In December 1999 Lucent was worth 285 billion dollars. I was running my first company at the time and watching that boom from close range.
I did not call the top. Almost nobody did. By May 2001 Lucent was worth about 32 billion, a merger with Alcatel valued at 23.5 billion had just collapsed over who would be in charge, and when the deal finally happened in April 2006 it was announced as a merger of equals, with Lucent going in at roughly nine billion against Alcatel’s thirteen.
Two hundred and eighty five billion to nine. And every strand of fibre those billions paid to bury is still in the ground, still carrying traffic, still doing the job it was laid down to do. The investors lost. The infrastructure stayed.
John Hempton has just made the same comparison about AI compute, which he calls fibre optics Mark II. His sum goes like this. Around seven trillion dollars of computing is set to be built by 2030. Chips last about five years and cost a great deal to run, so earning that back takes roughly two trillion dollars a year of AI revenue. That can only come from businesses replacing labour, and since nobody restructures a company to save ten percent, the savings have to be about twice what is paid. Call it four trillion a year, against a rich world wage bill of around forty trillion. One dollar in every ten paid to workers across thirty eight countries.
Take that much income out and you get a recession. In a recession nobody pays two trillion a year for software. The only path to earning the compute back runs through destroying the thing that would pay for it.
To give the two trillion a shape, it is close to a third of all IT spending on earth, and roughly fifteen times the combined run rates of the two largest AI labs.
I have an unusual seat for this argument. Hempton is an investor looking at capital spending, and most of the people repeating him are commentators. I sell workflow software into business process outsourcing firms, which puts me on the revenue side of the two trillion. If that money is ever going to exist, some of it has to come out of deals I sign.
So here is what he has right, and what he has wrong.
The arithmetic is right. The people arguing that cheaper chips make it go away have not done the sum. Price the hardware at cost, strip out Nvidia’s entire margin, and the same calculation still needs AI to save five and a half percent of all pay in the rich world. Cheaper compute halves the problem. It does not remove it.
What he has wrong is the door.
Hempton’s model needs a wave of firings, because firing is how you book a saving fast enough to justify the spend. Someone put that sum to the US data this week and the wave is not there. Layoffs ran at 1.1 percent of jobs a month in the year to September, below the 1.2 percent of 2019. There are five million more jobs than when ChatGPT launched.
What changed is the other door. Hiring fell to 3.3 percent of the workforce a month, the slowest since 2013, against 3.9 in 2019. The number of people out of work for six months or more rose by 58 percent. And the saving is visible where you would expect it if this were real. Each hour of work now produces 10.6 percent more than it did in 2022, while pay per hour after inflation rose 6.1 percent. Labour’s share of what American business produces is the lowest in 319 quarters of records, which is to say the lowest since 1947.
That is what I see. A BPO that buys this does not fire anybody. They stop replacing leavers. These are businesses with high natural attrition, so they do not need to make a decision at all. They simply do not reopen the position, and the saving arrives six months later without a single difficult conversation.
Which is why I think the paradox may not close the way Hempton expects. His recession needs income taken out of pockets that currently hold it. A hiring freeze does not do that. It takes income from people who would have been hired, who are mostly young, mostly recent graduates, and entirely invisible in the numbers that make headlines. There is no wave, no announcement, no photograph. The cost is real and it lands on people nobody counts.
The honest thing about this data is that it does not prove AI caused any of it. Hiring began slowing in mid 2022, after the first rate rise and before ChatGPT existed. Technology firms over hired through the pandemic and some of this is that unwinding. The category that holds call centres has been shrinking since 2016, and labour’s share has been drifting down since 2000. The pattern is the one the AI sum requires. It is also the pattern a rate cycle produces. Anyone telling you which it is, including me, is guessing.
Here is what I actually think will happen:
The compute bill probably cannot be paid. Some very large numbers are going to become very small ones, the way 285 billion became nine, and a lot of people who are certain right now are going to be wrong in public. None of that touches the other thing. The work itself is already changing, from the bottom up, process by process, in companies where nobody has announced anything. That change does not need the bubble to survive, any more than the internet needed Lucent to survive.
The fibre stayed in the ground. This will stay too. We are among the people putting it there.
