Computers can generally be characterized into two types: ones that are designed to have more than one user attached and those intended for a single user. In the beginning almost all computing was done on large multi-user machines, partly due to their expense, which precluded their use to all but large institutions or wealthy corporations. Mainframes ruled this era and excelled at their role: providing a reliable computing platform for hosting databases, transaction servers, and centralized applications. The interaction was through character-based screens that, while providing fast and efficient green screen access, was to be their Achilles heel.

At the other end of the scale are personal computers. PCs have two major benefits over mainframes: a lower cost per unit and the ability to host operating systems with graphical user interfaces. GUI applications make use of event-driven user interfaces that can respond to fine-grained mouse, keyboard, timer and paint requests. This provides the framework on which everything from shoot-em up 3D games, WYSIWYG word processors, business presentations with embedded video, and a plethora of powerful desktop programs reside.

Most of the problems over the 20 years in IT have occurred because one of the two ends of the computing spectrum has tried to venture into the other's domain. PCs tried to become multi-user servers and big iron boxes attempted presentation logic.

The computing section of my local science museum has an exhibit showing black and white photographs from the 1970s with reel-to-reel tape drives, floor-standing disk platters, and wardrobe CPU units filling an operations room. Next to this is a display case with an Altair 8800, the sign teaching us how the smaller machine replaced the room-filling mainframe by matching its computing power at a cheaper cost. The analogy drawn is to that of the dinosaurs, where the large and inefficient behemoths couldn't cope with extreme climate change and died out while smaller and nimbler mammals arose to rule the world in their place.

The rise of PCs is a huge phenomenon where, for most of the 1980s and 1990s, there were more new PCs sold per year than the entire installed base. The prediction by George Moore in 1965 was that the transistor density of semiconductor chips would double every 18 months. This largely held true for the next 40 years, benefiting PCs that continued to double in speed while halving in cost. For those who were in the game of downsizing from mainframes, this enabled them to create server farms by simply daisy-chaining PCs together.

The big iron server guys have always wanted to challenge the rise in PCs, fueled by resentment at the insults of "dumb screen" and "legacy system" that were being thrown at them. The Intern et gave them an opportunity to do this, by enabling them to reinvent themselves as hosts for Web application servers dishing up HTML to clients in place of 3270 or 5250 datastreams.

What has occurred is that PCs have scaled up to become servers and servers have become controllers of presentation through HTML. Both are poor compromises and I think have hurt usability, resilience, and general IT efficiency. The trend in many social systems can be characterized as a pendulum that swings between two extremes, politics being a prime example - once policies become attempted, they fall short of promise and expectations, allowing the previous failed polemic to regain popular traction.

For the fast, nimble PCs that now fill rooms, their fate, ironically, is to be replaced with smaller and faster modern mainframes that outperform them in terms of speed, price, and simplicity. Because of advanced workload management techniques, mainframes can be driven harder, often running at 70-90% utilization, while Wintel boxes typically only manage 5%. While Moore's law held fast for the past 40 years, the runway has run out, as physical laws governing thermal flux prevent any further significant miniaturization. A modern Pentium consumes 100 watts of power and generates more heat per square inch than exists inside a nuclear power station's reactor core. The scalability of a virtualized mainframe is huge, with benchmarks showing that up to 20,000 copies of Linux all running Web servers can co-exist happily inside a single box.

For the PC, what we're seeing now is a growth in applications that exploit its capabilities as a first-class client desktop, rather than a rendering engine for dumb HTML. For Google, the bastion of all things Web, two of their most impressive applications are Google Desktop, which indexes all of a PC's files and provides a set of integrated functions such as chat, to-do lists, phone clients, and for mapping Google Earth offers the ability to walk the earth in 3D making use of the PC's native graphics functionality through DirectX.

What this should spell is a new era in which the two poles of computing go back to basics and rediscover what they're best at is doing what they were designed for. Big multi-user servers will continue to grow in terms of their capacity to become application hosting giants, while PCs will enjoy a period of rich applications that fully exploit their graphics capabilities and provide a high-usability end point. Over the past 20 years the server and the client have fought wars where each has tried to replace the other. What we need for the next 20 is for each to excel at what they're best at, and for users to benefit from faster, easier, and richer software.