🎥 Watch the latest 7-minute briefing · Subscribe on YouTube

The AI trade's two engines ran in opposite directions overnight: Meta shipped the first real consumer hardware stack for an AI agent and Google signalled Gemini 4 is weeks away, while the 10-year Treasury punched through 5.15% — its highest since 2007 — and took the Nasdaq's record with it.

🔥 Today's Top Stories

🔥 Meta: Zuckerberg Ships the Entire Muse Hardware Stack in One Night — $1,299 VR Glasses, a 5G Pendant, Camera-Free Ray-Bans, and Muse Wired Into Walmart Checkout

Meta Connect 2026 turned Muse from a chart-topping app into a product line. Zuckerberg unveiled $1,299 Meta VR Glasses (the "Project Phoenix" mixed-reality device, lighter than Quest and leaning on an external compute puck), the Muse Charm — an AirPods-case-sized pendant with a ≈2-inch touchscreen and built-in 5G that reaches Muse without a phone — and Meta's first camera-free smart glasses, codenamed Luna, which swap cameras for a six-microphone array. For investors the hardware is the sideshow: the monetisation news is that Muse now transacts with Walmart and Instacart, which is the first time Meta has attached a revenue mechanism to the agent that added roughly 900,000 downloads in six days and put 11% on the stock last week.

• Meta VR Glasses at $1,299; Ray-Ban Meta Audio (camera-free) at $349, shipping October 13; Muse Charm ships in the December holiday window
• The camera-free design is a direct concession to the privacy backlash — lighter, longer battery, no lens
• Muse retail partners at launch include Walmart and Instacart, moving the agent from engagement to commerce

  📰 Source: Mark Zuckerberg debuts $1,299 Meta VR Glasses and Muse Charm pendant as part of AI agent push — CNBC (Tier 1) 📰 Source: Meta wants Muse to be part of your everyday life. It just announced a bunch of gadgets to get there — CNN Business (Tier 1)

🔥 The Tape: The 10-Year Hits 5.15% — Highest Since July 2007 — and the 30-Year Takes Out a 2004 High as Nasdaq Futures Drop 1%

The AI complex just ran into its discount rate. Wednesday's session broke the Nasdaq 100's record with a 0.8% retreat, the S&P 500 fell 0.7% and the Dow shed 351 points, with chip profit-taking leading: NVIDIA −1.5% to $225.51, Broadcom −2.6%, Micron −2.2%. This morning it got worse — the benchmark 10-year yield surged to 5.15%, a level unseen since July 2007, and the 30-year touched 5.446%, its highest since June 2004. The trigger is inflation, not growth fear: S&P Global's PMI showed strong new orders with input and output price inflation at multi-year highs, and a weak Treasury auction plus rising crude did the rest. Markets are now pricing another Fed hike this year, on top of last week's first increase since 2023.

• Sept 24 pre-market: Nasdaq-100 futures −1.0%, S&P 500 futures −0.6%, Dow futures −0.3%; QQQ the day's laggard at −1.13%
• Sept 23 close: NVDA −1.47% to $225.51 · AVGO −2.6% · MU −2.2% — a pause after six straight up sessions for semis
• Rate math matters more for AI than for anything else in the index: these are the longest-duration cash flows in the market, and the capex is debt-financed

  📰 Source: Nasdaq futures fall 1% after soaring Treasury yields trigger market sell-off — CNBC (Tier 1) 📰 Source: Stock market today: Dow, S&P 500, Nasdaq futures slide as bond sell-off troubles markets — Yahoo Finance (Tier 2)

🔥 Koray Kavukcuoglu (Google DeepMind): Says Gemini 4 Is Already in Post-Training and Will Ship "Much Earlier" Than Year-End — His First Public Words as CEO

In his first media appearance since replacing Demis Hassabis at the top of Google DeepMind in August, Kavukcuoglu said Gemini 4 has already cleared pretraining and is in the early stages of post-training, and that Google wants an early post-training checkpoint out the door "much earlier" than the end of 2026. The strategy he described is explicitly anti-perfectionist: ship an early version because the results so far are encouraging, then iterate fast, rather than wait for a matured checkpoint. Sundar Pichai said in July that Google had begun its "most ambitious pretraining run yet" for the model — this is the first timeline attached to it, and it lands while Anthropic and OpenAI are cutting flagship prices at each other.

• Kavukcuoglu is SVP and Google's chief AI architect, reporting to Pichai; he owns Gemini models, frontier research, and the Gemini app and developer teams
• This is Google's first next-generation flagship since the Gemini 3 series — the anticipated Gemini 3.5 Pro was skipped
• For Alphabet holders it is the cleanest catalyst on the board: a competitive Gemini 4 validates the TPU capex story; a weak one revives the "Google is behind" discount

  📰 Source: Google nears release of Gemini 4 AI model, DeepMind head tells The Information — Investing.com, carrying The Information's reporting (The Information = Tier 1 original) 📰 Source: Google's Gemini 4 Could Launch 'Much Earlier' Than Year-End, Says DeepMind Exec — Benzinga (Tier 2)

🔥 Scott Bessent: The Busan Agreement Gets Extended Through January 10, 2027 — the First Hard Number Out of a Summit That Starts in Washington in Hours

Ahead of today's bilateral and state dinner, Treasury Secretary Scott Bessent confirmed the US and China will extend the Busan Agreement — the tariff-and-rare-earths truce struck last October — through January 10, 2027. That is the concrete deliverable; everything else remains open. CNBC's read this morning is that China's recent run of domestic chip and model launches gives Xi a talking point rather than leverage: Huawei's latest SuperPoD links far fewer processors than originally planned and each chip still delivers roughly half of NVIDIA's compute. The AI incident hotline agreed by Bessent and Vice Premier He Lifeng last weekend now has a follow-up venue — Shenzhen, in about two months, alongside APEC.

• State dinner guest list confirmed: Musk, Bezos, Huang, Altman, Pichai, Michael Dell and Citi's Jane Fraser, plus Apple executive chairman Tim Cook
• Xi's first White House visit in 11 years; export controls on advanced NVIDIA silicon remain explicitly off the negotiating table
• The truce now has an expiry date inside the 2027 calendar — that is a tradeable event, not a communiqué

  📰 Source: China's AI chip blitz arms Xi with a message for Trump: 'You can't choke us off' — CNBC (Tier 1) 📰 Source: Trump-Xi summit: Leaders meet at White House amid tensions over Taiwan, trade and AI — NBC News (Tier 1)

🌏 Asia Desk — the AI supply chain overnight

SK Hynix (NVIDIA's lead HBM supplier): ₩1,863,000 (+1.2%) · Samsung Electronics (HBM challenger and foundry #2): ₩286,500 (+3.24%) · Hanmi Semiconductor (HBM thermal-compression bonders): ₩243,000 (+1.04%) — all from the Sept 23 Seoul session, the latest close in today's feed. TSMC: NT$2,475 (−1.0%). USD/KRW: 1,369.2 (+1.39%).

Samsung's 3.24% outperformance is the tell, and it cuts against the Qualcomm headline below: Seoul bid up the foundry story even as Qualcomm confirmed TSMC took both 2nm flagship sockets and Samsung Foundry got priced out. What Seoul is actually buying is memory — HBM4 is sold out through calendar 2027 across the industry, and Hanmi moving with the two memory names says the bid is in equipment, not handsets. TSMC slipping 1% on the day it won an exclusive is the cleaner signal: Taipei is trading the same duration problem New York is, not the order book. With the won down 1.39% against the dollar and the US 10-year at 5.15%, the Asia AI complex is now importing a rate shock at the open — watch whether NVDA, MU and AVGO can hold the line after six sessions of gains and one of profit-taking.

👤 Key People Updates

📌 Jeremy Legg (AT&T CTO): Says the Company Will Not Hold Its Current Headcount Five Years From Now — and AT&T Scrambles to Deny the 85,000 Number

AT&T's technology chief told WIRED plainly that automation and the wind-down of legacy landline infrastructure mean a smaller workforce by 2031. WIRED extrapolated the current pace to roughly 85,000 employees by 2030; AT&T has since said that figure is not a target, while not disputing the direction. The functions Legg named — service disconnections, tower siting, maintenance triage, writing code, service-problem management — are exactly the middle-management and junior-developer layers the AI-displacement thesis has been pointing at for two years. This is the first time a Fortune 50 CTO has said it on the record with a time horizon attached.

• Named automation targets: disconnection workflows, tower site selection, maintenance flagging, code generation, service issue management
• WIRED reports middle-management and junior-developer roles are expected to disappear; no AI-attributed job-loss count was given
• AT&T rejects the 85,000-by-2030 figure as a target but confirms it expects fewer employees

  📰 Source: Original reporting by WIRED (Tier 1), "AT&T Is Automating Away Jobs — and Its Old Telecom Empire" (Sept 23); WIRED blocks direct linking, so the verifiable writeups are AT&T Expects More Job Cuts but Rejects Reported 85,000-Worker Target and AT&T plans workforce cuts as it automates operations and shifts to fiber — Crypto Briefing (Tier 2)

📌 Sam Altman and Dario Amodei: The Actual Asks From the UN Security Council Session — a Bioweapons Carve-Out First, Common Loss-of-Control Tests Second

The substance, now that transcripts have circulated: Amodei argued for starting narrow rather than comprehensive — a ban every member state can sign, specifically on using AI to build biological weapons — then layering on common global standards for testing models for loss-of-control and misuse risk, plus a notification system for globally significant AI incidents. Altman's framing was about autonomy drift: models become more capable and more autonomous, and the line to defend is that humans remain in control of decision-making. Two companies that spent this week undercutting each other on price agreed on the regulatory architecture.

• Amodei's sequencing — narrow bioweapons agreement → shared eval standards → incident notification — mirrors the US-China incident line Bessent just set up bilaterally
• Altman: "Anything this powerful is also intimidating"
• No formal UN regulatory body emerged from the session

  📰 Source: Sam Altman, Dario Amodei urge UN Security Council to adopt international AI standards — CNN Business (Tier 1) 📰 Source: OpenAI, Anthropic CEOs call for global AI regulation at UN — Al Jazeera (Tier 2)

🏢 Big Tech Updates

📌 Qualcomm: Splits the Flagship in Two on TSMC 2nm — First Phone CPU Past 5GHz, 30 Billion Parameters Running Fully On-Device, and Samsung Foundry Priced Out of Both Sockets

Qualcomm launched the Snapdragon 8 Elite Gen 6 and a new top SKU, the 8 Elite Extreme Gen 6, both on TSMC's N2 node. The Extreme runs a 2+3+3 Oryon cluster peaking at 5.11GHz — the first smartphone CPU past 5GHz — against 4.4GHz for the standard part, with a 35% faster GPU at 40% better efficiency, and it will hold a 30-billion-parameter model entirely on device with no cloud call. Chris Patrick confirmed on stage that all 2nm components announced are manufactured by TSMC; CEO Cristiano Amon said Samsung Foundry cleared technical qualification on its own 2nm process but the two sides could not agree on price. That is the story for the supply chain: N2 volume is real, and Samsung's qualified-but-unsold 2nm line is still looking for its first flagship customer.

• Two SKUs by design — Amon's framing is a shift from "phone-centric" to "agent-centric" computing, with the agent completing tasks instead of the user switching apps
• Motorola, Xiaomi and ZTE are the launch OEMs
• Follows MediaTek's Dimensity 9600 Pro (Sept 15), which claimed the first mobile processor on N2 — Qualcomm's claim is the first mass-market 2nm mobile silicon

  📰 Source: Qualcomm Taps TSMC for All Snapdragon 8 Elite Gen 6 Chips — The Elec (Tier 2) 📰 Source: Snapdragon 8 Elite Gen 6 Lands on TSMC 2nm: First 5 GHz Mobile CPU Built for AI Agents — Tech Times (Tier 2)

📌 Microsoft: Its Digital Crimes Unit Seizes EvilTokens — an AI Phishing Service Sold on Telegram for $1,500 Plus $500 a Month That Took 12,000 Inboxes Across 10,000 Organisations

Microsoft obtained authorisation from the US District Court for the Eastern District of Virginia, seized 50 websites running the service and disabled more than 150 associated domains. EvilTokens launched in February 2026 and sold a device-code phishing kit where the AI did the work that previously required a skilled operator: analysing compromised inboxes, selecting targets, deciding who to impersonate and drafting the lures. British police arrested two men aged 32 and 38, both released on bail. This is the DCU's 40th court-authorised disruption in nearly two decades — and notably, OpenAI, Cloudflare, Coinbase, Railway, SpyCloud, TRM Labs, Shadowserver and Health-ISAC all joined the action.

• Pricing tells you the economics: $1,500 initiation plus $500/month put frontier-adjacent social engineering inside a hobbyist budget
• 12,000+ inboxes compromised across 10,000+ organisations worldwide since February
• OpenAI participating in a Microsoft takedown is a quietly significant piece of industry cooperation

  📰 Source: Microsoft, partners disrupt EvilTokens, AI-powered phishing service — Axios (Tier 2) 📰 Source: UK cops arrest 2 EvilTokens suspects, Microsoft seizes 50 phishing kit websites — The Register (Tier 2)

📌 NVIDIA and Palo Alto Networks: Two Different Bets That Agents Are Now the Product — Isaac ROS 5.0 Puts Coding Agents Inside Robotics Workflows, Unit 42 Rents Out Claude and GPT for Continuous Pentesting

NVIDIA shipped Isaac ROS 5.0, adding AI-agent skills to its GPU-accelerated robotics stack — coding assistants that configure environments and migrate software, with ROS 2 Lyrical support and broader Jetson deployment. Separately, Palo Alto launched Unit 42 Continuous Frontier AI Defense, a service that runs Anthropic and OpenAI models against customers' web apps, APIs and cloud infrastructure on a continuous basis, explicitly multi-model because different vulnerability classes respond to different models. Both are the same trade in different clothing: the agent is no longer the demo, it is the line item.

• Isaac ROS 5.0 targets the developer-workflow bottleneck in robotics, not the robot
• Palo Alto's multi-model design is an implicit admission that no single frontier model covers the vulnerability surface

  📰 Source: NVIDIA Isaac ROS 5.0 and Palo Alto Unit 42 launches, Sept 23 roundup — Tech Startups (Tier 2)

📊 Market & Investment Signals

💬 Comma.ai: NHTSA Opens an Investigation Into Its Hands-Off Driving System After Five Crashes, Three Deaths and Eleven Injuries

NHTSA's Office of Defects Investigation opened PE26007 on September 21, with the story breaking Sept 23. The failure mode named in the probe is the classic one — not detecting slow or stopped vehicles in the same lane. What makes this case different from the Tesla precedents is the open-source ecosystem around openpilot: third-party modifications complicate any attribution of fault, and that ambiguity is itself the regulatory question. One of the two fatal crashes — a 2022 Toyota RAV4 that struck a stopped first-responder vehicle in Ascension Parish, Louisiana in February — was running FrogPilot, a third-party fork, and NHTSA has said it will explicitly evaluate whether forked software was involved. A finding here would be the first US safety ruling to grapple with an open-weights driving stack.

📰 Source: Comma's hands-off driving tech under investigation after 2 fatal crashes — TechCrunch (Tier 1) 📰 Source: NHTSA Office of Defects Investigation, PE26007 opening resume — NHTSA (primary document)

💬 FBI: ShinyHunters Claims 2–3TB From the FBI Jobs Portal — Including Agents' Home Addresses and SSNs — via an Unconfirmed Oracle PeopleSoft Zero-Day

Names, contact details, addresses and Social Security numbers, allegedly lifted from FBIjobs.gov and the Special Agent Application Portal, both of which went offline Tuesday. ShinyHunters claims 2–3TB pulled via Amazon GovCloud, entered through an Oracle PeopleSoft zero-day — which if confirmed puts every large enterprise HR deployment in scope, not just the Bureau's. Nothing in the claimed attack path has been confirmed by the FBI, Oracle or Amazon; the Bureau says only that it is investigating. The group says the motive is not financial — it wants the FBI to retract a May advisory describing its harassment and swatting tactics. The AI angle is indirect but real: identity data at this scale is the raw input for exactly the kind of AI-drafted impersonation EvilTokens was industrialising.

📰 Source: FBI investigating claims that a major cybercrime group stole sensitive personnel data — Axios (Tier 2) 📰 Source: Cybercriminal group claims it stole FBI personnel and applicant data — CBS News (Tier 2)

💬 Arista and Google: A CVSS 10 VeloCloud Zero-Day Under Active Exploitation, and Chrome Patches 108 Flaws With 11 Rated Critical

Arista warned that CVE-2026-93952 in VeloCloud Orchestrator — a perfect-10 severity score, meaning unauthenticated remote access to privileged internal functionality — is being exploited in the wild. Chrome 154 landed the same day with 108 vulnerabilities fixed, 11 critical, 32 of them found by external researchers, across ANGLE, WebGL, GPU and ServiceWorker. CISA added it to the Known Exploited Vulnerabilities catalog Tuesday and gave federal agencies three days to patch under BOD 26-04; hosted instances are fixed, but on-prem VCO at 6.1.3.7 and below and 7.0.0.2 and below is still waiting. Chrome 154 landed the same day. The volume itself is the story: 32 external findings in a single release is what a maturing AI-assisted vulnerability research pipeline looks like from the outside.

📰 Source: Arista Urges Immediate Patching of Exploited VCO Zero-Day — SecurityWeek (Tier 2) 📰 Source: Arista patches actively exploited VeloCloud Orchestrator zero-day — BleepingComputer (Tier 2)

💬 Micron: Reports FQ4 After the Close on September 30 — Guidance Is $50B in Revenue, an 86% Gross Margin, $31.00 EPS and $30B+ of Free Cash Flow in a Single Quarter

Worth restating only because of where the bond market went this morning. HBM4 has already cleared $1 billion in revenue and the 12-high ramp is outpacing HBM3E; HBM3E and HBM4 are fully booked through calendar 2027. FY2026 capex is now ≈$27B and FY2027 goes above the mid-$40B range, mostly construction. That is the exact profile — enormous, debt-adjacent, multi-year construction capex — that a 5.15% 10-year reprices hardest.

📰 Source: Micron Technology Q3 FY2026 prepared remarks and FQ4 guidance — Micron investor relations (Tier 2, company filing)

🎥 Prefer to watch? This brief becomes a daily video on The Daily Alpha — AI × markets in ~7 minutes, every weekday. ▶ Watch the latest briefing 📺 Subscribe to the channel

Generated by Alan · The Daily Alpha. Editorial, not financial advice.