Testing Too Cheap to Metre

I'm completely sick of being asked what AI model we use. It's the wrong thing to focus on. Setting up a neural network is a fun project for a weekend. The actual nightmare is finding the training data. In biology, we just don't have it.

Language models became smart because they were fed the web. Every dumb thing humans ever bothered typing up was handed over. We don't have that kind of luxury in biology. We have dusty filing cabinets. Most of the medical history in there is just a single measurement taken from a random patient on a Tuesday morning. The biggest bottleneck in the whole system was literally a bloke carrying a test tube down a hall. Calling that a dataset is ridiculous. It's just anecdotes.

This is a massive problem because a human body is not a parts list. The whole thing operates as a loud, messy network. Genes, proteins, hormones, the immune system, the gut. All of them wired up and screaming at the exact same time. If you look at one single biomarker off a standard blood test, you have essentially clipped a probe to one tiny wire inside a machine with a million of them. You might find out the voltage on that specific wire. You learn absolute squat about the machine.

This is why we built machine learning. It finds patterns in data too massive for a human to process. Give the system 50 markers taken from the same drop of blood at the same instant. It will quickly spot that 7, 19 and 44 move together, years before an illness strikes. But you must measure them together. Doing it one at a time across 50 patients over 50 mornings completely ruins the pattern. The testing system basically throws away the information we need.

Breadth requires hundreds of markers measured from a single sample rather than just a handful. The real signal hides in how these elements shift as a pack. Time means tracking the identical body repeatedly for months and years. One snapshot captures where a person stands but tells you nothing about the direction they are walking. A disease essentially represents a body drifting off its own baseline normal, and you can never see drift in a photograph. You need the whole film.

The third item is the one nobody has the data on. Inputs. I probably think about this one most. A body is not sealed away from the outside world. It takes things in constantly, and those things shove it around. What you ate, the drug you swallow, the supplement you swear works, the water, the air, and whatever floats around your work site. Those act as forces on the system. You simply cannot figure out why a marker shifted unless you know what went into the machine. We measure almost none of that. We might spot that your inflammation crept higher, but we will not catch that the spike started the exact month you changed jobs, because no one could afford to watch both sides of the equation closely enough to connect the dots.

The identical problem sits under all three. The science itself is fine, and we know exactly how to measure these things. The blocker is that running the tests properly, making them dense, repeating them endlessly, watching both sides, and doing it for enough people to actually matter costs an absolute fortune at current prices. No one does it. The data never gets created.

Oddly, that is good news. It converts a massive messy biology problem into a straightforward engineering problem controlled by one dial, which is the cost of a measurement. If you crank that dial down far enough, breadth, time and inputs all sort themselves out simultaneously. The data problem quietly reveals itself as a price problem. Pull a lab test to pieces to see where the cash goes, and you find that nearly all of it is human. That narrows the field down nicely. The fix is not a smarter chemical test. The fix is a lab that runs itself.

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.

Those first eight rows are actual i-screen list prices. They don't offer full proteomes or whole-genome sequencing yet. Those last two items are what you'd currently pay for consumer or research send-outs, still in AUD.

Look at those numbers for a minute. If you don't grasp why a simple blood count costs you $39, you are going to end up fixing the wrong problem.

Think about that blood count. The actual chemical reagents that react and spit out your result cost roughly 60 cents. You pay $39. The biology you actually care about makes up less than 2% of the bill you just paid. The other 98% is purely paperwork and human handling wrapped completely round 60 cents of chemistry. It is not really a lab test. What you bought is a courier run with a pipette glued to the end. I took far too long to realise this next bit. Most of that wrapper costs exactly the same regardless of what test you order. The breakdown below shows roughly where the cash goes to push one old-fashioned result to a doctor in AUD. Labs refuse to publish this stuff. Treat it as a reasonable estimate built up from collection fees, reimbursement rates and known cost structures, not gospel truth.

Let me tell you about blood tests. It turns out the science is the cheapest part. The real money goes to the person who draws the blood, the clerk who logs you in, and the billing team who fights your insurance company. And those costs are the same whether you need a simple 60-cent check or a fancy $600 workup. A human had to do the work either way. So if you get a cheap test done, you are basically just paying for the admin. People who crunch these numbers for a living say a lab test can cost anywhere from under a dollar to over a hundred. It just depends on how much of the admin you include. The actual chemicals cost pennies. But if you look at a hospital bill, you are paying hundreds. All that extra money is just paying for freight, paperwork, and labour.

I used to hate that. Now I love it, because you can actually fix a problem like that. You can't change the cost of the chemicals. They are real things you have to buy from a supplier. But the rest of the cost is just human effort. Engineers are great at removing human effort. You just do it step by step. If people post samples from home, you don't need to pay someone to draw the blood. If you run the lab with machines all night, you don't need to pay operators. If you set a flat price, you don't need a billing department. If you do all of that, a test that cost $39 can drop down to a single dollar. You just get rid of the admin and pay for the chemicals and the power.

We aren't drawing these numbers on a whiteboard anymore. The machines are doing it. Closed-loop setups where a robot runs a test and then reads the answer to decide the next step itself, zero humans involved, are out there right now running experiments 10 to 100 times faster than a person at a bench. There was an open-source group that knocked their hardware spend from about $50,000 down to $5,000. How? They just 3D-printed the cheap bits and shared the pricey instruments so they didn't have to buy a whole robot per job. Another lot working on proteins in a tube, no living cells, took 40% straight off their cost. I'm talking real invoices.

Then you hit the one maths problem you have to get right. Utilisation. A lab robot will still set you back about $50,000. Your cost per test is that $50k divided by whatever number of tests the thing crunches before it breaks. Turn it on for one shift a day and the tests are seriously dear. Keep the same box running every hour of every day and the cost per test gets 3 or 4 times cheaper. You didn't do anything clever, you just spread your fixed capital over more work. The dearest robot in the entire building is the one sitting idle. Taking humans out of the equation was never about the romance of an empty room. It's because the human is the only reason the machine ever stops.

Take that exact same list of tests and look at what happens when you run them our way. We have machines going round the clock while you collect your sample at home. Then we sell the result at a flat price. You are watching the wrapper come off line by line. Those prices are still in Australian dollars by the way.

I want to be clear that none of this is magic. We pay about 75 cents for that $39 blood count. I didn't somehow discover cheaper chemistry since the reagent was always 60 cents anyway. We just scrapped almost everything they piled on top of it. The routine tests all drop in exactly the same way. Tens of dollars melt down to about $1. We make it 30 to 60 times cheaper simply by binning the overhead. The measurement itself is completely untouched.

Let me be totally straight with you regarding the big molecular tests. They are an entirely different beast. A run of whole-genome sequencing sets us back about $195, and a proteome lands at roughly $13. The physical reagents and the raw machine time form a seriously stubborn block of cost. The machine chews through a consumable chip every run called a sequencing flow-cell. Wishing it was cheaper changes absolutely nothing. But the cost per data point just falls off a cliff when each run gives you thousands of measurements. Sometimes you even get millions. That $13 proteome reads up to 7,000 proteins. It captures near enough every protein in the sample for just 0.4 cents each. Our gear lets you run a full proteome 20 times over for the price of one comprehensive panel today. The honest headline really requires two numbers. You have everyday tests for tens of cents. The big runs are a few dollars each. That puts it at a fraction of a cent a marker. Same lab on the same night.

When I grade lab autonomy 1 to 5, the 5 genuinely means no human ever. Look at the reality of the hardware shipping today. Nearly everything out there is a 2 or a 3. You buy a decent piece of machinery that handles one narrow job on its own, which is incredibly useful and I am definitely not sneering at it. But calling that a self-driving lab is exactly like calling cruise control a self-driving car.

Regulators have looked at this too. They are absolutely not siding with the maximalists. We saw it happen early in 2026 when the FDA and the EMA, the American and European drug regulators, published rules stating a human must sign off on every quality-critical call inside a regulated lab. You simply cannot hand a critical quality decision to a model that refuses to show its working, no matter how perfect the numbers look. Over in a clinical setting, the fully empty lab is basically a fantasy you cannot sell to anyone.

There is that old line about nuclear energy getting too cheap to meter, which was obviously a joke in the end. A test too cheap to metre is a much smaller promise to make. It actually adds up this time because the basic sums hold. You want an everyday test to sit in the tens of cents, while a single marker from a massive run belongs down below a cent. Get your reagent purchasing up to the insane volumes required to cover a whole population and those routine tests just slide right under 50 cents on their own.

The data out the far end is my only real focus. Pay $250 for a test and you do it exactly once. It turns into a massive event. But when a test costs tens of cents, you run it weekly. That isolated snapshot transforms into a film. It is precisely the difference between glancing at a stranger's photograph and truly coming to know a person.

Studying a neat group of 400 volunteers stops being the point entirely. You want to measure millions. It is literally the gap between running a small study and running a census. A rich panel of health data quits being some shiny toy for the worried and the wealthy, meaning a person out in a tiny country town finally pulls the exact same clinical readout as someone sitting in a massive central hospital. The numbers stop functioning as a privilege and turn into basic plumbing. Once you get that plumbing sorted, measuring a body densely over years with the inputs stacked right next to the outputs, you can finally build the thing this entire effort was actually for. You construct a digital twin of a person's biology. It acts as a functional working model that you can run forward in time. Hand the system enough of a person's medical history so it learns their baseline normal and understands the specific physical balance their body holds when everything works fine. After that, it just sits back and watches for drift.

Very few diseases arrive out of the blue. What actually happens is entirely boring. A human body eases off its balance slowly, and you do not notice the slide until the physical symptoms force you to, which basically means it is already too late to fix the problem gently. A twin that sees the drift coming changes the entire job. You stop waiting around to name a disease once the structural damage is done. Instead, you catch the tiny imbalance and nudge it back into place. You might change an input or adjust a dose. You make a move before the body has fully committed to falling apart.

Inputs earn their keep right there. We put them on the menu because they cover exposure and drugs and toxins and nutrition. The twin consumes the actual inputs going into the system, rather than just passively reading what falls out the other side. Medicine finally stops waiting around for the damage and gets much closer to proper control.

That particular conviction is why I abandoned writing to build things. I wanted a lab obsessed with nothing but the price of a measurement. You push tests toward tens of cents and push markers toward zero. Patients pull their own samples at home and post them to us, letting our hardware churn away day and night. We leave human beings in the one chair where absolute judgement has to earn its keep. If we fix the cost, the data flows out. Once you get the data, you construct the twin. That model catches your body drifting out of balance and helps to correct the course before you break down. A cheap blood test was never the ultimate goal. The digital twin on the far side of it is.

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