GoM · Resource & Power
The same answer, at tens of watts.
The same answer as the best model you can run locally — measured, 29/29 on hard math for both —
from a 2.54 GB package on a laptop or phone at tens of watts, instead of a
multi-hundred-GB model behind a datacenter GPU cluster at hundreds of watts per accelerator. Every footprint, speed and cost figure below is a
logged run (on-device Apple-MPS, or a logged cloud-API latency/cost).
Recorded — real logged runs, not live
1 · Footprint
2.54 GB on a phone vs ~39 GB on a workstation vs a cluster
GoM-Brain is a ~40 MB trainable brain riding on the smallest viable base (qwen3:4b,
2.5 GB) — 2.54 GB total. That is ~15.5× smaller than the best model a person can
realistically run locally, while matching it on hard math (both 100%, 29/29). Versus
frontier cloud models, the trainable brain itself is 4–5 orders of magnitude smaller,
and the whole package sits a full infrastructure tier below — pocket hardware instead of racks.
GoM-Brain qwen3:4b + ~40 MB brain · edge
2.54 GBphone / laptop
Best-Local Qwen3.6-35B-A3B · single Mac
39.4 GB128 GB Mac, offline
Frontier (cloud) GPT-5.5 / Gemini / Grok / Opus
100s GB–TBest. · datacenter
Brain itself: ~40 MB (10.6M params)
vs Best-Local: ~15.5× smaller
tiny-GoM solvers: 132K–37M params
Hard math: GoM 100% = Best-Local 100%
Source: math_base_sweep.json (total_footprint_gb 2.54; brain_footprint_mb 40), math_local_data.json (size_gb).
2 · Speed & token delta
Exact on-device execution vs multi-second cloud round-trips
Where GoM exact-executes (op-graph arithmetic, Bayesian belief update) there is no
sampling distribution to retry and no budget to exhaust — it runs in microseconds, locally,
with no network. This is deliberately not a "GoM always wins" table: on heavy local
search (Hanoi at larger N) GoM is comparable or slower, and we say so per card.
Source: speed_data.json (frontier_median_ms = logged cloud-API latency; gom_median_ms = on-device MPS). Deltas as logged.
3 · The token tax GoM doesn't pay
Frontier spends the budget — and still fails — at the hard regime
| Task | Frontier outcome (logged) | GoM |
|---|
Tower-of-Hanoi N=11
budget to 160k+ tokens |
0% — all 6 models
illegal move @ 1280 (GLM-5.2) / 768 (Gemini) / 225 (DeepSeek) / 96 (Opus) / 65 (Grok); GPT-5.5 unparseable |
100/100 at N=11 under committee search (95% CI ≥ 96.3%) — see the reasoning note for the full result |
| Sudoku (17-clue extreme) |
truncation blow-up
DeepSeek-V4 truncated 30/30; GPT-5.5 truncated 18/30 and the run cost $7.73; best frontier (GLM-5.2 @64k) 60% |
90% (270/300) |
Source: comparison_data.json (Hanoi N=11 6/6 solved=False w/ logged illegal-move indices; Sudoku truncation + GPT-5.5 $7.73; GoM Sudoku 270/300).
4 · When power is the ceiling.
Move the workload off the datacenter, onto the edge
GoM runs fully offline on a single laptop/phone-class device — no cluster, no GPU rack,
no network. Because it uses only standard on-device primitives, the same artifact runs
on Apple Silicon, AMD and NVIDIA — it is not locked to one vendor's cloud.
⚡ GoM-Brain → edge
tens of watts
~2.54 GB, fully offline on a laptop/phone. $0 marginal cost, no API, no network round-trip.
☁️ Frontier → datacenter
100s of watts / GPU
Multi-hundred-GB model on datacenter GPUs (~350–700 W TDP band — vendor-published, est.), in racks, behind an API. $0.09–$30 per Mtok.
The macro hook. AI data-center electricity demand is climbing steeply, and the primary
sources agree on the direction: Berkeley Lab / DOE measured data centers at 4.4% of US
electricity in 2023 and project 6.7–12% by 2028; the IEA attributes about half of all
US electricity-demand growth in 2025 to data centers; and Columbia's energy-policy center
projects AI GPUs alone taking up to 27% of planned new US generation capacity by 2027.
When capability growth is throttled by power, the industry's answer is to buy more power —
GoM's answer is software: the same capability, executed exactly, at a fraction of the
footprint, at tens of watts. That decouples AI capability from datacenter power.
Sources: LBNL/DOE,
2024 United States Data Center Energy Usage Report; IEA, Electricity 2026;
Columbia SIPA Center on Global Energy Policy, Projecting the Electricity Demand Growth of
Generative AI.
The cheapest datacenter is the one you don’t have to build.
Provenance. Footprint / speed / cost figures are logged runs from
speed_data.json, comparison_data.json, math_base_sweep.json
and math_local_data.json, shipped alongside this page. They are
replays of real recorded runs, not live model calls. GPU TDP bands and closed-model
footprints are third-party estimates (est.); grid figures are from the primary
sources cited above. The full quantified write-up and methodology are available to qualified
partners under NDA — request access.