← Back to archive

What If AGI Never Happens: GPU Limits, Memory, Scaling Costs

12 December 2025 · 39 views · 0 likes · 0 reposts · 132 blog views

AI Infrastructure chatgpt coding philosophy product

What if AGI never happens?

I just found an interesting article, and the author makes points that are actually pretty close to what I think too.

GPUs have basically stopped improving

After years of huge leaps (tensor cores, low-precision formats, HBM), we’ve exhausted the major tricks. Future GPU gains will be minimal and come with trade-offs.

And yeah, I know people point to the B200 with its 4x training and 30x inference numbers. But ... the B200 is basically 2 chips glued together because of the reticle limit. The H100 was already pulling 700W, and now the B200 is gulping around 1,200W. That 30x speedup also depends on new ultra-low-precision formats, which are basically mathematical shortcuts.

Memory is the bottleneck (not compute)

You can add FLOPS endlessly, but without faster and closer memory, they become useless. Transformers are already close to physically optimal designs.

Linear progress now requires exponential resources

This pattern appears across physics, hardware, AI scaling. Easy gains are gone and refinements become extremely expensive.

Scaling laws still work but costs exploded

We used to get linear capability improvements for roughly linear investments. Now scaling costs grow exponentially while benefits shrink. We may have 1–2 generations left before hitting physical infeasibility.

AGI is impossible. Superintelligence is a fantasy.

True AGI would require handling physical tasks and resource constraints. Most important robotics problems are already solved or economically irrelevant. The rest are impractical to scale. Recursive self-improvement ignores physical limits: computation, memory, manufacturing, energy, and diminishing returns. You cannot escape exponential resource needs.

Future AI progress will be slow, physical, and tied to economics. More like steady, practical adoption shaped by hardware limits, costs, and real-world value.

Source: timdettmers.com/2025/12/10/why…

View on X

What If AGI Never Happens: GPU Limits, Memory, Scaling Costs

Originally posted on X

PS: I'm pretty active on 𝕏 if you'd like to follow along

Related posts

Archive

2026

2025

2024

2018