Julien Reszka
Singapore switched on a data center rack running on sixteen million living human neurons this month, and the number every outlet led with was power: about one kilowatt for the whole rack, against up to one hundred kilowatts for a comparable silicon AI server. That comparison is accurate, and it is also the wrong bill to read if you want to know whether the thing is actually cheaper to operate.
A CL1 unit is not a chip in the usual sense. Cortical Labs grows neurons from human stem cells and lays them on top of a silicon chip covered in electrodes, called a microelectrode array. The chip sends small electrical pulses into the culture and records how the neurons fire back, and over repeated cycles the culture adapts its firing patterns to whatever task it is being trained on, the same feedback loop that let an earlier Cortical Labs prototype learn to play Pong in 2022. A separate life support system keeps the culture at body temperature, feeds it nutrients, and exchanges gases, because it is living tissue, not etched silicon.
NUS Medicine, DayOne, and Cortical Labs built the Singapore rack from twenty CL1 units. Cortical Labs prices access to that same hardware through its own Cortical Cloud service: $2,200 per unit per month. Twenty units of capacity at that rate would run $44,000 a month, or $528,000 a year, regardless of whatever arrangement NUS Medicine itself has with Cortical Labs.
The same twenty units draw roughly one kilowatt total. That saves an estimated $190,800 a year in electricity against a comparable hundred kilowatt silicon rack, using Singapore's industrial tariff of about $0.22 per kilowatt hour.
$528,000 is not close to $190,800. Renting that much neuron capacity costs about 2.8 times what it saves on power, and the gap holds before counting what the price list does not itemize:
- Neuron cultures stay viable for about six months, then must be regrown from stem cells under sterile lab conditions
- Regrowing a culture needs a trained cell biologist, not a data center technician, the kind of specialist a stem cell lab pays $130,000 a year
- The rack also needs a continuous gas feed of carbon dioxide, oxygen, and nitrogen, plus fresh nutrients delivered every three days
None of this means CL1 is a bad product. Cortical Labs makes a different comparison in its own pricing pitch: $2,200 a month for CL1 access against about $4,300 a month for a comparable cloud AI chip. That comparison is price against price, and CL1 may well win it.
The headlines this week made a different comparison: CL1's power draw against silicon's power draw. That one favors CL1 too, until you check what CL1 access actually costs to rent, which is not electricity. $190,800 saved and $528,000 spent are both numbers Cortical Labs publishes. The coverage only used one of them.
That $2,200 figure deserves its own scrutiny:
- Cortical Labs is the one calling CL1 and a high end cloud AI chip comparable, and no published benchmark shows they do equivalent work per dollar
- The $4,300 baseline is a premium, on demand chip price, not the reserved or mid tier options that sit well below it
- Cortical Labs is reported to have about 20 paying customers, few enough that the price does not need to cover what a unit actually costs to run
The customer count matters most. With only about 20 paying customers, Cortical Labs does not need $2,200 a month to cover what a unit actually costs to run. And that real cost is biological: a trained cell biologist must regrow each culture every six months, in a sterile lab, with continuous gas exchange. That kind of labor does not get cheaper per unit the way silicon manufacturing does. Nobody has published what a CL1 unit actually costs Cortical Labs to run each month, so there is no way to confirm from public numbers whether $2,200 is a sustainable price or a loss leader.
Cortical Labs has raised $11.6 million across three funding rounds, a modest amount for a company running a wet lab at commercial scale. That does not settle whether $2,200 a month is a real price or a subsidized one, but it does mean any subsidy could not run at large scale for long.
I may be missing something in this comparison, a scale efficiency, a subsidy, a workload difference that changes the math. If a reader can show what I got wrong, I will correct it.
Discussion