Two Quiet Disasters: The Lab That Became a Business, and the Managers Nobody Hired

openai ai-strategy leadership workforce-planning

Most disasters arrive with a great deal of noise. Volcanoes, for instance, are admirably upfront about their intentions. The more interesting disasters are the quiet ones: the kind that can be seen coming for years, plotted on a chart, discussed at length in a meeting, and then allowed to happen anyway because the person who would have to fix them will have moved on to a different job by the time anyone notices.

I can see two of these from where I'm sitting. They are not the same problem, but they are cousins, and they share a family trait: both involve trading something slow and valuable for something fast and countable, and both will be very obvious in hindsight. The first is happening inside one company. The second is happening inside nearly all of them.

 

Problem One: The Lab That Became a Business

OpenAI was founded in 2015 as a non-profit research laboratory with a mission statement that, read today, has the slightly wistful quality of a wedding photograph. Its stated aim was to advance digital intelligence in the way most likely to benefit humanity as a whole, "unconstrained by a need to generate financial return." It published papers. It released code. It said, quite explicitly, that it would share its research openly. The name was not, at the time, ironic.

Then the bills arrived. Frontier AI is the most expensive science experiment since the Apollo programme, except that Apollo knew roughly where the Moon was. In 2019 OpenAI grew a "capped-profit" arm so it could raise money. In 2023, the GPT-4 technical report declined to say how big the model was, how it was built, or what it was trained on, citing "the competitive landscape" alongside safety. In 2024 a steady procession of senior researchers walked out of the building, including the co-founder who had led its research and the head of its long-term safety work, the latter observing on his way out that "safety culture and processes have taken a backseat to shiny products." In 2025 the company completed its restructuring into a public benefit corporation, with the original non-profit now holding a stake in the thing it was supposed to be supervising, like a referee who has been given shares in one of the teams.

And now the company is widely reported to be preparing for a stock market listing, at a valuation that has wandered past the GDP of most of Europe and is still looking for somewhere to sit down. The timing has reportedly slipped to 2027, but the direction of travel has not.

None of these steps, taken individually, is outrageous. Each one comes with a perfectly sensible explanation, usually involving the words "compute" and "billions." That is precisely the point. Culture does not change in announcements. It changes in a long series of reasonable decisions, each of which moves the centre of gravity a few centimetres, until one day you look up and discover that the building is now pointing in a different direction and nobody remembers voting for it.

 

What an IPO Actually Does to a Research Culture

I have sat on enough boards to know what happens in the eighteen months before a listing, and it is not a conspiracy. It is gravity. Bankers arrive. Lawyers arrive. Someone produces a slide titled "Equity Story," and from that moment every project in the organisation is quietly sorted into two piles: things that make the equity story better, and things that make it more complicated.

Curiosity-driven research is, almost by definition, in the second pile. It is slow. It fails a lot. Its best results are frequently things nobody asked for, and its most valuable output, a paper that lets a thousand other people build on your work, is from a shareholder's perspective a free gift to your competitors. Publishing your insights is wonderful for the sum of human knowledge and terrible for your moat. A pre-IPO company is a machine for building moats.

So the tell-tale signs are not dramatic. They look like this:

  • Fewer papers, more product launches. Research results that once appeared on arXiv now appear in a keynote, with a pricing page.
  • Metrics migrate. The questions in the corridor change from "is this true?" and "is this safe?" to "what does this do to monthly active users?"
  • Revenue lines appear in unexpected places. Advertising, for example, inside a product whose chief executive once described ads combined with AI as uniquely unsettling.
  • The people who joined for the mission leave, and are replaced by people who joined for the options. Both groups are entirely rational. Only one of them argues in meetings.

 

To be fair, which I try to be on alternate Tuesdays, OpenAI still does genuinely important work, released open-weight models in 2025, and employs a great many brilliant people who care deeply about getting this right. Its chief executive has reportedly cited safety as one reason not to rush a listing this year, which is either a sincere act of restraint or extremely good investor relations, and quite possibly both. And it is hardly alone: the whole frontier is under the same financial pressure, Anthropic included. But OpenAI is the one that started with the loudest promise to be different, and so it is the one where the drift is most instructive.

Why does it matter? Because a frontier lab is not a soft drinks company. If the most capable AI organisation in the world optimises for quarterly growth, then the questions that do not have a revenue line (interpretability, alignment, honest reporting of what the models can and cannot do) get funded in the way that the office plants get watered: when someone remembers. As I argued in The Frontier Is Getting Expensive, the economics of the frontier are getting harder, not easier. Hard economics plus public shareholders is a recipe that has only ever produced one dish, and it is called "cost discipline."

 

Problem Two: Where Do Managers Come From?

The second problem is far less glamorous, involves no trillion-dollar valuations, and will cost considerably more.

Across the industry, companies are quietly closing the front door. Graduate schemes are being trimmed, junior roles are being left unfilled, and the work that a 23-year-old used to do (first-draft research, basic analysis, reconciliations, test scripts, boilerplate code, meeting notes) is being handed to an AI that does not need a desk, a pension or a mentor. The evidence is still arriving, and some of it is contested, but the shape is consistent. Stanford's Digital Economy Lab found that employment among 22 to 25 year olds in the most AI-exposed occupations has fallen in absolute terms since late 2022, while older workers in the very same jobs carried on as before. Harvard researchers looking at hundreds of thousands of firms found that adopters of generative AI cut junior hiring, not senior headcount. In the UK, around one business leader in five tells the Open University they have reduced entry-level recruitment.

(There is a counterpoint worth noting: a fair proportion of hiring managers admit that "AI" sounds better in a board pack than "we ran out of money." Some of the decline is ordinary belt-tightening in a fashionable coat. This does not help the 23-year-olds very much, who are unemployed regardless of the explanation.)

On a spreadsheet, this is lovely. The cost line goes down, the output line stays flat, and the CFO is so pleased they consider smiling. The problem is that the spreadsheet only has a single year on it.

 

The Pipeline Has a Lag, and the Lag Is About Ten Years

Here is the thing about senior managers: nobody has ever found one in the wild. They are not hatched. They are grown, slowly, from junior staff who spent years doing the tedious work, making small mistakes cheaply, watching how good managers handled bad days, learning where the bodies are buried and which of the bodies are still technically on the payroll.

That grunt work was never really about the output. The output was the excuse. The real product was the person. The junior analyst reconciling invoices was learning how the business actually makes money. The junior developer writing tedious unit tests was learning why systems fail. The graduate taking minutes was learning, by osmosis, how decisions are really made, which is to say, mostly before the meeting starts. Automate the task and you keep the output, but you silently delete the apprenticeship that came free with it.

Now run the clock forward. A typical path from new graduate to someone you would trust with a team of fifteen and a budget is somewhere between eight and twelve years. Companies that stopped hiring juniors in 2024 and 2025 will therefore discover, in roughly 2033 to 2036, that their internal bench of promotable middle managers is mysteriously empty. Not thin. Empty. There will be a gap in the org chart shaped exactly like the cohort they did not hire, and it will move up through the organisation like a pig through a python, emerging at senior leadership level some time in the 2040s.

When that happens, every company that made the same decision will simultaneously try to hire experienced mid-career managers from outside. They will all discover, at the same moment, that the other companies didn't grow any either. Prices will do what prices do when demand is universal and supply was cancelled a decade earlier. And the managers they do manage to hire will be strangers to the culture, the systems and the customers, which is roughly the opposite of what internal promotion is supposed to deliver.

We have seen this film before, incidentally. Several countries cut medical training places in the 1990s to save money and spent the following twenty years importing doctors at great expense. I wrote about the human-capital version of this in The Buildings Come Back. The Doctors Don't. Expertise has a long supply chain, and you cannot expedite it by phoning the warehouse.

 

This Is Predictable, Which Means It Is Avoidable

The irritating thing about the manager shortage of the mid-2030s is that it is the most forecastable crisis in corporate history. You can calculate it with a pencil. It requires no economic model, just the ability to add ten to the current year. Which means there is still time to do something, and the something is not complicated:

  • Keep hiring juniors, but change the job. Hire fewer, perhaps, but stop hiring zero. A junior with good AI tools can operate at a level that would have taken three years to reach in 2015. That is an argument for training them faster, not for not training them at all.
  • Make the apprenticeship explicit. If the AI now does the grunt work, the learning that used to come hidden inside the grunt work has to be designed deliberately: rotations, shadowing, owning a small thing end to end, and being allowed to get it wrong while the stakes are low.
  • Put the pipeline on the risk register. If "key person dependency" belongs on the board's risk register (it does), then "no internal succession candidates in ten years" belongs there too, with a named owner and a number attached.
  • Measure the cost honestly. The saving from not hiring a graduate cohort is visible this year. The cost of hiring externally at a premium in 2035, and of the strategic errors made by leaders who never learned the business, is invisible until it isn't. Put both on the same page.
  • Treat AI as a tutor, not a replacement. The same tools that can do a junior's work can also explain it, critique it and accelerate it. Pair every junior with the machine and a human mentor, and you may well grow better managers than the old model ever did.

 

The Common Thread

Both of these problems are the same error wearing different outfits. In each case an organisation is consuming a long-term asset (open research in one case, its own future leadership in the other) to improve a short-term number. In each case the trade looks excellent for exactly as long as the people making it remain in post. And in each case the bill is not cancelled, merely forwarded to someone who has not yet been hired, or, in the second case, someone who will now never be hired at all.

Douglas Adams once observed that the problem with trying to make something completely foolproof is underestimating the ingenuity of complete fools. I would add a corporate corollary: the problem with optimising everything for this quarter is that eventually, inevitably, it is a different quarter.

The good news is that neither outcome is inevitable. A research lab can choose to keep publishing. A company can choose to keep hiring the people who will one day run it. Both choices are slightly more expensive today. Both are a great deal cheaper than the alternative. And both, unlike the volcano, give us plenty of warning.