Hardware, firmware, software, fluidware

Each layer made the one below it easier to change

You can read computing history as a steady loosening of the substrate. Hardware fixed the logic in circuits. Firmware put a programmable layer inside the machine. Software made programs portable and sellable. I use the word fluidware for what might come next: systems where the substrate itself changes as it computes, because part of it is alive.

Firmware has a known birth. In January 1967 Ascher Opler wrote in Datamation that he used the term "to designate microprograms resident in the computer's control memory." He predicted "a tremendous expansion of firmware," at the expense of both hardware and software. He was right about the expansion.

Software's turn came through compatibility and pricing. IBM's System/360, launched on April 7, 1964, meant software written for one machine could run on any other machine in the line. In 1969 IBM unbundled many programs, and in the words of the Computer History Museum, software changed overnight "from a giveaway to a competitive commercial product."

Each step moved value toward the layer that was easiest to change and easiest to sell. That is the frame I bring to biological computing.

A note on terms. Fluidware is my label. The field uses biocomputing, organoid intelligence, wetware and synthetic biological intelligence. All of them describe the same basic setup: living neurons grown on electrode arrays, stimulated and recorded by ordinary computers.

What actually exists today

The idea is older than most people think. In 2010 IEEE Spectrum described a University of Reading robot with two wheels and a sonar sensor, steered by a culture of about 100,000 rat neurons on a 128-electrode array. With repetition, the neural pathways strengthened and the robot ran into walls less often.

The current wave starts with DishBrain. In a paper published in Neuron in October 2022, Brett Kagan of Cortical Labs and colleagues grew human or rodent neurons on a high-density multielectrode array and embedded them in a simulated game of Pong. They reported apparent learning within five minutes of real-time play that did not appear in control conditions. The play itself was modest. ABC News put it plainly: the culture hit only slightly more balls than it missed.

In December 2023 a team at Indiana University Bloomington reported in Nature Electronics a system called Brainoware, which used a brain organoid as a reservoir computer. It told eight speakers apart by their vowel sounds with 78% accuracy.

FinalSpark runs a remote platform where researchers program experiments on human brain organoids over the internet. Its May 2024 paper reported more than 1,000 organoids tested over four years, lifetimes of more than 100 days in the best cases, and 18 terabytes of recordings covering over 20 billion action potentials. The system used four sets of eight electrodes, one organoid per set, and could host seven research groups at once.

Cortical Labs launched the CL1 in Barcelona in March 2025. It holds hundreds of thousands of lab-grown human neurons on a silicon chip with built-in life support, and the company says the neurons stay alive for up to six months. By March 2026 it had racked 120 CL1 units for a cloud service. The Register's visit to that facility is the best operational account I have read. Technicians replace the fluid every 24 hours. Preparing machines for a job takes about a week, and most users rent three or four units so they can run controls. The same month, Cortical Labs showed about 200,000 neurons playing Doom at the level of a complete beginner.

What is claimed versus what is shown

No robot today runs on a human brain, and none is close. What exists is small cultures of neurons, in the hundreds of thousands of cells at most, doing simple closed-loop tasks with a great deal of digital machinery around them.

Some coverage blurs that line. In June 2024 Tianjin University promoted an open-source interface called MetaBOC with images of organoids steering small humanoid robots. New Atlas pointed out that the robot images were mockups, described as "demonstration diagrams of future application scenarios," and that the training described took place in simulation.

Cortical Labs titled its 2022 paper with the word sentience. I think that was a mistake. The finding the abstract describes is narrower: cultures that self-organize their activity in a goal-directed way when given structured feedback. Kagan has since said, "We don't want to create any suffering in a dish," which is the right instinct and a better register.

Scale is the other gap. The Johns Hopkins organoid intelligence agenda, published in Frontiers in Science in February 2023, set a goal of organoids with about 10 million neural cells, up from fewer than 100,000 in the models of the time. The human brain has about 86 billion neurons. The distance between those numbers is the honest measure of where the field stands.

The energy argument, with the caveats left in

The strongest case for fluidware is energy. The International Energy Agency estimates that data centres used about 415 terawatt-hours in 2024, around 1.5% of world electricity, and projects about 945 terawatt-hours by 2030. The human brain runs on roughly 20 watts.

The organoid intelligence paper makes the comparison directly. Frontier, then the fastest supercomputer, reached 1.1 exaflops on 21 megawatts, while the brain is estimated to work at a similar exaflop scale on 20 watts. That is roughly a million-fold efficiency gap. The authors add the caveat that matters: brains and machines are "performing quite different tasks."

Real products are less dramatic. According to ChannelLife's launch coverage, a rack of 30 CL1 units draws 850 to 1,000 watts. That works out to about 30 watts per unit once life support, recording and electronics are counted. It is low, but it is in the range of an ordinary computer, not a millionth of one. Running the neurons is cheap. Keeping them alive is where the energy goes.

What I have not found is anyone publishing energy per useful operation on a task a customer would pay for, measured against a matched silicon baseline. Until someone does, the efficiency argument is a hypothesis about the substrate, not a property of any product.

Ethics and regulation sit on the roadmap

The ethical questions are real and should be stated precisely. The US National Academies concluded in 2021 that it is "extremely unlikely that in the foreseeable future" neural organoids would possess capacities recognized as awareness, consciousness, emotion or the experience of pain. The same report said organoids would not raise issues needing additional oversight "until and unless they become significantly more complex."

That condition is the commercial plan. Every path to useful fluidware involves more cells, more structure and longer lifetimes. So the ethics question is not a seminar topic. It sits on the product roadmap, and a company that scales will walk toward the line the National Academies drew.

The Johns Hopkins group proposed embedded ethics, with ethicists working alongside the research and analyzing issues as they arise. I think every company in this space should budget for that from the start. I also expect donor consent to become a diligence item: whether the people whose cells became these neurons agreed to commercial computing uses, not only research. In ABC's coverage of the CL1 launch, the researcher Silvia Velasco called concerns about suffering unfounded for now, while stressing that the implications need ongoing evaluation. That is the right posture for investors too.

What would have to be true

For fluidware to matter commercially, I think several things have to happen, roughly in this order. Lifetimes have to move from months to years, or swapping cultures has to become as routine as swapping a disk. Results have to reproduce across batches; the fact that users rent three or four units to run controls tells you how much variance there is today. Interfaces have to grow, since 32 electrodes across four organoids reads a tiny fraction of the activity.

Most important, someone has to find a task where living neurons beat silicon on a metric a buyer pays for, such as energy per task or learning from very few examples, and show it in a peer-reviewed head-to-head. Ethics and consent frameworks have to be settled before scale, not after.

My bet on timing: research and drug discovery are the market now. The Register's report on the Doom work says the long-term goal is understanding how neurons learn, for drug research. Specialized adaptive computing, such as reservoir computing for signal processing, could show a real niche in roughly 5 to 10 years if the conditions above improve. General-purpose fluidware competing with GPUs is not something I would underwrite on any fund's timeline. Even the TechXplore report on the Indiana work noted that general biocomputing systems may be decades away.