Matt Garman dropped a number at AWS re:Invent that should worry anyone betting on crypto miners' AI strategy. Inference now consumes two-thirds of all AI compute demand, up from one-third three years ago. That's not a gradual shift. That's the market rewriting itself.

For context: training is the expensive part, the months-long process of teaching models to think. Inference is what comes after, the actual thinking. The industry poured billions into training. Now the spending is moving downstream, toward inference. Simple math, but it breaks the business model that crypto miners were counting on.

The hardware mismatch

Companies like Core Scientific and Hut 8 spent the last year spinning up GPU capacity for AI training after Bitcoin margins got squeezed in the April 2024 halving. They repurposed data centers, signed deals, positioned themselves as AI infrastructure players. It made sense at the time. Training workloads are hungry, sustained, predictable revenue.

But Garman went further. He predicted 80 to 90 percent of enterprise AI value will eventually come from inference-powered agents, not chatbots but actual autonomous systems completing real work. AWS is already backing this with Trainium3 chips, purpose-built for inference instead of the training-optimized GPUs that have been printing money for Nvidia.

Here's the problem for miners: training and inference need completely different hardware. Training demands massive parallel processing across clusters running for months. Inference needs lower latency, higher throughput per query, often different chip architectures altogether. The gear that works for one doesn't work for the other.

What happens next

The "AI pivot" thesis that propped up mining stocks was always fragile. It worked as long as AI companies were burning cash on training. But if Garman is right, and he usually is, the demand curve is bending hard toward inference. That means the miners who invested in training infrastructure are holding the wrong tools. The ones who adapt fastest will survive. The ones who don't will be stuck with expensive capacity nobody wants to rent.

Watch for announcements. Which miners pivot to inference partnerships. Which ones quietly write down their AI assets. That's where the real story is.

This is informational material only, not financial advice. Crypto markets and AI infrastructure demand remain highly volatile and speculative.