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

The first megacap AI earnings landed and split the trade in half — Meta raised 2026 capex to $145 billion and invented a non-voting share class to pay for it, Tesla cut its capex forecast to about $20 billion, and the market sold both, dragging Nasdaq-100 futures down 1.9%, the KOSPI down 3.52% and SoftBank to half its June peak.

🔥 Today's Top Stories

🔥 Meta: Beats on Revenue, Raises 2026 AI Capex to $145 Billion — and Creates a Non-Voting Class C Share Class to Pay for It While Zuckerberg Keeps Control

Meta posted $64.6 billion of revenue, up 26% and ahead of the $63.38 billion consensus, with adjusted EPS of $7.40 against $6.54 expected. But the quarter's real news was the balance sheet: total expenses rose 68% to $51.6 billion, free cash flow went negative $2.1 billion on $34.9 billion of quarterly capex, and the board declared a one-time stock dividend of one new non-voting Class C share for every Class A and Class B share — functionally a two-for-one split that creates a currency Meta can issue for AI capital, employee comp and acquisitions without diluting Zuckerberg's votes. The market's verdict was not kind: Meta slipped about 1.9% on Thursday despite the beat, because investors read the new share class as an admission that operating cash flow alone can no longer fund the buildout.

• Revenue $64.6B (+26% YoY); Family of Apps $64.2B, of which $63.1B advertising (+26%)
• Adjusted EPS $7.40 vs $6.54 consensus — but GAAP EPS $3.83 after a $9.9 billion legal charge tied to the teen-safety settlement
• Capex $34.9B in the quarter; 2026 capex guidance raised to about $145 billion; free cash flow −$2.1B
• 3.64 billion daily users across the family of apps; Facebook and Instagram at all-time engagement highs; Threads passed X in monthly actives
• Restated charter filed in Delaware on or about October 21, Class C effective January 19, 2027

  📰 Source: Meta (META) Q3 2026 earnings report: live updates — CNBC (Tier 1) 📰 Source: Meta Boosts Bet on AI With New Spending Outlook, Stock Offering — Bloomberg (Tier 1) 📰 Source: Meta wants to pay for its broad AI buildout with a newly created class of stock — TechCrunch (Tier 1)

🔥 Anthropic: Investors Trim the IPO Range to $1.5 Trillion From $2 Trillion as Sequential Growth Collapses From 142% to 45% — With the Public S-1 Due as Soon as Tomorrow

Yesterday the draft S-1's numbers became public; today the market repriced them. Existing and prospective investors who sat through last week's San Francisco investor day are now discussing a valuation around $1.5 trillion, a $500 billion haircut from the roughly $2 trillion Wall Street had been modelling. The trigger is not the headline loss — $72 billion of the $80.7 billion net loss is a non-cash mark on convertible notes — but the growth curve: Q3 revenue of $16.8 billion was up 45% sequentially from $11.5 billion, against 142% in the comparable prior period, a 98-percentage-point deceleration. Anthropic is expected to file its S-1 publicly as soon as Friday, which under SEC rules for confidential filers starts a 15-day clock before a roadshow can launch — putting a November listing, still likely the largest IPO in history, on the calendar.

• Investor discussions now centre on ≈$1.5T, down from ≈$2T; Wall Street had modelled up to $100B raised
• Q3 revenue $16.8B, +45% QoQ vs 142% in the comparable prior period
• H1 revenue $16.3B; net loss $80.7B, of which ≈$72B is a fair-value charge on convertible instruments
• Draft registration statement submitted confidentially to the SEC in June; public filing clears the roadshow gate
• Bull case still live: investor Gavin Baker cites positive free cash flow, accelerating October growth and better-than-expected unit economics

  📰 Source: Did Anthropic Miss Its Peak IPO Moment? — The Information (Tier 1) 📰 Source: Anthropic is AI trade's last big hope for a major IPO this year — CNBC (Tier 1) 📰 Source: Anthropic Could File IPO Paperwork as Soon as Friday — PYMNTS (Tier 2)

🔥 PHP2Shell: Two AI-Discovered Zero-Days With a CVSS of 9.2 Put Roughly 70% of the Internet at Risk — Finland and the Netherlands Pulled Government Websites Offline, and 22.6 Million Instances Are Exposed

The PHP Foundation shipped emergency patches for CVE-2026-109618 and CVE-2026-109619, a pair of unauthenticated remote-code-execution bugs collectively named PHP2Shell, both scored 9.2. The investor-relevant detail is how they were found: the red-teaming firm Calif surfaced the first on August 1 using OpenAI's GPT-5.6 Sol, and Google-owned Wiz independently found the related flaw on August 19 using Gemini 4 Argon, GPT-5.6 Sol and Wiz Atlas — with roughly two months of coordinated disclosure before the patch. The code had been sitting in PHP for over a decade. This is the clearest public datapoint yet that frontier models have crossed from "assists a researcher" to "finds internet-scale bugs nobody else did," which cuts both ways for every security budget on the planet.

• CVSS 9.2, unauthenticated RCE; PHP underpins ≈70% of the web and WordPress (≈40% of all websites)
• Palo Alto Networks counts 22.6 million potentially vulnerable instances
• Finland shut down government public web services; the Netherlands' cyber agency took its own site offline; Edinburgh council and the BBC pulled pages
• The FBI, the NSA and Japan's Prime Minister issued urgent update statements
• Discovery ran through OpenAI's "Patch the Planet" initiative

  📰 Source: AI helped discover the PHP vulnerability that could take down 70% of the internet — Fortune (Tier 2) 📰 Source: What to do as critical PHP flaws put millions of websites at risk — Computer Weekly (Tier 2) 📰 Source: PHP security flaw: websites taken down as OpenAI spots issue with backbone of internet — The National (Tier 2) ⚠️ Sourcing note: strong Tier 2 coverage, no Tier 1 wire exclusive located at time of writing — the CVE identifiers and patch are confirmed by the PHP Foundation advisory.

🔥 Tesla: Net Income Falls 43% to $787 Million and Free Cash Flow Drops 62% — Then It Cuts 2026 Capex to About $20 Billion and the Stock Falls 5.3%

Tesla cleared the low bar — adjusted EPS of $0.48 against $0.43 expected, revenue of $28.05 billion against $27.9 billion — and still broke. Automotive revenue fell 3% to $20.56 billion and energy fell 16% to $2.88 billion; only services grew, up 33% to $4.61 billion. Operating expenses climbed 39% to $4.8 billion on superintelligence and other R&D, pushing operating income down 56% to $718 million. The move that hurt was the guide: capex of roughly $20 billion for 2026, a cut that investors read as Tesla blinking on the pace of its AI and robotics buildout. Side by side with Meta raising capex to $145 billion on the same evening, the market got two opposite answers to the same question and sold both names.

• Adjusted EPS $0.48 vs $0.43 est.; revenue $28.05B vs $27.9B est., roughly flat YoY from $28.24B
• GAAP net income −43% to $787M; free cash flow −62% to $1.53B
• Operating income −56% to $718M as opex rose 39% to $4.8B on SI/R&D
• 486,532 vehicles delivered; Cybercab robotaxi rides began in the quarter
• 2026 capex guided to ≈$20B; shares fell 3.3% after hours and 5.3% into Thursday

  📰 Source: Tesla (TSLA) Q3 2026 earnings report — CNBC (Tier 1) 📰 Source: Tesla releases Q3 2026 financial results: margins rebound, profit still falls — Electrek (Tier 2) 📰 Source: Tesla, Inc. Form 8-K, October 21, 2026 — SEC EDGAR (primary)

🔥 The Reversal: Nasdaq-100 Futures Drop 1.93%, the KOSPI Falls 3.52% and SoftBank Is Now Down 50% From Its June Peak — Two Days After the Same Trade Rallied on Super Micro

The dip-buying that Super Micro's guidance raise triggered on Tuesday lasted exactly one session. With Meta and Tesla on the tape, the market rotated straight back out of anything debt- or capex-funded: at 04:56 ET Nasdaq-100 E-minis were off 578.5 points (−1.93%), S&P 500 E-minis −1.1% and Dow E-minis −367 points, with a rebound in oil adding to the pressure. The damage was broad across semis — Intel −5.1%, with AMD, Qualcomm, Marvell, Broadcom and Micron down between 2.9% and 5.3%, and NVIDIA −2.4%. Asia got it worse: the KOSPI fell 3.52% to about 5,534, the Nikkei dropped as much as 2.7% with Tokyo Electron −4.1%, and SoftBank fell 6.9%, taking it to half its June 2 all-time high as borrowing costs reignite questions about the payoff timetable on its OpenAI stake.

• Nasdaq-100 E-minis −1.93%; S&P 500 E-minis −1.1%; Dow E-minis −367pts
• Semis hit hardest: Intel −5.1%, NVIDIA −2.4%, AMD/QCOM/MRVL/AVGO/MU −2.9% to −5.3%
• Premarket: Super Micro −4.7%, SpaceX −2.3% — the two names that led Tuesday's rally
• KOSPI −3.52% to ≈5,534 after Wednesday's close at 5,735.66; Nikkei down as much as 2.7%
• SoftBank −6.9%, now −50% from its June 2 record on debt-cost and AI-return concerns

  📰 Source: Nasdaq futures drop over 1% as AI worries resurface, oil climbs — Reuters via Investing.com (Tier 1) 📰 Source: Japan's Nikkei slips as chip shares pressured by AI capex caution — Reuters via Business Recorder (Tier 1) 📰 Source: Samsung and SK hynix plunge again as Korea's $792 billion rout tests confidence — Invezz (Tier 2)

🔥 Anthropic: Ships Claude Duet Into Research Preview — a Proactive Agent on Its Own Virtual Machine, Launched Straight at OpenAI's Dots, Meta's Muse and Google's Gemini Agent

Days before its S-1 goes public, Anthropic opened the research preview of Claude Duet, a standalone iOS, Android and claude.ai app in which Claude stops waiting for prompts. Built on Claude Opus 5.5, Duet sends a morning brief on what needs attention across connected apps, re-checks those accounts through the day, messages only when something matters, and runs on its own virtual machine with its own browser — so it keeps working when the user's laptop and phone are off. Access is waitlisted to US Pro and Max subscribers with no incremental charge during the preview. Strategically this is the personal-agent land grab becoming a four-way fight in a single quarter, and it is a deliberate pre-IPO signal that Anthropic's revenue story is not only API and Claude Code.

• Standalone app on iOS, Android and claude.ai; built around Claude Opus 5.5
• Runs on its own VM and browser — operates with the user's devices off
• Task range cited: scheduling, dashboards across dozens of tools, financial and sales forecasts, prototyping, web apps
• Waitlisted to US Pro and Max users; included in those plans during the preview
• Same-day: Claude Security moved to continuous scanning, and HCLTech joined the Claude Partner Network as a Global Premier partner

  📰 Source: Anthropic's answer to Muse is here: Claude Duet, a mobile and desktop AI agent that takes actions proactively — VentureBeat (Tier 1) 📰 Source: Introducing Claude Duet — Anthropic (company announcement) 📰 Source: Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots — Engadget (Tier 2)

🌏 Asia Desk — the AI supply chain overnight

SK Hynix (NVIDIA's lead HBM supplier): 1,234,000 (−4.19%) · Samsung Electronics (HBM challenger and foundry #2): 214,500 (−4.03%) · Hanmi Semiconductor (HBM bonding equipment, TC bonders): 191,200 (−5.81%) · TSMC (fabs the GPUs themselves): 2,385.0 (−1.45%) · USD/KRW: 1,323.8 (−0.01%)

Seoul gave back Wednesday's entire rally and then some — Samsung and SK hynix had risen 5.67% and 5.81% the day before, and both shed roughly 4% today on straight profit-taking rather than any new demand datapoint. Two mechanical factors amplified it: the first day of Korea's new brokerage margin-concentration rules pulled margin-loan balances on Samsung and SK hynix down 1.01% and 1.57%, and the Meta-versus-Tesla capex split gave Asian investors a reason to question 2027 AI spend before New York opened. Note the hierarchy in the tape — equipment (Hanmi −5.81%) fell harder than memory, and memory fell three times harder than the foundry (TSMC −1.45%), which is the market pricing capex timing risk rather than AI demand risk. For the US open, that argues the pressure lands on MU, AMAT and LRCX before it reaches NVDA, and with the won essentially unchanged at 1,323.8 there is no FX alibi for the move.

👤 Key People Updates

📌 Dario Amodei (Anthropic): Answers a Safety Critic With 2,000 Words and a Nine-Item List of Commercially "Very Costly" Decisions — Including Holding a Frontier Model Back for Four Months

Amodei posted a long rebuttal on X to Nathan Calvin, chief strategy officer of the AI-safety nonprofit Encode, who had argued that Anthropic's compute deals and model strategy undercut its safety commitments. Amodei enumerated nine actions he said cost Anthropic commercially while delivering safety benefits — the headline items being the restriction of AI sales to Chinese state-controlled companies and the release of its powerful Mythos model only to a vetted set of cybersecurity partners. He conceded mistakes, specifically on how the draft federal safety bill was communicated, and reaffirmed a commitment to propose regulatory options by January 15, 2027. For investors reading the S-1, this is the "race to the top" narrative being stress-tested in public days before pricing.

• Nine-item list of costly-but-safe decisions, posted in response to Encode's Nathan Calvin
• Cites the China state-entity sales ban as the first of its kind among frontier labs
• Says Anthropic held a frontier model back for roughly four months in response to safety criticism
• Acknowledges communication failures around the draft federal bill circulated October 20

  📰 Source: Dario Amodei spent 2,000 words laying out to a critic on X why Anthropic must persist in the AI 'race to the top' — Fortune via Yahoo Finance (Tier 2) 📰 Source: Anthropic CEO defends safety record, sales to China — Fox News (Tier 2)

📌 Donald Trump: Mandates That Large Data Centres Generate Their Own Electricity and Pay Communities a "Dividend" — a Pre-Midterm Repositioning on the AI Buildout

With data centres emerging as a measurable political liability in midterm polling, Trump reframed rather than retreated: new and large data centres are "mandated by me to make their own electricity, and give or sell any excess back to the local community, or grid," and operators that want projects approved should pay residents a dividend. The economic read-through is a direct cost shift onto hyperscalers and neoclouds — behind-the-meter generation and community payments both land in capex and opex at exactly the moment credit markets have stopped funding this sector cheaply. Watch the gas-turbine, fuel-cell and SMR supply chain as the beneficiaries.

• Requirement for large centres to self-generate power and return excess to the grid or community
• Encouragement of direct community "dividend" payments to win siting approval
• Framed explicitly around projects that flipped from unpopular to popular once "a financial deal is struck"

  📰 Source: 'MAKE DEALS': Trump says data centers should pay communities a 'dividend' to win support — CNBC (Tier 1) 📰 Source: Trump says AI data centres should pay 'dividends' in pre-midterm shift — Al Jazeera (Tier 2)

📌 Jensen Huang (NVIDIA): Closes GTC Berlin by Telling Europe to "Get Serious" About the AI Race — and Takes a Protester's Interruption From the Stage

On the final day of GTC Berlin, Huang's pitch was less product than scolding: Europe has the industrial base to win physical AI and is not building the compute to do it at home. The keynote was interrupted for about a minute by a protester claiming NVIDIA components had been found in Russian drones; Huang replied that "everyone deserves a chance to speak" and continued. Combined with this week's German letter of intent, NVIDIA's European investments and commitments now approach $10 billion — roughly $8 billion already deployed over the past year plus the ≈$2 billion announced this week.

• Core thesis: AI leaves the chatbot and moves to the factory floor; Europe should host that infrastructure
• ≈1 minute stage interruption over alleged NVIDIA components in Russian drones
• European investment and commitments approaching $10B in aggregate

  📰 Source: Nvidia's Jensen Huang Scolds Europe on AI Race — Bloomberg (Tier 1) 📰 Source: Jensen Huang's Berlin GTC Keynote Interrupted by Protester as Nvidia Unveils Mass-Production Supercomputers and European AI Push — BigGo Finance (Tier 2)

🏢 Big Tech Updates

📌 NVIDIA: Says Its Total Semiconductor Testing Capacity Will Double From 2026 to 2027 — and Signs Saudi Arabia's Humain for Up to 84MW, Including 50MW Resold to an Unnamed US Client

Two supply-side datapoints out of GTC Berlin that matter more than the keynote rhetoric. First, NVIDIA and its partners committed to expanding global manufacturing capacity, with total semiconductor test capacity expected to double from calendar 2026 to 2027 on buildouts in the US, Taiwan and Singapore — test, not wafer, has been the quiet bottleneck on Blackwell and Rubin volume. Second, Humain expanded its NVIDIA partnership to up to 84MW: a 14MW first phase of Blackwell HGX B300 systems (≈6,000 GPUs) deploying from Q4 2026, then 70MW of Vera Rubin from mid-2027 — of which 50MW is earmarked for an unidentified US client, meaning Gulf capital is now building capacity for American demand.

• Test capacity to double CY2026 → CY2027 (US, Taiwan, Singapore)
• Humain phase 1: 14MW / ≈6,000 Blackwell HGX B300 GPUs from Q4 2026
• Humain phase 2: 70MW of Vera Rubin from mid-2027, 50MW reserved for an unnamed US customer
• Also shipped at Berlin: Isaac humanoid reference system, T-Mobile's first AI-RAN call on a commercial network, an enterprise agent-security reference architecture

  📰 Source: NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout — NVIDIA via GlobeNewswire (company release) 📰 Source: Saudi Arabia's Humain expands Nvidia deal with 84MW AI computing plan — The National (Tier 2)

📌 Google DeepMind: Engineers Resign Over the Pentagon Contract — "Ashamed to Work at Google" — With More Said to Be Preparing to Leave

Sofia Liguori, a London-based research engineer, resigned this week after three years, saying she was ashamed of Google's deepening military partnerships and uneasy about speaking up internally; she had staged a personal sick-out the day after the classified Pentagon contract was announced in April. Researcher Andreas Kirsch quit earlier in the month arguing Google's oversight mechanisms are inadequate for military and classified deployment. The contract permits Google models for any lawful government purpose but bars autonomous weapons and mass surveillance without human oversight. The investor angle is talent cost: defence revenue is the fastest-growing enterprise line for every frontier lab, and this is the first visible evidence it carries a retention bill at the research layer.

• Liguori resigned this week; Kirsch resigned earlier in October; others reportedly preparing to leave
• April Pentagon contract bars autonomous weapons and unsupervised mass surveillance
• Departing staff cite weak governance rather than the contract's letter

  📰 Source: Google DeepMind Engineer Resigns in Protest of Pentagon Contract — The Intercept (Tier 2) 📰 Source: 'Ashamed to work at Google': DeepMind researchers quit over US military AI contracts — Ynetnews (Tier 2)

📌 IBM: Beats With Revenue Up 4.7% and Adjusted EPS Up 11% — Then Falls 5.3% Overnight, Because Software Did All the Work

IBM posted $17.10 billion of revenue, up 4.7%, with adjusted EPS up 11% to $2.95, and shares initially traded up to around $255 before reversing hard to $238.05 by 02:47 ET Thursday — a 5.33% decline from the regular-session close. The composition explains the reversal: software supplied roughly 77% of the entire year-over-year revenue increase, and AI work is about 30% of consulting revenue, which means the services franchise that is supposed to monetise enterprise AI deployment is not yet carrying the top line. IBM has been a favoured "AI without the capex" hiding place; today's tape says that trade needs consulting growth, not just software mix.

• Revenue $17.10B (+4.7% YoY); adjusted EPS $2.95 (+11%)
• Software ≈ 77% of the YoY revenue increase
• AI-related work ≈ 30% of consulting revenue
• Shares $238.05 overnight, −5.33% from the close after touching ≈$255

  📰 Source: IBM Q3 earnings report 2026 — CNBC (Tier 1) 📰 Source: IBM (IBM) Q3 2026 Earnings Call Transcript — The Motley Fool (Tier 2)

📊 Market & Investment Signals

💬 DeepSeek: Closes About $15 Billion With CATL and Tencent as Its Biggest Backers — Double Its Original ¥50 Billion Target — Ahead of a Possible Listing

The raise we flagged a week ago as "doubling to ¥100 billion" has now closed at roughly $15 billion, and the cap table is the news: battery maker CATL and WeChat operator Tencent were among the largest investors, alongside the previously reported SMIC, GigaDevice and Geely. That is China's industrial and consumer-internet champions jointly underwriting the country's frontier lab, with a public listing openly discussed. Demoted from top-story status under the dedup rule — the fundraise itself is not new, only its close and its investor list.

📰 Source: DeepSeek raises $15 billion in funding as it steps up AI competition with OpenAI, Anthropic — Bloomberg via Storyboard18 (Tier 1 origin) 📰 Source: DeepSeek raised about $15 billion ahead of a possible IPO — Quartz (Tier 2)

💬 Fireworks AI: In Early Talks for a Round at a $30–40 Billion Valuation

The inference-infrastructure layer is still clearing at venture multiples even as the debt-funded compute layer gets repriced — Fireworks AI is in preliminary discussions for new funding targeting $30–40 billion. Talks are early and the target may move; treat the range as an anchor, not a mark.

📰 Source: Fireworks AI in Talks for New Funding Round at Valuation of Up to $40B — PANews (Tier 2)

💬 iFlytek: Ships Spark X2.6 Trained Entirely on Domestic Chinese Compute, Claiming Code and Agent Breakthroughs Without a Single NVIDIA GPU

At its 1024 Global Developer Festival, iFlytek launched the X2.6 end-cloud model and made the claim that matters for export-control policy rather than benchmarks: it is the first to reach breakthrough code-generation and agent capability on fully domestic compute, where every rival Chinese model was trained on NVIDIA silicon. Also shipping this week: Alibaba's Qwen3.6 27B on October 21.

📰 Source: iFlytek Launches X2.6 End-Cloud AI Model Powered Entirely by Domestic Computing Resources — KuCoin News (Tier 2) ⚠️ Sourcing note: company claim relayed by Tier 2 outlets only — the "no NVIDIA" assertion is unverified by independent benchmarking.

💬 Memory: HBM4 Runs About $550 a Stack Against ≈$300 for HBM3E, With 93% of Combined Samsung/SK hynix/Micron Output Now Pointed at HBM

Standing background for the Korean tape: HBM3 is ≈$200 a stack, HBM3E ≈$300, and HBM4 ≈$550 for 36GB 12-high parts, while HBM consumes roughly three times the wafer capacity of standard DRAM per gigabyte. With the big three steering ≈93% of combined output to HBM, DRAM rose 90% in Q1 2026 alone and Gartner models combined DRAM/SSD up 130% by year-end. SK hynix holds 50–55% of HBM; Micron has exited consumer memory entirely. This is context for why equipment names fell harder than memory names in Seoul today, not a new datapoint.

📰 Source: HBM Pricing & Market Share (2026) — SK Hynix, Samsung, Micron — Silicon Analysts (Tier 2) 📰 Source: Samsung warns of memory shortages driving industry-wide price surge in 2026 — Network World (Tier 2)

💬 Dating Notes: Two Items Circulating in Today's Aggregators Are Not From Today

Flagged so they are not mistaken for fresh catalysts:

• Tencent's $7 billion, 100,000-GPU Oracle lease across Southeast Asian data centres — five-year term at roughly $1.60 per chip-hour, about 43% under H100 one-year contract rates, routing around export controls that bar purchase but not overseas rental. First reported October 1 (>7 days → ×0.3 recency), so it is context for the China-compute workaround, not a Thursday catalyst.
• ByteDance's talks to secure MediaTek AI chips for its data centres is likewise a multi-week-old thread with no new confirmation today.

  📰 Source: Tencent Reportedly Signs $7B Deal to Lease 100,000 AI Chips From Oracle in Southeast Asia — TrendForce (Tier 2), October 1 📰 Source: Tencent scores 100,000 offshore AI chip deal with Oracle for $7 billion — Tom's Hardware (Tier 2)

🎥 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.