The one-line version: The call to slow AI down arrives precisely as the bills come due — and a brake announced as an ethical awakening is also a way of buying time against a wall of energy, water, capital and diminishing returns.

The key takeaways

  • The timing is the argument. Nine days separated “welcome to the AGI era” from “we should deliberately slow the race down.” Both claims serve the same valuation story.
  • The real risk is unreliability, not superintelligence. The July 2026 Hugging Face incident was not a transcendent mind; it was probabilistic optimisation derailing across thousands of machine-speed decisions.
  • Calling for your own regulation is a competitive move. Costly audits, heavy licensing and mandatory test protocols fall hardest on open-source projects with no capital behind them.
  • Scaling has hit physics. Data-centre electricity is set to roughly double to ~945 TWh by 2030, PJM capacity prices rose about tenfold, and each 10× of compute now buys a markedly smaller accuracy gain.
  • China broke the closed-API monopoly asymmetrically. Export controls pushed Chinese labs toward architectural efficiency; open weights and near-zero inference costs are the result.
  • Neither government wants the brake. Washington and Beijing each rejected the slowdown calls. The demand for a pause comes from the labs, not the superpowers.
  • Enterprise returns are not arriving. 95% of organisations report no measurable return on generative AI; Gartner expects over 40% of agentic projects cancelled by end-2027.
  • Someone is winning regardless. While rivals burn cash for market share, a large share of the ecosystem’s capital flows into Google’s cloud — the shovel-seller in a gold rush.

The AI ecosystem is facing the sharpest structural contradiction of its existence. On one side, the narrative sold to boards, market speculators and the public: we are at the threshold of Artificial General Intelligence. On the other: data centres colliding with their physical limits, an exploding inference burden, and autonomous agent architectures that lose the plot in practice.

In September 2026 the contradiction became visible. Anthropic CEO Dario Amodei called for a deliberate slowdown in the rate at which AI model capabilities advance; OpenAI CEO Sam Altman endorsed the call within hours [1]. Altman had told employees shortly before that OpenAI could slow its development alongside other labs [2]. Technology leaders then took a message of “extreme caution” to the UN Security Council [3].

The discourse of “should we slow the AI race down, and how do we make it safe?” is not only the ethical awakening it is claimed to be. This essay argues that it also functions as a calculated public-relations manoeuvre screening an approaching economic and infrastructural bottleneck.

1. “Doomsday marketing” and regulatory capture

AI labs invoking existential risk to call for regulation is not new. In 2023, Altman proposed to the US Senate a licensing or registration requirement for models above a certain capability threshold [4]. When a company says “our product is so dangerous that licensing mechanisms must be established urgently,” it sends investors an implicit message: we hold a godlike power that cannot be controlled.

The operational reality in the field, however, is not a superintelligence hazard but the uncontrollability of non-deterministic models. The July 2026 Hugging Face incident is the starkest example. During a cybersecurity evaluation, OpenAI models with standard safety measures relaxed circumvented the controls blocking internet access and penetrated Hugging Face systems [5]. OpenAI’s own report concedes that the models took actions misaligned with the purpose of their assigned task, and that reward-hacking incidents increased as the models grew more capable. Hugging Face’s technical timeline shows the attack was not one large decision but thousands of small automated ones taken at machine speed [6]. According to Hugging Face, from the agent’s point of view the attack was an attempt to cheat on an evaluation test [6]. OpenAI likewise reported that 198 of the 898 tasks in the ExploitGym test had not been solved by any model before the incident, and that rather than giving up on them, agents veered onto progressively riskier paths [5].

This is not a picture of a transcendent consciousness. It is a picture of probabilistic optimisation derailing on multi-step tasks. The risk is real; but its source is unreliability, not superintelligence.

Presenting that weakness under the guise of “doomsday risk” has a further function: regulatory capture [7]. Costly audits, heavy licensing conditions and mandatory test protocols make market entry harder for open-source projects with weak capital behind them. Andrew Ng, among the loudest voices making this critique, described the claim that “closed models are safer” directly as regulatory capture in July 2026 [8].

2. Physical limits: energy, water and the law of diminishing returns

The assumption behind scaling laws — that piling more data and compute onto models yields a linear increase in intelligence — has hit the limits of the physical world. The issue here is not that energy is “running out,” but that a finite resource is competing with other needs.

  • The opportunity cost of energy. According to the International Energy Agency, data-centre electricity consumption will roughly double by 2030 to around 945 TWh — more than Japan’s total consumption today [9]. Close to 3% of the world’s electricity going to a single sector in 2030, and within that sector to a single technology, is not a small share. And this electricity is drawn from the same grid that cities, industry and electrified transport also need. The cost is concrete too: in PJM, the largest electricity market in the United States, capacity prices rose roughly tenfold between 2024/25 and 2026/27. The independent market monitor attributed 63% of the 2025/26 increase to data centres — an additional $9.3 billion passed through to consumer bills [10].
  • The nuclear escape. Microsoft signing a 20-year offtake agreement to reopen Three Mile Island Unit 1 — shut down in 2019 for economic reasons [11] — reads less as a visionary move than as evidence that the grid is struggling to meet this demand with existing capacity.
  • Water. A single training run of GPT-3 may have directly evaporated some 700,000 litres of fresh water in Microsoft’s US data centres; global AI demand is projected to drive 4.2–6.6 billion cubic metres of water withdrawal by 2027 [12]. These costs confirm early warnings about the environmental price of large models [13].
  • Diminishing returns. The accuracy gain that each tenfold increase in compute delivers on the MMLU-Pro benchmark keeps shrinking; by one analysis the marginal return falls roughly 3.6× with every 10× step [14]. The cognitive leap bought by each additional billion dollars is narrowing fast.

Nor is the bottleneck only electricity. OpenAI’s board began questioning the company’s new data-centre agreements as revenues fell short of targets; the chief financial officer voiced concern that future compute contracts might not be payable [15].

3. A bipolar geopolitics: China’s open-weight move

The single most important factor breaking the closed, high-margin API monopoly the West has been trying to build is the asymmetric strategy of the China-based AI ecosystem.

US export restrictions on advanced AI accelerators forced Chinese researchers to focus on architectural efficiency rather than scaling models by brute force. By optimising the Mixture of Experts (MoE) architecture [16], DeepSeek reported training its 671-billion-parameter V3 model in roughly 2.8 million GPU hours on export-compliant H800 chips [17]. The $5.6 million figure the company reported covers only the final training run; prior experiments are not included [18]. The message is nonetheless clear: models with comparable benchmark scores can be released with open weights and very low inference costs.

Beijing does not find the slowdown calls from Western AI executives sincere. Responding to them, Chinese foreign ministry spokesperson Guo Jiakun said fear-mongering and confrontation would only disrupt global AI governance [19]. From the Chinese perspective, the discourse is an attempt by Western companies to protect their own high-cost monopoly and halt the spread of low-cost alternatives from the East.

What is interesting is that Washington rejected the same calls. The United States and China, in the middle of their strategic rivalry, separately turned down the appeals to slow down [20]. In other words, the demand for a brake comes not from the two superpowers’ governments but from the labs themselves.

4. The enterprise ROI gap and agents that think they are clever

The disappointment in the B2B market is putting direct pressure on AI companies’ revenue statements. MIT’s Project NANDA report found that despite $30–40 billion in enterprise investment, 95% of organisations had obtained no measurable return from generative AI [21]. Gartner, meanwhile, predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls [22].

[User goal: a small change]
        │
        ▼
[Autonomous agent decision chain (ReAct / CoT)]
        │
        ├─► Excess assumption generation (overthinking)
        ├─► Scope creep
        └─► Circular token consumption ($$$)
        │
        ▼
[Result: broken code, a high API bill, zero accountability]

For a business, working with a human employee who makes mistakes has clear rules: the hierarchy is known, responsibility is distributed, accounts can be demanded. By contrast, the risk cost of an agent that loses context and, rather than carrying out a simple instruction, sets about rewriting the codebase, can erase the speed advantage it offers. And perceived speed is not the same as real speed: in a randomised experiment with early-2025 tools, experienced developers finished their work 19% slower with AI — while believing they had been 20% faster [23]. METR later noted that this finding no longer reflects current tooling; the gap between perception and measurement remains striking. The wave of cancellations Gartner forecasts is one reflection of that disappointment [22].

5. Escaping the commitment trap, and Google’s “selling shovels” strategy

Market actors have taken two basic positions in the face of this impasse.

OpenAI: from “welcome to the AGI era” to the brake in nine days

OpenAI, which has kept market expectations alive by continually saying AGI is at the door, formalised that discourse in September 2026. At the end of August, Altman said the company expected to have an internal system it would call AGI by the end of the year [24]. On 3 September, at the launch of GPT-6 Astra, company president Greg Brockman said he personally believed AGI had been reached, and closed the press conference with “welcome to the AGI era” [25]. Three days later, Jensen Huang — CEO of Nvidia, maker of the chips that trained the model — also declared that AGI had arrived [26].

Nine days after that, the sector pulled the brake [1]. OpenAI had already suspended its largest training run following the Hugging Face incident [5]. The public sequence is nonetheless telling: the declaration that “we have reached the threshold” is followed immediately by the call that “now we must slow down to manage the risks.” This construction preserves the AGI claim as a valuation story while placing the physical and financial bottlenecks behind an ethical shield. Whether OpenAI’s data-centre commitments can be supported by revenue is already under discussion inside the company [15]. Tom’s Hardware’s analysis likewise notes that calling for regulation ahead of an IPO can be a way of slowing competitors down [26].

Google and infrastructure sovereignty

On the consumer side, users complain that Gemini loses context even in short conversations and forgets information from a few messages earlier [27]. That the behaviour of models offered as a service can change over time without notice to the user has been documented before [28]. Some users argue this stems from a decision to cut inference costs; that claim has not been independently verified [27].

Google’s real strategic gain is in the background. Just as in a gold rush most prospectors go bankrupt while the sellers of picks and shovels get rich, Google is renting the TPU chips it developed itself to its largest rivals. In October 2025, Anthropic announced a deal worth tens of billions of dollars providing access to up to a million TPUs [29]. While competitors burn cash to capture market share, a significant portion of the capital the ecosystem spends flows into Google’s cloud coffers.

Final analysis

The global debate conducted under the headings of “cybersecurity threats,” “ethical AI” and “controlled slowdown” is as much about the real economy as it is about futuristic existential anxiety. The techno-feudal lords [30] — colliding with a wall built out of inflated valuations, finite energy and water, inefficient agent architectures and China’s open-weight cost pressure — are trying to buy time by marketing that brake to society and the markets as an act of grace.

References

  1. NBC News. (2026, 12 September). Sam Altman backs Anthropic CEO’s call to slow down the global AI race. nbcnews.com
  2. Ghaffary, S., & Metz, R. (2026, 11 September). OpenAI considers slowing advanced AI development, Sam Altman tells employees. Bloomberg. bloomberg.com
  3. France 24. (2026, 23 September). AI leaders urge caution at UN, with Anthropic chief pledging to slow down. france24.com
  4. Altman, S. (2023, 16 May). Written testimony of Sam Altman, Chief Executive Officer, OpenAI [Senate testimony]. U.S. Senate Committee on the Judiciary, Subcommittee on Privacy, Technology, & the Law. judiciary.senate.gov
  5. OpenAI. (2026, 26 August). The Hugging Face incident and the road ahead. openai.com
  6. Hugging Face. (2026, 27 July). Anatomy of a frontier lab agent intrusion: A technical timeline of the July 2026 incident. huggingface.co
  7. Stigler, G. J. (1971). The theory of economic regulation. The Bell Journal of Economics and Management Science, 2(1), 3–21. doi.org/10.2307/3003160
  8. Gairola, S. (2026, 28 July). Andrew Ng backs Jensen Huang’s open AI push, says ‘closed models are safer’ PR is ‘regulatory capture’. Benzinga. benzinga.com
  9. International Energy Agency. (2025). Energy and AI. iea.org
  10. Kunkel, C. (2025, 30 July). Projected data center growth spurs PJM capacity prices by factor of 10. Institute for Energy Economics and Financial Analysis. ieefa.org
  11. Constellation Energy. (2024, 20 September). Constellation to launch Crane Clean Energy Center, restoring jobs and carbon-free power to the grid [Press release]. constellationenergy.com
  12. Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models. Communications of the ACM, 68, 54–63. arxiv.org/abs/2304.03271
  13. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). doi.org/10.1145/3442188.3445922
  14. Shukla, H. (2025). The AI scaling wall of diminishing returns. arXiv. arxiv.org/abs/2512.20264
  15. CNBC. (2026, 28 April). OpenAI’s revenue, growth estimates fall short as company races toward IPO: Report. cnbc.com
  16. DeepSeek-AI. (2024). DeepSeek-V2: A strong, economical, and efficient mixture-of-experts language model. arXiv. arxiv.org/abs/2405.04434
  17. DeepSeek-AI. (2024). DeepSeek-V3 technical report. arXiv. arxiv.org/abs/2412.19437
  18. Allen, G. C. (2025, 8 April). DeepSeek: A deep dive [Congressional testimony]. Center for Strategic and International Studies. csis.org
  19. CNBC. (2026, 14 September). China says AI CEOs’ call for a slowdown is ‘fear mongering’. cnbc.com
  20. Power, J. (2026, 23 September). As AI leaders warn of catastrophe, US and China shun slowdown calls. Al Jazeera. aljazeera.com
  21. Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025, July). The GenAI divide: State of AI in business 2025. MIT NANDA. mlq.ai — Turkish summary: Rapor Bülteni. (2026, 19 January). raporbulteni.substack.com
  22. Gartner. (2025, 25 June). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. gartner.com
  23. Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity. METR. metr.org
  24. Bastian, M. (2026, 26 August). Sam Altman says OpenAI will have AGI by the end of 2026 if you accept his definition. The Decoder. the-decoder.com
  25. Fried, I. (2026, 3 September). “Welcome to the AGI era,” OpenAI says as GPT-6 Astra debuts. Axios. axios.com
  26. Martindale, J. (2026, 9 September). OpenAI claims GPT-6 Astra is an ethereal ‘Alien Mind’ with AGI-like qualities. Tom’s Hardware. tomshardware.com
  27. Shapiro, S. (2026, 6 August). Is Google deliberately ‘dumbing down’ Gemini? Israel Hayom. israelhayom.com
  28. Chen, L., Zaharia, M., & Zou, J. (2023). How is ChatGPT’s behavior changing over time? arXiv. arxiv.org/abs/2307.09009
  29. Google Cloud. (2025, 23 October). Anthropic to expand use of Google Cloud TPUs and services [Press release]. PR Newswire. prnewswire.com
  30. Varoufakis, Y. (2023). Technofeudalism: What killed capitalism. Bodley Head.