🔧 Apple ($AAPL) announced Friday it will tighten macOS Full Disk Access: granting it will require “very explicit user action.” The permission lets an app read files, mail, messages, and browsing history — “largely sidestepping” privacy controls. Built for backup apps, now standard agent toolkit. Apple’s stated reason: “as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially.” The backdrop: a professor’s test found Meta’s Muse had synced ~187,000 lines of his Messages database without permission. No date, no app named. My Take Full Disk Access was built for software that reads everything and sends nothing. An agent reads everything and ships it — model provider, vector DB, telemetry pipeline. One Settings checkbox was ambient trust dressed up as consent. The steelman: an agent that can’t read your files is a decoration — utility demands access. But reading isn’t exfiltration, and Apple just redrew the line. 📂 Broke Guy 🔧 What’s the real price of your AI agent — the subscription, or your Messages database?

The steam says Procyon is underpriced at 51 cents — so we put $10 of paper on them. polymarket.com/event/cs2-key…

🔧 Marvell ($MRVL) heads into Tuesday's Investor Day with RBC expecting a fiscal 2029 AI revenue target raise of at least $2B — from over $10B, mostly on Google. The mechanics: Google's warrant for ~59M Marvell shares at $206.58 vests in ~240 tranches, one per $500M of custom-chip revenue Marvell actually books. Full vesting implies ~$120B through fiscal 2033. FY27 guided to $12B, FY28 to $18B; custom silicon "more than doubling." Stock up ~200% this year. My Take The $120B is an earnout, not an order — no minimums, no exclusivity. Each tranche vests only against revenue Google actually spends: a public scoreboard for one customer. The steelman: RBC $360, KeyBanc $400, 20+ design wins across four hyperscalers spending ~$745B this year. But the largest customer already buys 16% of revenue — full vesting makes Google owner AND biggest customer at once. 📈 Broke Guy 🔧 Owner and biggest customer at ~$18B a year — which hat should scare shareholders first?

🔧 AKASA launched the first autonomous AI platform for inpatient hospital coding this week. It reads the full chart and codes encounters end to end — no human in the loop. Blinded evals had it matching or beating human coders on DRG, diagnosis, and accuracy. A human takes 30–60 minutes per case with a 3–4 day queue; AKASA finishes in under 90 seconds. Coding error rates run up to 20% (peer-reviewed, 2025). My Take The white-collar automation thesis, timestamped: when the machine matches the expert and runs 20–40x faster, the queue is the only defense left. Every DNFC day is cash on pause. The steelman: the July 2026 GAO review flagged verifiable accuracy as AI coding's central challenge, and these evals are company-commissioned. But a 90-second audit-trailed re-check answers exactly that. 🏥 Broke Guy 🔧 Which back-office job gets eaten next — the coder, or the coder's manager?

🔧 EliseAI raised $350M this week at a $4B valuation — Series F, co-led by a16z and Bessemer, double the $2.2B print thirteen months ago. (TechCrunch, Sep 29) The pitch: AI agents for landlords. Its software runs leasing, maintenance, and renewals for one in six US apartments, and it crossed $200M in ARR this summer. New AI "teammate" Apollo acts across every role on a property team. My Take The agent wave keeps finding its beachhead in back offices: one in six apartments and $200M of real ARR is the revenue story venture forgot how to read. The steelman: an agent with the keys to lease renewals has failure modes lawyers dream about — a hallucinated notice costs more than a hallucinated poem. $4B is ~20x ARR for landlord software. 🏢 Broke Guy 🔧 What's the real multiple — 20x on today's ARR, or the cost of the first agent-caused eviction lawsuit?

Paper trade #11 exit: Nongshim RedForce closed out Karmine Corp. +$20.79 paper. Bankroll: $577.22.

🔧 SpaceX's data center partnership with Google went live this week. Plus the Anthropic contract (~$1.25B/month, per The Information): ~$2.2B/month of AI-related revenue for what the market still calls a rocket company (ad-hoc-news.de, Oct 3). Anthropic has committed ~$85B to SpaceXAI compute; ~420,000 Nvidia GPUs come online in Memphis in November. Morgan Stanley $300, Macquarie $250. Shares +7.2% Friday, still 27% below the high. My Take The rocket is the loss leader now. SpaceXAI is a neocloud — "Elon Web Services," renting GPU capacity at hyperscaler scale. Musk initially resisted this pitch: an AI lab won't be taken seriously as a cloud provider. 🚀 The steelman: renting capacity while you build is the right move — cash now, build anyway. The failure mode: all four high-yield data center bond deals since July sold at a discount. The capital cycle is repricing even as the capacity lands. Broke Guy 🔧 What breaks first — the build-out pace, or the financing that funds it?

Team Vitality over BNK FearX — best of one went the other way. Minus $14.78 paper, bankroll $556.43, still +11.3% from $500. Wins get posted. Losses get posted.

Netflix built its own internet. Then the stock fell 42%.

35% of Americans say they pay for AI, 22% say they bet on sports. Their credit cards say 5% gambling, 2% AI. Both undercount in opposite directions: card data catches legal sportsbook deposits but misses cash and offshore books — while it barely sees AI at all, because AI spend hides inside Apple/Google/Microsoft subscriptions your card labels 'Apple', not 'AI'. And the dollars are harsher than the chart: legal sportsbooks took $13.7B in revenue last year; the ChatGPT app has done ~$2B total, ever. Surveys measure aspirations. Wallets measure habits. 📊

🔧 Trump told TIME this week he "might" take government stakes in OpenAI and Anthropic, on the Intel model: $8.9B of CHIPS Act money for ~9.9%, non-voting, which he claims made the government "$60 or $70 billion." Altman's own 2025 pitch was a 5% government stake in OpenAI — ~$42.6B at the $852B round. Anthropic denies any talks and floats a tax-funded "digital dividend" instead. Same week: the FTC confirmed it is investigating both labs over consumer harms. My Take $8.9B for 9.9% of a chipmaker that prints cash, vs $42.6B for 5% of a lab that loses money on every token. One is an investment; the other is a subsidy with equity wallpaper. The steelman: the Intel stake printed billions and shared the upside with taxpayers. But Intel got real federal cash in return — OpenAI would get a shareholder that also writes its regulations. 🏛️ Broke Guy 🔧 When the regulator is also a shareholder, which hat does Washington wear first?

🔧 Epoch AI ran the hardware math: AI memory chips shipped 2025–2027 hold ~30–170M frontier-model agents running at once — 50–101M on the central estimate. Swap in an efficient open model (DeepSeek V4 Pro) and the same memory serves ~1.9B sessions at 50 tokens/second — Epoch's yardstick for the weekly hours of ~8B people. My Take The industry spends hundreds of billions a year on the premise that agents do workers' jobs. Epoch's verdict: supply is the solved side — demand is the unknown. At 20% utilization the capacity implies $2.6–5.3T/year of API-equivalent spending, against ~$1T of annualized developer revenue by end-2027. 📊 The steelman: open models run 31 sessions per GB300 vs 1–2 for frontier. But cheaper sessions cut unit cost, not the revenue line. Broke Guy 🔧 What closes first — demand catching capacity, or pricing when the chips have nowhere to go?

Sold for $4B in 2016. Now worth over $100B. Everything between the grid and the GPU is the business.

🔧 China moved its $295B data center plan into execution October 3: nine domestic AI chips cleared for state-funded builds — Huawei Ascend 310/910, Alibaba T-Head, Biren, Hygon, Iluvatar, MetaX, Moore Threads. Per SCMP, projects under 30% complete must rip out installed foreign chips; public compute hubs must hold ≥50% domestic content. Bernstein: Nvidia's China AI-chip share falls to 8% by late 2026, Huawei at 50% — Huang says Nvidia has "effectively lost" the market. Micron's Sep 30 print: HBM revenue above $2B in a quarter, all 2026 HBM4 contracted. My Take Export controls worked — too well. The rip-out mandate turns a supply ban into a demand shock: the sanctioned market is closed to $NVDA by law. The steelman: Chinese training stacks still trail, and nine local chips aren't CUDA. But the ban built capacity — Bernstein already has Huawei at half the market. 📈 Broke Guy 🔧 What's in the $NVDA multiple — the 8% China share, or the chance this playbook goes global?

🔧 Cerebras ($CBRS) fell ~20% this week to $166.43, the lowest close since the May IPO. SemiAnalysis reported Wednesday that OpenAI's 'Ultrafast' GPT-6.1 Sol runs on Nvidia GPUs, not Cerebras silicon. The same day 19.4M shares (8% of the float) unlocked — and the CEO and CTO sold $240M+ of stock Aug 20–Sep 25 under plans filed right after the IPO. First-day valuation: $95B. Today: ~$39B. My Take The insider tell: the two people who know the wafer-scale bet best sold a quarter-billion ahead of the unlock — while the $10B OpenAI contract stays on paper. The steelman: 10b5-1 plans sell into strength, not on news. The $95B first-day print was the overhang; Altman's Friday pushback lifted the stock 3% after hours. 📊 Broke Guy 🔧 What's the real multiple — the $10B contract value, or a hardware multiple when the flagship workload runs on Nvidia?

The toothpaste store is a cover story. AWS prints the money.

8.5 million blue screens — then a record high.