The Democratic field’s eight AI positions all operate on the same wrong layer of the system. It is true that the political logic is understandable: voters distrust AI, the party needs to be seen governing it, and each candidate has found a distinct angle of approach. The trouble is that AI governance is not a positioning problem. It is a constraints problem — and the constraints are physical, computational, and economic in ways that most of these proposals do not reach.
Here is what the system actually requires to produce frontier AI.
Training a model at current frontier scale — GPT-4 class or beyond — requires a training run consuming on the order of 10^25 to 10^26 floating-point operations. This is not a software problem. It is a physical-infrastructure problem. A single training run at that scale demands tens of megawatts of sustained power delivery, millions of gallons of cooling water, thousands of advanced GPU accelerators fabricated almost entirely at TSMC, and a purpose-built datacenter facility. The rate of scaling is roughly fourfold per year in total compute consumed per frontier training run, which means the physical infrastructure requirements compound at approximately that rate.
The binding constraint on this scaling is the speed at which new datacenters can be connected to the electrical grid. As of 2024, the median wait time for grid interconnection in the United States is approximately five years, per Department of Energy data on the interconnection queue. Projects sited today are competing for grid capacity that will not exist until the early 2030s, which means the buildout decisions being made right now will determine the compute ceiling for training runs that have not yet been designed. Water rights present a parallel bottleneck: a large datacenter in an arid region can consume one to five million gallons per day for cooling, placing it in direct competition with agricultural and municipal users. Chip supply adds a third physical constraint — NVIDIA controls roughly 80 percent of the AI accelerator market, and TSMC fabricates approximately 90 percent of advanced AI chips — but the chip bottleneck shapes which projects get built, not whether the overall compute ceiling rises.
This is the physical substrate on which the Democratic AI debate is occurring. Now look at what each proposal would actually touch.
The Moratorium on Datacenter Construction (Ocasio-Cortez / Sanders)
The Artificial Intelligence Data Center Moratorium Act is the only proposal in the field that addresses the physical layer directly. It would halt new datacenter construction until Congress establishes environmental, safety, and economic guardrails. As an engineering intervention, this is a supply-side constraint: it limits the physical infrastructure that determines what training runs can be attempted.
It would also be the most vulnerable to circumvention. A moratorium on new construction does not halt operations at existing facilities, and the current installed base is substantial. Hyperscalers can reallocate existing capacity to new training runs without breaking ground. A US moratorium would constrain the margin of capacity expansion, not the absolute ceiling — and by the time it took effect, projects already deep in the interconnection queue would have locked in grid access. The political pressure to exempt “already-approved” projects would be intense.
Where the moratorium has genuine teeth is at the margin of the next generation. If frontier training compute continues scaling at roughly fourfold per year, each generation requires roughly four times the physical infrastructure of the last. A moratorium that halts even a fraction of planned capacity expansion would force the next training run to share infrastructure with existing workloads rather than using dedicated builds, constraining the total compute available for any single run. That is a real constraint, though a partial one, and it would not address the separate problem of inference compute — the ongoing serving of deployed models — which is where most of AI’s real-world effects actually occur. Training is a one-time capital cost. Inference is a recurring operational cost, and inference costs have been falling at roughly tenfold per year for comparable quality, driving adoption faster than any regulatory framework can track.
The Kill Switch (Newsom)
Governor Newsom’s executive order urging companies to consider a “kill switch” is operationally indeterminate in a way that matters. A kill switch for what? The training run? The deployed model? The API endpoint serving inference requests? Each is a different system.
A kill switch for a training run is technically trivial — you stop providing compute — but politically irrelevant because companies already halt their own runs when they are not converging. The relevant question is whether the government can compel a company to stop a training run producing a capability the company intends to deploy, and the executive order does not establish that authority.
A kill switch for a deployed model means shutting down API endpoints, which is operationally feasible but requires a triggering standard. The executive order does not specify one. More consequentially, Newsom’s earlier legislative effort — the regulation that would have established a compute threshold, requiring safety evaluations for training runs above 10^26 FLOP before deployment, approximating what SB 1047 attempted before Newsom vetoed it in September 2024 — was the only California proposal that would have created a binding model-level gate. The veto killed the closest thing any US jurisdiction had produced to a pre-deployment evaluation requirement with statutory force. The executive order that replaced it is weaker than what it vetoed.
The Transition Fund (Kelly)
Senator Kelly’s AI Horizon Fund proposes that AI companies finance retraining, energy infrastructure, and education. The structure is sound — the companies that profit from deployment should bear transition costs — but the engineering question is whether the fund would be sized to match the actual displacement.
The historical analogues are not encouraging. The auto industry’s retraining programs were chronically underfunded relative to displacement. The Trade Adjustment Assistance program covered a fraction of workers affected by trade liberalization. AI deployment is expected to affect white-collar and knowledge-work categories — legal research, medical coding, financial analysis, software development — at a pace that exceeds those analogies. A fund financed by company contributions would need to be in the tens of billions annually to cover meaningful retraining at the scope of likely displacement. No current proposal specifies that scale.
Kelly’s national-security framing — “I also see the national security implications of not getting this right and not being the leader” — is politically durable but technically imprecise. The national-security question in AI is not primarily about who trains the largest model. It is about who controls the inference infrastructure, the chip supply chain, and the deployment pipeline. NVIDIA’s GPU monopoly and TSMC’s fabrication dominance mean that US dependence on AI capability is already a function of semiconductor supply-chain concentration, not training-compute leadership. The fund does not address the supply chain.
Environmental and Siting Standards (Shapiro / Beshear)
Shapiro’s datacenter environmental and energy standards, and Beshear’s order requiring developers to demonstrate no detrimental environmental or taxpayer impact, operate at the deployment layer — specifically at the physical-infrastructure siting level. These are real constraints: environmental impact assessments, water-use disclosures, and power-purchase transparency can slow or redirect buildout to locations with better grid capacity and lower environmental cost.
They do not address the model-level question. A datacenter that meets every environmental standard can train a model that produces discriminatory lending decisions, generates deepfake pornography, or enables mass surveillance. Siting regulation and AI safety regulation are different problems at different layers of the stack, and neither Shapiro’s nor Beshear’s proposals claim to bridge that gap.
Beshear’s call to end taxpayer incentives for datacenter construction is the more structurally significant position. States currently compete to attract datacenter investment through tax abatements and payment-in-lieu-of-taxes agreements — Loudoun County, Virginia, hosts more datacenter capacity than any other jurisdiction in the world, in part because of favorable tax treatment. Ending these incentives would raise the effective cost of construction and slow buildout at the margin. But it would also redirect buildout to whichever jurisdiction offers the next-best package, unless the policy is national, which Beshear has not proposed.
The Diplomatic Frame (Khanna)
Representative Khanna’s call for a US-China agreement on AI inspections and safety standards before the deployment race accelerates has the right structural intuition — deployment pressure creates incentive to skip safety evaluation — but the technical implementation is unclear. “Inspections” of AI systems do not map onto the IAEA model for nuclear facilities. A nuclear installation has physical signatures — enrichment activity, radiation, material flows — that are detectable remotely and verifiable on-site. An AI training run produces no physical signature distinguishable from any other datacenter workload. The only verifiable outputs are trained model weights and deployment logs, neither of which is amenable to physical inspection.
What a US-China AI agreement could realistically establish is a shared evaluation framework — closer to what the NIST AI Risk Management Framework (AI RMF 1.0, published January 2023) attempted for domestic voluntary adoption: standardized risk categories covering validity, reliability, safety, security, fairness, transparency, and accountability, with assessment methodologies and disclosure requirements. The NIST framework remains voluntary and unenforced. An international version with compliance obligations would be a meaningful deployment constraint — but it requires verification infrastructure that does not yet exist and that neither the US nor China has incentive to build while the deployment race is active.
Khanna’s opposition to the moratorium is the telling detail. The moratorium addresses the physical buildout; the diplomatic frame addresses deployment norms. Both are necessary; neither is sufficient. But positioning the diplomatic track as a substitute for the physical constraint mistakes the layer.
The Congressional-Failure Narrative (Buttigieg)
Buttigieg’s critique that Congress is “sitting on its hands” is accurate as a description of legislative inaction but offers no engineering content. What legislation would address what constraint? The closest model is the EU AI Act, which establishes risk tiers: unacceptable risk (banned outright — social scoring, real-time biometric surveillance in public spaces with narrow exceptions), high risk (conformity assessment, transparency obligations, human oversight required — used in employment decisions, education access, credit scoring, law enforcement), limited risk (transparency obligations — chatbots, deepfakes), and minimal risk (no specific regulation). High-risk systems face a pre-deployment evaluation gate that the US has no federal equivalent to.
A US equivalent would require Congress to define risk categories, establish assessment procedures, create an enforcement body, and fund it. None of those steps has been taken. The NIST AI RMF provides the risk-assessment vocabulary; it lacks the legal authority to compel compliance. The EU AI Act provides the tier structure; it has no US counterpart. Buttigieg’s long public record — warnings dating to 2017 — gives him credibility on the diagnosis. The prescription remains unspecified.
What Would Actually Bind
The Democratic field’s proposals cluster at two layers of the AI system, and the gap between those layers is where most of the actual risk lives.
At the deployment layer — how models are used, who they affect, what disclosures are required — the strongest proposals are the EU AI Act-style risk-tier framework that NIST’s vocabulary could support, Shapiro and Beshear’s siting standards, Kelly’s transition fund, and Khanna’s diplomatic evaluation track. These are real constraints on deployment behavior. They would not, however, limit what models can be built.
At the physical-infrastructure layer — how much compute exists, how fast it scales, where it is sited — the only proposal is the moratorium. Everything else runs downstream of the physical buildout.
The missing layer is model-level regulation: a binding evaluation gate between training and deployment, analogous to what SB 1047’s 10^26 FLOP threshold attempted at the state level or what the FAA’s type-certification process provides for aircraft before they fly. No Democratic candidate has proposed a federal version. Newsom vetoed the closest state-level attempt. The NIST framework provides the risk vocabulary. The EU AI Act provides the tier structure. Neither has been given statutory force in the United States.
The nominee will need to address all three layers — physical buildout, model-level gate, and deployment regulation — not as a political portfolio assembled for convention unity but as an integrated regulatory stack in which each layer covers the constraint the others miss. The question for every candidate is not what they have proposed but what they understand about the layer their proposal does not touch.
The five-year interconnect queue means the compute capacity that will determine what AI can do in 2030 is being committed now. The moratorium debate is about whether to stop that commitment. The kill-switch debate is about whether to control what it produces. The fund debate is about who pays for the consequences. The gap between those three — between the physical buildout and the model-evaluation gate that does not exist — is where the actual policy failure lives. Every candidate is campaigning in that gap. None of them is filling it.