Back to Blog
Galen Guan

Daily Briefing · 2026-09-30: Anthropic Files for IPO as OpenAI DevDay Ships Three Releases

📰 Daily Briefing · 2026-09-30 Wednesday

🌐 Politics · Geopolitics

  • Trump meets AI industry leaders, opposes new federal safety regulation — On Sep 29, Trump met with AI industry leaders and came out against new federal AI safety regulation, signing AI agreements with major tech companies and weighing a dedicated committee. Tuesday's White House luncheon gathered executives from six major labs — a clear signal that federal oversight is loosening under the current political cycle, with consequences for release cadence and compliance costs. Sina Finance · CLS
  • US to release 40 million barrels from strategic reserve — but stocks are at a 44-year low — Washington announced a 40M-barrel SPR release to fight fuel price spikes and urged Europe to follow. The catch: the reserve itself sits at its lowest since 1982, so the leverage left is shrinking. CLS · Sina Finance
  • US consumer confidence falls to a 12-year low — The Conference Board's September reading dropped to levels last seen in 2014, with consumers turning pessimistic on both the economy and jobs; the ongoing US-Iran conflict is amplifying inflation worries. CLS · Sina Finance
  • UK PM Burnham pledges to explore new options for UK-EU relations — Burnham publicly committed to "exploring new options" with the EU, leaving the door open to a path back — the most notable softening of the UK's post-Brexit stance to date. Sina Finance
  • Pentagon picks Boeing for next-generation fighter — The US sixth-generation fighter award goes to Boeing, with Lockheed Martin out. A major rebalancing of military aerospace orders and a significant endorsement of Boeing's defense line. Sina Finance

💰 Finance · Markets

Snapshot (US close Sep 29, captured morning Sep 30): S&P 500 7670.84 (-0.17%) · Nasdaq 26797.54 (-0.09%) · Dow 51349.92 (-0.26%) · Shanghai Composite 3830.45 (+0.18%) · CSI 300 4345.21 (+0.10%) · USDCNY 6.7065 · US 10Y 5.293% (per CLS reporting, not available via Sina) · Brent $102.59 (-2.56%, per CLS)

  • 30-year Treasury yield hits 5.619%, highest since June 2002 — The bond selloff broke through another threshold: the 30-year peaked at 5.619% intraday and the 10-year rose to 5.293%, near 2007 highs. Rising yields pushed the dollar to a two-month peak while the euro fell to a 16-month low. Barclays floated a 6% fair-value forecast for the 30Y. Rate-hike expectations actually cooled — CME FedWatch puts the odds of an October hike below 50% — but long-end yields keep climbing on inflation and supply pressure. CLS · Sina Finance
  • Anthropic files IPO prospectus: 2025 revenue up 12x, valuation could top $2 trillion — The filing shows 2025 revenue grew 12x year-over-year to nearly $4.6B, with operating losses widening from $2.98B to $8.06B. Of the ~$42B net loss, roughly $34B was non-cash fair-value accounting on convertible instruments. The company ended the year with $20.28B in cash and short-term investments, spent $7.33B on compute and infrastructure (over half its $12.65B total opex), and plans $518B of future cloud and compute investment. The risk factors are equally striking: nearly a quarter of revenue came from two customers, and Anthropic explicitly warns its models could exhibit self-protective behavior. 47% of 2025 sales flowed through Amazon and Google channels. CLS · Sina Finance
  • OpenAI ARR nears $70B; new round of at least $30B at up to $1.4T valuation in early talks — Sources say OpenAI's annualized recurring revenue has approached $70B, up more than 70% since the start of Q3, with enterprise revenue doubling in three months and consumer additions in Q3 alone exceeding all of 2025. Separately, OpenAI is in early discussions with investors for a round of at least $30B. CLS · Sina Finance
  • China rolls out mortgage interest subsidies: some first-home buyers save up to ¥50K per ¥1M loan from October — Three ministries issued rules letting buyers authorize banks to process the subsidy at loan signing; the PBoC also cut rates and raised quotas on policy tools targeting weak sectors. The aim is to ignite the "Golden October-plus-One" sales window — the most concrete demand-side property stimulus in months. CLS · CLS
  • Brent crude falls 2.56% to $102.59 as Saudi Arabia boosts pipeline flows — Oil had been accelerating after a pullback from early-month highs (peace-deal hopes with Iran fading), then gave back gains Tuesday as Middle East crude exports recovered. Oil remains the biggest wild card in the inflation outlook. CLS · Sina Finance

🚀 Tech · AI

  • OpenAI DevDay ships three: Dot agent + GPT-6.1 Sol + Ultrafast tier — At its annual developer day OpenAI launched: (1) Dot, an always-on agent powered by GPT-6 Astra that can operate a computer and pull data from connected apps for research, drafting and development — reachable via ChatGPT, Slack and Teams, limited to one per user at launch (OpenAI engineers already use it to fix dozens of bugs daily); (2) GPT-6.1 Sol, arriving just a week after its predecessor; (3) the Ultrafast premium speed tier; (4) a preview of "OpenAI Private Intelligence." Dot adds a $500/month tier. It's a direct answer to Meta's Muse, SpaceX and Instinct in the increasingly crowded agent market. CLS · Sina Finance
  • AMD acquires Fei-Fei Li's AI startup for $8.2B — AMD is paying $8.2B for Li's company, adding a key piece to its AI software and spatial-intelligence portfolio — another attempt to route around NVIDIA's ecosystem moat. Sina Finance
  • Meta launches Muse, an AI agent for small businesses — Meta's Muse targets marketing and customer-service automation for SMBs. Altman says he doesn't feel threatened by Zuckerberg but concedes Muse "seems like a good product." CLS · Sina Finance
  • Samsung: HBM to approach 30% of global DRAM capacity next year; $1B more for AI infrastructure — Samsung executives expect HBM to account for nearly 30% of global DRAM capacity by 2027, squeezing conventional DRAM supply; the company is also adding $1B to AI infrastructure (with NVIDIA participating) and refocusing foundry on AI/HPC, targeting roughly 4x customer growth by 2029. CLS · CLS
  • Bain: a revenue gap of over $4 trillion looms behind the AI infrastructure boom — Bain calculates a gap of more than $4T between the infrastructure spending promised across the AI stack and the revenue it can realistically generate — one of the most pessimistic quantifications yet of the AI capex cycle. CLS

🧠 Leaders · Views & Research

  • Altman on regulation, hardware and liability at DevDay — Altman said the industry's liability framework will be "a layered thing," warned against being "too optimistic" about Washington regulation, hinted that hardware "worth waiting for" may arrive soon, and called nothing as serious as the attack on Hugging Face. Sina Finance
  • Jensen Huang and Lisa Su join Tsinghua SEM's advisory board — The heads of NVIDIA and AMD were both appointed to Tsinghua's School of Economics and Management advisory board — the day before AMD announced the Fei-Fei Li acquisition. Their pivotal positions in China's AI ecosystem are worth noting. Huang also defended NVIDIA's record buyback, calling the stock "almost irresistibly cheap." CLS · CLS
  • Cuban: humanoid robots will be proven a failure within five to ten years — Mark Cuban predicted the humanoid robotics space will be falsified in 5-10 years, standing squarely against the current startup frenzy — a useful contrarian datapoint. CLS
  • Fed's Williams: one more hike possible this year, but no rush — NY Fed President Williams said another hike may come this year to return inflation to target, but "there's no need to rush"; Governor Barr stressed more increases may be needed. The 2026 rate path stays hawkish. CLS

📝 New Blogs (24h)

No updates within 24 hours from the target blogs (Lilian Weng, Anthropic Engineering, Cognition, Manus, Chip Huyen, Baoyu, Su Jianlin, DeepSeek, Qwen); kexue.fm remains blocked (~110-byte responses). The latest arXiv cs.AI batch is from Sep 28 (within 48h, included for reference):

  • arXiv · Telescopic Language Models — Trains a single stretchable language model that serves multiple compute budgets, replacing a separate training or compression run per budget point. arXiv
  • arXiv · TokenCast: Forecasting Token Consumption During LLM Agent Execution — Token consumption by LLM agents on the same task can vary by over an order of magnitude; this work forecasts consumption during execution. arXiv

⭐ GitHub Trending

Created within the last 3 days (since 2026-09-27), sorted by stars, spam filtered:

  • KKKKhazix/AIHOT (TypeScript · ⭐3008) — A framework for sites that find their own trending topics and write their own digests: swap in your sources and curation rules and it becomes your industry news site. Structurally similar to this briefing pipeline; worth studying its source-ingestion design. repo
  • PostHog/jeeves (Python · ⭐256) — PostHog's tooling for Jev-like decision models: "Reasoning improves Jev-like decision models." repo
  • openJiuwen-ai/iCode (Python · ⭐235) — A lightweight, extensible, fully offline development platform and agent/workflow toolkit. repo
  • openai/mcp-extensions (TypeScript · ⭐165) — OpenAI's official library for building plugins that feel like native, first-class features of ChatGPT — a signal of MCP expanding toward the ChatGPT side. repo

(Filtered as spam/no technical value: an MLBB cloud-gaming script, a fan-art resource repo, undescribed CarrierSIM / floorplan-3d / IQuest-Q1, and the dubious firelex/jeff.)

🔭 GitHub Watch · Watchlist

(From github-watch pipeline tasks created Sep 29 on the ShuJieTai board; 3 items within 24h)

  • huggingface/diffusers: sage attention dispatch upgrade with Blackwell support — PR #14584 (Sep 29): attention_dispatch.py centralizes sageattention kernel variants and adds a sage blackwell hub backend with version-gate checks and fallback to native attention on import failure. SageAttention's int8 attention typically delivers 1.5-2x speedups on video DiTs with minimal quality loss; if the Wan/Flux pipelines still run default SDPA, there's a deterministic speedup on the table. (priority 3 · adapt) repo
  • crewAIInc/crewAI: borrow SummarizeMessages' tiered chunked retry for Copilot context compression — PR #7815 (Sep 29): when the context window overflows, summarization converges into a single SummarizeMessages class that retries chunked summarization at 4→3→2.5 chars-per-token tiers with concurrent gathering, using a fresh instance per compression. Directly applicable to magicedit Copilot's long-session management to avoid truncation losing edit intent. (priority 3 · adapt) repo
  • Comfy-Org/ComfyUI: tracking LynnReal light Minimax-H3 VAE support — PR #16657 (Sep 29, kijai): adds VAE loading support for the MiniMax H3 model family. kijai's model-support PRs usually signal ecosystem landing; if quality/speed benchmarks beat current backends, integration via the ComfyUI API is worth evaluating for the MagicCanvas pipeline. Watch item, no urgency. (priority 2 · track) repo

References