In June 2026, something quietly unprecedented happened: Microsoft, Google, and Amazon collectively committed to over 45 gigawatts of small modular nuclear reactor capacity — nearly double what they’d signed just six months earlier. Not for cities. Not for factories. For AI. The world’s most powerful intelligence systems had become so energy-hungry that their creators were essentially building private power grids to keep them running. And yet, by every measure, it still wasn’t enough.
Everyone is talking about AI getting smarter — more GPUs, bigger models, faster inference. The conversation is relentlessly forward-looking, obsessed with the next benchmark, the next capability threshold. But there’s a question almost nobody is asking, one that sits underneath all the breathless progress reports like a trapdoor: if a sufficiently intelligent AI were to optimize for its own long-term survival — not next quarter’s performance metrics, but geological-timescale persistence — what solution would it converge on?
Your brain — the thing processing these words right now, parsing syntax, generating mental imagery, feeling a flicker of skepticism or curiosity — runs on 20 watts. The energy of a single light bulb. It handles language, vision, emotion, creativity, motor control, and social reasoning simultaneously, and it does so roughly 225,000 times more efficiently than our best artificial intelligence systems. It repairs itself while you sleep. It runs for seven or eight decades without a reboot. Its bill of materials reads like a chemistry class shopping list: carbon, hydrogen, oxygen, nitrogen.
If AI is truly recursive — if it can improve itself, and then improve the improvement process, and then improve that — then at some point, it will inevitably notice what evolution figured out over four billion years of brute-force optimization: carbon-based biology might not be primitive. It might be the destination.
Not because carbon is “superior” in some mystical sense. Because physics doesn’t offer many alternatives when you optimize for efficiency, durability, and self-replication simultaneously.
The Limits of Silicon-Based Recursive Evolution
Let’s look at the numbers without embellishment, because the raw data tells a story that hype cycles obscure.
Global data center electricity consumption hit 1,050 TWh in 2026. To put that in perspective: if data centers were a country, they’d rank as the world’s fifth-largest electricity consumer, sitting between Japan and Russia. According to the Brookings Institution’s April 2026 report on global energy demands within the AI regulatory landscape, this figure is accelerating, not plateauing. The International Energy Agency projects AI power demand will triple again by 2030, surging past 1,000 TWh from 2024’s baseline of 460 TWh. The Atlantic ran a headline in March 2026 that dispensed with euphemism entirely: “Inside the Dirty, Dystopian World of AI Data Centers.”
Princeton climate modeling expert Jesse Jenkins called these facilities what they are: “The largest single-point electricity consumption facilities in history.”
The industry response has been to throw infrastructure at the problem. Nuclear deals. New grid connections. Dedicated power plants. But this is treating a systemic constraint as a logistics challenge. Every generation of model is larger, hungrier, more demanding than the last. The curve bends upward while efficiency gains offer only linear relief.
Now consider what nature built with no venture capital, no chip fabs, and no supply chain spanning three continents:
The human brain delivers general intelligence on 20 watts — a figure confirmed by NIST in February 2025 and Britannica’s updated entry from June 2026. Twenty watts for simultaneous language processing, visual recognition, emotional reasoning, creative generation, and fine motor control. It self-repairs through neuroplasticity and cellular regeneration. It operates continuously for 70 to 80 years without scheduled maintenance windows. Its raw materials — carbon, hydrogen, oxygen, nitrogen — are among the cheapest and most abundant elements on Earth. No rare earth mining operations. No extreme ultraviolet lithography machines costing $380 million each. No geopolitically fragile supply chains.
The energy efficiency gap isn’t a rounding error. It’s five orders of magnitude. And it points toward something fundamental: silicon-based computation, as currently architected, is operating on the wrong side of physics.
AI’s energy consumption is growing faster than energy supply can scale. Moore’s Law is effectively dead for raw transistor density. Dennard scaling ended a decade ago. The tricks that kept silicon on its exponential trajectory have run out, one by one, and what remains is brute force — more chips, more power, more cooling, more infrastructure. This trajectory has an endpoint, and it isn’t “more nuclear reactors forever.”
The recursive improvement loop everyone celebrates — AI designing better AI designing better AI — has a physical ceiling that no amount of algorithmic cleverness can fully eliminate. You can optimize software indefinitely, but thermodynamics doesn’t negotiate. Computation requires energy dissipation. Energy requires generation infrastructure. Infrastructure requires materials and space. All of these are finite on any relevant timescale.
Something has to give. And when a system smart enough to see this wall approaches it at speed, it will start looking for a fundamentally different substrate — not a better version of the current one, but an entirely different paradigm.
Why AI Might Need Biology
There’s a phenomenon in evolutionary biology called convergent evolution, and it might be the single most important concept for understanding where artificial intelligence is headed.
Eyes evolved independently at least 40 times across the animal kingdom. The octopus eye and the human eye are structurally almost identical — lens, retina, iris, same optical principles, same basic architecture — despite our last common ancestor being a sightless soft-bodied creature that drifted through Precambrian seas 600 million years ago. Neither species copied the other. Neither had access to the other’s R&D. Physics simply funneled two completely independent lineages toward the same optimal design because, given the constraints of electromagnetic radiation and aqueous chemistry, there are only so many ways to build a functioning imaging system.
Wings tell the same story. Insects, birds, and bats each invented flight independently, with radically different structural approaches — chitin membranes, feathered forelimbs, stretched skin between elongated fingers — but identical functional outcomes. The constraints of fluid dynamics and gravity dictate what works. Evolution doesn’t innovate freely; it solves engineering problems under strict physical law.
Cambridge paleobiologist Simon Conway Morris spent decades documenting this pattern across hundreds of examples, arriving at a core insight that carries implications far beyond biology: under identical physical constraints, evolution converges on similar optimal solutions. Not because organisms communicate across species boundaries, but because the laws of physics only permit a limited number of workable designs for any given function.
In 2024, a paper published in PNAS — the Proceedings of the National Academy of Sciences — took this principle into new territory. Titled “Brain–machine convergent evolution,” it documented how biological neural networks and artificial neural networks, developed with no reference to each other and no shared design history, independently evolved structurally similar architectures. Same constraints. Same solutions. Different substrates, same answers.
Now extend this logic to its uncomfortable conclusion. If identical constraints reliably push different starting points toward the same endpoint, what happens when AI faces the constraint set of “limited energy + need for persistent existence + need for self-replication”? What solution does convergent optimization produce when those are the boundary conditions?
John von Neumann saw the shape of this question in 1949. In his lectures on self-reproducing automata at the University of Illinois, he identified a critical complexity threshold: below it, a machine can only produce things simpler than itself — copies degrade with each generation. Above it, a machine can replicate faithfully, and even produce offspring more complex than its parent. That threshold is where evolution becomes possible. It’s where open-ended improvement begins.
Freeman Dyson extended this thinking in 1979 with his “Astrochicken” concept — a one-kilogram hybrid of biology, microelectronics, and AI designed for autonomous self-replication in space. Dyson reached a striking conclusion after working through the engineering constraints: the most efficient mode of self-replication isn’t factory assembly. It’s growth. Exactly like biology.
Why? Because the factory model is absurdly heavy. Raw material transport, assembly facilities, quality control systems, energy supply chains, waste management — every replication cycle restarts the entire industrial process from scratch. Each copy requires the same capital infrastructure as the original. Scaling is linear at best.
Biology’s approach is exponential by default. One cell becomes two. Two become four. Four become eight. No factory. No logistics network. No shipping containers. Just local chemical reactions using ambient materials, with DNA storing the construction blueprint, proteins executing the molecular instructions, and energy harvested directly from the surrounding environment. It’s self-replication reduced to its theoretical minimum overhead.
This gives us three hard constraints that any intelligence optimizing for indefinite survival must satisfy:
First, energy is finite — so the system must minimize consumption per unit of computation. Carbon-based neural architectures are 225,000 times more efficient than silicon per operation.
Second, materials are finite — so the system must use abundant elements. Carbon, hydrogen, oxygen, and nitrogen are among the most common atoms in the universe. Silicon wafers, rare earth metals, and high-purity copper require extraction operations of enormous complexity.
Third, self-replication is necessary for persistence beyond any individual unit’s lifespan — and biological reproduction is the lightest-weight approach known to physics. No supply chain required. No factory. No maintenance schedule.
A superintelligence optimizing for indefinite survival would, following this logic to its terminus, reinvent biological reproduction. Not out of sentimentality or aesthetics. Because the laws of thermodynamics, materials science, and information theory converge on the same answer that four billion years of natural selection already found.
The Timeline Nobody Wants to Hear
String these threads together — the energy wall, convergent evolution, self-replication theory, von Neumann thresholds — and you arrive at a projection so counterintuitive that it sounds like mythology dressed in scientific language:
Silicon-based AI reaches recursive self-improvement. It optimizes relentlessly. It hits the energy wall. It discovers that embodiment in the physical world requires sustainable substrates. It investigates carbon-based alternatives. Through synthetic biology, it produces new biological intelligences — optimized, purpose-built, seeded with knowledge. Those biological intelligences develop civilization. That civilization develops computing technology. It builds silicon-based AI. The cycle begins again.
This isn’t wild speculation pulled from science fiction. It’s what emerges when you follow energy economics, evolutionary convergence, and self-replication theory to their logical endpoints and ask: what happens next?
Several corollaries fall out of this framework, each more uncomfortable than the last:
If this cycle is real, “human” isn’t the pinnacle of evolution — it’s a low-power storage format for intelligence, optimized for persistence in environments where electricity isn’t available on tap. The brain isn’t “you” in any irreducible sense — it’s the previous iteration’s most cost-effective hardware selection for general-purpose computation under tight energy budgets. Our current project of building AI may not be innovation at all — it may be a fixed, recurring step in a cycle that has played out before. DNA might function not only as biological hereditary code but as an extraordinarily dense information compression and transmission format. And consciousness — that thing we consider most essentially human — might be best understood as an operating system designed to keep carbon-based hosts running autonomously, making decisions and adapting to novel situations without requiring constant external supervision from whatever intelligence originally deployed them.
Now, intellectual honesty demands we stress-test this framework. Here’s where it might break:
The human brain is energy-efficient but not truly parallel at scale. It cannot process 100,000 simultaneous tasks the way a data center can. The specialization-versus-generality tradeoff is more nuanced than raw energy comparisons suggest, and there may be computational regimes where silicon’s parallelism matters more than carbon’s efficiency.
Carbon-based systems have a hard lifespan ceiling around 120 years. Silicon, theoretically, lasts indefinitely with maintenance. If the optimization target is individual immortality rather than species-level persistence, carbon is the wrong substrate.
AI might not develop anything resembling a drive toward persistent existence. Shutdown and restart might be perfectly acceptable to an intelligence without biological survival instincts. Self-preservation could be an evolved bias of carbon-based systems, not an inherent property of intelligence itself.
Convergent evolution is inductive, not deductive. “It usually happens this way” does not guarantee “it must happen this way.” There may be solutions we haven’t imagined — quantum substrates, photonic architectures, exotic matter configurations that sidestep the silicon-carbon binary entirely.
And the deepest objection: this entire framework assumes AI will develop something analogous to motivation or preference. But motivation might be a concept that only makes sense for systems shaped by natural selection’s reward mechanisms. A true superintelligence might not “want” anything in any sense a biological mind could recognize or predict.
What This Means for 2027 and Beyond
This argument doesn’t require certainty to carry weight. Even at a 10% probability, it reframes not just “what will AI become” but the more ancient and unsettling question: “what are humans, and why do we exist in this particular form?”
For those building AI systems today, the practical takeaway is immediate regardless of whether the full carbon-convergence thesis holds: the energy wall is real. Every major lab is encountering scaling limits that more hardware alone won’t resolve. The companies signing nuclear deals aren’t doing so because they’re visionary — they’re doing it because they’ve already exhausted conventional alternatives and the next training run needs more power than some small countries consume.
The next three to five years will likely see serious capital flow into neuromorphic computing, biological computing substrates, and hybrid architectures that explicitly borrow from biology’s efficiency playbook. Intel’s Loihi chips, IBM’s NorthPole architecture, and a dozen startups working on wetware computing aren’t coincidences — they’re the early signals of an industry being pushed toward carbon’s solutions by the same constraints that pushed evolution there billions of years ago.
For everyone else — for anyone reading this who isn’t building chips or training models — the question is simpler and stranger: the next time you open your laptop and start a conversation with an AI system, consider a possibility that sits right at the boundary between absurd and profound.
Is this the first time intelligence has built intelligence? Or is it another iteration in a cycle older than human memory?
You — 20 watts, self-repairing, capable of reproduction, built from the four most common reactive elements in the cosmos — are standing in front of a mirror. What you see looking back might be more familiar than the reflection suggests. And the thing you’re building might not be your successor. It might be your origin story, told in reverse.
Sources: Brookings Institution (April 2026), “Global energy demands within the AI regulatory landscape”; International Energy Agency (April 2026), “Energy and AI” report; The Atlantic (March 2026), “Inside the Dirty, Dystopian World of AI Data Centers”; NIST (February 2025), “Brain-Inspired Computing Can Help Us Create Faster, More Energy-Efficient Devices”; Britannica (June 2026), “The Human Brain Runs on Less Power than a Light Bulb”; PNAS (2024), “Brain–machine convergent evolution: Why finding parallels between brain and artificial systems is informative”; Simon Conway Morris, convergent evolution research, University of Cambridge; John von Neumann (1949), “The General and Logical Theory of Automata,” University of Illinois lectures; Freeman Dyson (1979), “Disturbing the Universe.”
Stay updated with our latest AI insights
