From Leopold Kohr to IBM to Artificial Intelligence
A “State of State” of Affairs in Computer Industry
October 2, 2026
This morning I asked my AI companion a simple question.
Why are they suddenly building so many enormous data centers across America?
The answer made sense. Artificial intelligence requires enormous computing power—not only to train the models but, increasingly, to operate them. Every question asked, every document analyzed, every image created requires computation somewhere. Multiply that by hundreds of millions of people and businesses, and computing power begins to look less like software and more like an industrial commodity.
Data centers become factories.
But that answer immediately raised another question in my mind.
If artificial intelligence is supposed to approximate some of the functions of a human brain, what happens when that “brain” becomes scattered among hundreds of thousands—or eventually millions—of processors?
They have to communicate.
They have to coordinate.
They have to synchronize.
And the larger the system becomes, the greater the amount of energy and processing capacity that must be devoted not to producing intelligence, but simply to keeping the system working together.
At first I thought the enemy was distance.
Then I realized that wasn’t quite right.
The enemy is SIZE.
And suddenly I was transported back more than thirty years.
The Corporate Empires
In 1992, I urged the CEO of EDS to consider breaking up his company into smaller, more autonomous and efficient units.
Three years later, in 1995, I made essentially the same argument about IBM.
These were not fashionable ideas at the time. Corporate America worshipped size. Bigger revenues, bigger organizations, bigger market shares, bigger headquarters, bigger empires.
And, naturally, bigger executive responsibilities.
The prevailing assumption was that scale produced efficiency.
I argued that beyond a certain point it could produce precisely the opposite.
As organizations grow, an increasing amount of their energy is consumed internally. More management. More meetings. More reporting. More controls. More committees. More people coordinating other people who are coordinating still other people.
Eventually an organization can reach the absurd point where a significant portion of its resources is devoted to managing its own complexity.
The CEOs weren’t interested in dismantling their empires.
Efficiency may have favored decentralization.
Ego favored centralization.
We Have Been Here Before
There was another precedent staring us in the face even then.
Computing itself.
In the early days of information technology, computing power was concentrated in centralized mainframes. They were magnificent machines for their time, and mainframes remain extraordinarily effective for certain kinds of work.
But eventually the industry discovered that not every problem was best solved by making the central machine bigger.
Smaller computers became dramatically cheaper and more powerful. Networks improved. Client/server computing emerged. Distributed computing allowed large problems to be divided among many machines. Later came the internet, server farms and eventually cloud computing.
Instead of asking one enormous computer to do everything, we increasingly learned to let many smaller computers divide the work.
And something remarkable happened.
The whole could become more powerful, more flexible and often more economical without requiring every component to become gigantic.
That history now makes the AI boom particularly fascinating.
Because, in one sense, AI grew from the embryo of distributed computing. Modern artificial intelligence would be impossible without the ability to divide enormous computational workloads among many processors.
Yet we are now using distributed computing to construct systems so enormous and so tightly integrated that we are beginning to recreate the very thing from which computing once escaped:
the giant centralized computer.
Except today’s mainframe isn’t a large IBM box sitting behind glass.
It is the building.
Or perhaps an entire campus.
Leopold Kohr Walks Into the Room
But the underlying idea wasn’t mine.
I had been lecturing about it since the 1980s, drawing partly upon the work of Austrian economist and social philosopher Leopold Kohr.
Kohr published The Breakdown of Nations in 1957.
1957.
Nearly seven decades before ChatGPT.
In my April 1995 essay Drang nach Osten, I quoted his extraordinary observation:
“It is always bigness, and only bigness, which is the problem of existence — social, as well as physical.”
Kohr believed that organisms and institutions could simply outgrow their natural limits.
His remedy was equally straightforward: reduce them to a scale at which they could function effectively again.
At the time, I was applying those ideas primarily to nations, governments and corporations.
Today I find myself wondering whether Kohr may also have something to say about artificial intelligence.
The New AI Factories
The technology industry is now engaged in perhaps the greatest computational construction boom in history.
Gigantic data centers are being planned and built. Vast quantities of electricity must be secured. New power plants, transmission lines and substations are required. Tens of thousands of sophisticated processors must be connected by extraordinarily fast networks.
Every additional processor theoretically adds computing power.
But every processor also adds something else:
complexity.
Another processor must communicate.
Another cluster must be coordinated.
Another rack must be connected.
Another building must be powered and cooled.
Another layer of software must decide where the work should go.
At first:
MORE COMPUTERS = MORE INTELLIGENCE
But eventually:
MORE COMPUTERS = MORE COMMUNICATION + MORE COORDINATION + MORE ENERGY + MORE COMPLEXITY
And somewhere between those two equations may lie a point of diminishing returns.
The marginal computer could eventually contribute less additional intelligence than the resources required to integrate it into the whole.
But How Do 50,000 Computers Agree on Anything?
That led me to another question.
Who makes all these GPUs? And how can tens of thousands of them work together so seamlessly?
I remembered the computer industry of my day. Getting equipment from competing manufacturers to communicate was often anything but seamless. Compatibility could become an industry unto itself.
The answer contains another irony.
The problem hasn’t entirely been solved. It has partly been avoided.
Today’s largest AI clusters generally don’t consist of a random collection of competing processors from NVIDIA, AMD and everyone else, somehow persuaded to cooperate harmoniously.
The highest-performance systems tend to be much more homogeneous.
Thousands of GPUs may come from the same technological family. They communicate through carefully engineered high-speed interconnections. Above them sits an enormous software architecture that distributes calculations, moves data and synchronizes the work.
In other words, rather than persuade an unruly United Nations of incompatible computers to cooperate, engineers increasingly construct something resembling a single technological ecosystem.
That greatly improves coordination.
But it also brings us back to centralization.
The network itself becomes part of the computer.
The data center becomes part of the computer.
Eventually, the entire campus begins behaving like one enormous machine.
Which sounds suspiciously familiar.
The Mainframe Returns
There is almost a circular quality to the history.
CENTRALIZED MAINFRAME
became
DISTRIBUTED COMPUTING
which became
CLIENT/SERVER
which helped create
THE INTERNET
which produced
CLOUD COMPUTING
which has now helped create
THE AI SUPERCLUSTER.
We spent decades escaping the limitations of centralized computing—
only to reinvent the mainframe on an astronomical scale.
The old mainframe occupied a room.
The new one may occupy hundreds of acres.
And once again we confront the ancient trade-off.
Centralization can improve control, speed and coordination.
Decentralization can improve flexibility, resilience and efficiency.
Technology pushes the practical boundary between them back and forth.
But technology does not eliminate the boundary.
Eventually size itself begins imposing costs.
And every engineering breakthrough that allows us to connect another thousand or ten thousand processors solves the problem of bigness by adding another layer of machinery whose job is to manage the bigness.
There is nothing inherently wrong with that. The additional computing capability may be worth vastly more than the additional complexity.
But eventually somebody ought to ask:
At what point does the machinery required to make the machine bigger become the principal burden of the machine itself?
Nature’s Embarrassing Comparison
Nature provides an uncomfortable benchmark.
The human brain weighs roughly three pounds and operates on about the power consumption of a dim light bulb.
It doesn’t occupy 500 acres.
It doesn’t require its own power station.
It doesn’t need thousands of cooling towers.
And my brain in Belgrade does not have to synchronize itself with somebody else’s brain in Arizona before I can finish this sentence.
Several billion years of biological experimentation produced something extraordinarily compact, distributed and energy-efficient.
Perhaps nature discovered something Silicon Valley is now rediscovering:
Intelligence is not synonymous with size.
Indeed, the ultimate achievement in artificial intelligence may not be constructing the largest possible machine.
It may be discovering how to accomplish more with less.
And perhaps the next great advance in AI will therefore not come from adding another 50,000 GPUs.
It may come from discovering that we don’t need them.
1957 → 1992 → 1995 → 2026
Which brings me back to Leopold Kohr.
In 1957, he was writing about nations and societies.
In 1992, I applied the principle to EDS.
In 1995, I applied it to IBM—and quoted Kohr in Drang nach Osten.
And now, in 2026, sitting at Blackjack Ranch and watching humanity spend staggering sums constructing ever larger artificial brains, I find myself asking essentially the same question I was asking corporate executives more than thirty years ago.
How big is too big?
There is a delicious irony here.
Artificial intelligence represents one of humanity’s most sophisticated attempts to create intelligence.
Yet the industry developing it may be rediscovering one of the oldest lessons of both nature and human organization:
Growth creates power.
Growth creates complexity.
And eventually complexity begins consuming the power that growth created.
Perhaps that is the central paradox of the AI revolution.
We are building bigger and bigger machines in order to make them smarter.
But if increasing size eventually requires an increasing proportion of their intelligence merely to manage themselves—
Leave a comment