In a recent Atlantic piece, correspondent Alex Reisner argues that the rapid deployment of large language models is contributing to a global shortage of high-end computer memory, driving up the cost of memory, storage, and affordable computers while straining the electricity supply. His broader claim is more provocative still: that generative AI, as currently built, represents an engineering failure rather than a triumph.

Below are key highlights from the Reisner’s article:

• The cost is landing on everyone. Reisner writes that models like ChatGPT and Claude are so resource-hungry that tech companies may be buying up to 70% of the world’s supply of high-end memory, causing a shortage. Prices are climbing as a result: hard drives he bought two years ago for $350 now cost $800 or are out of stock, some laptops have risen as much as 50%, and one forecast suggests affordable entry-level computers could “disappear by 2028.” The memory is flowing into data centers, which firms are expanding so fast that some are repurposing jet engines to power them.
• An engineering failure, not a triumph. Reisner’s central thesis is that modern language models are fundamentally inefficient. He argues the industry has chosen brute-force scaling, pouring in ever more capital and silicon, over the harder work of designing elegant, efficient systems.
• No economies of scale. Reisner contrasts LLMs with conventional software, where serving each additional user gets progressively cheaper. With generative AI, he argues, every increase in usage demands proportionally more compute, memory, electricity, and data-center capacity. As tech critic Gary Marcus put it in sharing the piece, by economic and engineering measures, generative AI “might be the worst technology ever deployed” on that score.
• The technical root: quadratic scaling. Reisner identifies a core flaw in the transformer architecture behind today’s models: the computation and memory they require grow disproportionately as the amount of text being processed increases. This “quadratic scaling,” he writes, is exactly the kind of characteristic computer scientists usually try to engineer away.
• Bigger, not better. Reisner argues that the faith in “scaling laws” has pushed companies to answer hard problems by building larger models rather than smarter ones. He notes that frontier models have reportedly ballooned from roughly 175 billion parameters in 2020 to well over a trillion today, even as each size increase yields smaller improvements and demands disproportionately more compute just to keep pace.
• A road not taken. Reisner contrasts today’s approach with earlier AI research, which tried to encode reasoning and human thought directly rather than learning by statistical imitation from massive datasets. That older tradition, he suggests, was far more computationally efficient, even if less capable. Meanwhile, he argues, work on smaller models and hybrid approaches receives far less investment and attention than the race to scale.
• It doesn’t have to be built this way. Reisner’s conclusion is that the industry’s current path is a choice, not a necessity, and that engineering effort should shift from simply expanding model size toward designing genuinely more efficient systems. The essay reportedly drew pushback online, with some AI researchers arguing it misreads why the field embraced LLMs and understates the real-world results they have delivered, but even critics largely engaged with its central question: whether today’s approach is the right one.

Reisner’s argument lands as a pointed challenge at a moment of soaring AI valuations: that scale and progress are not the same thing, and that the most important engineering work in AI may still lie ahead, in making these systems dramatically leaner rather than merely larger.

Acknowledgement: This summary was generated using AI writing tools as part of my experimental learning process. For more information, check out the original article listed below.

Related Resource Link

• Generative AI Is an Engineering Disaster | The Atlantic