the data-center operator can simply allocate the idle resources to the purveyors of lingerie.

Rackspace Managed Hosting in San Antonio, Texas, and Atlanta-based Interland (Nasdaq: INLD) have already started to make changes to their business models to overcome the limitations of today's data centers. "The build-to-order model popularized by Dell Computer [Nasdaq: DELL] is the only Web-hosting model that works in our industry," says Joel Kocher, who is slated to become CEO of Interland. Interland and Rackspace have each built their own software to achieve that, but other companies will have to depend on products from startups like Ensim, PlateSpin, Sphera, and Xevo. These companies have developed software that allows data-center operators to set up virtual servers and remotely monitor application usage and billing.

Without strong demand and growth, most data centers face a catch-22: they don't have the capital to invest in the much-needed hardware and software necessary to meet the capacity and performance that corporate customers require. Worse yet, most of these data centers are already heavily in debt.

It all boils down to basic economics, and it's not a straightforward proposition for Internet data centers to profit. A typical data center running at full capacity needs to generate revenue of $808 per square foot to break even in its first year, according to Tier 1 Research, a market research firm. That's a far cry from the annual $216 per square foot these centers can obtain today. "Everyone was focused on getting more customers and not lowering costs," says Rosen Sharma, who cofounded Ensim for precisely those reasons. "Today it takes almost 60 days to get up and running in a data center, and as long as customers aren't up, they can't be charged."

But new technologies can help a data-center operator spread the total cost of ownership among hundreds of customers. By doing so, operators can make a more compelling economic argument to large corporations, which have so far been buying millions of dollars of new gear every year just to support their basic operations.

The good news is that analysts expect demand for these services to grow during the next few years. In fact, Wall Street analysts estimate that in 2005, data centers will pull in $28.5 billion in business, a seven-fold increase from $3.5 billion in 2000. Most of that money is likely to be spent by large Global 2,000 companies, which are expected to siphon off portions of their technology infrastructure to data centers.

That's also good news for startups like Abeona Networks, Datasweep, Ensim, Inkra Networks, Mellanox Technologies, and NetScaler, which are vying to use technology to fix data-center problems.

BACK TO THE FUTURE The data-center business isn't new. In the heyday of mainframe computers, companies like IBM (NYSE: IBM) and EDS (NYSE: EDS) promoted the concepts of time-share and data centers heavily and profitably. The rise of client/server computing saw the data center vanish into the safe cocoon of a corporation's four walls. Then it came back, as Netscape's vision of having businesses use Web browsers to get information gained acceptance in the marketplace. Demand boomed for services offered by the likes of Exodus Communications (Nasdaq: EXDS). "Fundamentally, when we first built a data center, we built high-quality real estate and a really fast network," says B.V. Jagadeesh, a cofounder of Exodus who now runs NetScaler, a company focused on content acceleration. And as e-commerce becomes part of doing business and corporations adapt real-time computing architecture, the demand on data centers will only increase. This has the VC community aggressively funding startups focused on upgrading the performance and reliability of data centers.

A key technology being touted for creating it is InfiniBand, short for "infinite bandwidth." InfiniBand is a switched-fabric input/output technology that ties together servers, network devices, and storage devices. Currently, devices are connected to storage systems, processors, blades, and networking equipment using a shared bus known as a peripheral component interconnect (PCI). InfiniBand uses a switched network fabric -- more a network of roads than PCI's single-lane highway -- to transfer data from one part of a computer to another. In doing so, the data center starts to look like a giant computer.

Nothing is more impressive than the money and time IBM is throwing at this concept of virtualization. Using the moniker "cellular system architecture" (CSA), it's a major movement within the company. "The new opportunity is in data delivery," says Mark Dean, an IBM fellow and vice president of the company's systems research. And for IBM, that means CSA -- simply microservers, mostly on small, thumbnail-size chips, bound to each other by optical links. Like human cells, every time one dies or becomes cancerous, another takes over. This is the fail-proof way of building the next-generation servers, and when that happens, chances are that even IBM will outsource its IT infrastructure to a firm like Intira. Let the upgrades begin.

Write to om.malik@redherring.com.

0 Assessing the Net's structural integrity

Coming soon: ethernet as business model

New protocol aims to score always-on goal

Infrastructure upgrades loom for data centers

Explained: How Internet core routers deliver

 – Architecturally impaired

VC Whispers: Optical components shine in dreary market