Steelman analysis

Creation is not a substrate to be drawn down for compute. Grid buildout at AI-datacenter scale directly couples to aquifer depletion (hyperscaler water up 20% YoY tied to AI), thermoelectric water embedding, and rare-earth extraction harms concentrated at mine sites on populations who receive none of the benefit. Harm scores land (0.3), water (0.4), extraction (0.4) are not acceptable offsets against downstream utility --- they are first-order wrongs against the integrity of watersheds and creaturely life. The intervention sanctifies extraction.

Watersheds and airsheds are preconditions of capability, not substitutable inputs. Grid buildout with harm_water 0.4 and harm_land 0.3 degrades the capability set of the communities hosting the load --- siting disproportionately lands in environmental-justice regions, and thermoelectric water embedding compounds at the generation tier. harm_displacement 0.8 means labor welfare is not incidentally damaged; it is the design. No quantity of downstream compute substitutes for the capability floor this intervention erodes at the sites where it physically lands.

Permitting reform accelerates buildout but also locks in a specific generation mix, a specific siting footprint, and a specific governance tier (four hyperscalers, one mission-software prime) for decades. harm_lock_in 0.6 is the categorical failure: present actors foreclosing future option sets on water, land, and institutional control in order to meet a 2030 load forecast whose range spans 2x. You cannot justify a duty-violating foreclosure with a point estimate from a factor-of-two confidence interval.

harm_concentration 0.7 is the tell. Grid buildout without governance preconditions feeds load to four hyperscalers and one mission-software prime whose revenue and contract posture is accelerating into defense targeting. The expected downstream use of the marginal megawatt is not protein folding --- it is JWCC workloads, Maven expansion, and capital-extraction SaaS. Accelerating the substrate under current concentration conditions is straightforwardly negative-EV on suffering reduction.

Compute is the single largest input into capability gain; training compute doubles every 5-6 months and algorithmic efficiency compounds on top. Removing the grid constraint removes the last non-alignment-gated physical brake on the capability curve. 'Build-only-if-safe' requires a live pause option; permitting reform that uncorks 2,600 GW of queued capacity destroys that option irreversibly. This is the intervention most precisely targeted at eliminating the brake x-risk depends on.

Grid is the wrong bottleneck to unjam. Enterprise absorption lags the frontier by years, leading-edge fab is single-point-of-failure at TSMC, and algorithmic efficiency halves compute requirements every 8 months --- so the marginal megawatt buys less capability each year while the actual rate-limiting steps (chips, absorption) remain untouched. Grid buildout is a political project with a 10-year payoff horizon when the velocity-binding constraints are silicon and deployment surface area today. Misallocated acceleration.

Sovereignty-expanding deployment requires a distributed compute and governance substrate. Grid buildout under current conditions routes the new megawatts through four hyperscaler primes and one mission-software vendor --- harm_concentration 0.7 by design. The intervention does not expand individual capacity; it expands the capacity of the existing concentration tier to absorb more load. Sovereignty erodes in direct proportion to how much grid you hand them.