In short
Today's AI is shaped by contingent material conditions, not technological inevitability. Transformers dominate because they suit GPUs, not because they're optimal. US shale gas slashed electricity costs just as deep learning scaled, while post-2008 quantitative easing kept gas cheap and funded capital-intensive labs. The result is a sociotechnical lock-in around transformer-scaling that crowds out efficient alternatives like Mamba and neuromorphic chips.
If “What are Generative Futures?” makes the case that AI is contingent, this essay shows you the receipts.
It argues that the specific shape of today’s AI, transformer models scaled with enormous compute, was made possible by a particular alignment of material conditions: GPUs that happened to suit transformers, fracked shale gas that collapsed electricity prices, and a decade of cheap money that funded capital-intensive labs over efficient ones. None of these were features of intelligence itself. They were the economic and energy substrate the technology grew in.
That matters because the same forces that produced the boom can lock us into it. CUDA, the frameworks, the talent pipeline and the venture money all now point one way, much as fossil-fuel infrastructure does. But the existence of far more efficient alternatives (Mamba, neuromorphic chips, DeepSeek’s frugal training run) shows the lock-in is a choice, not a law of nature.
Key takeaways
- Neural nets won partly because they 'play well with hardware'; transformers suit GPU parallelism, especially after FlashAttention.
- US shale gas (2008–2012) collapsed energy prices exactly as deep learning scaled, making brute-force compute viable.
- Post-2008 quantitative easing kept cheap shale alive and funded capital-intensive scaling over efficiency research.
- CUDA, frameworks, hiring and VC all optimise for transformers, creating fossil-fuel-like sociotechnical lock-in.
- Alternatives exist: neuromorphic chips use ~1000x less energy; DeepSeek trained for ~$5.6M, scaling isn't inevitable.
Read the full piece
This is a summary. Read the complete essay, with all the sources and argument, on Substack.