AI Morning Brief · Oct 2, 2026

SoftBank Group officially announced it executed the third and final $10B tranche of the $30B follow-on investment announced Feb 27, via Vision Fund 2; the funds came from foreign-currency senior notes issued Sept 24; cumulative investment reaches $64.6B for roughly 13%; it also cancelled the remaining $10B undrawn portion of the $40B March 27 bridge loan facility, which is now fully repaid.

Filing disclosures: the Broadcom–Anthropic relationship spans compute supply, equipment leasing and financing; the $42B loan is convertible into Anthropic equity and could cover about a third of its $125.2B five-year TPU lease commitment; Anthropic warns of a potential conflict of interest in Broadcom acting as both supplier and creditor; Anthropic could become Broadcom's largest chip-design customer next year. Rothschild called it "a classic case of reciprocal spending for Wall Street's AI skeptics".

Bans employers from firing or disciplining workers solely on automated-decision-system outputs ("No Robo Bosses Act" SB 947, effective July 1 2027), bans using AI to infer worker emotions from biometric data (AB 1883, also bans neural-data collection), and requires written notice for AI-driven mass layoffs; Newsom also signed an executive order directing state agencies to keep saying "artificial intelligence", rejecting Trump's directive that diplomats call it "superintelligence". Newsom publicly criticized the lack of comprehensive federal AI regulation.

One of the most comprehensive state-level AI laws in the US: AI subscription auto-renewals require written notice of key terms and written consumer consent; frontier-model company employees can anonymously report "catastrophic risks" (defined as a single event causing 50+ deaths/serious injuries or $1B+ in losses) with whistleblower protection; employers using high-risk AI in hiring/promotion/firing must do risk management and impact assessments, and "we used a third-party algorithm" doesn't transfer discrimination liability.

Newsmax reported: DeepSWE v1.1 77.9% (vs 74.1% for GPT-6 Astra), LVBench 91.7%, CWE-bench v1 68%, AutomationBench 51.3% (about 9 points above Claude Opus 5.5); Artificial Analysis Intelligence Index 53, tying GPT-6 Astra; Wiz used Argon under its "Scan for Good" program to find a critical vulnerability in a global hospital-software platform (exposing sensitive patient data); promotional pricing $2/$10 with 95% cache discount, rising to $4/$20 after promotion. Community skepticism: with no public API, it isn't really a launch.

A TypeScript extension mechanism: mods can run before/after or instead of events, rewrite prompts sent to the model, intercept/retry tool calls, approve or reject permission requests, redact sensitive data from tool outputs, and replace UI components; available in Claude Code CLI and desktop. Note: with great power comes great risk — a mod has full agent-level access to the user's machine (runtimewire relaying the X announcement, single source).

Barclays expects half its developers on Claude Code by year-end and full coverage by end of 2027; Claude is already used for legacy-platform modernization; the Global Markets business processes 120,000 customer emails a day; the Colleague Knowledge Assistant has served 16,000 UK staff and handled over 1M queries.

Needham analyst Laura Martin keeps Apple at Hold, saying the biggest investment risk for Apple shareholders is Meta (or another AI-first company) building an "AI agent + hardware + monetization" stack that bypasses the iPhone (tracing back to Apple's 2021 ATT policy, which cost Meta $10B in ad revenue); BofA's Wamsi Mohan (Sept 29 note) warns Apple could keep every phone sale yet lose the discovery/referral/transaction-initiation entry point. Apple closed down 0.8% at $330.48; Meta edged up 0.1% to $726.09.

Could list before the Nov 26 Thanksgiving break; prospective investor valuation range $1.8–2T; 2025 revenue ~$4.6B (vs $386M a year earlier), net loss near $42B. Note the contrast with Altman's no-IPO-in-2026 line (Aju Press, Oct 1: Altman says "no specific timeline", safety first).

BigGo Finance reported on Oct 1 that new tier names "PLUS / PREMIUM / SUPER / ULTRA" appeared in X app code, with an early unified-subscription build visible on Android; Barron's published a pricing-strategy analysis (SPCX ~$151, above the $135 IPO price). The four-tier plan itself remains a single Bloomberg-internal-document source, unannounced.

① SoftBank's final $10B tranche closes the story (see Top Story 1, officially confirmed). ② "Follow-up" on the $30B raise: CoinDesk on Oct 1, citing Bloomberg, says the annualized revenue run rate topped $40B in the summer, up 70% since July (secondhand, unverified). ③ No new developments in-window on the GPT-6.1 Astra cancellation (WSJ/CNN, Sept 28, already covered).

① Claude Code mods shipped (see Top Story 6). ② Barclays partnership deepened (see Top Story 7). ③ IPO roadshow timeline (see Top Story 9). ④ "Follow-up" on Claude for Government GA: the-decoder reported on Oct 1 it is targeting civilian agencies, while the Pentagon fight continues (the military still bans Claude). ⑤ A brief Claude platform credit-balance-delay glitch on Oct 1 was resolved in ~3.5 hours (pingoru, single source, minor).

Gemini 4 Argon follow-up (see Top Story 5); nile1 reported on Oct 1 that new lead Koray Kavukcuoglu said Gemini 4 is in post-training and already used internally on the Antigravity coding platform (aggregator relaying an interview, single source). No other new launches.

① Muse-vs-Apple analyst notes (see Top Story 8). ② The Muse macOS zero-day is confirmed as disclosed Sept 21 and hot-patched Sept 22 — outside this window, not repeated. ③ Muse for Small Business was already covered (Sept 29).

Grok+X subscription revamp follow-up (see Top Story 10); no new model releases in-window.

No significant additions in-window (the "negligence" remarks were covered Sept 30). Footnote: its early investor Headline closed a €400M European early-stage fund EU VIII on Oct 1 (tech.eu); Mistral is its flagship holding.

Follow-up analysis of the Huawei Ascend tooling open-source (webpronews, Oct 1) — the six modules (TileLang/DeepGEMM/DeepEP/TileKernels/FlashMLA/DeepSelect) mirror the Nvidia versions one-to-one; NYT's telling has TileLang as the core tool of DeepSeek's AGI research (secondhand); DeepSeek says every TileLang operator used in its training has a high-performance Ascend implementation (official claim, not independently verified).

Kimi K3 and GLM 5.3 joining OpenAI's B2B marketplace (ITHome, single source) — no new sources in-window, not repeated under the dedup rule; likewise no new developments on the Mindgard jailbreak (no BBC original or official response).

No significant additions in-window (National Day holiday factor in China).

- Rumored "NDRC ¥2T national AI datacenter plan with 80% domestic-chip quota" and "CAC banning unregistered generative-AI services from Sept 30": kimkj blog, single source, no credible origin — listed as unverified.

Unpacks why "$64.6B / ~13%" can't simply be divided to derive a valuation (cumulative cost basis ≠ valuation) — key to understanding the deal structure.

Analyzes the distribution logic behind SpaceXAI's unified four-tier subscription — Grok native to X is a distribution channel rivals don't have; the pricing war is really about early market share.

A survey of the regulatory era opened by the FTC rogue-agent probe, plus macro angles including the Bank of England's "AI asset-price correction" warning and Bain's "$6T of revenue needed by 2031 to fund data centers".

OpenAI's new $30B round at a $1.4T pre-money valuation: still a single Bloomberg source (secondhand via CoinDesk/CoinCentral, Oct 1); "annualized revenue topped $40B in summer, +70% since July" from the same source, unverified.

SoftBank pledged its OpenAI equity for a $10B margin loan (with margin-call triggers on sharp share-price falls): Aju Press, Oct 1, single source.

SpaceXAI's four-tier subscription plan: still a single Bloomberg-internal-document source; app-code tier names and the Android early build are BigGo Finance single-source.

Rumored "NDRC ¥2T national AI datacenter plan + 80% domestic-chip quota" and "CAC banning unregistered generative-AI services from Sept 30, 988 registered": kimkj blog, single source, no credible origin.

Micron FY2026 Q4 "quarterly revenue of $54.2B": kimkj's figure is clearly distorted (Micron's quarterly revenue is in the ~$10B order of magnitude); IBD Live on Oct 1 only noted strong guidance, without verifying specifics.

Manus raising $500M at a ~$4B valuation: still a single cryptobriefing source.

"A startup applied to build 456 reactors on Utah public land to power data centers": mentioned only in the Century Report, single source, no other origin.

Yesterday's unverified items on the Kimi K3/GLM-5.3 B2B entry, the Mindgard jailbreak BBC original, Muse's "5M downloads", Meta's R&D tax credits and Huawei's early Ascend 960DT: no new sources in-window, not repeated under the dedup rule.