01
Centralized scaling is encountering infrastructure constraints.
Power availability is becoming a binding constraint, while water use, pollution, transmission costs, and local opposition move beyond the data center fence line.
Centralized AI is consuming more power, water, capital, and public patience. A different path starts with a different question: not “what word comes next?” but “what operation is worth trying next?”
Power availability is becoming a binding constraint, while water use, pollution, transmission costs, and local opposition move beyond the data center fence line.
Consumer hardware and enterprise fleets contain idle capacity. Folding@home demonstrated that coordinated volunteer machines can compute at extraordinary distributed scale.
Richard Sutton and Yann LeCun argue that intelligence needs goals, experience, world models, and consequences, not just larger text prediction systems.
Gartner’s estimate, up from 447 TWh in 2025.
GARTNER ↗Rystad Energy’s estimate of direct cooling water. Its risked central case reaches nearly 644B liters by 2030 without adaptive measures.
RYSTAD ENERGY ↗From $28.92 per MW-day in 2024/25 to $333.44 per MW-day in 2027/28.
PJM ↗A Brookings Papers on Economic Activity paper projects average investment equal to 3.63% of U.S. GDP per year across buildings, power, networks, and chips. This is one paper’s model projection, not committed spending or an institutional forecast.
BROOKINGS ↗TWH · ZERO BASELINE · 2025 TO 2026
TWH · ZERO BASELINE · 2025 TO 2027
Both series are Gartner estimates reported in June 2026. The 2027 AI-server value is a projection.
In an April 2026 lawsuit, the plaintiffs alleged that 27 unpermitted methane-gas turbines at a Southaven, Mississippi power site serving Colossus 2 in Memphis had the potential to emit more than 1,700 tons of nitrogen oxides, 180 tons of fine particulate matter, and 19 tons of formaldehyde each year.
PLAINTIFFS’ ALLEGATIONS IN ACTIVE LITIGATIONThe earlier proposed design was reported to require 7.6 million liters per day, about the daily household needs of 55,000 people. The current design uses air cooling. Google says its watershed assessment did not meet the responsible-use threshold, and Uruguay’s government confirms the revised air-cooled design. This evidence does not establish that public opposition caused the change.
GOOGLE ↗ URUGUAY GOVERNMENT ↗For the 2025/26 auction, Monitoring Analytics estimated data-center load’s contribution in a counterfactual analysis. These are modeled sensitivity estimates, not directly metered causal charges. IEEFA reported the modeled contribution as about 63%, or $9.3 billion, with average household impacts of $16 per month in Ohio and $18 in western Maryland.
MONITORING ANALYTICS ↗ IEEFA ↗Folding@home reported 2.43 exaflops in April 2020 and described its volunteer network as bigger than the top 500 supercomputers combined. The separate figures of 4.63 million CPU cores and about 430,000 GPUs date from late March 2020, when the project was reported at around 1.5 exaflops. Distributed x86-equivalent throughput is not directly comparable to TOP500 HPL benchmark scores. This is evidence of distributed scale, not a benchmark win over supercomputers.
FOLDING@HOME ↗ MARCH 2020 COUNTS ↗A five-facility sample of about 4,000 servers found 30% had delivered no compute services for six months or more. The estimates of roughly 10 million servers worldwide and at least $30 billion in idle capital were extrapolations, not a 2026 measurement.
ANTHESIS / KOOMEY ↗NRDC reported that average server utilization had remained at 12% to 18% between 2006 and 2012. This is historical evidence, not a current utilization estimate.
NRDC ↗No defensible aggregate measure of utilization across all household PCs, phones, and consoles was found in this research. Distributed home hardware is also not a drop-in replacement for tightly coupled AI clusters: interconnect, latency, memory, security, energy mix, and heterogeneous devices matter. Folding@home proves possibility, not equivalence.
“A world model would enable you to predict what would happen. They have the ability to predict what a person would say.”
“We’re never going to get to human-level intelligence by training LLMs or by training on text only. We need the real world.”
Their alternatives differ in implementation but converge on experience, goals, consequences, and models of the world. Neither says that a larger next-token predictor is enough.
AIEN uses this descriptive term for a proposed system that would accumulate inspectable procedures and use learned search guidance, while external verification and machine evidence retain authority over correctness. It is not an established scientific taxonomy.
AIEN_0 is planned as a verification-guided synthesis model, more specifically a learned search policy for program synthesis. “Synthesis Policy Model” is AIEN’s own working descriptive label, not an established scientific taxonomy. The planned model would see the current problem state, remaining objective, and semantics of possible next operations, then rank which action OMEGA should try next. It is planned to learn from OMEGA-generated traces, not language prediction. Verification would remain outside the model.
Omega searches for programs, executes them, verifies them, measures them, and stores successful procedures and abstractions in a governed library. Capability accumulates as inspectable programs, not hidden weights.
AIEN_0 is planned to learn from synthesis traces: candidate transitions, verifier outcomes, machine results, costs, successes, failures, and pruning decisions.
The planned supervision comes from synthesis traces. The eventual objective has not yet been frozen; candidate approaches include supervised/pairwise ranking, imitation learning, value learning, and potentially offline RL.
Read the stack from the bottom up. Each layer has a separate job. The machine exists today. The planned AIEN_0, a learned guide inside the top layer, has not been trained.
TOP: INTELLIGENCE AND SEARCH
Thinks, searches, and invents. The planned AIEN_0 is described by AIEN with the working label “Synthesis Policy Model”: a small neural search guide planned to learn from OMEGA traces, not a language model.
Defines, synthesizes, verifies, and realizes programs. Working programs become part of a governed library.
The trusted substrate. It authorizes what is allowed to run on the bare metal.
The seed. It awakens the machine through verified boot.
BOTTOM: VERIFIED BOOT
The solid green track is the machine and its early work. Gold begins at M20, where the roadmap remains planned. The planned AIEN_0 belongs to that future training era.
The point where the system can design, build, and verify its own successor.
AlphaZero-style learned search guidance, but over program construction rather than board moves. The planned neural part would not need to know the answer. It would need to identify the most promising next operation. OMEGA’s external verifier would determine validity.
Intuition about where to search.
Durable procedural knowledge.
Remembered empirical facts and outcomes.
What computations mean.
What actually worked on a machine.
Does useful work per joule remain favorable after coordination, redundancy, networking, and device cooling are counted?
Can procedures run reproducibly and securely across different CPUs, GPUs, memory limits, and network conditions?
Do governed program libraries produce reusable abstractions beyond the tasks that created them?
Will the planned AIEN_0 improve synthesis search under equal verification budgets compared with unguided or hand-designed heuristics?
Can a distributed system avoid exporting power, water, noise, pollution, and financial risk to the same communities it claims to help?