daily notes

At 50 tokens/s/user P90 interactivity (excluding time to first token), vLLM reports 7,567 versus 6,083 total tokens/s/GPU for MI355X and B300. SemiAnalysis’s modeled TCO converts that to 18.16 million versus 9.69 million tokens per dollar (+87.4%). The full first-party account traces the result through cache reuse and offload, MoE, memory traffic, MLA/KDA scheduling, and speculative verification; it compares complete serving stacks, not isolated GPU kernels.

This 2019–26 history distinguishes deposit tokens, bearer instruments settled by on-chain ownership, from tokenized deposits, which remain bank-ledger claims and typically require interbank settlement. Compared with the Dallas Fed’s Aug 26 duration/liquidity analysis, the new mechanism is that instrument design changes the settlement path and regulatory perimeter, including deposit-insurance visibility and AML oversight. Large banks appear to favor tokenized-deposit networks, but implementation remains uneven.

updates 2026-08-26 · prior evidence ↗

Using 2025 SHED survey responses, the Kansas City Fed finds 37.0% of financially distressed consumers reported increased borrowing, versus 7.8% of other consumers. Among those who borrowed more, 50.4% of distressed consumers used alternative financial-services credit or overdrafts, versus 20.8% of non-distressed borrowers. These are observational, self-reported associations, not proof that inflation caused the borrowing.