The Atlantic · AI Watchdog

Generative AI Is an Engineering Disaster

A shockingly inefficient trillion-dollar project — large language models scale quadratically, not logarithmically, driving a global memory shortage and price spikes across computing.

By Alex Reisner · Published July 14, 2026 · Original article  ·  KG curated by kg-generator, rdf-infographic-skill, and Claude Sonnet 5 on behalf of Kingsley Idehen

Illustration for Generative AI Is an Engineering Disaster
Memory & Hardware Scarcity

AI Firms May Be Buying 70% of the World's High-End Memory

70%

Share of the world's high-end computer memory supply AI companies may be purchasing, per The Wall Street Journal.

$350 → $800

Price rise for hard drives over two years, now out of stock, per the article's own reporting.

+50%

Laptop price increases reported by Business Insider, hitting low-cost computers hardest.

Planned multiplication of total U.S. data-center capacity over the next few years, per CleanView.

Some forecasts suggest affordable entry-level computers may disappear by 2028, and the shortage is expected to last years. Demand for electricity at AI data centers is so great that some companies are repurposing jet engines to power them.

The Core Problem

Why Generative AI Does Not Scale

Efficient software scales logarithmically — added input costs proportionally less over time. Large language models instead scale quadratically: the bigger the input, the disproportionately more time and memory each additional word costs. Model sizes have grown from 175 billion parameters in 2020 to more than 1 trillion today, chasing diminishing returns under the industry's belief in scaling laws.

"Maybe with 10 gigawatts of compute, AI can figure out how to cure cancer."Sam Altman, OpenAI CEO
Industry Response

Vague Fixes vs. Genuinely Efficient Alternatives

Chatbot companies claim vague "compute multiplier" efficiency gains with no public evidence that quadratic scaling has been solved. Meanwhile, Microsoft researcher Alexia Jolicoeur-Martineau won a $50,000 prize for a tiny recursive model that solves logic problems without massive-scale training.

"It's a bit insane. At some point you have to learn to be a bit more efficient."— Alexia Jolicoeur-Martineau, Microsoft
"[Brute force] gives you a very low-risk way of investing your resources."Ilya Sutskever, co-founder & former chief scientist, OpenAI
Everywhere, All at Once

AI Is Being Bundled Into Everything

LLMs were integrated into Windows and macOS in 2024–2025, and added to Adobe Photoshop and Microsoft Word — raising the computing power needed for basic tasks even as Moore's Law slows due to molecular-scale limits on shrinking chip components.

Skepticism at the Top

Even AI's Own "Godfathers" Are Doubtful

"LLMs are not a path to superintelligence or even human-level intelligence."Yann LeCun, one of AI's "godfathers," to The New York Times

The article argues a near-religious conviction that something mindlike could arise from LLMs persists despite their inability to recall basic facts, lack of common sense, and dissimilarity to a biological brain.

Framework Applied

The Blitzscaling Lens on Generative AI's Inefficiency

Reid Hoffman and Chris Yeh's Blitzscaling framework — prioritize speed of growth over operational efficiency, on the assumption that efficiency can be retrofitted after scale is captured — maps onto frontier labs' brute-force scaling bet in ways that clarify what's actually happening, and one important way it doesn't.

The Brute-Force Bet Is a Blitzscaling Choice

Sutskever's framing of brute-force scaling as "low-risk" mirrors blitzscaling's core logic: prioritize speed-to-capability over efficient engineering, with scaling laws serving as ideological cover for unresolved inefficiency.

Externalized Cost — Onto the Whole Computing Market

Classic blitzscaling externalizes cost onto competitors or investors. Frontier labs are externalizing theirs onto a shared physical input market — the memory shortage and hardware price spikes documented above.

Rising, Not Falling, Marginal Cost

Blitzscaling works when scale reduces unit cost through network effects. Quadratic LLM scaling inverts that: each added parameter costs more, not less — the structural break this article documents.

Blitzscaling's Own Stated Failure Mode

Hoffman describes blitzscaling as "assembling a plane while falling." Vague compute-multiplier claims mean the trillion-dollar bet can't retrofit efficiency later if capital access tightens.

Reader's Guide

How to Evaluate Claims About AI Infrastructure Efficiency

  1. Check whether model size and scaling behavior are disclosed

    Undisclosed sizes, as with ChatGPT and Claude, make efficiency claims hard to verify.

  2. Distinguish logarithmic from quadratic cost growth

    Efficient systems scale logarithmically; unscalable systems like current LLMs scale quadratically.

  3. Treat vague "efficiency breakthrough" claims skeptically

    Absent public verification, assume the underlying quadratic-scaling problem is unresolved.

  4. Track component and hardware price trends

    Rising memory, storage, and laptop prices are a real-world proxy for AI's resource inefficiency.

  5. Look for smaller, task-specific alternatives

    Check whether a small, efficient model can perform the same task before assuming a massive LLM is necessary.

  6. Weigh AI demand growth against Moore's Law-limited hardware

    Chip improvements have slowed, so AI's exponential resource curve is increasingly unsustainable to meet with hardware gains alone.

FAQ

Frequently Asked Questions

Glossary

Key Terms

Large Language Model
A statistical language-generating AI model trained on huge quantities of text, such as ChatGPT or Claude.
Scaling Law
The industry belief that AI model performance reliably improves by making models bigger.
Quadratic Scaling
A growth pattern where resource use increases faster than input size, as seen in LLMs.
Logarithmic Scaling
An efficient growth pattern where added input requires proportionally less additional resource.
Moore's Law
The historical trend of microchips becoming faster, smaller, and cheaper, now slowed by molecular-scale limits.
Compute Multiplier
Dario Amodei's vaguely defined term for techniques claimed to improve AI model efficiency.
Tiny Recursive Model
A small, resource-efficient AI model that solves logic problems without massive-scale training.
AI Watchdog
The Atlantic's ongoing investigative series covering the generative-AI industry.
AI Bubble Economy
Concern that generative-AI valuations are inflated relative to unproven profitability and high operating cost.
Data Center
A facility housing computing infrastructure; U.S. capacity is being expanded roughly eightfold to serve AI demand.
Blitzscaling
Reid Hoffman and Chris Yeh's growth framework: prioritizing speed over efficiency in winner-take-most markets, assuming efficiency can be retrofitted after scale is captured.
About

About This Page

This knowledge graph companion was generated from The Atlantic's article "Generative AI Is an Engineering Disaster" by Alex Reisner. The source content was transformed into an RDF/Turtle Knowledge Graph using the kg-generator and rdf-infographic-skill AI Agent Skills, modeling the article, its author, quoted experts, organizations, key statistics, glossary terms, FAQ, and a reader's How-To guide as schema.org-aligned entities. This HTML infographic was then rendered from the validated RDF.

Technology Stack: