Back to Blog
Galen Guan

Daily Briefing · 2026-10-29: Fed holds steady, Google TPU goes external, AI infra financing hits a wall

📰 Daily Briefing · 2026-10-29 (Thursday)

🌐 Politics · Geopolitics

  • Fed holds rates steady as expected; two voting members push for a hike — The federal funds target range stays at 3.75%-4%, in line with consensus. Lisa Cook and Beth Hammack dissented, preferring a 25bp increase at this meeting. Chair Warsh said the September hike is starting to work but inflation remains too high — "there is no soft inflation target." He also announced the Fed will restructure its meeting calendar from 2027, with some sessions no longer carrying rate votes. No forward guidance on the path ahead; December's statement wording is the next watchpoint. CLS
  • Bank of Canada hikes for the first time in over three years, to 2.5% — The overnight rate rises to 2.5%, the first hike in 3+ years, as expected; the bank signaled further tightening may follow. The stated driver: oil price gains from the Middle East conflict feeding through to broader goods and services prices, with inflation not expected back at 2% until mid-2028. Central banks globally are pivoting from easing to fighting energy inflation. CLS
  • Houthi attack on Saudi Arabia's largest refinery; Brent up to $115.70 — Brent crude rose 1.62% to $115.70/barrel after Wednesday's strike; Saudi officials reported intercepting three ballistic missiles and four drones, while damage remained unconfirmed. Trump's latest comments on the Middle East and Israel's election result both hint at further escalation. US media report Washington has told several Gulf states to prepare for "a brief but intense military operation" after the midterms (Nov 3), and a Patriot battery is being moved from Texas. CLS
  • European natural gas tops €100/MWh for the first time in nearly four years — The benchmark contract broke €100 on Wednesday, a first since December 2022; storage sits at ~75%, the lowest for this time of year on records dating to 2009. Europe's LNG import spend will rise at least $12 billion (+18%) this year. Middle East supply disruption plus low inventories are reviving winter energy-crisis fears — and adding to ECB hike pressure. CLS
  • Trump: "AI" is a terrible term and will be banned, in favor of "SI" — On Tuesday Trump said he will ban the term "artificial intelligence" from "all relevant documents" and everywhere else, criticizing the WSJ for not using his preferred "superintelligence (SI)." No details on enforcement; the practical policy impact is questionable, but it shows the White House continues to politicize the AI narrative. CLS
  • Meta and Snap face a landmark teen-addiction trial — The two companies head to a key Los Angeles trial over alleged harm to teen mental health; the case was chosen as a "bellwether." The verdict could shape litigation strategy and settlement talks across hundreds of similar cases — a landmark test of platform liability. CLS

💰 Finance · Markets

Snapshot (US close Oct 28 EDT; A-shares Oct 28 close): S&P 500 7743.98 (+0.28%) · Nasdaq 26549.85 (+0.27%) · Dow 52475.10 (+0.19%) · Shanghai Composite 3681.47 (+0.91%) · CSI 300 4083.87 (+0.85%) · USDCNY 6.6660

  • Oracle guides to free-cash-flow positive in 2H FY2028; stock jumps 10.73% — Oracle expects FCF to turn positive in the second half of FY2028 and to exceed $30B in FY2029, far above expectations, while capping new financing through FY2030 at $20B. The guidance directly eased fears that AI capex would swallow the company's cash, lifting AI-cloud names broadly; Microsoft's and Google's cloud results, also out the same day, beat as well. CLS
  • AI infrastructure financing hits a wall: bond yields into double digits as Wall Street buyers retreat — Yields on bonds and loans funding AI data centers keep spiking into double-digit territory, and banks find it increasingly hard to offload risk to insurers and other investors. Citadel Securities warned clients that even the largest tech firms could get dragged into the funding squeeze. This is the first systemic signal of the "AI spending wave meeting a financing bottleneck" — worth tracking closely. CLS
  • Anthropic IPO filing: $16.3B H1 revenue, up 1,300% YoY, but Q3 growth slows — Per the filing, Q3 revenue growth decelerated to 47% quarter-over-quarter. OpenAI's momentum may be accelerating, with a consumer business (ads + hardware) forming a differentiated moat; Anthropic concedes a consumer gap but doesn't plan to chase it. A commercial divergence between the two leading labs is taking shape. CLS
  • Vertiv: revenue up 61%, organic orders double YoY, full-year guidance raised — The liquid-cooling leader beat on both Q3 revenue and adjusted EPS; the CEO says product development cycles are compressing from years to months to serve an accelerating AI buildout. Liquid cooling and power gear are the "picks and shovels" of AI capex — doubled orders are hard data confirming the boom. SK Hynix separately announced a KRW 100-trillion shareholder-return program. CLS
  • Gold defends the $4,000 level; State Street still sees $5,500, with lower odds — State Street Global Advisors maintains its Q1-2027 forecast of $4,750-$5,500/oz but acknowledges the probability has fallen. Mideast hedging demand and central-bank buying support gold, but if the Fed pivots back toward hikes on energy inflation, rising real rates will cap upside. CLS

🚀 Tech · AI

  • Google TPUs go on external sale at scale: Oracle becomes the first external hyperscaler — Oracle will offer Google TPUs with deployments expected to start in 2027; analysts project $62.5B-$99B in Google TPU sales for 2027 at ~20% operating margins. Anthropic has committed to using up to 1 million Google TPUs, and OpenAI is reportedly planning purchases too. NVIDIA's near-monopoly on AI compute faces its first structural competition. CLS
  • Amazon ships five AI home devices at once; 16-inch Echo Show 16 takes on Apple's HomeView — The Alexa-based lineup spans speakers, clocks, digital frames and smart displays; the flagship Echo Show 16 costs $499.99, going head-to-head with Apple's $299 HomeView. Smart home is becoming the next big-model battleground, with screen size and AI capability escalating together. CLS
  • US close: indices edge up, AI cloud leads — S&P +0.28%, Nasdaq +0.27%, Dow +0.19%; Oracle's cash-flow guidance lifted AI cloud stocks, Microsoft's and Google's cloud results beat, and memory/liquid-cooling names extended the tech rebound. The combination of steady rates and beats is, for now, outweighing funding-side worries — but the surge in AI infra bond yields is the medium-term hedge signal to watch. CLS

📄 Papers · Research

(72h window; latest arXiv cs.AI batch is Oct 27)

  • ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning — Proposes letting VLA models reinforce chain-of-dynamics reasoning via "world imagination," easing generalization bottlenecks from sparse action-label diversity. Directly relevant to modeling "dynamical causality" in video generation and embodied AI; useful for architecture decisions in video understanding pipelines. arXiv
  • Understanding the limits of latent world models — A systematic evaluation of where latent world models break down on long-horizon temporal and spatial reasoning, arguing current benchmarks over-weight task performance over reasoning range. A sober empirical calibration for the "can world models be agent brains" question. arXiv
  • VideoLVR: Latent Visual Reasoning over Grounded Video Evidence — Points out that language-based chain-of-thought offers no verifiable video evidence, and proposes latent-space, evidence-grounded video reasoning. Relevant to video-generation verifiability/evaluation — a candidate idea for video quality assessment pipelines. arXiv
  • Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops — Models relation composition in attention via lattice gauge theory, where entity relations interact only through "Wilson loops," explaining order sensitivity in multi-hop reasoning. Heavily mathematical, but offers a geometric lens on why LLMs fumble multi-hop reasoning — good blog material. arXiv
  • Timestral — A family of four time-series foundation models (5M-209M params) setting a new SOTA in probabilistic forecasting. Not core to the beat, but the "small model family + probabilistic forecasting" open-source play is worth noting. arXiv

🧠 Thinkers

  • No new posts in the past 24h from the watchlist (Lilian Weng / Chip Huyen / Baoyu / Zhang Junlin / Mu Li): Lilian Weng's latest remains "Harness Engineering for Self-Improvement" (2026-07-04); Baoyu's latest is the Oct 20 NotebookLM Video Overview system-prompt teardown; Chip Huyen's feed still ends in Jan 2025. Reported honestly: no new viewpoints within 24h.

📝 New Blogs (24h)

  • Strictly within 24h: no hits. Closest over-window item for reference: Cognition · FrontierCode 2.0 (published Oct 27 PDT, ~44h ago) — a new generation of their coding agent; and Devin Security (Oct 26, ~69h) on automatically finding and fixing critical vulnerabilities. Both fall outside the window and are excluded from the formal list. Cognition Blog
  • Anthropic Engineering's latest visible date remains Oct 20 (SPA page, nothing within 24h); Manus's last update is late July; DeepSeek's blog is an SPA with no evidence of new posts on Oct 27; no reliable new stories extracted from Jiqizhixin or Shanghai AI Lab within 24h.

⭐ GitHub Trending

(Search API, 3-day window with stars:>100, strictly hand-filtered)

  • anhao/ai-pose-gallery-zh (HTML · ⭐1818 · created 10-27) — A fully localized Chinese AI pose reference gallery: 1,439 sets / 2,878 images, zero-dependency static site. A practical offline asset library for composition control in AI image/video generation; 1,800+ stars in two days shows real demand. repo
  • andyhuo520/digital-human-koubo (TypeScript · ⭐1299 · created 10-26) — Generate AI digital-human talking-head videos from one photo: Renoise MCP generation + 5 editing styles + 48 creator-grade packaging templates. Digital-human avatars are among the fastest-monetizing AI video niches; the MCP toolchain shape is worth studying. repo
  • Motionsites-AI/claude-skill (Python · ⭐866 · created 10-26) — A free Claude Skill that makes Claude build award-level, cinematic landing pages. The Skill ecosystem is shifting from "productivity tools" to "aesthetic output" — structurally similar to Copilot Skills. repo
  • patch-gg/anymove (Python · ⭐340 · created 10-27) — Motion capture for any creature from a single AI video: two cameras in one frame, a DINOv2 joint reader, automatic rig fitting. Single-video mocap addresses a real downstream need in 3D pipelines and composes naturally with ComfyUI video workflows. repo
  • konlu/story-studio (Python · ⭐375 · created 10-26) — An all-in-one AI short-video creation Agent Skill, from topic selection to JianYing (CapCut) drafts. The "Agent Skill + editor draft" landing shape maps closely onto real video-production workflows. repo
  • yaojingang/GEOInflow (TypeScript · ⭐688 · created 10-26) — Self-hosted Generative Engine Optimization analytics: track visitors and AI-search referral sources. The SEO→GEO shift now has a toolchain; content site owners (blogs included) should start watching AI-citation traffic. repo

(Another ~60% of results were filtered as low-value: the wildrig-multi GPU miner, a descriptionless .github repo, a calendar design template, the watson OSINT people-search tool, etc. — spam or off-topic.)

📡 Radar

  • No new terms captured today (ledger unchanged).

🔭 GitHub Watch · Watchlist

  • Comfy-Org/ComfyUI: evaluate integrating BFL Flux Video Upscale V2's target-resolution mode into the video super-resolution chain (priority 3 · adapt before adopting) — PR #17394 (Oct 28) adds a target_resolution DynamicCombo to FluxVideoUpscaleNodeV2: fixed 1080p/2K/4K target boxes replacing the Legacy node's 1.5-3.0x factor-only mode; auto orientation switching for portrait/landscape, auto_downscale when fit<1.5, scale clamped to 1.5-3.0, ~14.4MP per-frame cap, 24fps output. WHY/IMPACT: users think in "output to 1080p/4K," not "upscale 2.3x," and target-resolution mode handles portrait natively. If the video post-processing/export chain has any scale-factor-based logic, it should converge to target resolutions instead. Recommend adapting and integrating. repo
  • The other 7 watchlist repos (crewAI, diffusers, langgraph, langfuse, autogen, promptfoo, dspy) each had 1 commit yesterday; no new board tasks.

References