15/09/2026
Here is a simple rule of modern technology investment: if you take a stream of numbers that used to live quietly inside a physical object, and you pass those numbers through a very large transformer architecture, you no longer have a physical object company. You have a "Foundation Model Ecosystem."
If you run a continuous glucose monitor company, your historical business model was:
Make a tiny plastic needle.
Charge insurance $150 a month for it.
Hope the Bluetooth connection doesn't drop when the user walks into a microwave beam.
That is fine! It is a honest, hard tech, device-manufacturing business.
But in 2026, a device manufacturing business trades at 4x EBITDA, whereas an "Ecosystem Foundation Model" trades at 40x forward ARR, mostly because people like saying the word "ecosystem."
So the move here—and it’s a brilliant, inevitable move—is to reframe glucose readings as token streams.
A glucose monitor samples interstitial fluid every 5 min. 288 data points a day.
To a medical device engineer, that’s a time-series line graph.
To an AI founder, 288 numbers a day isn't a graph; it's a prompt! It's a context window!
If you feed five years of 288-daily-tokens into a Transformer, your CGM isn't just telling you that you shouldn't have eaten that second croissant at 2:00 PM.
It is predicting that because your glucose spiked at 2:00 PM, your heart rate variability (ECG model) will dip at 11:00 PM, which means your sleep score (wearable model) will collapse, which means you will order a double espresso tomorrow at 8:00 AM, which means... well, you get the idea.
Everything is connected? All human biological output is just one giant, multi-modal token stream waiting for a GPU cluster to explain it to us.
I spent a lot of time thinking about how this applies to physical infrastructure and deeptech startups across global cities—from Bengaluru to Fort Worth.
If you are a founder building deeptech hardware or municipal sensors, you spend half your life being told by software VCs that your margins are too low and your sales cycles are too long.
"Why can't you just be a SaaS platform?" they ask, while sitting in a room cooled by a massive HVAC system that nobody has upgraded since 1998 because no one built an LLM for municipal chilled-water loops yet.
So the playbook gets clearer:
Don't sell a sensor; sell the foundation model of the sensor's physical domain.
If you monitor water pressure in city pipes, you aren't a utility vendor; you are training the Hydrological Foundation Model.
If you run corporate-startup matching across 30 fast-growing city hubs, you aren't doing business development; you are running the Archetype Matching Neural Network for municipal capital.
Is every continuous stream of sensor data actually suited for a 70-billion-parameter foundation model? Probably not
Sometimes a simple linear regression—or even a threshold alert that screams "Hey, blood sugar is low, eat an apple"—is strictly better, faster, and cheaper than running inference on a cluster of H100s.
But linear regressions don't get you a press release, and they certainly don't get you listed as a "Generative Health Partner".
Anyway. Expect every wearable company on earth to announce its own "Large Medical Model" by Q4. The devices will still break when you go swimming, but at least the predictions about why they broke will be remarkably eloquent.
