This material reflects my opinions and not those of my employers.
The data for this essay was produced with research help from Claude
We are engineering something that has never existed. Not one AI agent, thousands. Not thousands with a human in the loop, but thousands with an orchestrator that is yet to be built at scale. The models are getting better. The hardware is getting faster. The frameworks are multiplying. But the question nobody has a satisfying answer to is simpler and harder than all of that: how do you make thousands of relentless agents, each one programmed never to quit, work together efficiently, at speed, without burning everything down?
Orchestrating three agents is a demo. Orchestrating thousands is the unachieved algorithmic Nirvana.
Before you try to answer that, look at the one system that already did.
It weighs about three pounds.
The human brain coordinates roughly 86 billion neurons organized into circuits that span regions, layers, and timescales simultaneously. There is no central orchestrator. No single region that issues commands and waits for acknowledgment. Recent neuroscience research has established something that should give pause to anyone building multi-agent systems: complex cognition, including memory, attention, and decision-making, emerges from distributed patterns of activity, not from any one module in charge.
The brain actively and deliberately forgets associations that no longer serve it, through dedicated neuronal mechanisms in the striatum. It runs all of this on twenty watts. The power draw of a dim light bulb. Now look at what we are building.
The scale of intelligence — power consumption as a lens
Single neuron
~0.00000000002 W
(20 picowatts)
One computational unit. Fires, resets, fires again. No concept of "done."
Human brain
20 W
86 billion neurons. No central conductor. Distributed, self-regulating, actively forgets what it no longer needs.
Stanford campus
(~36,000 people, neural energy only)
720 kW
The combined brain energy of an entire research university — every student, professor, and staff member, thinking simultaneously.
Stanford community
(full metabolic, food energy)
3.84 MW
92,160 kWh/day
Total body energy for 36,000 people. Enough to power ~3,178 average US homes for a day.
Enterprise AI
data center
100 MW
2,400,000 kWh/day
26× the metabolic energy of the entire Stanford community. Equivalent to ~82,800 US homes daily. Still no built-in stopping mechanism.
Hyperscale AI
facility
500 MW
12,000,000 kWh/day
130× Stanford. Powers more homes than exist in the entire city of Denver. One voltage disruption in Northern Virginia in 2024 produced a 1,500 MW surplus from 60 facilities going offline simultaneously.
Global AI fleet
IEA 2030 projection
~108 GW avg
945 TWh / year
The equivalent of Japan's entire annual electricity consumption, dedicated to computation. The orchestration problem, unresolved, running at planetary scale.
Bars use a logarithmic scale spanning ~22 orders of magnitude (20 pW to ~108 GW). Human energy = metabolic (food). AI energy = electrical. Sources: IEA Energy and AI 2025; EIA 2024; Stanford IRDS 2024–25; WHO dietary reference values.
What the numbers are actually saying.
The gap between a human brain and a hyperscale data center is 25 million to one in power consumption. That is not a gap of degree. It is a gap of kind. We are spending orders of magnitude more energy on systems that still do not have a reliable mechanism for knowing when to stop, a problem evolution solved in biological neural circuits hundreds of millions of years ago.
I ran a simple experiment: I asked Claude Code to build an application, then repeatedly prompted it to review and improve what it had just built. Every single iteration found something to improve. It never came back and said "this code is perfect, my work here is done". That is not a flaw, it is the nature of these systems. They are trained to keep optimizing, and they will, indefinitely, unless something external imposes a stopping condition.
Now multiply that by ten thousand agents working concurrently on interdependent tasks. The questions cascade: when does an agent abandon a failing path? When does another take a different approach? How does that decision propagate across agents sharing context? How does the system balance efficiency against accuracy when both are moving targets? How does it balance the equation between thought and cost? Gartner documented a 1,445% surge in multi-agent system inquiries between 2024 and 2025. Almost two-thirds of enterprise leaders name orchestration complexity as their top barrier. Everyone senses this is where the next wall is. Nobody has climbed it yet.
The blueprint already exists.
The most recent neuroscience research is converging on something architecturally significant: the brain does not solve coordination by building a smarter top-level controller. It solves it through layered, self-regulating networks where local circuits handle micro-decisions, global states modulate everything from above, and the entire system actively prunes what is no longer needed.
Distributed. Adaptive. With a natural stopping condition built into every circuit. And it runs on 20 watts, less energy than a single GPU server rack uses every eight seconds. But also, when overwhelmed can completely collapse.
The orchestration challenge in agentic AI is not primarily a software problem. It is a complexity problem that biology already solved, at a fraction of the energy cost, with no engineering team and no product roadmap. The new models are getting us closer to the raw capability needed to attempt orchestration at scale. The question remains who will break through the architectural part of the equation, and whether, when they do, we will even be able to understand how their orchestrator is actually doing it.
Whoever cracks it will not do so by adding more compute. They will do it by building something that looks less like a software architecture and more like the thing evolution spent 540 million years refining.
It already fits in the palm of your (more realistic, Shaquille O’Neal’s) hand. It already runs on 20 watts. That is the bar.
References
Stanford population & staffing
Stanford Facts (official), Autumn Quarter 2025 — students and faculty headcount. https://facts.stanford.edu/
Stanford Common Data Set 2024–2025 — instructional faculty breakdown. https://www.scribd.com/document/846789118/stanford-cds-2024-2025
Stanford Institutional Research & Decision Support (IRDS) — staff headcount methodology and staff-to-student ratio data. https://irds.stanford.edu/data-findings/staff-headcounts
Stanford Daily, March 13, 2024 — “Behind Stanford’s doubled staff-to-student ratio,” reporting the 0.94 staff-per-student ratio figure used to estimate non-faculty staff. https://stanforddaily.com/2024/03/13/behind-stanfords-doubled-staff-to-student-ratio/
LeadIQ (December 2025) — cited in the original version as the source for 27K employees; flagged and discarded during the accuracy review as an unreliable third-party scraper. https://leadiq.com/c/stanford-university/5a1d89162400002400629cba/employee-directory
Human metabolic energy
WHO/FAO dietary energy guidelines — basis for the 2,000–2,500 kcal/day adult range; 2,200 kcal used as the midpoint estimate. Standard reference, not a single URL.
Unit conversion: 1 kcal = 4,184 J; 1 Wh = 3,600 J → 1 kcal = 1.162 Wh. Standard physics constant, no single source required.
AI data center energy
International Energy Agency (IEA), Energy and AI report, 2025 — global data center consumption (~415 TWh in 2024), hyperscaler household equivalent (100,000 homes), and projected doubling to 945 TWh by 2030. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
Pew Research Center, October 2025 — “What we know about energy use at US data centers amid the AI boom,” citing IEA figures including the 100,000-household hyperscaler benchmark and 183 TWh US data center consumption in 2024. https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/
Congressional Research Service (CRS), Report R48646, January 2026 — “Data Centers and Their Energy Consumption: Frequently Asked Questions,” including the 25.3 MW training load figure and 176 TWh US data center figure. https://www.congress.gov/crs-product/R48646
Lawrence Berkeley National Laboratory (LBNL), 2024 report — US data center annual energy use ~176 TWh in 2023, projected growth to 325–580 TWh by 2028; cited through CRS and Belfer Center.
Deloitte, Unlocking exponential value with AI agent orchestration, 2025–2026 predictions — data center electricity consumption projections and agent orchestration context. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/ai-agent-orchestration.html
Belfer Center for Science and International Affairs, Harvard Kennedy School, February 2026 — “AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment,” including the Northern Virginia 1,500 MW disruption event and LBNL projections. https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
World Economic Forum, December 2025 — “How data centres can avoid doubling their energy use by 2030,” including the 1,500 MW Boston power-demand comparison. https://www.weforum.org/stories/2025/12/data-centres-and-energy-demand/
Household energy baseline
US Energy Information Administration (EIA) — average US household consumption ~865 kWh/month = ~29 kWh/day (2024 data). https://www.eia.gov/tools/faqs/faq.php?id=97&t=3
EIA, Electricity use in homes — average annual household consumption ~10,500 kWh. https://www.eia.gov/energyexplained/use-of-energy/electricity-use-in-homes.php
Denver household count (used to verify and correct the city comparison)
Data USA, 2024 — Denver, CO: ~335,000 households. https://datausa.io/profile/geo/denver-co/
US Census Bureau / Census Reporter, ACS 2024 — Denver city population ~719,000, corroborating the household count. http://censusreporter.org/profiles/16000US0820000-denver-co/
Nebraska housing units (used as the corrected 1,000 MW comparison)
US Census Bureau housing unit estimates for Nebraska — ~820,000 housing units; used as the corrected anecdotal comparison for the 1,000 MW tier. Standard Census reference.