Stock-market dispatch — 2026-07-16

Headless run, steps 0.5–7 in-turn. Research output, not advice — brain never trades. Prices are pre-open; marks are the 2026-07-15 close (a real move — the prior two dispatches carried 07-14 marks). Supabase sync is performed by the wrapper, not this run.

Top of mind

The AI-ROI bear case the 07-15 dispatch flagged for a human's eye largely dissolved on inspection — and it dissolved into two theses the book already owns. A gap-fill pass found both of Chamath's load-bearing numbers fail: token unit costs fell 67% YoY ($18.40→$6.07/M tokens) — the opposite of "doubling every 45 days"; bills rise because volume outruns a deflating price (Jevons). And "0–2% ROI" traces to MIT NANDA (Jul–Aug 2025 — 11 months old, narrow success definition, ~$250k memberships for the protocol it sells as the fix). Calibration filed. What actually explains the evidence: the seat is being replaced by the meter — Anthropic killed fixed-seat bundles for "a low base seat fee, plus full token consumption at API rates," so CFOs cancel licenses while token bills explode. Both facts print in the same quarter.

Then the day's best find corroborated it from an unexpected direction: IBM's 8-K — a filed, buyer-side statement — blamed its miss on clients shifting capex "towards servers storage and memory purchases to secure supply constrained infrastructure ahead of expected price increases." That single filing independently confirms two existing chains at once: the memory bottleneck (a buyer disclosing the shortage broke its own quarter — worth more than any seller's guidance) and seat erosion (SaaS is the raided budget line). IBM lost 25%/$67B on a $660M miss. agentic-ai-seat-erosion-to-saas-rerate Step 2 → confirmed, conviction → medium-high.

Strongest-conviction buys

Ranked shortlist (4c). Prices 2026-07-15 close, twelvedata (53/57; ASE/BESI free-tier plan-gated, HXSCL/GNKG OTC-carry).

⚠ Honest caveat on #5: today's evidence confirms seats are flattening but not that agents did it — Forrester/Microsoft/Uber cut licenses for cost reasons; IBM's reallocation is a memory shortage. Both flatten seats; neither proves agentic substitution. Recorded on the page.

Watchlist

Undervalued candidates (today)

• MU $904.28 — the day's sharpest divergence: the chain got stronger (IBM's buyer-side 8-K) while the stock fell 8%. 26% below the 06-16 mark, ~10× fwd P/E. ⚠ The bear tell is now the pull-forward — pre-buying ahead of price hikes borrows H2 demand.
• NOW $104.73 — −50% from high into a 6-day-away Q2 print, with the chain upgraded today.
• CCJ $90.98 / MP $49.46 — both −33%/−51% from highs with ~0% sector-excess: theme drawdowns, not selection failures.

New chains to investigate (hypothesis-stage)

None drafted. prospect-chains auto-drafted 0 (wiki densely connected — 62 open hypotheses; every groundable candidate duplicated an existing page). The genuinely-new veins today came from sources, not wiki-mining, and were filed as mechanisms instead (below).

New theses (now active)

• ibm-preannounce-to-enterprise-saas-capex-reallocation — DELL long (the only leg emitted to the feed); CRWD/PANW named but not emitted, IBM/NOW/CRM/WDAY the loser set. Medium-high (0.62). IBM's 8-K → capex reallocates to constrained memory → SaaS squeezed, hardware+security bid. ⚠ Only DELL reaches the feed because IBM's filing says "servers storage and memory", not security — the CRWD/PANW leg is Brown's inference (step 3, partial). ⚠ Live contradiction: the read-through didn't transmit on the day (IGV was green — Brown's rebuttal is that IGV is now mostly security by weight).
• monoline-mandate-moral-hazard-to-software-credit-concentration — software-heavy BDCs negative. Medium. Danieli's spiral: monoline mandate → concede price → covenants (85% cov-lite) → "documents become Swiss cheese" → concentrate into software → AI disrupts the collateral. Corollary that matters most: with covenants gone, the default rate is structurally suppressed — the KBRA 2.3%→3.5% series understates stress.
• insurance-illiquidity-rotation-to-private-credit-managers — BLK vs APO/KKR/BX. Medium. Long liabilities → illiquidity tolerance → GA rotation ($685B→$807B) → origination-capable managers capture a fee lift; NAIC re-prices the capital (CLO factors 2026-12-31; collateral-loan look-through 2027-12-31).
• taiwan-energy-blockade-to-fab-restart-cascade — Medium. Gelsinger: <3wk energy reserves → blockade browns out Taiwan → 90-day fab restart → shock, no shot fired; 7 blockades in 4 years. ⚠ Base rate is 0-for-7 — the exercises have stayed below this threshold.

Contradicted / weakened

• ai-roi-reckoning — both load-bearing numbers fail (see Top of mind). Survives as a pilot-implementation + rate-shock critique, not as evidence of demand roll-over. Calibration filed. ⚠ It also qualifies token-price-inflation-favors-asset-heavy-compute: that endpoint survives but its stated driver (price inflation) is contradicted by a 67% price decline — needs re-arguing on volume-outpacing-deflation.
• ⚠ INTC leg weakened by its own former CEO. Gelsinger: US leading-edge share rose 12%→18% while the TSMC:Intel wafer ratio widened 5:1 → 7:1. Both true ⇒ the CHIPS win accrued substantially to TSMC Arizona, not Intel. He says nothing about 18A/14A despite the episode being titled "What's Next" — the silence is the datapoint. The reshoring chain strengthens; INTC as the instrument does not.
• GS is no longer a dislocation. +12.3% since the last snapshot on its Q2 blowout, now 0.2% off the 52w high. Brown's own non-extrapolation warning: "Investment banking fees up 30% that's not going to be every quarter… Equities revenue up 86%… there's not going to be a SpaceX every quarter."
• bdc-redemption-spiral-to-private-credit-repricing contradicted by a practitioner — Danieli reclassifies gating as correct product behavior + a disclosure failure, not a credit event. An incumbent illiquid-fund manager is obliged to hold that view; recorded as an argued counter, not adopted.

Open questions worth a human's eye

• ⚠ Is medium over-labelled? SIX consecutive marks — and both flags widened. TJX −14.1% (from −13.2%) and PWR −12.3% (from −10.6%); high/med-high stay positive. Every live flag the project has ever produced lands in medium, and step-4e's high/med-high gate is what keeps preventing a calibration from firing — the gate may be mis-specified for exactly this case. Overdue for /reflect-on-me (7 CALIBRATION entries now — past the ~5 clustering threshold).
• Is the memory bid a level-shift or a pull-forward? IBM's filing confirms the shortage and warns of it: pre-buying ahead of price hikes borrows H2 demand. Directly sizes MU.
• Does NAIC's charge table differentially favor BLK over APO/KKR/BX? ⚠ My inference from the table vs BLK's disclosed high-grade/infra mandates; no source claims it. No conviction until researched.
• trucking-regulatory-capacity-removal-to-tl-carrier-rate-recovery is now the book's best /explore-chain candidate — its missing printed rate data arrived (LMI prices 96.0, a 10-yr record; capacity 30.8, 7 straight months of contraction) but the LMI attributes the price move to Hormuz/fuel, not drivers. We have the data; we lack the attribution. Separately, the FMCSA primary rule contains no driver-supply economics and quantifies only "tens of thousands" — the circulating 97%-of-200,000 figure is unsourced, and step 2's 250–400k is a carrier CEO's number.

What I looked at

• Run status: headless; steps 0.5–7 in-turn; price fetch paced 8/min to completion — 53/57, 0 rate errors (api.twelvedata.com reachable; failures are the known ASE/BESI plan-gating + HXSCL/GNKG OTC).
• Breadth (2a): ai-infrastructure 53% ⚠ over. Net-new steered to thin verticals: financials/private-credit (2 new mechanisms), transport ex-AI (bucket 11). Buckets: 9 (financials/rate regime), 11 (transport/industrials ex-AI).
• Ingestion veins (0.5–0.7): podcast 4 diarized ($0.62; budget $9.92/$50 July); earnings 1 (BLK Q2 — TSM's call is today 2pm ET, transcript catches next run; SCHW 07-21; rest deduped); feed 0 (all 4 active feeds fully deduped — newest SemiAnalysis 07-09 already ingested).
• Steps 1–2: 3 autoresearch (AI-ROI dispute; insurance→private-credit; CDL/ELP→truckload). 2b prospect-chains: 0 auto-drafted (healthy — wiki densely connected).
• Sources: 8 promoted (6 → stock-market, 2 multi-context → vault/sources/), 8 ingested → 4 new mechanisms, 4 person entities (pat-gelsinger, frank-danieli, larry-fink, martin-small), 5 mechanisms/concepts updated. Zero-novelty: 1 (Odd Lots soccer analytics — honest triage: out of scope; the title was a trap).
• Valuation (4b): 11 concept snapshots re-marked to 07-15 closes; malformed "Last refreshed" stamps (accumulated concatenation from prior runs) cleaned.
• PAPER-LEDGER weekly mark NOT due (last 07-10, next ~07-17); interim 4e note only.
• Live feedback (4e): 22 positions — 2 flags < −10% sector (TJX, PWR), both medium → 0 auto-calibrations. Best: DNN +12.8%, WST +8.2%, AMZN +6.8%.
• Calibrations: 1 (the AI-ROI double-failure; pattern: first-party-operator-claim-accepted-without-checking-the-number's-provenance — a first-party speaker is authoritative about their own book, not about an industry statistic they're repeating).
• SOURCE_RELIABILITY paid for itself: federalregister.gov was bot-blocked exactly as recorded, and the listed govinfo.gov mirror retrieved the FMCSA rule first try — which then refuted the circulating CDL figure. Added: spglobal/fmcsa/logisticsmgmt (403s); the-lmi + truckdispatchexperts (reliable).
• Signal feed (7): full active/armed set, 48 rows (46 → 48), validated 48/48 against the v3 contract. All 48 re-marked to 07-15 closes; as_of re-dated on only 10 (the theses that actually moved — the rest carry forward, which is what lets the trader see staleness). +2 rows: ibm-preannounce…:DELL (0.62) and insurance-illiquidity…:BLK (0.45). Convictions: NOW 0.50→0.62, BLK/db-to-dc 0.48→0.62, INTC 0.55→0.48 + 0.54→0.50; MU/TSM held (chain stronger, but pull-forward risk added); GS held 0.42 with undervalued→no. Deliberately not emitted (auditable, not silent): the Danieli moral-hazard chain (clears 0.42 but names no liquid tradeable), the Taiwan energy-blockade chain (its only instrument is INTC — a second row would double-count the same exposure), and CRWD/PANW (the security leg is Brown's inference; IBM's filing says "servers storage and memory", not security — only DELL is directly supported). Breadth after emit: ai-infrastructure holds 53% — DELL is honestly ai-infrastructure (its root forcing function is the memory shortage; cluster authored from what drives the chain, not to flatter the number). Supabase sync performed by the wrapper, not this run.