Testing Too Cheap to Metre

People keep asking which AI model we use. Fair enough. The model is the fun part. You spin one up on a laptop and feel like a scientist by dinner. Almost nobody asks the boring question sitting underneath it: where does the data to train the thing come from? In biology that's the only question that matters, and the honest answer is that there mostly isn't any.

Think about how the language models got good. We handed them the internet, everything anyone had ever typed, near enough, for free, already in a form a computer could read. Biology got nothing like that. Biology has a filing cabinet. Most of what's in it measured one thing, in one person, on one day, and the slowest step in the whole job was a person carrying a tube of blood down a corridor. That's not a dataset. That's a pile of anecdotes with a centrifuge attached.

And it matters for a plain reason. A body isn't a parts list. It's a network: genes, proteins, hormones, the immune system, the gut, all of it wired together and talking at once. Measure one biomarker, one number off one blood test, and you've put a probe on a single wire in a machine that has a million. You learn about the wire. About the machine, nothing.

This is the exact job machine learning is built for: find the pattern in a pile of numbers too big for any human head. But it only works if you give it the numbers. Hand it 50 markers from the same drop of blood at the same moment and it'll tell you that 7, 19 and 44 move together in a way that shows up years before anyone gets sick. Same drop, same moment. That's the whole trick. Measure those same 50 things one at a time, in 50 people, on 50 mornings, and the pattern is gone before you start. Our testing system is built, almost perfectly, to throw away the one thing we're trying to see.

So the shopping list is short. Three things. Breadth: hundreds of markers off one sample, not a handful, because the signal is in how they move together. Time: the same body, over and over, for months and years. A single snapshot tells you where a person is standing and nothing about where they're walking, and a disease is mostly a body drifting off its own normal. You can't see drift in a photo. You need the film.

The third one nobody has data on, and it's the one I chew on most. Inputs. A body isn't sealed off from the world. It's taking things in all day that shove it around. What you ate. The drug you're on. The supplement you swear by. The water, the air, whatever's floating round your work. Those are the forces on the system, and if you want to know why a marker moved, you have to know what went in. We measure almost none of it. We might catch that your inflammation crept up. We won't catch that it crept up the same month you changed jobs, because nobody could afford to watch both sides closely enough to join the dots.

Same thing sits under all three. It isn't the science. We know how to measure this stuff. It's that doing it properly, densely, over and over, both sides, for enough people to count, costs a fortune at today's prices. So nobody does it, and the data never gets made.

Which is good news, oddly, because it turns a sprawling biology problem into an engineering one with a single dial: the cost of a measurement. Turn that down far enough and breadth, time and inputs all come good at once. The data problem quietly becomes a price problem. And when you pull a lab test apart to see where the money actually is, nearly all of it is human. That narrows things right down. The fix isn't a cleverer test. It's a lab that runs itself. So let me start with what a test costs today, and where that money really ends up.

Every figure in this piece is in Australian dollars. The prices below are what i-screen, an Australian test-direct-to-you service, charges right now.

The first 8 are i-screen's live prices. They don't sell whole-genome sequencing or a full proteome, so those two are current consumer and research send-out prices instead. Still Australian dollars.

Sit on those numbers a second, because until you get why a blood count costs $39, you'll fix the wrong thing.

Take the blood count. The reagents, the chemicals that do the actual reacting and give you the number, cost about 60 cents. The test sells for $39. So the thing you came for is under 2% of the bill. The other 98% is people and paperwork wrapped round 60 cents of chemistry. 60 cents of biology, $39 of everything else. That isn't a lab test. It's a courier run with a pipette on the end of it. And here's the bit that took me too long to see: most of that wrapper costs the same no matter which test you run. Below is roughly where the money goes to get one old-fashioned result back to a doctor, in Australian dollars. No lab publishes this, so take it as a fair build from the collection fees, the reimbursement rates and the cost structures people do report. Not gospel.

Look at the blood-draw line, the admin line, the line where a human signs the result off. They barely move between a 60-cent blood count and a $600 proteome, because a person drew the blood either way, a person booked it in either way, and a billing department fought an insurer over it either way. On a cheap test that wrapper is the whole bill. The people who cost this for a living say it plainly: a lab test costs anything from under $1 to over $100, and where you land is just how much of the wrapper you decide to count. Count the reagent alone and it's pennies. Read it off a hospital price list and it's three figures. Everything in between is labour, freight, paperwork and billing, and none of it is biology.

That used to depress me. Now it gets me up in the morning, because a cost that's almost all wrapper is a cost you can go after. The reagent has a floor. It's a real molecule with a mass and a supplier, and I won't pretend otherwise. But the 98% around it is just human steps, and a human step is exactly the kind of thing an engineer takes out. So take it out, one bucket at a time. Post the sample in from home and the blood draw goes to zero. Run the lab with nobody in it, round the clock, and the operator goes to zero. Sell it at a flat, published price instead of feeding every test through the insurance machine and the billing department goes to zero. None of those moves is clever. Stack them up and $39 becomes about $1. What's left when the wrapper is gone is the molecule and the power bill.

And this isn't a whiteboard dream any more, which is the exciting part. Closed-loop labs, where the machine runs a test, reads the answer and picks the next one itself, no human in the loop, already run experiments 10 to 100 times faster than a person at a bench. One open-source group got their hardware from about $50,000 down to $5,000, just by 3D-printing the parts and sharing the pricey instruments instead of buying a whole robot per job. A team making proteins in a tube, no living cells anywhere, took 40% off their cost. Real invoices, all pointing the same way.

There's one thing you have to get right or the whole sum falls over, and it's utilisation. A lab robot still costs about $50,000. Its cost per test is just that price divided by how many tests it runs before it dies. One shift a day and it's dear. Every hour of every day and the same box gets 3 or 4 times cheaper per test, for nothing, because you're spreading one fixed cost over far more work. An idle robot is the dearest robot in the building. Taking the human out was never about the romance of an empty room. It's that the human is the reason the machine ever stops.

So here's that same list of tests, run our way, machines going round the clock, sample collected at home, result sold at a flat price. The wrapper coming off, line by line. Still Australian dollars.

None of it is magic. That $39 blood count costs us about 75 cents, not because we found cheaper chemistry, the reagent was always 60 cents, but because we scrapped nearly everything piled on top of it. The routine tests all drop the same way, tens of dollars down to about $1, 30 to 60 times cheaper, and every cent of that comes from binning overhead, not from touching the measurement.

The big molecular tests are a different animal, and I'll be straight about it. Whole-genome sequencing still costs us about $195 a run, a proteome about $13, because there the reagents and the machine time are a real, stubborn chunk of the cost. You can't wish a sequencing flow-cell, the consumable chip the machine chews through every run, any cheaper. But each run hands back thousands of measurements at once, sometimes millions, so the cost per data point falls off a cliff. That $13 proteome reads up to 7,000 proteins, near enough every protein in the sample, which is 0.4 cents each. For the price of one comprehensive panel today, you could run a full proteome 20 times over on our gear. So the honest headline is two numbers, not one. Everyday tests, tens of cents. The big runs, a few dollars each and a fraction of a cent a marker. Same lab, same night.

Now let me argue with myself, because I've sat through enough demos to distrust anyone who won't.

The hard part was never the hardware. The hardware's cheap and it works. The hard part is the software, and there's a lot more of it than the pitch decks admit, driving the instruments, keeping an honest simulation of what the machines are doing, running the workflow, the decision engine, and a layer of human oversight to catch a bad step before it torches a week of reagents. Most labs waving the autonomous flag have built one or two of those pieces and are quietly doing the rest by hand. The marketing is miles ahead of the reality. Grade lab autonomy 1 to 5, where 5 is genuinely no human ever, and nearly everything shipping today is a 2 or a 3: a decent machine doing one narrow job on its own. Useful. I'm not sneering. But calling that a self-driving lab is like calling cruise control a self-driving car.

The regulators have looked at this too, and they're not with the maximalists. Early in 2026 the FDA and the EMA, the American and European drug regulators, put out rules that a human has to sign off every quality-critical call in a regulated lab. You can't hand a quality call to a model that can't show its working, however good the numbers look, so in a clinic the fully empty lab just isn't a thing you can sell. And there's a quieter point the empty-room crowd skips: the scientist is often the thing keeping the data honest. A good one knows when a result is off, when the machine has latched onto an answer that looks right and isn't, when the simulation has wandered away from what's physically happening. That judgement isn't friction to automate away. Half the time it's the only thing between you and a very efficient machine for making wrong answers at scale, and a wrong answer is worse than none, because the model learns it and repeats it a million times.

So here's where I've ended up, and I'm surer of it than when I started. The autonomous lab is the key. It's the one thing that takes the human wrapper off, and the wrapper is the cost. What's changed is which autonomous lab I mean. The pitch-deck one, the empty room with nobody in it, is a badge, not a goal. Chase it and you'll burn 3 years on the last 5% of autonomy the regulator won't let you use anyway, fighting to sack the one person catching your mistakes. The one that works automates everything that's only labour: the drawing, the pipetting, the running, the reading, the billing, and keeps a human in the one chair where judgement earns its keep, watching the machine and signing what matters. That's still an autonomous lab. It still runs round the clock on a sample you posted in. It still takes the wrapper off. It still drops a test to tens of cents. It just doesn't pretend the last human is the enemy. Build that and the cost falls. Chase the empty room and you never ship.

That's the number I'll put my name to, because a vision with no number is just a mood. Tens of cents for an everyday test. A fraction of a cent for a marker inside the big runs. Our own menu already gets the cheapest tests near 75 cents with the machines running all night and the sample turning up in the post, and buying reagents at population scale takes the everyday ones under 50 cents. Not a fantasy. Arithmetic. The wrapper comes off and the molecule is what's left. The old nuclear line was power too cheap to meter, and it never showed up. A test too cheap to metre is a smaller promise, and a more honest one, and this time the sums actually close.

And what drops out the far end is the part I actually care about. At $250 a test you test once and treat it like an event. At tens of cents you test every week, and the snapshot turns into a film. That's the difference between a photo of a stranger and actually knowing someone. You stop studying 400 volunteers and start measuring millions, which is the difference between a study and a census. A rich panel stops being a toy for the worried and the well-off, and someone in a country town gets the same readout as someone in a city hospital. The data stops being a privilege and turns into plumbing.

Once that plumbing's in, a body measured densely, over time, with what's going in logged right next to what's coming out, you can build the thing this was all for. A digital twin of a person's biology: a working model you can run forward in time. Feed it enough of someone's history and it learns their normal, the balance their body holds when things are fine. Then it watches for drift. Most disease doesn't come out of nowhere. It's a body easing off balance slowly enough that nobody notices until the symptoms turn up, which is usually too late to fix gently. A twin that sees the slide coming changes the whole job. Instead of naming a disease once it's here, you catch the imbalance while it's small and nudge it back, change an input, adjust a dose, move before the body's committed. That's where the inputs earn their keep, and why the menu measures them too: it already covers exposure, drugs, toxins and nutrition, so the twin gets fed what's going in, not just what's coming out. Medicine stops waiting for the damage and starts doing something closer to control: hold the body near its balance point, and steer it back when it wanders.

So that's the bet, and it's why I stopped writing about this and started building. One lab, pointed at one number: the cost of a measurement. Drive a test toward tens of cents and a marker toward zero. People collect their own samples and post them in. The machines run day and night. A human keeps the one chair where judgement still earns its place. Get the cost right and the data shows up. Get the data and you build the twin, the model that catches a body drifting off balance and helps you set it right before it breaks. The cheap blood test was never the point. The twin on the other side of it is.

That's what we're building at Wallace. We call it Robolab. We're building it now.