Updated 2026-06-24 with the final version of the paper — added GPU-generation comparisons, a physics-based ceiling on silicon scaling, an AGI timeline, and a note on the economics.
I just finished a paper and wanted to write it up here in plain language, because the core idea is simple even though the numbers behind it aren’t.
The question
Everyone talks about the energy cost of training large models as if it’s mostly a software problem — better attention mechanisms, lower precision, smarter sparsity. Those help, but they’re all optimizations within the same hardware paradigm: a GPU, where memory and compute are physically separate chips that have to constantly shuttle data back and forth between them.
I wanted to know: how much of the energy cost is actually that architecture, and not the algorithms running on top of it?
The hypothesis
My hypothesis is that the human brain gives us a clean way to measure this, because it’s the only general-purpose intelligence we know of that solves comparable problems, and it does it on about 20 watts.
So I built an energy-equivalent comparison: take the total metabolic energy a human consumes from birth to age 25 (~17,500 kWh), and ask what a commodity GPU cluster could compute if it burned exactly that same amount of energy. Then compare outputs.
The result: a gap of about 50,000×. The brain delivers roughly 50,000 times more useful computation per unit of energy than the GPU cluster does, over the same energy budget.
I deliberately used a commodity GPU (the RTX 3090) for that number, because it’s the most conservative comparator I could pick. Run the same comparison against the H100 — the actual chip GPT-5 was trained on — and the gap widens to roughly 112,000×. Even against NVIDIA’s newest B200, it’s still about 22,000×. The 50,000× figure is the floor of the claim, not the ceiling.
What’s interesting is that this isn’t just one ratio — it’s two ratios, derived completely independently (one from raw compute output, one from energy efficiency per FLOP), and they converge on the same number. That convergence isn’t a coincidence, it’s mathematically guaranteed once you fix the energy budget — which means the figure isn’t noise, it’s a real signature of something structural.
Where the gap comes from
I decomposed it into three independent, multiplicative sources, all rooted in architecture rather than algorithms:
- The memory wall — moving data between separate memory and compute chips burns up to 90% of a GPU’s energy. In the brain, a synapse is both the memory and the compute event. No bus, no fetch cycle. (~100–1,000× gain)
- Dense vs. event-driven activation — GPUs fire every core every cycle. Real neurons fire rarely (under 1% active at any moment) and cost almost nothing when idle. (~100–1,000× gain)
- Digital vs. analog compute — a digital multiply-accumulate costs orders of magnitude more energy than a biological synaptic event. (~10–100× gain)
Multiply the minimums together and you get a ceiling of ~100,000×, which comfortably brackets the observed 50,000× gap with no leftover unexplained.
The neuromorphic hardware that already exists backs this up directionally: IBM’s NorthPole, Intel’s Loihi 2, and recent spiking/co-located designs have shown gains from 400× up to 5,600× over the last decade, tracing a clear trajectory toward that ceiling.
Why GPU progress alone won’t close it
The obvious objection is: won’t NVIDIA just engineer its way there over the next decade? I checked. GPU energy efficiency has been improving at about 1.34× per year. At that rate, closing the gap to brain-level efficiency takes roughly 28 years — and that’s optimistic, because the rate is already slowing (H100 to B200 delivered less improvement than the historical trend would predict). Worse, there’s a hard physics ceiling: CMOS transistor switching energy can’t drop below the thermal noise floor without the chip becoming unreliable, which caps GPU efficiency at roughly 5,000 TFLOPS/W — still an order of magnitude short of the brain. You cannot get there by making better GPUs. You can only get there by changing the architecture.
When does this become AGI-relevant
If you define “AGI-class compute” as roughly 100× the brain’s total lifetime computation, current frontier training runs are about three orders of magnitude below that line. Training compute has been scaling at roughly 2× per year, which puts the crossing point around 2032–2036 — but only if someone is willing to spend GPU-paradigm energy at state-actor scale to get there (something like 10,000 GWh for a single training run). On the neuromorphic path, the same milestone costs roughly 200 MWh — within reach of an ordinary research lab. The architecture question isn’t just about electricity bills. It’s about who is even capable of building the next tier of AI at all.
Why this matters
If even a fraction of this gap closes, the implications are large: a GPT-5-class training run currently costing tens of millions of dollars in electricity could drop to a few hundred dollars. Inference cost per query could fall enough that frontier models could run on a hearing-aid-sized power budget. The concentration of frontier AI in a handful of companies is, in part, a side effect of this architectural energy tax — not an inherent law of how intelligence has to be built.
There’s also a less obvious economic angle. Roughly $7.6 trillion in AI data-center capital spending is projected for 2026–2031, almost all of it betting on the GPU paradigm continuing. But GPUs have a functional economic life of only 2–3 years before the next generation makes them uncompetitive, while that capital is typically depreciated over 4–6 years on the books. If a credible neuromorphic alternative reaches commercial maturity before that capex cycle finishes paying for itself — and the physics above says it eventually must — a large share of currently-deployed GPU infrastructure could become economically stranded well before it’s fully written down. That’s the setup for a repricing event across GPU manufacturers, the neocloud operators renting out GPU capacity on multi-year contracts, and the still-small neuromorphic chip sector, roughly in that order. I’m not making any specific predictions about individual companies here — just noting that the architecture question above isn’t only a research curiosity, it’s sitting underneath one of the largest capital allocation bets in industrial history.
None of this requires new physics to discover — only to engineer around. It requires building hardware that doesn’t separate memory from compute, doesn’t fire idle units, and doesn’t insist on digital precision everywhere. The brain has been proving this works for 500 million years on 20 watts.
Read the full paper
I go through all the math, the references, and the case against the “it’s just a software problem” counterargument in the full writeup:
Download: The 50,000× Gap (PDF)
Happy to hear pushback on the assumptions, especially around the brain compute estimate, which is the most contested number here.