The New York Times just gave AI training compute the visual essay treatment, but the real story is what's happening to the economics underneath the hype.
The Summary
- Epoch AI's training compute estimates got the NYT visual treatment, spotlighting how rapidly compute demand is reshaping tech infrastructure
- AI costs are dropping 13-fold annually even as total spending approaches $1 trillion, creating a paradox of plummeting unit costs and skyrocketing aggregate investment
- Copper and aluminum prices are retreating as investors question whether AI infrastructure spending can sustain current levels
- The divergence between falling per-unit costs and rising total expenditure signals either a massive buildout that will reshape compute infrastructure, or a bubble in the making
The Signal
Epoch AI's research on training compute just got mainstream attention through the New York Times, but the visualization obscures something more interesting than growth curves. While the Times focused on the exponential trajectory of compute demand, two other data points paint a stranger picture of where AI economics actually sit right now.
AI training costs are collapsing at a 13-fold annual rate, according to Epoch AI's own numbers. That's not a typo. The cost to train a model is dropping faster than solar panels did during their steepest decline. Meanwhile, total industry spending is approaching $1 trillion. Those two facts should not coexist in a rational market.
"The cost to train a model is dropping 13x annually while total spending approaches $1 trillion."
Either someone is building infrastructure at a scale that will fundamentally reshape computing, or a lot of capital is chasing efficiency gains that will evaporate the business case for all this investment. The distinction matters. In scenario one, we're witnessing the buildout of Web4's backbone. In scenario two, we're watching a pile of subsidized compute that will get commoditized into oblivion before anyone figures out how to monetize it.
The commodities markets are already hedging their bets. Copper and aluminum prices are pulling back as investors question whether AI infrastructure spending can maintain its current pace. That's the canary. When the inputs to building data centers start retreating, it means someone is worried about demand growth keeping up with supply expansion.
Key dynamics at play:
- Training compute demand growing exponentially
- Per-unit costs falling even faster
- Total capital deployed still climbing toward $1T
- Raw material prices softening on sustainability concerns
This creates a strange inversion. The cheaper it gets to train models, the more money gets spent doing it. That works until either the models stop getting meaningfully better, or someone figures out that most of this compute is building features no one will pay for. The metals market seems to be pricing in the latter scenario before the tech industry admits it's even possible.
The Implication
Watch what happens when training costs hit some threshold where the marginal improvement in model capability no longer justifies even the reduced expense. That's when the $1 trillion in infrastructure either becomes the backbone of an agent economy that actually runs on this stuff, or it becomes a massive write-down. The commodities traders are already positioning for volatility. If you're building in this space, the question isn't whether compute will be cheap enough. It's whether there will be enough demand to fill the capacity being built.
For builders: the window is now. Cheap compute is an arbitrage opportunity, but only if you're solving problems people will actually pay to automate. For investors: the divergence between falling unit costs and rising aggregate spending is unsustainable. Something has to give. Follow the metals.