What AI can do is set by the stack underneath it. Who it reaches, and who it puts at risk, is set by what we build on top.
I chose chemical engineering over sociology at 17 and never let the sociology interest go. This is where the two finally plug back together.
Model launches, benchmark scores, funding rounds. The conversation everyone else is already having. I don't add to it.
Seven layers of infrastructure below the application layer, and what each one costs you. Running now, one layer a week.
Two series on what changes once intelligence is cheap: The Rewired Organization for how companies run, The Larger Reckoning for everything else.
Running alongside all of it, because what's happening to people's jobs won't wait for chapter three.
The Capital StackOne standalone piece on who actually absorbs the loss if the build-out's financing assumptions break.
Pick the line closest to what you're actually asking. It reorders this page and tells you where that thread is running right now. Nothing gets hidden, and you can change it any time.
Weekly-verified historical pricing across flagship and budget model tiers. Watch the cost curve in real time.
Everyone talks about AI apps. Almost nobody talks about what makes them viable — the six infrastructure layers underneath: custom silicon, memory hierarchies, model architecture, optical interconnects, energy, and the decode stage that accounts for roughly 60% of every inference cost. This is the map.
SpaceX filed to launch one million orbital data center satellites. Google signed a $920M/month deal. Space compute is currently 3–4× more expensive than ground — but that may not be the point.
Enterprise AI runs at 5% GPU utilization, measured across 23,000 production clusters. The industry's answer is more data centers. A VSM analysis of all 8 pipeline stages and the optimizations that cut inference cost 50–70% without new hardware.
You can negotiate token prices, switch to a cheaper model, and still watch your AI margin get worse. What decides it is the architecture around the model: what context the workflow carries, how each step is routed, and what happens when an output is wrong.
How this gets published
Every piece publishes on this site first and stays here. Short posts three to five times a week point back to whichever one is new, so following there means you don't have to keep checking this page.