Every generation gets its own absurd sentence.
Ours may be this: the future of intelligence is wetware.
Not metaphorically wet. Not “biology-inspired.” Not neurons on a slide or proteins in a logo. I mean literal fluid, molecules diffusing through solution, chemistry pushing information around the way voltages do in silicon. A kind of programmable broth. A sentient soup.
That phrase sounds unserious on purpose. Most important technologies sound ridiculous before they become infrastructure. A room-sized calculator sounded ridiculous. A global network of computers sounded ridiculous. Talking to a machine in natural language sounded ridiculous. Then the world reorganized around each of them.
The question is not whether a “sentient soup” sounds silly. The question is whether we are nearing the limits of the dry alternative.
Today’s AI boom is built on an extraordinary substrate: silicon chips, power-hungry datacenters, cooling systems, transmission lines, supply chains, foundries, and staggering capital expenditure. We have gotten used to calling this “software,” but that has always been a convenient lie. Modern intelligence is industrial. It is carved into wafers, pumped through cables, and paid for in megawatts.
That machine has been astonishingly successful. But success has a way of exposing its own bottlenecks. Bigger models require more compute. More compute requires more power. More power requires more capital, land, cooling, permits, geopolitics, and patience. Training the next generation of models increasingly looks less like writing code and more like building railroads.
When a technology becomes constrained by physical infrastructure, the natural question is no longer “how do we optimize it?” The natural question becomes: is there another substrate?
This is where biology enters the room.
Biology has been computing for billions of years. Cells sense gradients, integrate signals, estimate hidden state, store memory, regulate behavior, repair damage, and make context-dependent decisions. They do all of this without GPUs, without CUDA, without floating point units, without datacenters in Oregon. They do it with chemistry.
Of course biology does not do matrix multiplication in any familiar sense. It is not “running transformers” in the way Nvidia does. But that is the wrong level of analysis. The important point is simpler: matter can process information. Molecules can represent state. Reactions can transform it. Concentrations can encode quantities. Feedback can implement control. Dynamics can implement inference. Training, at a deep enough level, is just a physical process that changes a system so that future inputs produce better outputs.
Once you see it that way, the boundary between “computation” and “chemistry” starts to feel suspiciously historical.
For the last half century, we have assumed that serious intelligence belongs to electronics. That assumption may turn out to be contingent, not fundamental. Silicon won because it was engineerable, fast, compact, and manufacturable. Not because the universe signed an exclusivity agreement.
And chemistry has some strange advantages.
It is massively parallel by default. It operates directly in the physical world rather than through sensors and interfaces bolted on afterward. It is native to biology. It can combine sensing, memory, computation, and actuation in the same substrate. It can run where electronics are awkward: inside fluids, inside cells, inside reactors, inside messy environments where “reading the world” and “changing the world” blur together.
If we ever learn to engineer these systems with the same seriousness that we engineered digital computers, the result will not be a niche scientific curiosity. It will be a new branch of computation.
That does not mean tomorrow’s ChatGPT will live in a tube of goo on your desk. The first real breakthrough will probably look much stranger and much smaller. Maybe it will be a molecular system that can adapt to noisy conditions without explicit reprogramming. Maybe a biochemical controller that learns optimal dosing in a local environment. Maybe a diagnostic that does not merely detect molecules, but reasons over them in situ. Maybe a training primitive so cheap in energy that it changes how we think about scaling.
At first, the systems will seem unimpressive compared to digital AI. They will be slower, narrower, harder to debug, and painfully alien to our current software stack. That is normal. Early computers were toys compared to human clerks. Early neural nets were toys compared to expert systems. The first airplanes were toys compared to trains. New substrates do not begin by dominating the old one. They begin by being weirdly good at one thing.
The real opportunity is not “replace GPUs.” That is too blunt, too early, and probably too unimaginative. The real opportunity is that as intelligence becomes more expensive in capital and energy, and as the world demands it in more physical contexts, the premium on alternative substrates goes up dramatically.
If digital intelligence is the story of electrons in rigid lattices, molecular intelligence could be the story of information in motion.
And if that sounds speculative, good. It is speculative. But there are two kinds of speculation. One is fantasy with no contact to reality. The other is taking a hard look at where the current curve breaks, then asking what becomes plausible on the other side.
The current curve has some obvious cracks. Compute is concentrating. Power is becoming strategic. Scaling is no longer just an algorithmic problem, it is a civilizational one. We are heading toward a world where the ability to train advanced systems depends not only on ideas, but on access to energy, land, regulators, fabrication, and grid upgrades. That is an unstable foundation for something as important as intelligence.
The more serious AI becomes, the more we should expect substrate innovation.
And chemistry is sitting there, underexplored, under-engineered, and slightly embarrassing to talk about in polite company.
That is usually a good sign.
The biggest mistake is to frame this as “can soup become conscious?” That is science fiction bait. The interesting question is much more practical: can chemistry become trainable? Can we build physical systems whose molecular state can be shaped through iterative feedback so that they perform useful computation, control, or prediction? Can we move from beautiful one-off experiments to reusable motifs, compilers, simulators, calibration loops, and eventually programmable molecular stacks?
If the answer is yes, even partially, the consequences are large.
Not because we discovered life in a beaker.
Because we discovered that intelligence does not have to remain trapped in chips.
The future may still belong to silicon. It may be that digital hardware keeps outrunning every challenger and chemistry remains what it has mostly been so far: brilliant, frustrating, and hard to scale. That is possible.
But there is another possibility. A quieter one.
That intelligence slowly escapes the datacenter.
That it leaks into materials, fluids, assays, therapeutics, reactors, and adaptive matter.
That training no longer happens only in server racks, but in physical substrates that sense and compute in the same breath.
And decades from now, when this seems obvious, people will laugh at how absurd it once sounded.
A sentient soup.
Ridiculous.
Until it isn’t.