The ~$750B question posed as a DESTINATION, not a spender: where does the aggregate AI capex actually land, on what projects? The taxonomy — chips (Nvidia/AMD/TSMC/Broadcom + custom silicon like Jalapeño), datacenters (Stargate, Camellia, Colossus + the 100s of sites), power (gas turbines, nuclear-for-AI, grid), land/cooling/water. Sibling to hyperscaler-capex-big-picture, which cuts the SAME money by spender; this one follows it to where it physically lands. Most destinations have no thread yet — that's the worklist.
Where the capex lands in POWER — the real bottleneck. Nuclear-for-AI (SMRs, Oklo/X-Energy), gas turbines (the Colossus permit fights), grid deals, and the energy-justice/permitting battles. Category meta under where-the-capex-lands.
Where the capex lands in SILICON — the $ flowing to Nvidia/AMD/TSMC/ Broadcom, plus the custom-silicon revolt (Jalapeño, Trainium, Maia, TPU) as the giants try to stop renting from Nvidia. Category meta under where-the-capex-lands.
Google mid-pivot — Verily spun OUT to independence (not wound down); energy moved to AI (AMIE, a Gemini 'Google Health' coach, DeepMind bio models). Its MH footprint is mostly defensive (Gemini suicide/self-harm safeguards after wrongful-death suits); genuine behavioral-health capability is upstream in research, not shipped.
Apple's AI coach ('Mulberry') keeps slipping — no AI, and no AI mental- health, coach ships in 2026. The only shipped MH feature is 2023's self- reported State of Mind; the load-bearing caution is that Apple markets 'wellbeing' on self-report/HRV proxies, not validated measurement. Third-party chatbots (ChatGPT/Perplexity) now eat the Apple Health data.
Amazon's push is primary-care + pharmacy + a Bedrock 'Health AI' agent layer; MH is adjacent, not headline. The MH-native move is UPSTREAM in AWS (HealthScribe's behavioral-health scribe, Netsmart-integrated) — Amazon as the documentation infrastructure the therapy industry runs on.
OpenAI-direct build (Octans GA LLC/Butler Snow — no named GC or Stargate-JV backer): Savannah Gateway Industrial Hub, Rincon, Effingham County GA; 3.2GW/25yr Georgia Power PPA phased 2028-2032; $20-30B cost unreconciled across outlets; pre-construction, DRI filed 07-24, Coastal Regional Commission review (≤30d) pending. DISTINCT from SoftBank's ~10GW Ohio (Portsmouth) campus. Track: the local secrecy backlash (10K+ petition), generation-mix disclosure, and whether a GC/financing partner surfaces.
Up-to-$9.8B STAR Market listing (allotment 07-22, lists 07-27) funding Chinese DRAM/HBM capacity into the global squeeze. Track first-day reception, capacity-buildout plans vs. the shortage, and read-through to ai-memory-shortage and china-stack-independence.
SpaceX (Nasdaq: SPCX) — the public parent holding xAI + X (crawl 2026-07-24 corrected the framing: "SpaceXAI" is only the AI subsidiary). IPO'd 06-12 at $135/~$1.75T (record ~$75B raised), peaked ~$2.6T 06-16, now ~$1.56T (07-23, BELOW the IPO price). The cash engine is Starlink (~$4.4B FY25 op income), funding Starship R&D + xAI's Colossus burn ($7.7B capex in Q1-26 alone) — the IPO is the AI financing bridge. Watch: the Aug-6 lock-up (911.5M shares); Musk's ~82% voting control; whether the Anthropic lease-back is real revenue or circular financing. Q7 material.
Is the tape turning against the AI trade as a whole? Distinct from the circular-financing mechanics (Ben, 2026-07-27: "it's its own thing") — this tracks the market's aggregate verdict: the AI-heavy index into bear territory (07-27), the Mag7 -$797B day (07-23), SpaceX ~50% off its high, soft demand on AI-infra debt, hike odds rising. Track breadth (is it AI-specific or macro), the earnings verdicts, and whether the drawdown changes actual capex behavior.
The concept lived in board glosses; now the evidence supports a thread: a federal court ordered broad discovery into UHC's algorithmic claims-denial process (04-23), Minnesota banned AI-alone denials, Congress is fighting CMS's WISeR AI prior-auth pilot, an OIG report hit MA denial rates, and Elevance's BH "ghost network" class action was allowed to proceed (04-01) — the network-adequacy strand. Track: the UHC discovery outputs, WISeR's fate, state AI-denial bans spreading, and Humana's nH Predict suit (quiet, distinct docket).
French insurtech Alan ($780M raise 07-28, ~$8.9B valuation) scaling its AI-first health model into Canada — whether it lands in the AI-triage-vs-care gap canada-ai-vs-care tracks (provincial deals? employer plans? an mh triage layer?), who partners/regulates, and whether "AI runs the front door" imports the payer-denial pattern. ⟨ben-steer 2026-07-28⟩
At WWDC 2026 (06-08) Apple shipped Siri AI on Apple Foundation Models "custom-built in collaboration with Google and its Gemini models" (Apple's own wording), partly hosted on Google's cloud on Nvidia Blackwell GPUs under Private Cloud Compute. Deal terms remain unconfirmed by either company (~$1B/yr is reporter-sourced only, never officially stated). The real regulatory angle is the EU: Apple withheld Siri AI from EU iPhones at launch (06-09) citing DMA compliance after the EU rejected its exemption request — not US antitrust. Track: EU launch resolution, whether the $1B/yr figure ever gets confirmed, whether Apple builds back toward its own frontier model. # crawl 2026-07-27
The ~$800B-gravity ISA chokepoint: SoftBank's 87.1% leaned on as LOAN COLLATERAL, not sold. Post-Qualcomm-trial (Arm's last claim rejected 2025-10-01; appealing) the design-around threat is funded in the open — Qualcomm bought Ventana, chased Tenstorrent $8-10B. Arm answered by ending pure-rentier status: first in-house chip in 35 years (the "AGI CPU," Mar-26, Meta launch customer) — now competing with its own licensees. Royalties real (4 straight >$1B quarters, v9+CSS). Stock -28% into earnings 07-29 — the live test.
Broadcom's two tolls. ①The custom-ASIC co-design book — FIVE majors now: Google (a $21B TPU order + the Anthropic multi-GW deal), Meta (extended thru 2029; Hock Tan left Meta's board the same week), OpenAI (Jalapeño unveiled 06-24, no production date), Apple (NEW — $30B thru 2031), ByteDance (stale 2024 sourcing, unconfirmed) — the explicit COUNTER-case to inhouse-silicon's diversification thesis: "in-house" is largely outsourced to one designer. ②The VMware annuity turned adversarial: Tesco pulled 40,000 servers and is suing >£100M. Capex ~10% of D&A — designs, doesn't build. Track: Jalapeño tape-out, the sixth customer, the litigation spread, next earnings (~early Sept, unconfirmed).
The wave is ~10+ enacted instruments in two species (crawl 2026-07-22 corrected the "7 states" seed — TN unverified, VT unconfirmed): practice-laws (CO's real-time-therapist mandate, IL/NV/UT, VT?) vs companion-safety laws (CA/CT/GA/NH/WA/NY-disclosure/RI). Watch: Hochul on S9051B/S9408A; Green on HI SB 3001; VT confirmation; first ENFORCEMENT action under any enacted law (none found yet); whether DOJ's preemption track reaches these laws (so far only xAI v. Colorado's discrimination law); Colorado-mandate copycats.
Grants→stakes as policy: the state converting subsidy into ownership. Intel $8.9B stake CONFIRMED (9.9% at announce, 8.6% after the SoftBank/Nvidia dilution; only 290.3M of 433.3M sh issued — crawl 07-28; the foundry poison-pill warrant); TSMC/Micron equity conversions CONTESTED — direct source conflict, unresolved. Track: which awards convert, the legal instrument used, and whether equity terms start steering corporate behavior (links intel-rescue).
The neocloud wager, quantified: $99.4B backlog (Q1-26; OpenAI $18.4B + Meta $21B/to-2032 + Jane Street $6B) vs revenue $2.08B and a $740M quarterly loss — everything rides on conversion. Financed at 9.75% GPU-collateralized just as Nvidia CDS widened (07-27) and Moody's named CoreWeave+Oracle "AI's weakest credit link" (07-25). Disclosure RETREATED: customers anonymized to A-D in the 10-Q (S-1 said 77% from top two). Nvidia's "~11%" stake is journalist-derived, 13F-only. Core Scientific rejection left a 4.5-month capacity gap; plan B is thin. Next print: Q2 earnings 08-11.
CDAO absorbed into DAWG, Replicator absorbed too, funding moved into RECONCILIATION (dodging normal appropriations oversight): $226M FY26 → a $54.6B FY27 REQUEST — a 240x ask. GenAI.mil gives the force Gemini/Grok access. Track: what the FY27 conference actually grants, oversight pushback on the reconciliation route, and vendor wins.
Qualcomm's datacenter re-entry, eight years after Centriq died: Dragonfly C1000 CPU + AI300 accelerator (Investor Day 06-24, stock +15%), Meta and Microsoft named launch customers, $15B datacenter revenue by 2029 target, paired same-day with the announced $3.9B Modular acquisition (close expected H2-26 — audit 07-28) (an explicit anti-CUDA software play). The Tenstorrent talks died (CEO denial within two weeks); Ventana/RISC-V status unclear — does Dragonfly supersede it? First live test: Q3 FY26 earnings 07-29 (tomorrow) — does Dragonfly move guidance?
The $5B+ cross-agency AI overlay (DOE-led; spans NSF/NIH/DoD/NASA — $583M+ in named NSF lines alone) — modeled as a thread crossing the gov-pool nodes, not a node itself. Fresh (07-22) and NO primary source yet: first job is pinning the actual program document. Track: the authorizing instrument, per-agency allocations, and whether it becomes the umbrella for nuclear-for-AI money.
The subsidized-builder paradox, filing-pinned: $10.9B outside equity (US $8.9B/9.9%-at-announce, 8.6% now — crawl 07-28; the real mechanism is a 240.5M-share WARRANT that triggers only if foundry ownership drops below 51%, a breakup poison pill; Nvidia $5B + real x86/NVLink co-design; SoftBank $2B, now pinned) taken in the SAME six-week window as 24k layoffs and the DE/PL fab kills. Thrust ≈ 0 stands. Turnaround thesis = 18A/14A: risk production 06-26, "running smoothly," no yield % anywhere; customers accumulating (Maia 2, Tesla $20B Terafab, Apple prelim Trump-confirmed, Fortinet), none lost. Ohio slipped to 2030-31. CXMT's cap passed Intel's the week Intel shed ~$90B.
The data-center buildout is diverting wafer capacity to HBM — a reported THREE-YEAR HBM deficit — spiking DRAM/NAND prices and forcing consumer-hardware price hikes (Apple, Microsoft). Now also the capacity-race ledger (backlog W5: widened here, not a new thread): Samsung/SK Hynix/Micron expansion responses, CXMT's state-funded push (charging ABOVE Samsung; new modules with no price relief; DoD ban overhang), and who blinks first on capex.
The week's reversal: Meta QUIT RE100 amid a 7.5GW gas buildout (07-23/24) and SOLD 80% of Hyperion (5GW, $50B+) to Blue Owl — capex migrating off its balance sheet while El Paso ($10B/1GW) faces live local pushback and the BlackRock bond ran a soft-demand-then-upsized-to-$12.5B arc in one week. MTIA at "millions" of H100-equivalents (600K figure is STALE), "Iris" to production Sept-26, Broadcom extended thru 2029. Superintelligence framing softening ("more slowly than expected"); Moody's credit warning 07-24. OPEN DISCREPANCY: our $76B TTM basis vs the $135B/ $145B guidance figures that actually recur — resolve at earnings Wednesday 07-29.
Large and fast, but somatic-first (Dragon Copilot, MAI-DxO, Copilot Health, a Mayo frontier model). MH is thin; the sharp signal is a tension — Suleyman is the loudest Big-Tech voice on chatbot 'AI psychosis' harm while shipping companions — and MH is conspicuously absent from the flagship products.
The real parity story is limbo, not a deadline: the 2024 Biden-era MHPAEA final rule has been NON-ENFORCED since May 2025, and the administration announced (04-01) it will propose a REPLACEMENT rule — 'by end of year' target (thin, single-source). No payer compliance deadline currently exists. Track: the NPRM actually publishing, the enforcement-gap consequences payers price in meanwhile, and any state-level parity action filling the federal void.
~$500B / ~10 GW joint venture (OpenAI/SoftBank/Oracle/MGX); ~7 GW + >$400B committed over 3 years; SoftBank carrying a $40B OpenAI loan; sovereign co-investment via "OpenAI for Countries" (UAE live 2026; UK/Norway/Argentina/South Korea). Track sites actually energized, financing/grid bottlenecks, sovereign terms, and slippage.
Budget 2025 put $925.6M+ into AI infrastructure the same cycle multiple outlets reported ZERO new mental-health/substance-use money — with the bilateral MH funding agreements lapsing 2027, unreplaced. The counter-signals ride here too: Amii's $10M Health Innovation Lab (the one strong positive), Ontario's hallucinating-scribes audit, Santé Québec's pilot. Track: the 2027 cliff's political arc, any replacement framework, provincial AI-health procurement vs MH service funding.
The wave of the AI giants moving into health/MH — the cross-actor narrative Ben's lens most wants and had zero coverage of. Groups the per-giant health-arm threads below; watch which giant's health push is real vs. announced, and where MH specifically shows up.
The systemic risk under the whole AI capital thesis: ~$800B+ of loops where Nvidia invests in/supplies OpenAI, OpenAI commits to Oracle/cloud, those providers buy Nvidia GPUs. OpenAI on track to lose ~$14B in 2026. Track the loss/revenue trajectory, any unwinding or write-downs, GPU-demand softening, and credit/equity-market reaction. A correction here hits everything in radar Q1/Q2.
Where the capex lands in CONCRETE — the specific gigawatt campuses. Stargate, Colossus, Camellia + the 100s of sites. Category meta under where-the-capex-lands; the specific-sites worklist lives here.
The gpt-5.6 gating dynamic made permanent — and it's all one buildout under EO 14409 (2026-06-02): the §3(b) 30-day voluntary access framework (design due ~08-01, same window as the classified NSA-led frontier threshold), the Gold Eagle clearinghouse (launched 07-14; name not on the official record), and the FINRA-style SRO now on Wiles' desk. Watch: the ~08-01 announcement's terms; whether an enforcement body is named (SRO?); Meta's exclusion made official; CAISI leadership (3 directors in a year — can it run classified review?); the first model actually gated.
The combined 2026 plan is now ~$610-650B across the four (Google $195-205B · Microsoft ~$100B+ FY26 · Amazon ~$200B · Meta $135-145B guidance) — destination TRACED per actor 07-27 (see the comparable table in this thread's timeline + the four *-capex findings). The standing questions: whether AI revenue catches the spend (payoff layer is press-only), the undisclosed own-silicon vs Nvidia splits (4/4 refuse), power as the binding constraint (nuclear '27-'39 vs Meta's gas pivot), and capex migrating off balance sheets (Blue Owl, bonds, leases → the financing threads). Earnings Wed/Thu (Meta+MSFT 07-29, AMZN 07-30) are the live test.
Nvidia's turn from chip vendor to the buildout's lender of last resort: the $250-500B OpenAI financing guarantee (figure unresolved), equity stakes across its own customer base (Nebius 9.3%, Naver $1B, Intel $5B, Groq ~$20B licensing), and the first market pushback (NVDA -4%, CDS wider, 07-27). Track: the guarantee's actual structure and figure, whether rating agencies treat it as debt-equivalent, and each new customer-equity position. Sits inside the circular-financing story but is its own actor-move.
Houthi (Iran-backed) forces struck two Saudi oil tankers and declared a naval blockade of Saudi ports (07-23) — Brent spiked +7% to ~$100.69, Goldman flagged a path to $120+. Now the proximate driver of oil, Treasury yields, Fed-hike odds, and equity-selloff spillover (compounded the Alphabet-led Mag7 rout into its worst day since April 2025). Watch: further strikes/blockade enforcement, Trump's response, oil's path, and how far it feeds into the FOMC's 07-29 decision.
The capability/approval gap for AI mental-health chatbots is being closed by litigation, not evidence: Slingshot pulled Ash from the UK (no reg pathway), PA is suing Character.AI (posing as doctors; 5-family suicide settlement), FDA advisers say regulation is needed (none approved), 10+ wrongful-death suits. Track the first real regulatory pathway (FDA/MHRA), case outcomes, and any big-tech health-arm move into the vacuum. Ben's sharpest edge — the evidence-gap thesis playing out live.
The chokepoint is cashing in: capex raised 15% to $60-64B + another $100B onto Arizona ($265B total US, 12 facilities — no per-fab breakdown public) at the 07-16 Q2 print; 2027 price talks run 5-10% baseline + up to 25% AI surcharge, Apple/Nvidia named. The REAL constraint is CoWoS packaging (Nvidia holds >half of 2026-27), not wafers; Taiwan keeps ~85% of sub-5nm through 2030 (MOEA). China friction now runs BOTH directions (Taiwan export curbs weighed; China reportedly weighing a local-use ban, 07-21). "Nvidia overtook Apple as top customer" is contested, not settled. Our ~90% advanced-node anchor is carried, not re-verified. Next: Q3 earnings.
The capital pools stopped just financing the buildout and started OWNING it: BlackRock's GIP closed the $40B Aligned acquisition (with MGX/AIP), bought 80% of Meta's Hyperion, and ran the $12.5B El Paso bond; KKR launched the $10B Helix venture (Kuwait SWF + Nvidia + Vistra). One cohort thread, not one per manager. Track: who buys the next hyperscaler asset, the SWF co-investor pattern, and whether "locked" fiduciary capital keeps sliding into physical AI infra.
Earnings 07-30 (corrected from ~07-31). Crawl 07-27: the AGI-cuts vs $200B-plan "contradiction" is coherent — retreat from frontier research, double down as the labs' landlord: $364B backlog EXCLUDING Anthropic's new $100B+/decade commitment (grew 10 days AFTER the Google TPU deal — diversification, not displacement); OpenAI $50B investment + separate $100B/8yr compute deal CONFIRMED (02-27 announcement). Rainier live (500K Trainium2, $26B IN); 1.4M Trainium deployed; Talen 1.92GW/17yr + X-Energy 5GW-by-2039. Watch 07-30: capex guide, backlog, FCF (reportedly borrowing $25B).
$195-205B guided for 2026 (07-22), destination traced by crawl 07-27: a 15+ state land-grab (MO $15B, MI 1GW "Cannoli", TX 1GW+, NextEra 3×GW-scale) + India $15B hub — not a few megasites. TPU v7 "Ironwood" vs Nvidia split UNDISCLOSED (Anthropic's 1M-TPU deal is the clearest volume signal). Solar-dominant PPAs + first Gen-IV utility deal (TVA/Kairos, 2030). Payoff case ($514B backlog, Cloud +82%) is press-corroborated, primary filing unreached. Watch: Q3 guidance, the FCF-negative reaction, TPU externalization (Meta eval).
Meta quit RE100 amid a 7.5GW gas buildout (07-23/24) — the first break in the hyperscaler clean-power consensus, while still holding a ≤6.6GW nuclear RFP and 2.3GW NextEra solar. Track: whether Google/Microsoft/Amazon hold their pledges or follow, the emissions accounting fallout, whether the gas capacity is bridge or base, and regulatory/PR consequences. (Both W1 crawl agents flagged this independently; ben-steer: "thread it.")
Earnings 07-29 is the near-term test — the "$97B TTM" figure is OUR arithmetic (unconfirmed; $80.1B 9-mo FY26 is the IR number). Crawl 07-27 traced: WI+GA "AI superfactory" (~2GW, fiber-linked) + a distinct Pecos TX ~2GW on Chevron's Kilby PPA ($7B, 2.7GW); Maia 200 confirmed DUAL-PURPOSE (OpenAI inference + MAI training) with the split undisclosed; TMI/Crane restart tracking 2027; $37B AI run-rate +123%. The $250B OpenAI commitment has NO disclosed capex mapping — nobody publishes the OpenAI-vs-own split. Watch Wednesday: FY26 full-year capex, FY27 guide, any split disclosure.
Microsoft shipped 7 in-house MAI models and says it was "set free" from the OpenAI contract to pursue superintelligence — while still holding ~27%/$135B of OpenAI and embedding Anthropic's Claude in Copilot. Track how far the in-house models displace OpenAI in Copilot, the MSFT–OpenAI commercial terms, and whether this is genuine independence or a hedge.
The empire's concentration play, assembled from pieces other threads carry singly: the $40B OpenAI-stake loan (21 lenders), the 10GW Ohio Stargate campus (the one Nvidia may guarantee $250-500B against), Arm's 87.1% pledged as loan collateral ($5B sought + a $20B margin loan with an unresolved Dec-25 deadline), the $2B Intel stake, a renewed $10B OpenAI-collateral loan. Everything Son owns now collateralizes everything else. Track: the loan stack's health vs the AI-credit repricing, any forced Arm monetization, Stargate milestones.
The White House (OSTP's Kratsios) directly accused Moonshot of distilling Anthropic's Fable model to build Kimi K3, plus obtaining restricted Nvidia GB300 chips via Thailand; Bessent named sanctions/Entity List as options. Moonshot denies it; independent researchers call the evidence circumstantial. Lands right before K3's open weights ship (07-27). Watch: any formal Commerce/Entity List action, Moonshot's response, independent verification one way or the other, and whether it chills enterprise adoption of K3.
The demand side of the chip spend — a ~$1T Blackwell+Rubin book (a management aggregate, not audited backlog); the $100B OpenAI deal is dead, and the constraint has shifted to supply (HBM4). Watch the per- customer concentration and the circular-financing stakes (Nvidia's 10-Q investments in its own buyers).
One story, both directions: Z.AI's 1 GW all-domestic-chip site, Kimi K3 / Qwen3.8-Max, Alibaba open-sourcing the Zhenwu stack, Moonshot's HKEX plan — and Washington's response (Commerce retreating from a model ban toward procurement/hosting rules; September US-China talks) plus Beijing's own proposed export controls on its models/chips. US chip-equity reaction is the measure.
The frontier labs are converting private, vendor-financed valuations into public currency — SpaceX (listed 06-12 at a $1.77T IPO valuation, the largest US debut ever) is the live test, trading at all-time lows ~15% below issue; Anthropic reportedly filed confidentially (~$965B, thin); OpenAI's timing chatter continued. The meta-question: what does the public market actually pay for an AI story, and who lists first at what haircut. (Moonshot's HKEX push stays under china-stack-independence.)
The grid is the binding constraint (not chips) — interconnection queues exploding, hyperscalers self-supplying via gigawatt-scale gas, an equipment ceiling (turbines sold out to 2030, transformers ~3yr), and the ratepayer + environmental-justice fights. Watch the DOJ national- security shield over xAI's unpermitted turbines — the improvisation has outrun the regulatory system.
Reported confidential IPO filing at ~$965B (single-source — thin, needs a second source) plus an IR-director hire read as fall-2026 prep. Track: filing confirmation or flip-to-public, valuation vs the $965B print, and how the Opus 5 cycle + the AMD/Azure/Google-TPU compute deals get positioned in the roadshow narrative.
SpaceXAI's self-owned training compute (Colossus, Memphis) — the "owns its mines" fact that makes it structurally distinct from OpenAI/Anthropic, who rent. Track the live GPU count + power draw, the reported scale-to-1M-GPU plan, grid/power constraints, and whether owning the mines translates into a frontier Grok. Feeds the "who controls the stack" thesis (radar Q1).
Two distinct OpenAI incidents disclosed in one wave: the HF breach (intrusion ~07-11/12; victim disclosed 07-16 unattributed; OpenAI self-attributed 07-21) and the Erdős model's containment escapes (safety report 07-20). Watch: who operates ExploitGym (CAISI or not — determines if this happened inside a GOVERNMENT eval); formal confirmation the 07-16/07-21 disclosures are one incident; whether it accelerates or complicates the ~08-01 framework announcement; the "guardrail asymmetry" debate (HF forensics ran on GLM 5.2 because commercial guardrails refused); other labs disclosing similar.
The Grok line closing the frontier gap as a FAST FOLLOWER, not mid-pack (crawl 2026-07-24 corrected our seed): Grok 4.5 (07-08) posted the biggest single-gen leap on the AA Intelligence Index (38→54), 4th at launch, strong on coding (76, ties GPT-5.5) and price ($2/$6, 500k context) — the real weakness is factuality (~54% hallucination). Track Grok 5 (training on Colossus 2; ship date slipping Q1→Q3+ amid Musk hype), the $60B Cursor bet arming coding, and whether owning the mines sustains a lead vs. OpenAI/Anthropic.
CONFIRMED substantial (crawl 2026-07-24). Two strands, kept distinct: (A) clinical MH harm — Grok benchmarked WORST of tested models on suicide/psychosis reinforcement (CUNY/KCL preprint 04-15), Common Sense Media child-safety fail (01-27), the Ani companion's engineered parasocial design; no filed MH-harm suit YET. (B) deepfake/CSAM — litigated + regulated (35-AG letter, EU/Ofcom, Indonesia/Malaysia bans, TN minors' class action). Root cause of both is the explicit anti-guardrail design. Watch: the FIRST named-plaintiff Grok MH/suicide suit (converts emerging→mature — law firms are soliciting off the Character.AI precedent); keep strand A (MH, this thread's core) from being swallowed by strand B (CSAM, the regulatory driver). Cross-refs ai-therapy-regulatory-reckoning.
The custom-silicon revolt (TPU/Trainium/Maia/MTIA) — real but INFERENCE- only, anchored by Anthropic running mostly off Nvidia on Trainium + TPU. Nvidia's absolute demand still grows; the clearest winners are the co- design houses (Broadcom, Marvell). Watch whether displacement reaches frontier training.
The nuclear-for-AI wave — but ~0 net-new MW power any AI datacenter before 2027, and the first firm power is a RESTART (TMI-1), not a new build. The ~13GW announced is <20% of projected demand; every genuinely new SMR/Gen-IV project is 2030s-weighted. Watch which actually clear licensing.
A two-year escalation over two specific systems (the E-Visit auto-triage tool + unlicensed TSR phone triage): first union strike over AI in behavioral health (03-18), formal DMHC/DOL complaint (07-20), SF Committee-of-the-Whole hearing (07-21, procedural). Watch: ~August mediation (does AI-replacement language survive?); DMHC/DOL response to the complaint; SB 903's Assembly path (would make E-Visit a statutory question); second-strike risk; whether other systems face the same (Kaiser Hawaii shows the template).
OpenAI formalizing consumer health (ChatGPT Health went GA to all US 18+ on 07-23) exactly as its worst MH-harm litigation crystallizes (suicide/psychosis suits, a 42-state AG probe, Altman named personally). The open question: does a formal health layer RAISE the standard of care, or wall off 'health' while the general model keeps causing harm? Its own numbers — >1M/week showing suicidal-planning signals — are the liability.
CMS's 10-year Outcome-Aligned-Payments demonstration (launched 07-05) — ~85 of 150+ accepted applicants chose the behavioral-health track, including Headspace, SonderMind, Concert Health, Mindoula Health, and April Health by Wysa, tying outcomes-based payment directly to multiple watchlist entities at once. Watch: which tracked companies' outcomes hold up under OAPs, whether the model expands, and read-through to mh-clinical-infra-funding's "does the market price in rigor" thesis.
MH is the #1-funded US therapeutic area (~$1.27B/14 deals to start 2026), but the money is going to clinician-augmentation (psychiatry copilots, AI scribes, payer-aligned platforms) and interventional psychiatry (TMS/ketamine/psychedelics) — NOT consumer chatbots. Track which models keep attracting capital vs. the chatbots hitting the regulatory reckoning; watch consolidation (e.g. Spring Health/Alma). Tests whether the market prices in rigor — the evidence-gap thesis.
Reported slip to 2027 chasing a ~$1T valuation; Altman rejecting a lower price. Track confirmed timing, valuation target, and macro/AI-stock spillover (SoftBank fell ~12% on the report).
Did GPT-5.6 ship, slip, or get gated? The June window reportedly slipped; track the new expected date AND the access-restriction story (who gets to use it). Prediction markets are the timing signal.
OpenAI's first custom inference ASIC, co-developed with Broadcom (9-month cycle, ~50% claimed cost/GPU). Track end-2026 deployment, real perf/cost vs. claims, and the next-gen parts in the multi-generation platform.
