tleitch.bsky.social

1/ New research note on the AI capex chain in credit. Over three months the single-name CDS tape repriced the builders one by one — Oracle 37→318 bp, Broadcom 31→156, Nvidia and Meta given CDS markets they didn’t have — before equities moved on Wednesday and the index two days later.

3/ Why the names moved first: the collateral. I built residual-value models and the insurance wrap for aircraft leases. Aircraft are prime: decades of data, a secondary market, maintenance reserves. GPUs are subprime: 3–5yr life set by the maker, no residual history, worthless unpowered.

1/ New research note on the AI capex chain in credit. Over three months the single-name CDS tape repriced the builders one by one — Oracle 37→318 bp, Broadcom 31→156, Nvidia and Meta given CDS markets they didn’t have — before equities moved on Wednesday and the index two days later.

We benchmarked local LLMs vs cloud APIs as part of SD-AI (github.com/UB-IAD/sd-ai), open-source AI tooling for System Dynamics modeling at UB-IAD. Local stack won. 91% vs 89% best single cloud model. But the energy finding is the real story. 🧵

At ~30% enterprise utilization: → H100 cluster: ~37 Wh/query → Mac Studio array: ~19 Wh/query GPU infra only wins above ~85% utilization. Most enterprises never get there. Nobody is doing the unit physics on this. We started.

We benchmarked local LLMs vs cloud APIs as part of SD-AI (github.com/UB-IAD/sd-ai), open-source AI tooling for System Dynamics modeling at UB-IAD. Local stack won. 91% vs 89% best single cloud model. But the energy finding is the real story. 🧵