Aaron Ginn: US export controls are backfiring — China is demolishing the US on AI efficiency because we gave them Nerf chips

Summary

Aaron Ginn argues that US export controls on AI chips have produced a self-defeating outcome: by restricting China to lower-spec hardware, Washington inadvertently forced Chinese AI developers to become world leaders in computational efficiency. 'China is absolutely demolishing us on the efficiency trend,' Ginn contends, framing the nerf-chip policy as a strategic own goal that insulates American frontier labs from the competitive pressure that would otherwise drive down inference costs domestically.

The H20 Deal and Its Constitutional Architecture

The Trump administration's recently announced 15% revenue-share arrangement on Nvidia H20 chip exports to China is read by Ginn as both pragmatic dealmaking and a constitutional workaround. Export taxes are prohibited under the US Constitution, so the structure has to be voluntary. In practice, it is Dell or Super Micro, not Nvidia, that applies to the Bureau of Industry and Security for export licenses, because they are the OEMs assembling full server racks for Chinese buyers. Nvidia's component pricing is never directly visible in the BIS filing. The revenue-share mechanism is therefore the only legally viable path to extracting a fee from the transaction.

Ginn estimates Nvidia's gross margin on H20s runs in the high 50% range, meaning China-bound shipments remain highly profitable even under the new arrangement. He notes TSMC and Intel foundry constraints mean restarting production takes months, so near-term H20 supply will largely come from existing inventory.

Export Control Critics as 'AGI Zealots'

Ginn dismisses what he calls 'faux neo-China hawks' — a policy class he believes is less motivated by genuine national security concerns than by a quasi-religious conviction that AI will produce a god-like AGI, and that America must build it first. He argues their policy prescriptions — blocking chip exports, restricting model weights, limiting tokens — form an incoherent list that only holds together if you accept the AGI-as-deity premise, which is politically unsaleable in Washington. The same donor class, he notes, was heavily involved in the Section 230 / Title II content moderation debates and has a pattern of wrapping metaphysical anxieties in techno-nationalist framing.

His practical line on controls is more targeted: restricting lithography equipment to prevent China from reaching 5nm, 3nm, or 2nm process nodes is reasonable because that equipment is physically enormous, costs around $200 million per unit, and is manufactured by ASML and Japanese suppliers. Conflating that with Nvidia GPU exports reflects a fundamental misunderstanding of the supply chain — Nvidia designs chips but does not fabricate them, making it roughly analogous to attacking Costco's Kirkland brand rather than the underlying manufacturer.

Intel: Nationalization Likely, Strategy the Real Problem

On Intel, Ginn is cautious about personality-driven criticism of the current CEO, focusing instead on strategy. He expects some form of de-facto government stake — structured through existing federal lending vehicles as a warrant — similar to the 15% equity position the federal government recently took in a rare earth minerals project in Nevada. He sees leading-edge domestic fabrication as economically unviable and would redirect that ambition to allied nations such as Brazil or Argentina, while preserving a US trailing-edge foundry capability. The TSMC Arizona expansion is complicated, he suggests, by workforce culture gaps that became apparent once TSMC engineers were on the ground.

GPT-5 and the Shift Toward Efficiency

Ginn's read on the GPT-5 launch is that the meaningful signal is strategic, not technical. Sam Altman is positioning OpenAI as a consumer application company, evidenced by UX improvements, downside-scenario mitigation, and a deliberately modest context-window expansion that reflects inference capacity constraints rather than a capability ceiling. Benchmark obsession misses the point — consumer and enterprise value delivered is the only metric that matters commercially.

He frames the broader AI industry trajectory over the next two years as a rotation from effectiveness (raw capability, a B2B competition) to efficiency (cost reduction, mass adoption). Chinese labs, forced onto constrained hardware, have already made that rotation by necessity. American labs are beginning to follow as competitive exposure increases — OpenAI's move toward cheaper models is cited as early evidence. The underlying economic law that compute costs fall is described as 'bulletproof,' with Jevons' paradox the expected outcome: cheaper inference drives higher consumption, not demand destruction.

Data Centers More Vulnerable Than GPUs

Ginn draws a sharp distinction between GPU supply and data center supply. GPU inventory is tightly controlled — effectively a monopoly through Nvidia — limiting bubble dynamics on that side. Data center lending, by contrast, is aggressive because lenders underwrite the asset on a 15 to 20-year useful life. If power-density profiles are misconfigured or GPU availability tightens as Nvidia scales back production, operators could face facilities with no viable tenants and debt they cannot service. He sees data center overbuilding as the more probable bubble, not GPUs.