AI/Biz

AI Is Not the New Electricity

Andrew Ng has been saying that AI is the new electricity, and it has become the line everyone repeats. I understand why. It’s vivid, it’s optimistic, and it comes from someone whose technical credentials are beyond argument.

I also think it’s wrong in a way that is costing companies real money, and I want to explain why.

What the Analogy Claims

The electricity comparison says: a hundred years ago electrification transformed every industry, not by doing one thing, but by becoming a general-purpose input that everyone could use. Factories, homes, transport, agriculture. You didn’t need to understand generation to plug something in.

The implication is that machine learning will do the same. It’ll become a general capability that any business can draw on, and the companies that adopt it early will pull ahead of those that don’t.

The first half of that might be right eventually. The second half is being used to justify spending, right now, on a premise that doesn’t hold yet.

Where It Breaks Down

Electricity had a standard interface. A socket. Two or three pins, an agreed voltage, and any appliance from any manufacturer works. The interface hid all the complexity of generation and transmission.

There is no socket here. Every application requires collecting data specific to that problem, labelling it specific to that question, training a model specific to that data, and building a pipeline specific to that context. Almost nothing transfers between two problems in the same building.

Electricity was the same everywhere. The current arriving at a factory in Manchester was identical to the current arriving in Birmingham. Your data is not like anyone else’s data. The model that works brilliantly for a competitor may be useless for you because your customers behave differently or your process records different fields.

Electricity didn’t degrade. Once wired, a factory stayed wired. Models decay. The world shifts, behaviour changes, an upstream system alters a field, and performance quietly erodes. Every model you deploy is a permanent maintenance obligation, not an installation.

“Electricity was infrastructure you installed once. This is livestock. It needs feeding, it gets sick, and if you stop paying attention it dies quietly and you find out from a customer.” — Sameer Gupta

Electricity failed loudly. The lights go out and you know. A degraded model keeps producing confident output that is gradually less correct, and nothing alerts you unless you built the monitoring, which most people don’t.

What the Analogy Is Doing to Budgets

Here’s the practical harm, and I’ve now watched it happen several times.

A board hears the electricity framing. It concludes this is infrastructure. Infrastructure gets a capital budget, a programme, a completion date, and a steering committee.

Then reality arrives:

  • The pilot works and doesn’t generalise. Six months on one use case, a good result, and none of it transfers to the second use case. The board expected a platform and got a bespoke solution.
  • There’s no completion date. The model needs retraining, monitoring, and re-evaluation forever. There is no point at which it is installed and you move on.
  • The running cost was never budgeted. Everyone modelled the build. Nobody modelled the two people it takes to keep it alive, indefinitely, per model.
  • The second and third projects don’t get cheaper. Some tooling is reusable. The expensive parts, understanding the problem and assembling the data, are not.

“The board approved a capital project and bought an operating liability. Both can be worth it. They are not the same conversation and they should not have the same shape.” — Sameer Gupta

A Better Analogy

If I had to offer one, I’d say this is closer to hiring specialists than to installing utilities.

A specialist has to be recruited for a particular need. They need context about your business before they’re useful. They need ongoing management. They can be excellent at one thing and useless at the adjacent thing. And you would never say “we’ve hired an accountant, so finance is solved.”

That framing gets the economics right. It’s per-problem, it’s ongoing, it requires management, and the value depends entirely on picking the right problem.

It also gets the strategy right. You don’t hire specialists speculatively because it’s the future. You hire them for a specific, valuable job you can describe.

What I’d Say in the Boardroom

  • Fund problems, not platforms. Name the decision that’s currently made badly. If nobody can name one, the honest answer is that you’re not ready and no amount of infrastructure fixes that.
  • Budget the running cost from day one, and make it visible. A model in production is a permanent line item.
  • Expect the second project to cost nearly as much as the first. If someone’s business case assumes steep economies of scale across use cases, ask them to show their working.
  • Measure in business outcomes. Not accuracy, not models deployed. Money or hours.

Final Thoughts

I’m not arguing this technology is overrated. I’ve spent five years writing about why it’s substantial, and I think the research coming out right now is remarkable.

I’m arguing that the electricity framing tells executives it’s an infrastructure purchase with a completion date, when it’s a set of individually-justified, individually-maintained capabilities that never finish.

Those two beliefs produce completely different budgets, different governance, and different definitions of success. One of them survives contact with reality.

Get the metaphor right and the rest of the planning follows. Get it wrong and you’ll spend two years discovering it the expensive way.