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

Europe finally ships a trillion-parameter open-weight model, Google locks 3.6 gigawatts of PJM power, and four frontier labs tell New York City lawmakers under oath that they cannot quantify the risk of a catastrophe — while the 30-year Treasury prints its highest yield since 2002 and futures pull back from Tuesday's records.

🔥 Today's Top Stories

🔥 Mistral: Ships "Le Chonk" — Mistral Large 4, a 1.05-Trillion-Parameter MoE With 49B Active, Trained on 3,800 Grace Blackwells in Its Own European Data Centres, Open Weights Later This Month

Mistral released Mistral Large 4 — internally and now publicly nicknamed "Le Chonk" — a 1.05-trillion-parameter mixture-of-experts model with roughly 49 billion active parameters, multimodal input and reasoning. The company claims it "significantly outperforms" any open-weight model developed in the US or Europe and trades blows with the Chinese open-weight frontier (DeepSeek, Alibaba, Moonshot, Z.AI). For investors, the load-bearing detail is the training footprint: ≈3,800 Grace Blackwell GPUs across ≈52 NVL72 racks, inside Mistral's own European data centres — the first credible demonstration that a European lab can train at frontier scale on sovereign infrastructure rather than renting US hyperscaler capacity.

• 1.05T total parameters, ≈49B active in the MoE routing — chosen explicitly to keep serving cost low; API went live Oct 6, open weights land on Hugging Face at the end of October (Oct 27–31 reported)
• Trained on a corpus spanning 160+ languages; Mistral pitches it hardest at cyber, coding, manufacturing, finance and multimodal work
• ⚠️ Independent measurement is cooler than the marketing: Artificial Analysis ranks it between DeepSeek V4.1 Flash and OpenAI's GPT-6 Luna — strong, but not the outright open-weight lead Mistral claims
• Mistral's own framing on cyber is a direct swipe at closed-model gatekeeping: "provider-level refusals can block legitimate vulnerability research and incident response," while open deployment "gives organizations both the capability and the autonomy to run advanced security work under their own policies"

  📰 Source: Mistral unveils new AI model it says rivals best open systems from China — CNBC (Tier 1) 📰 Source: European AI flag bearer Mistral's new open weights model is 'Le Chonk' — The Register (Tier 2) 📰 Source: Mistral unveils le Chonk, says "significantly outperforms" any open-weight models developed in US or Europe — Tech.eu (Tier 2)

🔥 Google: Signs 3,590 MW With Constellation — 890MW of Nuclear on a 20-Year PPA That Funds $4.3B of Uprates at 11 Reactors, Plus 2.7GW of Merchant PJM Supply for 15 Years

Google contracted 3,590 MW from Constellation Energy inside PJM, America's largest grid. The structure matters more than the headline gigawatts: an 890MW, 20-year power purchase agreement that underwrites $4.3 billion of Constellation capital spending to uprate 11 existing nuclear units in Illinois, New Jersey and Pennsylvania — plus a separate 15-year, 2.7GW agreement that is not tied to any particular plant and simply keeps Constellation's existing fleet delivering energy and capacity into the PJM market. Google is no longer just buying clean electrons; it is financing capacity expansion at an incumbent generator and backstopping merchant supply in the grid where AI load growth is most acute.

• The $4.3B of uprate investment is the real asset creation — new megawatts from existing reactors, with the first upgraded plant delivering extra power in 2028
• PJM spans 13 states including Virginia and Pennsylvania, the two densest data-centre markets in the US, and expects up to 30GW of peak load growth through 2030
• The 2.7GW unsited tranche is a quiet admission about the constraint: there is not enough new firm generation to contract, so hyperscalers are now paying to keep existing plants from retiring
• Constellation is now the counterparty on back-to-back hyperscaler deals — Amazon took 690MW from the same company days earlier

  📰 Source: Google enters massive 3.6-GW power deal with Constellation Energy — Yahoo Finance (Tier 2) 📰 Source: Google signs 3.6GW deal with Constellation to secure power in PJM service area — Data Center Dynamics (Tier 2) 📰 Source: Google Signs 3.6 GW Power Deal With Constellation Energy, Including 890 MW of Nuclear Power — Technology.org (Tier 2)

🔥 OpenAI, Anthropic, Google and Meta: Testify Under Oath at New York's City Hall and Refuse to Guarantee Their Agents Will Obey Safeguards — While Musk's SpaceXAI Simply Defies the Subpoena and the Council Heads to Court

All 51 members of the New York City Council sat as a Committee of the Whole on October 5 and put four frontier labs under oath. Every one of them declined to guarantee that their AI agents will always comply with safety guardrails, telling members that eliminating risk or promising perfection is not possible. The exchange that will be quoted for months: after repeated non-answers on legal liability for financial harm or death, a council official said, "I'm going to take it then that neither of the four of you, no company here, can quantify the risk of something cataclysmic happening." This is the first time the industry has been forced to say that on a public record, under oath — and it lands six weeks before Anthropic's IPO.

• The labs sent policy staff, not principals: Anthropic's Frontier Red Team head Logan Graham, OpenAI policy development head Morgan Dwyer, Google AI policy director Alice Friend, Meta AI legislation policy director Shane Cahill
• SpaceXAI did not appear at all — in direct violation of a subpoena issued Sept 28. Speaker Julie Menin says the Council is pursuing enforcement in New York State Supreme Court
• Three municipal bills are live: a whistleblower incentive programme, a private right of action letting New Yorkers sue for AI-caused harm, and mandatory independent third-party validation before deployment
• Outside witnesses were harsher than the companies: an ex-OpenAI researcher told members that current safety fixes may be "duct tape that will fall off," and a whistleblower argued it is "more likely than not that humanity loses control"
• Investor read: a private right of action at the city level is the tail risk here — it converts model behaviour from a regulatory matter into open-ended civil exposure, and New York is where the financial counterparties live

  📰 Source: Anthropic, OpenAI, Google, Meta execs testify NYC Council AI hearing — CNBC (Tier 1) 📰 Source: Ex–OpenAI researcher tells NYC lawmakers AI safety fixes may be 'duct tape that will fall off' — Fortune (Tier 2) 📰 Source: Elon Musk's SpaceXAI violates NYC subpoena: Speaker Menin — PIX11 (Tier 2)

🔥 Lambda: Raising up to $4B at a $14.5B Pre-Money Led by Blackstone and Coatue — After Its Order Backlog Went From $15B in June to $50B in September on a Single $35B Anthropic Commitment

NVIDIA-backed GPU cloud operator Lambda is closing a final pre-IPO round of up to $4 billion at a $14.5 billion pre-money valuation, led by Blackstone and Coatue, with a 2027 listing targeted in a letter to limited partners. The number that should stop investors is the backlog: $15 billion in June, $50 billion by September. Most of that delta is one customer — a reported $35 billion cloud commitment from Anthropic. A neocloud whose contracted revenue triples in a quarter because one pre-IPO AI lab signed a take-or-pay is the cleanest expression yet of how concentrated the AI infrastructure credit chain has become.

• $4B raise at $14.5B pre-money, Blackstone and Coatue leading — both now repeat players across AI compute and data-centre credit
• Backlog $15B → $50B in roughly three months; the Anthropic commitment alone is reported at $35B
• IPO targeted for 2027, explicitly contingent on execution and market conditions
• The counterparty question is the whole investment case: Anthropic has not yet listed, booked a $41.97B net loss in its leaked prospectus, and has over $518B of compute obligations. Lambda's backlog is only as good as that credit

  📰 Source: Lambda raises up to $4 billion ahead of planned 2027 IPO, WSJ reports — Investing.com, reporting WSJ (Tier 1 origin) 📰 Source: Nvidia-Backed Cloud Computing Firm Lambda Looks To Raise $4B In Final Round Ahead Of IPO — Yahoo Finance (Tier 2)

🔥 Lisa Su (AMD): Says From Taipei That 2027 Brings a "Substantial" Supply Increase, Admits Demand Still Exceeds What AMD Can Ship, and Reveals AMD Now Plans Capacity 3–5 Years Out Instead of 1–2

AMD's CEO said in Taipei that the company will substantially increase chip supply in 2027 while conceding that demand is still running ahead of shipments, with "very, very high demand for the next several years." The structurally important disclosure is the planning horizon: AMD has moved from booking wafer and packaging capacity one to two years ahead to three to five years ahead. That is a supply chain where the binding constraint has shifted from demand visibility to physical CoWoS and advanced-packaging allocation — and it is the same signal Micron sent with its jump from $22B to $32B of long-term supply commitments in a single quarter.

• EPYC Venice server CPUs are reported sold out through 2027 with orders pushing into 2028
• Su called CoWoS and advanced packaging "absolutely critical"; AMD committed over $10B to its Taiwan supply chain in May and says that figure will grow
• Third Taiwan trip of the year — meetings with TSMC, Foxconn, Quanta, Acer and ASUS
• She flies to South Korea later this week to sit with Samsung and SK Hynix on memory — directly into Samsung's Oct 8 preliminary print and the HBM4 pricing question

  📰 Source: AMD plans to substantially increase supply in 2027, CEO says — Reuters via Euronext (Tier 1) 📰 Source: AMD CEO Lisa Su Says 2027 Will Bring a "Substantial" Supply Increase — TechPowerUp (Tier 2) 📰 Source: AMD CEO Lisa Su Says AI Demand Will Stay 'Very, Very High' for Years — Benzinga (Tier 2)

🌏 Asia Desk — the AI supply chain overnight

SK Hynix (NVIDIA's lead HBM supplier): 1,715,000 (−3.87%) · Samsung Electronics (HBM challenger + foundry #2): 269,000 (−1.47%) · Hanmi Semiconductor (HBM TC bonding equipment): 264,000 (−5.38%) · TSMC (fabs the GPUs themselves): 2,585.0 (0.00%) · USD/KRW: 1,339.2 (−0.34%)

Seoul sold the memory complex for a second straight session, with the KOSPI closing down 1.98% at 6,803.90 — and it sold the equipment harder than the chipmakers, with Hanmi down 5.4% against SK Hynix's 3.9%. Korean press reported foreign investors dumping a net ₩2.60 trillion of KOSPI shares on the day, absorbed almost entirely by retail. That ordering tells you this is positioning ahead of Samsung's preliminary Q3 print tomorrow, not a demand signal: equipment names get cut first when the market fears HBM4 ramp timing slips, and nothing in the session contradicted AMD's message from Taipei a day earlier that memory demand still exceeds supply. TSMC was unchanged in Taipei.

For the US open: this is a Seoul-specific earnings-risk trade, not a read-through to NVIDIA or Broadcom — but it is a live negative for Micron and Western Digital/SanDisk sentiment, since Korean memory weakness has repeatedly transmitted to US memory names in the same session this year. Watch whether Lisa Su's Seoul meetings later this week produce any HBM4 commentary; she is the one buyer whose public tone can reset that pricing debate before Samsung's full report in late October.

👤 Key People Updates

📌 Sam Altman (OpenAI): Says the World "Should Accept Some Bad Things Happening" for AI's Benefits — and Draws Bipartisan Fire From Bernie Sanders and Ron DeSantis in the Same Week OpenAI Is Marketing a $30B Round

In an October 4 interview with Politico's Decoded, Altman said "one of the biggest differences between us and some of the stricter, let's say, AI safety people, is we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency." The backlash crossed the aisle within 48 hours. He later clarified that he does not accept "the really catastrophic risks," including a serious loss of control to AI — but the clarification is not what is circulating, and the timing is unfortunate: it landed one day before four labs told New York's City Council they cannot quantify catastrophic risk, and in the same week OpenAI is pitching a $30B round.

• Florida Gov. Ron DeSantis: "And a handful of tech oligarchs get to make that decision for the rest of us? No dice."
• Sen. Bernie Sanders publicly asked what "bad things" the public is supposed to accept
• Gary Marcus: Altman is "saying the quiet part out loud: suck it up, so you can make us rich and powerful"
• Investor relevance: this is the sentence that gets read into the record at the next federal hearing, and it raises the odds that the FTC's civil investigative demands against OpenAI and Anthropic expand in scope

  📰 Source: Sam Altman says 'some bad things' will happen with AI. That doesn't mean we should stop it. — Reason (Tier 2) 📰 Source: 'Freaking Insane': Sam Altman Sparks Horror by Saying We Should 'Accept Some Bad Things Happening' for Sake of AI — Common Dreams (Tier 2)

🏢 Big Tech Updates

📌 Microsoft and NVIDIA: Jensen Huang Joins Satya Nadella on Stage in San Francisco in Hours to Launch RTX Spark and the Surface Laptop Ultra — Pitched on Running 120B-Parameter Models Locally at 80 Watts

Microsoft holds its first dedicated Windows and Surface launch since the 2024 Copilot+ PC event today at 10am Pacific, with NVIDIA's CEO on stage alongside Satya Nadella and Windows/Surface chief Pavan Davuluri. The product is RTX Spark — Grace Arm CPU cores fused to Blackwell RTX graphics with up to 128GB of unified LPDDR5X — inside the Surface Laptop Ultra. The strategic claim is the one to watch: that a laptop drawing up to 80 watts can run 120-billion-parameter models and agentic workflows entirely on-device, with no cloud round trip. If that lands, it is the first serious commercial argument that inference demand is not purely a data-centre line item.

• RTX Spark top configuration: 20 Arm CPU cores, 6,144 Blackwell GPU cores, up to 128GB unified memory, ≤80W
• Surface Laptop Ultra is positioned at AI developers and content creators — an explicit MacBook Pro attack, and Microsoft's first genuinely NVIDIA-silicon Surface
• ⚠️ Everything above is pre-event reporting; the keynote has not happened at publication time, so treat specs and framing as expectations rather than announcements

  📰 Source: What to expect at Microsoft's Windows, Surface and AI event on October 7 — Engadget (Tier 2) 📰 Source: Microsoft confirms October 7 Windows and Surface event with Jensen Huang, RTX Spark launch expected — VideoCardz (Tier 2)

📌 Anthropic: Opens Mythos-Class Cyber Capability to Verified Red Teams With Reduced Restrictions — Three Tiers, and Red Team Access Performs "On Par With Unrestricted Use" at 34 of 50 Assessed Tasks

Anthropic merged Project Glasswing and its earlier Cyber Verification Program into a single three-tier CVP that gives vetted security professionals access to Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 with cybersecurity guardrails deliberately loosened. Defense Access covers SOC work, incident response, malware reverse engineering and vulnerability analysis; Red Team Access covers authorised penetration testing; Specialized Access is for verified organisations testing systems where failure threatens lives or markets. The disclosed benchmark is the notable part — Red Team Access completed 34 of 50 assessed tasks, which Anthropic says is on par with unrestricted use. That is the company deliberately shipping near-unrestricted offensive capability nine days after publishing a threat analysis warning that Zhipu's open-weight GLM-5.3 does the same thing with safeguards that break 64–100% of the time.

• Access is conditioned on data retention — pen testers trade log visibility for capability
• Models covered: Claude Mythos 5.1, Opus 5.5, Sonnet 5.5, plus new models as they ship
• Strategic logic: if open-weight Chinese models already provide the capability without guardrails, refusing verified Western defenders costs safety without buying any
• Sits awkwardly beside the Pentagon's statement last week that it has "ceased the use of Anthropic products" — six weeks before the IPO

  📰 Source: Anthropic on X: expanding our Cyber Verification Program — Anthropic (company, Tier 2) 📰 Source: Anthropic Opens Claude Access to Red Teams and Verified Cyber Defenders With Reduced Restrictions — Cybersecurity News (Tier 2) 📰 Source: Anthropic Opens Mythos-Class AI to More Defenders — Technology.org (Tier 2)

📌 OpenAI: Abu Dhabi's MGX Leads a UAE Consortium Weighing up to $10B of the $30B Round, With BlackRock in Talks — and the $1.4T Pre-Money Is Being Presented as a Fixed Price With No Lead Investor

OpenAI is in talks with Abu Dhabi's MGX and other UAE funds, which are discussing forming a consortium to write as much as $10 billion of the $30 billion round, with BlackRock also in discussions. The structural detail that matters: OpenAI is presenting the $1.4 trillion pre-money as a fixed price despite not having settled on a lead investor — a term sheet posture only available to a company that believes capital is more abundant than allocation.

• Last priced round was March: $122B raised at an $852B valuation. $1.4T is a ≈64% mark-up in roughly seven months
• Sovereign wealth is now the marginal buyer of frontier AI equity — the same pattern as SoftBank's $64.6B position
• IPO remains pushed to at least 2027 on stated AI-safety grounds, which is precisely why the private round has to be this large

  📰 Source: OpenAI In Talks For US$30 Billion Round With UAE Funds, BlackRock — Business Today, reporting Bloomberg (Tier 1 origin) 📰 Source: OpenAI in Talks With UAE Funds, BlackRock for $30 Billion Funding Round — The Elec (Tier 2)

📌 AWS: Files to Build 36 Data Centres Across 1,100 Acres at Pennsylvania's Homer City Campus — Powered by a 4.5GW Gas Plant on the Site of a Retired Coal Station, Phased From December 2026 to 2036

Amazon filed a Preliminary Land Development Plan for 36 two-storey data centre buildings at 225,000 sq ft each across six sites on 1,100 acres in Indiana County, Pennsylvania, roughly 50 miles east of Pittsburgh. Power comes from a 4.5GW natural gas facility under construction by Homer City Redevelopment on the footprint of a 2GW coal plant that ran from 1969 to 2023. Construction runs in four phases from December 2026 to a January 2032 final start, with full completion targeted June 2036 — a ten-year build that tells you how hyperscalers are now underwriting power rather than waiting for the grid.

• 36 buildings, two storeys, 70 feet tall, 225,000 sq ft each, plus substation and redundant on-site generation
• Gas, not nuclear — the honest counterpoint to the same day's Google–Constellation nuclear headline
• 1,500–2,000 permanent jobs at $80,000–$100,000 average salaries, per the filing

  📰 Source: AWS plans 36-building data center campus in Indiana County, Pennsylvania — Data Center Dynamics (Tier 2)

💬 Anthropic: Gives Startups a Free Year of Claude Team With Five Premium Seats Plus $1,000 in API Credits — Launched at SF Tech Week, Six Weeks Before the IPO

Anthropic expanded Claude for Startups with a free year of Claude Team (up to five premium seats), $1,000 in API credits and access to Claude Marketplace for building plug-ins. Read it as pre-IPO developer-funnel spend: the cheapest way to show a listing roadshow that seat growth and API adoption are both compounding.

📰 Source: Anthropic is giving startups a free year of Claude Team and $1,000 in credits — TechCrunch (Tier 1)

📊 Market & Investment Signals

💬 Samsung Electronics: Reports Preliminary Q3 Tomorrow Into Consensus of ₩106.1T Operating Profit — a Near Nine-Fold Jump — With Goldman Flagging Oct 8 as a Quadruple-Convergence Volatility Event

Samsung's preliminary Q3 lands tomorrow with consensus operating profit around ₩106.1 trillion (≈$79B) against ₩12.17 trillion a year earlier, on revenue near ₩201.7 trillion. Crossing ₩100 trillion would be a South Korean corporate record. The asymmetry is unkind: a print at or just above consensus can still disappoint if there is any softness on memory pricing or HBM4, and the preliminary release carries only headline revenue and operating profit — no segment detail, no HBM commentary, no call until the full report late in October. Goldman flags October 8 as a volatility flashpoint where preliminary earnings, ETF rebalancing, options expiry and buyback completion all land together.

📰 Source: Samsung Q3 earnings preview: record profit expected but margin outlook in focus — Investing.com (Tier 2) 📰 Source: Samsung Stocks Face October 8 Earnings Test After Record Memory Profit — Yahoo Finance (Tier 2)

💬 The Tape: The 30-Year Treasury Prints 5.70% — Highest Since 2002 — and Nasdaq-100 Futures Give Back 0.6% the Morning After the S&P Closed at a Record 7,818.93, With September Fed Minutes Due Today

Dow futures slipped 0.5%, S&P futures 0.2% and Nasdaq-100 futures 0.6% on Wednesday morning as global bonds fell, one session after the S&P 500 set a record close at 7,818.93 and the Nasdaq Composite logged its 23rd record of 2026. The 30-year yield rose to 5.70%, the highest since 2002, ahead of minutes from the September FOMC — the meeting at which policymakers raised rates to fight inflation. Megacap tech has been absorbing higher long yields all week on the argument that it is comparatively hike-resistant; today tests whether that holds when the minutes spell out the hiking bias in the committee's own words.

📰 Source: Stock market today: Dow, S&P 500, Nasdaq futures fall after tech rally as Treasury yields rise — Yahoo Finance (Tier 2) 📰 Source: KOSPI slides nearly 2% as Samsung earnings caution weighs — Aju Press (Tier 2)

💬 Waymo: Upsizes Its Debut Private Loan From $3B to $5B With Pimco, Blackstone and Sixth Street at 525bp Over the Benchmark — Its First Real Move Away From Equity Funding

Alphabet's robotaxi unit took its first-ever private credit deal from an initial target above $3 billion to $5 billion, with commitments from Pimco, Blackstone and Sixth Street, arranged by Goldman Sachs and priced at 5.25 percentage points over the benchmark. Waymo raised $16 billion of equity at a $126 billion valuation earlier this year; the shift to debt for a capital-intensive international expansion is the signal — private credit is now funding the physical AI buildout on the same terms it funds data centres, and at spreads that assume the growth holds.

📰 Source: Waymo Taps Pimco, Blackstone for $5 Billion Loan to Accelerate Growth — Bloomberg (Tier 1) 📰 Source: Waymo upsizes debut private loan to $5 billion to fund expansion — Investing.com (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.