The Substrate Revolt: DeepSeek's Huawei Pivot and Virginia's Ratepayer Bill Meet on the Material Layer of AI

Two stories broke on Wednesday 16 April 2026 that at first glance had nothing in common. In Richmond, Governor Abigail Spanberger sent SB 253 and HB 1393 back to the General Assembly with amendments that stripped out the language — "capacity auction and infrastructure costs on high-load users" — that would have forced data-centre operators to stop shunting grid upgrades onto residential ratepayers (Virginia Mercury, WHRO, 16 April 2026). Dominion Energy said the same afternoon that the amendments "undermine" the legislation's goals. Hours later in Washington, Nvidia chief Jensen Huang told reporters that DeepSeek's V4 — expected to launch in late April trained natively on Huawei's Ascend 950 PR — would be "a horrible outcome for the United States," because "if AI models are optimised in a very different way than the American tech stack, China will become superior to" the US (SCMP, Dataconomy, 16 April 2026).

The first story is American grid politics; the second is Sino-American chip geopolitics. Read through Kate Crawford's Atlas of AI (Yale, 2021) they are the same story told at opposite ends of the supply chain. The material substrate of AI — silicon, copper, power purchase agreements, the ratepayer — is where the political economy of the "AI race" is being negotiated in April 2026. What Spanberger's red pen and Huang's public warning share is a single admission: the infrastructure under AI is no longer a given. It is politically contested, at the level of chips and at the level of kilowatt-hours.

What happened this week

The timeline is tight. On 7 April TrendForce reported Huawei's Ascend 950 PR — SMIC-fabricated on N+3, 112 GB HiBL memory at 1.4 TB/s, 1 PFLOPS FP8 / 2 PFLOPS FP4 — had launched in Q1 and sat "between Nvidia's H100 and H200 in capability." The same week Tom's Hardware reported on a DeepSeek paper finding the older 910C delivered roughly 60 per cent of H100 inference performance — the first peer-visible benchmark showing Chinese silicon inside the range where substitution becomes rational. On 16 April three developments converged: (a) Reuters reporting, summarised in Dataconomy and TechBriefly, that DeepSeek had "ordered hundreds of thousands" of Ascend 950 PR units for V4, with months of work rewriting CUDA-dependent code against Huawei's CANN framework; (b) Huang's "horrible outcome" interview reported by SCMP; and (c) Spanberger's amendments landing, with a new return-on-equity cap at 9.3 per cent, down from Dominion's recently approved 9.8 per cent, excess returned to customers (WHRO, 16 April 2026).

Underneath both headlines runs a quieter set of official numbers. The IEA's Energy and AI report puts global data-centre electricity at roughly 415 TWh in 2024 — 1.5 per cent of global consumption — rising to around 945 TWh by 2030. "Accelerated servers," driven by AI, grow 30 per cent per year and account for nearly half of the net increase; US data-centre load rises around 240 TWh by 2030, a 130-per-cent jump. Ireland's CRU, in its March 2026 implementation deadlines, forces new large sites to provide dispatchable on-site generation matched to their Maximum Import Capacity and source 80 per cent of annual demand from renewables on a six-year glide path; CRU projects 5.8 GW of additional data-centre demand in the pipeline. CSIS's December 2025 analysis "If Compute Is the New Oil" treats the Gulf grid — Saudi HUMAIN, UAE G42 — as "asymmetric interdependence" with the US, contingent on continued H200-class export-licence approval.

The primary documents, read against each other

The BIS interim final rule of 15 January 2026, "Revision to License Review Policy for Advanced Computing Commodities," is the policy document that binds the chip half (Federal Register, docket 2026-00789). It moves H200-class silicon (TPP under 21,000, DRAM bandwidth under 6,500 GB/s) from "presumption of denial" to "case-by-case review" for China and Macau, subject to four gating conditions: domestic-supply sufficiency, no diversion of foundry capacity, aggregate China/Macau shipments capped at 50 per cent of US-bound shipments, and customer security-certification. The rule is the legal frame inside which Huang's "horrible outcome" warning lives. He is not warning that export controls are failing; he is warning they are succeeding at the wrong thing. They have bought enough friction to push DeepSeek to port its entire training pipeline off CUDA. If V4 ships in late April on Ascend 950 PR, every subsequent Chinese frontier model has a viable non-American path, and every subsequent BIS rule will be drafting for a market that has learned to do without.

The Virginia primary document is the amendment language itself. The original SB 253/HB 1393, passed by the General Assembly in March 2026, imposed explicit capacity-market and distribution-cost assignment on GS-5 rate-class customers (loads above 25 MW, effective January 2027) and targeted an estimated $5.52 per month in savings for residential Dominion customers (Virginia Mercury, 17 March 2026; WHRO, 16 April 2026). Spanberger's amendment replaced the mandatory cost-shift with advisory SCC language — "ensure that GS-5 customers are not being subsidised by other customers of the utility" — and raised the opt-out employee threshold from 200 to 10,000, narrowing the affected population to large industrials. The amendment also cut the incremental cap on Dominion's line-burial programme from 4 per cent to 2 per cent of the distribution rate base, against the $1.4 billion already invested and the $4.88 per month residential customers now carry for it. Del. Destiny LeVere Bolling, the HB 1393 patron, told reporters the governor's office had not consulted before the amendments landed. Dominion, which has separately admitted it "cannot meet" peak data-centre demand without large new gas generation, supports the amendments for a reason (DCD, March 2026).

The Crawford frame: AI as a mining operation

Kate Crawford's Atlas of AI opens with the observation that an artificial-intelligence system is not a computational abstraction but "a vast planetary network of extraction" — lithium pits in the Atacama, cobalt in Katanga, rare earths in Bayan Obo, and the electrical substations that turn stranded-asset coal and curtailed-wind MWh into training runs. "AI," Crawford writes in Chapter 2, "is made from natural resources, fuel, human labor, infrastructures, logistics, histories, and classifications." When the compute layer is stable, that substrate is invisible. What happened on 16 April is that the substrate made itself visible at both ends of the pipeline at once.

On the hardware end, the Huawei/DeepSeek pivot is the inflection Crawford's framework predicted: if compute is an extractive industry, extractive-industry logic produces import-substitution whenever the sanctioning power's leverage gets expensive enough. The 2022 BIS controls were designed to "control the height of the technology ladder by a couple of rungs" — former commerce undersecretary Alan Estevez's phrase, quoted in CSIS reporting. What happened instead is what happens in any sanctioned commodity market: a parallel vertical supply chain was funded. SMIC's N+3 node, Huawei's HiBL memory, Cambricon's MLU370, and the CANN software stack together describe a vertically integrated AI-compute supply chain independent of TSMC, Nvidia and CUDA. DeepSeek's V4 is the first frontier model built to consume it. Chinese hyperscalers are reportedly placing orders at a 20-per-cent Huawei price premium (TrendForce, 7 April 2026) — the subsidy-to-sovereignty calculus Crawford names in Chapter 4. Huang's warning is the first time the incumbent has publicly priced the loss.

On the energy end, the Spanberger amendment is the mirror image. Hyperscalers have quietly converted grid access into a take-it-or-leave-it relationship: Virginia's original SB 253/HB 1393 tried, briefly, to renegotiate its terms in the ratepayer's favour. ITIF's 7 April 2026 briefing, "Four Reasons New AI Data Centers Won't Overwhelm the Electricity Grid," landed within days of the amendment — the kind of industry-funded counter-argument that reaches legislators precisely when a policy threatens to redistribute infrastructure costs. Bernie Sanders and Ron DeSantis, quoted together in CNBC on the same fight in January, are an unlikely coalition precisely because the issue has escaped the usual ideological grooves (CNBC, 1 January 2026).

Why the Global South reader should care

Two structural consequences. First, the compute-geography of frontier AI is bifurcating faster than the policy literature assumes. CSIS's "Open Door" analysis (December 2025) argues Asian, African and Middle Eastern labs face an increasingly explicit choice between an American stack (Nvidia plus hyperscaler cloud plus Anthropic/OpenAI APIs, licensed through BIS) and a Chinese stack (Ascend plus Huawei Cloud plus DeepSeek/Qwen weights). The 20-per-cent Huawei price premium is the floor; the ceiling is whatever compliance cost a non-Western lab assigns to US end-use verification under the 15 January rule.

Second, the datacentre grid fight is not a US domestic issue. Ireland's data-centre load on course to 32 per cent of national electricity consumption by 2026, Saudi HUMAIN's Neom build, and South African fights over Johannesburg-area hyperscaler cooling are the same grid-socialisation argument in different jurisdictions. The workloads are globally mobile; the grid costs are locally fixed. Infrastructural decisions dressed as technical choices redistribute political power — the Virginia ratepayer paying $5.52 extra a month so a Loudoun hyperscaler can train the next model is the same structural figure as the Irish household paying rising bills while Dublin's data centres recover their substations.

What it means going forward

Three implications. First, the Anthropic-Pentagon supply-chain-risk saga is a contract-layer story; the Huawei/DeepSeek pivot is a silicon-layer story; the Spanberger amendment is an energy-layer story — and they are resolving in different directions. The contract layer is consolidating state power over US labs. The silicon layer is fragmenting into two stacks. The energy layer is socialising costs within each host grid. That is not a stable equilibrium.

Second, the "AI race" frame Huang deployed — the implicit threat that China will "become superior" if DeepSeek V4 runs natively on Ascend — is an argument for relaxing export controls, not tightening them. The incumbent supplier, whose margin depends on access to Chinese customers, recruits national-security framing to argue for the opposite policy the framing was deployed to justify. Watch whether BIS's next revision moves H200 review further toward case-by-case approval under lobbying pressure.

Third, the honest map is Crawford's. The "AI revolution" that reads on the front page as Dario Amodei walking into the West Wing or Jensen Huang giving a podium interview is, on the ground, a reorganisation of mines, substations and fabs. When Spanberger strikes out a paragraph of cost-assignment language, or when Huang names a model that has ported off CUDA, what is moving is not a talking point. It is the substrate. And the substrate does not negotiate back.

That is the story.