Software monetized while the builders paid: Thursday's only green layer was software (+1.9%), as semis (−3.0%) fell on doubts about AI-lab revenue and hyperscaler free cash flow printed −83% year over year. Rates are the headwind — 10-year real yield +46 bp in 20 days, borrowing costs on the rise.
Green means watching — we are tracking it and it looks normal. Amber means warning — it needs a closer look. Red means danger — time to act. The color never carries the meaning alone: every card says it in words.
So what? The AI buildout is running on borrowed money: Oracle, Broadcom and SpaceX all borrowed billions for AI chips this week while Big Tech's cash pile shrank 83%. If borrowing keeps getting pricier, the thing that slows AI spending won't be a chip shortage — it'll be the debt.
Big picture: borrowing costs are the headwind (day 1) — the 10-year real yield is up nearly half a point in 20 days; corporate borrowing spreads are widening too. Where the money is: software. It's beating the S&P 500 by 18.6 points over three months — the monetization phase of the cycle. Posture: no entry on file; a review is prompted on any regime or phase change.
What changed: (1) Two warning lights — Big Tech cash flow (free cash flow $6.7B, down 83% from last year) and borrowing costs (3.09%, up 0.38 points in 20 days). (2) Software +1.9% while semis −3.0% and hyperscalers −2.9% — a monetization-day tape. (3) It costs more to borrow in the US than in Canada by a widening margin; the Canadian dollar slipped to about 1.43 per US dollar.
Thesis: — (Anferny writes after reading.)
| Layer | 1D % | 5D % | 20D % | 3-mo RS vs S&P | % > 50-day | 1D z |
|---|---|---|---|---|---|---|
| US power (L1) | ▼ -3.02 | ▲ +5.43 | ▲ +2.57 | ▼ -13.31 | 62.5 | -1.26 |
| TSX power (L2) | ▼ -0.06 | ▲ +1.20 | ▼ -0.53 | ▼ -13.22 | 37.5 | -0.13 |
| Semis (L3) | ▼ -3.01 | ▼ -1.83 | ▲ +7.26 | ▼ -4.28 | 90.0 | -0.48 |
| Hyperscalers (L4) | ▼ -2.92 | ▼ -0.51 | ▲ +0.60 | ▼ -2.34 | 50.0 | -1.50 |
| Software (L5) | ▲ +1.85 | ▲ +0.89 | ▲ +1.61 | ▲ +18.56 | 50.0 | 0.39 |
| China tech (L6) | ▼ -2.47 | ▼ -1.62 | ▼ -4.80 | ▼ -17.87 | 0.0 | -0.95 |
| Physical (L1+L2) | ▼ -1.54 | ▲ +3.34 | ▲ +1.12 | ▼ -12.69 | 50.0 | -1.04 |
| US AI (L3+L4) | ▼ -2.96 | ▼ -1.17 | ▲ +3.96 | ▼ -2.93 | 75.0 | -0.78 |
| Spread | 20-day change (pts) | z20 | 1-year percentile |
|---|---|---|---|
| Physical vs Silicon | ▼ −5.9 | 0.23 | 5.2 |
| Software vs Semis | ▼ −5.4 | 0.21 | 18.7 |
| China tech vs US AI | ▼ −8.8 | 0.06 | 1.6 |
Software the only layer up on the day (+1.9%); China tech breadth sits at 0% above its 50-day. China tech vs US AI is at the 1.6th percentile of its 1-year range — near a 1-year low, though the move itself is not statistically unusual (z 0.1). Kill-switch: K2, K4 amber (see dashboard); K1 green with 0 cuts recorded — capex-guidance input pending from Anferny.
Triggered callouts: none fired. T2 z 0.23 (|z| < 1.5); T3: no confirmed China events, z20 0.06; T1 not wired in pilot; T4 n/a on the first issue.
News (headline, link and one-line tag only):
Earnings: none reported since the last issue. On the radar: ASML Oct 14, TSMC Q3 call Oct 15.
RVOL: EMA.TO (L2) 3.4× its 20-session average volume, −0.2% on the day — the only universe name at 2×+.
Gold finder: no C&C hits in the pilot — the primary-brief mapping for the v3.3 signal_hits table is still pending.
Ideas book: empty — awaiting Anferny's first ideas.
The daily lens from Competing in the Age of AI (Iansiti & Lakhani): in digital markets, value creation and value capture separate — the firms that create the value are not always the ones that keep it.
Today's tape is that separation made visible. Hyperscalers are creating enormous value — debt-funded data centers, with Oracle, Broadcom and even SpaceX tapping bond markets for AI chips — while capturing less of it: aggregate free cash flow fell 83% year over year as capex absorbed the cash. Chip stocks, meanwhile, now trade on AI labs' revenue rather than their own — creation without capture. The day's capture went to software (+1.9%): distribution and pricing power, the two-sided position the book argues wins. The question to carry into tomorrow: which bottleneck can't be multihomed — the electrons, or the interface?