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

The Fed opens a live meeting into a 5% ten-year and a chip bear market, while the AI safety fight stops being an essay and starts being an institution — and the President calls the whole thing a hoax.

🔥 Today's Top Stories

🔥 The Tape: The FOMC Opens With ≈93% Odds of the First Hike Since 2023 — the 10-Year Prints 5.041% and Money Rotates Out of Silicon Into Security

The two-day FOMC meeting began this morning with fed funds futures pricing roughly a 92–93% chance of a hike tomorrow — which would be the first increase in over three years. The 10-year Treasury printed 5.041% Tuesday morning after briefly touching 5% on Monday, its highest since 2023, and crude held $102–103 on the Saudi East-West pipeline closure and fresh Houthi attacks. For AI equities this is the first genuine discount-rate event of the cycle: the entire complex is priced off terminal-value cash flows that a 5% risk-free rate mechanically compresses.

• Dow futures −0.27%, S&P 500 futures −0.16%, Nasdaq futures −0.11% into the open
• Chairman Kevin Warsh's press conference and the dot plot are the actual event — the hike itself is already in the price
• The selling is not uniform: money left semis and went into cybersecurity as a block (see Market Signals below)
• Crude at $102–103 means the inflation print the Fed is hiking against is not yet rolling over

  📰 Source: Stock market today: Live updates — CNBC (Tier 1) 📰 Source: Dow, S&P 500, Nasdaq futures retreat ahead of Fed meeting amid AI safety fears — Yahoo Finance (Tier 2)

🔥 Donald Trump: Calls AI Risk "a HOAX" and a "SICK Conspiracy" — Then It Emerges He Met Altman Privately in Dallas

Yesterday this report carried the President's "whoever wins AI wins" line; today the position hardened into something markets have to price. Trump posted that "AI taking over the World, destroying Humanity, and all other things bad, is a HOAX," framed guardrail advocates as running a "SICK conspiracy" against AI and data centers, and phoned Jensen Huang live on stage at the All-In Summit on Monday to repeat it to a room of investors. Separately, it emerged that Trump met Sam Altman privately on the sidelines of the Republican midterm convention in Dallas last week — neither the White House nor OpenAI would describe the conversation.

• The explicit target is the Amodei–Altman–Musk pacing coalition: the White House reads a slowdown as a gift to China
• "Don't kill the Golden Goose" — the framing ties AI policy directly to the data-center capex cycle
• Huang has separately dismissed extinction scenarios as "fiction," putting the largest AI company on the President's side of the argument
• Investment read: federal guardrails are now unlikely this Congress, which pushes the regulatory risk premium onto state law (California SB 813) and onto private commitments like AEF-1

  📰 Source: Trump calls AI risks a 'hoax,' says there is a 'SICK conspiracy' against AI and data centres — Bloomberg (Tier 1) 📰 Source: Trump calls AI risks a 'hoax,' says there is a 'SICK conspiracy' against AI and data centers — ABC News / AP wire (Tier 2)

🔥 OpenAI: Buys Glass Imaging for More Than $300M — the Team That Built the iPhone's Portrait Mode Now Works for Sam Altman

OpenAI has acquired Glass Imaging, a computational-photography startup, for north of $300 million — roughly 3× the ≈$100M valuation it carried in its last round. Glass was founded in 2019 by Ziv Attar and Tom Bishop, both ex-Apple engineers who led the team behind Portrait Mode, and sells GlassAI, a neural image signal processor that does demosaicing, color reconstruction, denoising, sharpening and frame fusion in a single learned pass while modelling the specific lens it sits behind. Read alongside the $6.5B io Products deal and the subsequent Apple hiring raid, this is OpenAI assembling a camera stack, not buying an app.

• Valuation step-up: ≈$100M last year → $300M+ today
• GlassAI is trained per camera module — it is a hardware-coupled asset, which only makes sense if there is hardware coming
• OpenAI has not said what role Glass plays post-close
• Competitive angle: the buyer is hiring away the exact capability that differentiates Apple's phone franchise

  📰 Source: OpenAI Said to Buy Startup Glass Imaging for $300 Million — The Information (Tier 1) 📰 Source: OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says — TechCrunch (Tier 1)

🔥 Euclyd: Raises $230M With Samsung Co-Leading — a Two-Year-Old Dutch Startup Claiming 100× the Inference Efficiency of Vera Rubin, Chaired by ASML's Former CEO

Euclyd, founded in 2024, closed a Series A of more than €200M (≈$230M) co-led by Samsung, Somerset Capital Partners, and the EQT/Innovation Industries-managed Scaleup Europe Fund. Peter Wennink — who ran ASML through the EUV monopoly years — is chairman. The company is designing a non-GPU inference system, co-designing processor and memory architecture together, and claims 100× higher power efficiency for inference than NVIDIA's latest Vera Rubin generation. For a Series A that claim is unverified; what is verifiable is who is writing the cheque.

• Samsung is now funding an NVIDIA inference rival while selling NVIDIA the HBM that makes Vera Rubin work
• Wennink's involvement gives a two-year-old startup credible access to the European equipment supply chain
• This is the third memory-first, HBM-skeptical inference architecture to raise big this month (cf. Positron's $875M at $5B)
• The thesis across all of them is identical: inference economics are a memory-bandwidth problem, not a FLOPs problem

  📰 Source: Samsung backs Nvidia AI chip rival in $230 million funding round as GPU alternatives boom — CNBC (Tier 1) 📰 Source: Chip Startup Euclyd, Backed By Ex-ASML CEO, Raises More Than €200 Million — Bloomberg (Tier 1)

🔥 MediaTek: Ships the Dimensity 9600 Pro — the First Mobile Processor on TSMC's 2nm Node, and the First Real Proof N2 Is in Volume

MediaTek launched the Dimensity 9600 Pro this morning from Hsinchu, the first mobile SoC built on TSMC's 2nm process and the first company to get to that threshold in a shipping phone chip. The node step from N3 to N2 delivers a 14% performance gain and a 23% power reduction at the chip level; the CPU is a 2+3+3 all-big-core layout on Arm's C2 cores, with two C2-Ultra cores to 4.55GHz. MediaTek also introduced the Dimensity 9600M on N3 for the broader flagship tier.

• The investable fact is not MediaTek's — it is TSMC's: N2 is in volume production with a customer willing to put its flagship on it
• 23% lower power at constant work is the number that matters for on-device agentic inference, which is thermally bound
• MediaTek positions the part for 30B-parameter on-device models with ≈51% higher throughput than its predecessor
• Context: MediaTek is also NVIDIA's NVLink Fusion partner, making it one of the few names levered to both edge and rack-scale AI

  📰 Source: MediaTek Launches New Mobile Chip Using TSMC's Most Advanced Technology — Reuters wire, via Deccan Chronicle (Tier 1 wire, syndicated reprint) 📰 Source: MediaTek announces Dimensity 9600 Pro built on 2nm process and ARM's C2 cores — Gizmochina (Tier 2)

🌏 Asia Desk — the AI supply chain overnight

SK Hynix (NVIDIA's lead HBM supplier): closed 1,712,000 (+1.72%) · Samsung Electronics (HBM challenger + foundry #2): closed 250,500 (+0.80%) · Hanmi Semiconductor (HBM TC bonder equipment): closed 225,000 (−0.44%) · TSMC: closed NT$2,385.0 (+0.21%) · USD/KRW: 1,357.9 (+0.98%).

Seoul did not follow Wall Street. US semis sold off hard on the Amodei slowdown essay and the rate shock, and Korea's two memory names still closed green — the KOSPI complex is pricing physical scarcity rather than narrative risk, with DRAM and HBM tightness expected to persist into 2027 and pricing power intact regardless of what the pacing debate concludes. The weak spot was equipment: Hanmi's −0.44% against Hynix's +1.72% says the market is paying for existing HBM output, not for the next capex wave. TSMC's modest +0.21% came on the same day MediaTek validated N2 in volume, and the won weakening another 0.98% to 1,357.9 flatters the exporters' Q3 translation. The read-through for the US open: if the memory bid is real, Micron is the cleanest expression, and the divergence between memory (bid) and the broad semi tape (offered) is the trade to watch through the Fed.

👤 Key People Updates

📌 Bilal Chughtai (ex-Google DeepMind): Resigns From the Safety Team and Says Publicly That AI "Has the Potential to Kill Us All"

A DeepMind AI safety and alignment researcher, Bilal Chughtai, published an exit post on X and LinkedIn saying the technology he worked on "has the potential to kill us all." He has moved to BlueDot Impact, a nonprofit that trains people across fields in AI safety, and says he remains optimistic that the technology can be navigated safely. He is the second frontier-lab safety researcher to leave loudly in a week, after Anthropic's Jacob Coxon.

• Chughtai left DeepMind in July 2026; the public statement is what is new
• The pattern now spans all three US frontier labs — this is no longer a single-company culture story
• Alphabet is the only one of the three that has not made a public pacing commitment

  📰 Source: Google DeepMind Staffer Says AI May 'Kill Us All' in Exit Post — Bloomberg, via Yahoo Finance (Tier 1)

📌 Bernie Sanders and Steve Bannon: Share a Stage Today at the "Pro-Human Assembly" — and Jacob Coxon Pulls Out at the Last Minute

Sen. Bernie Sanders and former Trump adviser Steve Bannon appeared together today at the Future of Life Institute's "Pro-Human Assembly," backing a declaration calling for protection of human agency, avoidance of concentrated power, and corporate accountability. Ex-Anthropic researcher Jacob Coxon — whose resignation was covered here on Sept 10 — was expected and withdrew on Monday, telling Axios that appearing alongside Sanders and Bannon "could be easily misinterpreted." Bannon's line: "The Oligarchs are finally being forced to admit they've lied from the beginning about the dangers of acceleration with zero guardrails."

• A left-right coalition is the only configuration that moves AI legislation against a White House calling the risk a hoax
• The declaration's "concentrations of power" language is antitrust-shaped, not safety-shaped — that is the durable political threat to the hyperscalers
• Coxon's withdrawal shows the safety camp is actively trying not to be absorbed into partisan politics

  📰 Source: Scoop: Coxon, Bernie and Bannon unite for "pro-human" AI — Axios (Tier 2)

🏢 Big Tech Updates

📌 AI Evaluator Forum: xAI, OpenAI and Anthropic All Cosign AEF-1 — the Amodei Essay Converts Into a Standard in Under Four Days

Three frontier labs have signed AEF-1, the AI Evaluator Forum's "Minimum Operating Conditions for Independent Third Party AI Evaluations" — a voluntary standard covering access and resources, conflicts of interest, analytic autonomy, transparent publication, and handling of sensitive information. Alongside it, Anthropic made a unilateral commitment to embedded external reviewers with "desks in our offices, access badges, and company laptops," and permissions roughly comparable to internal risk teams, with the right to publish findings without Anthropic's editorial control. This is the first concrete, checkable artifact to come out of the pacing debate rather than another op-ed.

• AEF-1 itself dates to December 2025; the lab cosignatures are the new development
• METR, Transluce and SecureBio have already published against the checklist; the European AI Office has endorsed key provisions for GPAI Code of Practice compliance
• Investment read: voluntary standards signed by all three labs become the de facto compliance baseline regulators later codify — it front-runs the rulemaking the White House won't do

  📰 Source: Minimum Operating Conditions for Independent Third Party AI Evaluations — AI Evaluator Forum (Tier 2, primary) 📰 Source: AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign — Latent Space (Tier 2)

📌 Google: Gives Every Engineer Access to Claude Opus 5 Through Antigravity — After Years of Banning Outside Coding Tools

Google has opened Anthropic's Claude Opus 5 to all of its engineers via Antigravity, its internal development platform, reversing a policy that had barred outside coding tools like Claude Code and Codex outside a handful of DeepMind teams and priority projects. Access is quota-limited per user and Google says "Gemini remains our primary and foundational model for internal development." The subtext is unmissable: the company that builds Gemini is paying a competitor so its own engineers can ship code faster.

• Previously restricted to select DeepMind teams and a small number of high-priority engineering projects
• Engineers had been pushing internally on the grounds that Gemini trails frontier models specifically on programming
• Revenue angle: Alphabet is both Anthropic's largest strategic investor and now, internally, its customer
• The competitive signal for Gemini's enterprise coding pitch is worse than the dollar amount involved

  📰 Source: Google gives all of its engineers Claude Opus 5 access via Antigravity — Hugh Langley / Business Insider, via Techmeme (Tier 2)

📌 Apple: Ships iOS 27 With the Rebuilt Siri and On-Device Violence Detection in FaceTime

Apple released iOS 27 today, delivering the rebuilt Siri across platforms — the feature conspicuously absent from John Ternus's first keynote a week ago — plus automated gore and violence detection in FaceTime and expanded Safari parental controls. The safety features are shipped as on-device classifiers, which is Apple's standing differentiation argument against cloud-inference rivals.

• The Siri rebuild is the single deliverable Apple's AI narrative has been carrying since 2024
• On-device content classification in a real-time video call is a meaningful compute lift for the A20 Pro

  📰 Source: Apple releases iOS 27 with Siri AI across platforms — The Verge, via Techmeme (Tier 2)

📊 Market & Investment Signals

📌 Cybersecurity: CrowdStrike +14%, Qualys +15%, Zscaler +14%, Palo Alto +13% — the AI-Risk Trade Finds Its Long Side

While semis sold off, the security complex ripped as a block: CrowdStrike +14% to $236.61, Palo Alto +13% to $372.76, Zscaler +14%, Qualys +15%, Okta +11%, Fortinet +8%. The uniformity across names with very different fundamentals — CrowdStrike and Palo Alto both ~+98–99% YTD, Zscaler −18% YTD — says this is category allocation, not stock picking. The thesis: you cannot short the risk Amodei described, but you can own the companies enterprises pay when they believe it. CrowdStrike CEO George Kurtz publicly rejected the slowdown framing even as his stock rallied on it.

📰 Source: Cybersecurity Stocks Surge as AI Safety Warnings Spark Security Bid — 24/7 Wall St. (Tier 2) 📰 Source: Cybersecurity Stocks Rally as AI Safety Debate Heats Up — Benzinga (Tier 2)

💬 Z.AI (Zhipu): Raises $5B in Its Second Round in Two Months — With Zero-Coupon Converts Priced 25% Above the Placement

The GLM developer raised ≈$2B placing 21.97M shares at HK$714 (about 10% below the prior close) and ≈$3B in RMB 20.14B of convertible bonds due September 2027, zero coupon, yielding between −0.5% and zero, converting at HK$892.50 — roughly 25% above the placement price. Sixty percent goes to next-generation GLM training and inference infrastructure. This is the third equity raise since its January Hong Kong listing and follows a ≈$4B placement in July.

• Investors accepting a negative yield for the conversion option is a direct market vote on Chinese frontier-model capex

  📰 Source: China's Z.AI raises $5 billion through share, convertible bond sales — Reuters, via Business Standard (Tier 1 wire)

💬 Cornelis Networks: Raises $205M Led by IAG Capital and Launches "Active Compute Fabric" With Qualcomm — Another Bet Against NVLink

Cornelis closed $205M led by IAG Capital Partners alongside the launch of Active Compute Fabric, an architecture pairing programmable in-fabric compute with lossless transport to cut congestion in large accelerator clusters. It comes with a Qualcomm collaboration on rack-scale networking, and Cornelis is building its scale-up interconnect on the open UALink and ESUN standards rather than NVLink. The interconnect layer keeps repricing as an AI trade — the Corning/Verizon fiber deal a week ago made the same point at the optical layer.

📰 Source: Cornelis Networks raises $205M and scales up and out with its new Active Compute Fabric — SiliconANGLE (Tier 2) 📰 Source: Cornelis to Build Scale-Up Interconnect with UALink and ESUN — HPCwire (Tier 2)

💬 Exein: Raises $270M at $1.7B Led by Headline — Europe's Most Valuable Cybersecurity Scaleup, and Its Code Already Ships on 2 Billion Chips

The Rome-based firm raised an oversubscribed $270M at $1.7B led by Headline, with Sofina, Goldman Sachs, the EIB Group's ETCI, KfW Capital and T.Capital joining. Exein builds embedded security for "physical AI" — and its technology is already shipping inside roughly 2 billion chips, with NVIDIA, AWS, MediaTek, Arm and Leonardo as customers. The valuation has gone $400M (July 2025) → ≈$800M (December 2025) → $1.7B in fourteen months; first foundation models are guided to Q1 2027.

📰 Source: Exein raises $270M at $1.7B valuation, claims title of Europe's most valuable cybersecurity scaleup — Tech.eu (Tier 2)

💬 NVIDIA: SemiAnalysis Measures Vera Rubin NVL72 at ≈7× Blackwell's Token Throughput per Megawatt on Agentic Workloads

On replayed production-style agentic-coding sessions against DeepSeek V4 Pro's 1.6T model, Vera Rubin NVL72 delivered up to 7× better token throughput per megawatt than Blackwell on pre-release software — NVIDIA's own marketing claims up to 30× versus GB300 on the AgentX benchmark at 160 tok/s/user, and CoreWeave measured 10× on single-turn DeepSeek R1. The spread between those numbers is the whole story: efficiency gains are extremely workload-dependent, which is exactly the opening challengers like Euclyd and Positron are aiming at.

📰 Source: Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis — SemiAnalysis (Tier 2)

💬 China: State Media Calls the US Slowdown Push "Self-Serving" — the Second Official Rebuttal in Two Days

Chinese state media accused US AI executives of using pacing calls to blunt Chinese competition and preserve American dominance. It follows Monday's Foreign Ministry "fearmongering" line and adds no new policy content — but it confirms Beijing will treat any coordinated Western slowdown as a containment instrument, which caps how far the pacing coalition can realistically go.

📰 Source: Chinese State Media Dismisses 'Self-Serving' AI Slowdown Call — Bloomberg (Tier 1)

💬 Shanghai AI Laboratory: Ships Atria Dawn Preview With No Blog, No Paper and No Pricing — a 744B Agentic MoE That Tops Five of Sixteen Benchmarks, MIT-Licensed

InternLM/Shanghai AI Lab quietly pushed Atria Dawn Preview to Hugging Face: a 744B-total, 8-active-expert MoE built on the GLM-5.2 foundation, 256K context, text-only, BF16 and FP8 checkpoints, MIT licence — with no announcement post, no paper, no API and no pricing. Across 16 benchmarks spanning tool use, research, workspace tasks, software and terminal engineering, ML engineering and cybersecurity it reports the highest published score on five. The training method is a "Verifiable Experience Pipeline" that grounds every task in a live execution environment. An MIT-licensed 744B agentic model appearing unannounced is the practical argument against any unilateral Western slowdown.

📰 Source: internlm/Atria-Dawn-Preview — Hugging Face (Tier 2, primary) 📰 Source: Atria Dawn Preview: InternLM's 744B Agentic MoE Ships Quietly — OrcaRouter (Tier 2)

💬 Anthropic: Claude Code Weekly Limits Fall 17% as the Summer Boost Expires — Announced as a "Permanent 25% Raise"

From September 14, Anthropic set Claude Code weekly limits permanently 25% above the original baseline for Pro, Max, Team and Enterprise — which is a 17% cut from what users actually had, since a temporary 50% boost running since May (extended four times) has now lapsed. On a base of 100 the sequence is 100 → 150 → 125. Users criticised the framing for leading with the gain rather than the net reduction. Minor in isolation; notable as evidence that inference capacity is still the binding constraint on agentic coding products — at the company telling everyone else to slow down.

📰 Source: Anthropic is cutting Claude Code's current weekly limits by 17% — BleepingComputer (Tier 2)

💬 Google DeepMind: Publishes the AlphaGenome Atlas — Precomputed Predictions for All 9 Billion Possible Single-Letter Human DNA Changes

DeepMind released a roughly one-petabyte repository of precomputed AlphaGenome predictions covering every possible single-nucleotide change in the reference human genome, across thousands of molecular effects and hundreds of human and mouse cell types, with a single-number variant impact score. Free for academic use; commercial licensing routes through Google Cloud — which is the actual business model signal, and the reason it belongs in an investment report at all.

📰 Source: AlphaGenome Atlas: Molecular predictions for 9 Billion human DNA variants — Google DeepMind (Tier 2, primary) 📰 Source: New Google DeepMind atlas could transform our understanding of genetic diseases — Scientific American (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.