A founder asked me last month which model we build on. GPT-5, Gemini, Claude. He wanted to know which one we had bet the company on, as if the answer would tell him whether we were any good.
You have probably asked a vendor the same thing.
It is the wrong question.
Not a bad question. The wrong one. And everything I know about why came from a year and a half I once spent trying to buy an aeroplane.
What aviation knows and healthcare refuses to
Before Coherent, I spent about eighteen months in the world of electric aircraft. They barely existed then. They barely exist now.
But that world rewires how you see every hard engineering problem. Including this one.
Here is what surprises people about certifying an aircraft. The years of work are not spent on the exciting part.
The propulsion system, the thing everyone wants to talk about, is mostly an energy-density problem. And that problem lives somewhere else entirely. Someone else's lab, someone else's timeline. Batteries get better, or they don't, without you.
The engineering that actually eats the years is duller, and harder. How do you take that component, put it in an airframe, put people inside, and keep the whole thing in the sky?
You trace where every screw came from.
You build so that any failure, anywhere, can be reverse-engineered, because when an aircraft comes down you had better be able to say exactly why.
None of that glamour attaches to the battery. All of it attaches to the casing you build around it.
Now put a patient where those passengers are sitting.
Healthcare AI is the same picture, and we are all staring at the wrong half of it. If Anthropic or OpenAI is the battery, the raw source of power, then the question that decides whether any of this is safe was never which battery. It is: what is the casing? What do you build around that power so it can run a hospital without hurting the person on the table?
The battery is designed to change under you
Aviation and cars teach a second thing healthcare has not absorbed at all.
When Tesla built the early Roadster, they knew something uncomfortable. The battery they would ship in two years would not be the one they were designing around. Energy density was going to move. So you cannot hard-wire the car to a single power source. You design as if the most important component will change underneath you, and you leave enough margin to absorb it when it does.
Get that right, and a better battery is a welcome upgrade.
Get it wrong, and the day the battery changes you are not upgrading. You are rebuilding. In a regulated industry, that means re-certifying from scratch. Anyone who has done it tells you the same thing: you would far rather make a change and re-certify than tear the whole thing down and start over.
Now hold that next to what is actually happening in AI.
The frontier labs ship a materially new model every few months. Each one is a new battery. Different shape, different weight, different behaviour, sometimes a brand-new way to fail.
Build your product rigidly around the last model, and every release is a small crisis.
Build it properly, and every release is just a swap.
A new AI model lands every few months. A better battery takes decades. That gap, the sheer speed at which the power source changes under you, is the new engineering problem. It is one aviation never had to solve.
So the discipline healthcare needs is not "pick the best current model." It is this: build the thing around the model so that when the next one arrives, and it will arrive before you have finished, it slots in without touching the part that keeps patients safe.
Build for the swap, not the spec
Say it plainly and the whole "which model?" debate looks like buying a car by arguing over the chemistry of a battery that will be obsolete before the finance clears.
The model is a component. A remarkable, improving, genuinely astonishing component. I am not knocking it. But a component.
The work that matters is everything you wrap around it:
- A framework that lets you change the battery without touching the safety mechanics.
- A verification layer that checks the output before it reaches a patient.
- Guardrails that do not need rebuilding every time the thing behind them gets smarter, or stranger.
Build that, and you are never locked to one supplier's hardware. The battery improves, you plug it in, the casing holds. The patient never notices the swap.
That is the point.
This is the accelerator, not the brake
Be careful with the lazy reading, which is that I am arguing to slow down. I am arguing the opposite.
A car built to accept a better battery goes further, faster, sooner than one that has to be redesigned every time. Designed-in safety is not the tax on the exciting work. It is the thing that lets you do more of it, for longer, without hurting anyone.
The clinics that get the most out of AI over the next five years will not be the ones chasing whichever model topped a leaderboard last week.
They will be the ones who built a casing good enough that every one of those models, in turn, dropped straight in.
Fast and safe are not in tension. You only get both if you build for the swap instead of the spec.
So why does everyone still ask about the battery?
Because the battery is the visible, exciting, demo-able part. It tops the leaderboard. It stars in the launch video. It is what the founder on my call wanted to bet on.
The casing is invisible. Design reviews, verification layers, unglamorous plumbing. Nobody makes a keynote about the drivetrain.
In most industries, betting on the battery and neglecting the casing gets you a product that breaks, and a refund.
In healthcare, the cost of skipping the casing is not a refund. It is a patient.
Ask the uncomfortable version. When the model gets it wrong at two in the morning, who is standing between it and the patient? If the answer is that the dashboard was glowing green, you did not buy safety. You bought speed with a confident voice.
That is the whole reason I keep writing about this: about moving fast and breaking patients, about healthcare's autopilot moment. This piece is the engineering underneath both.
A brilliant model dropped onto fragile, decades-old healthcare infrastructure does not inherit that infrastructure's safety. It inherits its fragility, and hands it a faster, more confident way to reach a patient.
The dazzle of the battery is exactly what makes it dangerous. It pulls every eye toward the one part of the system that improves on its own, and away from the part that only improves if someone deliberately, unglamorously builds it.
What the casing actually is
If the battery is the model, the casing, for healthcare, is a short and unfashionable list.
- A human kept in the loop. Not as a temporary scaffold to remove once the model is "good enough," but permanently, because healthcare is the management of the unpredictable, and the human is what catches the case the model was never built for.
- A verification layer. It checks the output before anything consequential happens, because a system that can give two different answers to the same question cannot be trusted the way ordinary software could.
- Transparency about failure. How it fails, how often, and on whom. A clearance is not a proof.
- An honest account of the seam, where the new model meets the old infrastructure, because that seam is where most of the real risk lives.
None of that is exciting. All of it is the job.
And it is the part that does not change when the battery does. That stability is exactly what makes it safe to keep swapping in something better.
The answer the founder did not want
I did not tell that founder which model to bet the company on.
Because the honest answer is that you should not bet the company on any of them. You should build so that you never have to.
The battery will keep getting better. That is the one safe prediction in all of this. Most of the value we will see in the next few years has not shipped yet.
Wonderful. Build for it.
But build the car that can safely use whatever the battery becomes, not the one that was perfect for the battery you happened to have on the day you started.
So stop shopping for batteries.
Start building the thing that turns all that power into care, safely, no matter which one you plug in next.
The battery was always going to be the easy part.
The casing is the whole job. In healthcare, it is the only part that was ever going to keep anyone safe.
