Author: Nita Farahany Source: https://nitafarahany.substack.com/p/your-electricity-bill-is-about-to Published: September 21, 2025
This is the eighth installment (Class 8 of 27) of Nita Farahany's "Inside My AI Law & Policy Class" essay series and the second of its classes on compute. It argues that compute is the most governable layer of the AI stack and uses two contemporaneous strategies — the US government's equity stake in Intel and the European Union's regulatory posture — to illustrate competing approaches to compute governance.
Summary of argument
The class is anchored on the August 22, 2025 purchase by the US government of 433.3 million Intel shares at $20.47 per share, an approximately 10% ownership stake. Farahany characterizes this as the most significant industrial-policy intervention into a technology company since the early 20th century, and distinguishes it from a 2008-style bailout. From this anchor she develops a framework for why compute is a tempting governance lever, and contrasts a US "buy your way in" approach with an EU "become the referee" approach.
Key claims
Four properties that make compute governable
Farahany sets out four properties of compute that, in her account, make it a governance leverage point:
- Physical control is possible. GPUs carry serial numbers, sit in specific data centers, and cross borders; the argument is that one cannot ban math but can ban semiconductors.
- Big compute is traceable. She cites GPT-4 training as roughly 25,000 H100s running for months at 10–50 MW, an installation visible from satellites.
- FLOPs are countable. She points to the EU AI Act systemic-risk threshold of 10^25 FLOPs as an example of compute being used as a measurable regulatory trigger.
- Compute is concentrated. Roughly 12 companies control global AI compute: Nvidia (70%+ of design) with AMD and Intel; TSMC (90% of advanced fabrication); ASML (100% of EUV lithography); and AWS, Microsoft, and Google in cloud.
US "buy your way in"
Farahany frames the Intel stake as giving the government insight, steering, and leverage over a chipmaker, while flagging risks she associates with the approach: market distortion, the government picking winners, and politicization, including the question of what happens when AMD or Nvidia seek equal treatment.
EU "become the referee"
The EU is described as holding roughly 32,000 AI chips in total, against Microsoft alone at roughly 1.8 million, a 56:1 ratio. The argument is that, unable to compete on raw compute, the EU positions itself as a referee instead. Citing the ICFG 2024 report, Farahany lists five interventions: strategic compute allocation toward niches where compute alone does not determine success; government observers placed inside frontier-model training runs; triage enforcement (top 10% deep evaluation, middle 30% standard, bottom 60% self-certification); a foresight unit anticipating capabilities; and multilateral compute oversight.
On-chip monitoring
Farahany describes both the US and EU as converging on building monitoring into chips — tracking cumulative FLOPs, distinguishing training from inference, and reporting against thresholds. The class stops at the point where this becomes a personal-privacy question: that every ChatGPT query could be logged, with open questions over who would see the logs and whether they could be subpoenaed.
She offers the 2016 Apple–FBI dispute as an analog, asking what would have followed had the government owned 10% of Apple and whether Apple could then have refused the FBI.
Cascade scenarios and four "uncomfortable truths"
Farahany sketches cascade scenarios in which compute controls are evaded: a black market in pre-2025 unmonitored chips; companies optimizing to 10^25.9 FLOPs to stay just below the threshold; satellite-based training to escape jurisdiction; and a "splinternet for AI" splitting into US, Chinese, and everyone-else blocs.
She closes with four propositions she frames as uncomfortable truths: that every technical solution can be defeated theoretically; that the pace mismatch between technology and governance is accelerating; that governments will take actions she describes as previously unimaginable, with the Intel stake offered as an example that would have been unthinkable five years earlier; and that constraints shift rather than disappear.
Provenance
Single-source page summarizing one essay in Farahany's course series. As an essay, its framings (the four compute properties, the US-versus-EU contrast, the cascade scenarios) are presented as the author's positions rather than as established findings.
Relationships
- part-of: Nita Farahany intro course series (Class 8 of 27)
- related: Compute Governance, Intel, EU AI Act (Regulation 2024/1689)
- previous: Inside My AI Law & Policy Class 7: Why China Quit US Chips (Farahany, September 2025) next: Inside My AI Law & Policy Class 9: When AI Discrimination Happens 1.1 Billion Times (Farahany, September 2025)