6 Comments
User's avatar
The Optimist's avatar

I'd push the physical constraint one step further — Dario Amodei told Dwarkesh that being off by one year on revenue timing can be "ruinous," but the timeline for compute delivery runs at the speed of electrician apprenticeship completions and retirement curves, not venture capital deployment.

Veronica's avatar

The point about thermodynamic constraints being the actual bottleneck is genuinely underreported, and the framing of physical infrastructure as the real frontier is clarifying. I do wonder if calling ethics debates "largely noise" risks collapsing two different things together. The performative version is noisy. The quieter operational version, the one showing up in insurance exclusions you mentioned, is actually the market running its own ethics calculation in the absence of better frameworks. Those feel like different conversations.

Rock Howard's avatar

Slowing down? The ability of AI to produce massive code sets has exploded during the last month or so. Teams of agents can deliver working solutions in single shots. I have been totally gobsmacked by my own efforts. I'm also getting much better results in areas other than coding from the newly updated frontier models. From where I sit, AI has turned a huge corner already in 2026. I wouldn't be surprised if this moment is pointed to in future history books as the moment when the singularity began.

Max B's avatar

China already has 2x+ power generation today. So grid and infrastructure constraints is where US has to play catch up. And given that US haven't build anything worthwhile in past 40 years its unlikely they can do it fast.

But of course there is Elon Musk ... So that could be an x factor

Whether China has less compute overall is an open question. And its quite likely within 3 years they will achieve parity ( or surpass US) with domestically produced chips only

As of quality of their models. You must beem living under a rock. Their models are not in any way shape or form inferior to US models. Coding and multimodal capabilities of their recent models edge out strongest US models. For less price as well.

If you think opus 4.5 is the best ...well then lol. - you just repeating marketing propaganda

Jesse's avatar

Dave, you’ve identified the friction of 2026 as a lag between 'silicon and steel,' but the true bottleneck is systemic obsolescence. The friction exists because we are attempting to run a 21st-century intelligence on a 20th-century social battery.

​My Declaration of Universal Dignity and Economic Reciprocity is the turnkey solution to the 'Steel' problem. By mandating the strategic utilization of AI to cover labor gaps and prioritizing massive infrastructure rebuilding, this blueprint provides the physical and legal nervous system your 'friction' theory says is missing.

​Furthermore, by abolishing homelessness and hunger, we remove the societal resistance to automation. People only fight progress when it threatens their survival; my Declaration of Universal Dignity and Economic Reciprocity makes progress the guarantor of their survival.

​You’ve diagnosed the engine stall; I’ve provided the new chassis. It is time to move from discussing the friction to implementing the restoration. The blueprint is ready. Are you prepared to discuss the architecture of the solution?

Pawel Jozefiak's avatar

"75.8 gigawatts by 2026, up from 25 gigawatts in 2024" - the energy math is staggering. 130 nuclear plants worth of capacity in six years.

The 128-week transformer lead times and 5-7 year grid interconnection queues are the real bottlenecks. You're right that capital can't solve physics. Memory sold out through 2026 means even unlimited funding hits hardware walls.

What strikes me: the debates dominating AI discourse (safety, ethics, regulation) contribute "remarkably little actual friction." The constraints are mundane and physical. Data center cooling, chip packaging, DRAM supply chains.

I've been navigating similar infrastructure-vs-capability gaps in multi-model workflows: https://thoughts.jock.pl/p/multi-model-ai-workflow-2026-gpt-claude-gemini