Back to Blog
Galen Guan

Morning Briefing · 2026-08-28

📰 Morning Briefing · 2026-08-28 Friday

🌐 Politics & Geopolitics

  • Senate passes Graham's Russia sanctions bill — handing Trump sweeping tariff powers — The late Sen. Lindsey Graham's Sanctioning Russia Act cleared the Senate 86-11 and now heads to the House. The core mechanism: punitive tariffs of up to 100% on China, India, and other buyers of Russian oil and gas, aimed at strangling Moscow's war chest. Zelenskyy took a rare seat in the chamber for the vote. Next steps are murky: Speaker Mike Johnson's Republican majority is reluctant to wade into overseas affairs, and House Democrats led by Rep. Meeks attack the bill as "a massive backdoor authority" for presidential tariffs with no congressional check — more than 30 senators had backed a failed amendment to limit those powers. AP News

  • Russia says it may strike British military targets — On August 27 Russia's foreign ministry warned it could hit UK military targets "inside and outside Ukraine" in response to Kyiv's use of British long-range missiles, calling the UK "one step away" from complicity in terrorism and threatening "catastrophic consequences." The UK responded it stands "shoulder to shoulder" with Ukraine. The same night, Russia launched a massive barrage across 9 Ukrainian cities — 2 Oniks anti-ship missiles, 258 drones, and ballistic missiles; Ukraine says it downed ~90% of the drones but its Patriot interceptor stockpile is critically low, and Zelenskyy pressed European allies for more air defense. Five drones strayed into Moldovan airspace. Reuters

  • Qatar's PM shuttles to Tehran as the US-Iran "economic war" escalates — Qatar's PM and foreign minister visited Tehran on August 27 to revive a phased framework: a temporary joint Iranian-Omani shipping corridor through the Strait of Hormuz plus a joint mine-clearing operation. The White House says no negotiations are underway and pressure will continue — from "Operation Epic Fury" (military strikes) to "Operation Economic Outcast" (sanctions strangulation). The war is nearing its sixth month (began February 28) and Hormuz remains the chokepoint: the IMO counts 70 incidents and 19 seafarer deaths since the war began. Iran's Revolutionary Guards claim a revenue-sharing deal with Oman is done; a senior Iranian source says details are still being worked out. This is the single biggest overhang for global oil. Reuters via ThePrint

  • China domestic policy: no reliable new policy landings within the 24h window — No verifiable first-party policy releases from the State Council or PBOC surfaced for August 27-28 in our sources. Leaving this slot honest rather than backfilling old news. (Context: the PBOC pledged "timely new policy rollout" earlier in August, per Reuters.) Reuters

💰 Finance & Markets

Snapshot: S&P 500 6,439.32 (-0.43%) · Nasdaq 21,449.29 (-0.22%) · Dow 52,224.64 (up, led by Nvidia) · US 10Y 4.66% (8/26) · DXY ~99.1 · Shanghai Composite 3,951 (+1.13%) · CSI 300 4,630 (+0.86%) · USDCNY 6.7227 (8/26). VIX/gold: not reliably captured today.

  • Nvidia Q2 FY2027: revenue doubles to $96.2B, guides $108B next quarter — Revenue for the quarter ended July 26 hit $96.22B (+106% YoY, +18% QoQ), beating consensus by ~$4B. Data center revenue was $89B (+117% YoY) — 93% of the company. Non-GAAP EPS of $2.22 came in ~6% above estimates. Guidance: $108B (±2%) for the current quarter and ~70% revenue growth in FY2028 — a pace CFO Colette Kress says is supply-constrained, not demand-constrained. Jensen Huang's framing: "AI has reached its inflection point. Its tokens are productive and profitable. Now, compute is revenue." Notable details: Vera Rubin began production shipments in August and should be ~20% of next quarter's data center revenue (the fastest ramp in company history); supply and capacity commitments jumped from $119B to $279B in one quarter; Nvidia disclosed up to $105B in guarantees on a 20-year Ohio data center lease with OpenAI as tenant (~1.5M GPUs and $150-200B in revenue per infrastructure generation); $26B returned to shareholders in the quarter; China data center compute is assumed at zero in the forward outlook. The stock rose over 5% after hours and closed the next day up 8.7% near $228. Fortune · Quartz

  • A-share optical networking rips higher as Shanghai reclaims 3,950 — On August 27 the Shanghai Composite closed +1.13% at 3,951, with the Shenzhen Component +1.50%, ChiNext +1.71%, and the STAR composite up as much as 3.17% intraday. Half-day turnover hit 1.35 trillion yuan, expanding by 120.8B, with over 3,000 stocks up. The leading thread was compute/optical networking: Changfei Fiber hit the daily limit (426.82 yuan, 238B yuan market cap, +267% YTD), with Hengtong Optic-Electronic, T&S Communication, Montage Technology, and CXMT all surging; several semiconductor names hit limit-up. The AI capex boom is transmitting straight from Nvidia's print into China's optical module and fiber supply chain. Sina Finance · Eastmoney

🚀 Tech & AI

  • OpenAI's first custom chip Jalapeño: 1.5-1.9x more work per watt than Blackwell — At Hot Chips (Aug 25-26), OpenAI published the first official benchmarks for Jalapeño, its Broadcom co-developed inference accelerator, on SemiAnalysis's InferenceX benchmark (third-party-observed, but all figures supplied by OpenAI): 1.5-1.9x more throughput per watt and 1.7-3.6x lower end-to-end latency versus Nvidia's GB200/GB300; on DeepSeek R1 it exceeded 700 tokens/s per user at a concurrency of one. Specs: TSMC N3P, 700W TDP (sustained ≤550W in tests), 128 accelerators per rack, 1.7 exaFLOPS of 4-bit compute, 27.5TB of HBM4, just under 2PB/s of bandwidth. Taped out November 2025, B0 stepping in production, deploying into OpenAI's own infrastructure by year-end — and not for sale. The big caveat: the comparison targets Blackwell, not the newer HBM4-based Vera Rubin, and even SemiAnalysis calls it "somewhat incomplete and unfair." Altman's summary on X: "we made a chip and it is fast." Analysts see real margin pressure in inference but no threat to Nvidia's training and general-purpose moat. Nvidia's blowout earnings the next morning confirmed the market's read: same week, both stories, GPU demand unfazed. CNBC · MLQ

  • Apple launches M6 and M5 Ultra: 2nm and quad-die, aimed squarely at local AI — M6 is Apple's first 2nm chip (12-core CPU, 12-core GPU, dual 16-core Neural Engine, 170GB/s unified memory bandwidth) in the new Mac mini from $899. M5 Ultra uses next-gen UltraFusion to fuse two M5 Max dual-die chips into Apple's first quad-die design — up to 36-core CPU, 80-core GPU, 1.2TB/s bandwidth — in the Mac Studio from $5,499, configurable to 512GB of unified memory (ships late October), enough to run very large LLMs entirely on one box. The software story matters as much as the silicon: since macOS 26.2, Thunderbolt 5 supports low-latency interconnects, and Mac Studio now officially supports RDMA clustering with memory pooling across machines — four clustered Studios deliver ~3x the inference throughput of one. A new Core AI framework joins MLX. This is the first time Apple has written distributed desktop inference into its product narrative — a genuinely new option for local-AI developers. Pre-orders opened Aug 25; availability Sept 22. Apple Newsroom · gHacks

  • DeepMind's talent grip slips: elite researcher retention lags rivals — Zeki data (tracking 20,900 research and advanced-engineering staff across 10 companies, compiled August 2026) puts Anthropic's elite AI talent retention at 22:1 and OpenAI's at 5.7:1 — with Google DeepMind significantly behind. Context: Gemini 3.5 Pro is months late, Demis Hassabis moved from CEO to chairman/Alphabet chief scientist in early August, and multiple Gemini co-leads have left for competitors. Signal for practitioners: the AI talent market is still repricing fast; elite-lab mobility is the norm and big-company halo premium is shrinking. Fortune

🧠 Thinkers & Research

  • Sam Altman on Bedrock Managed Agents: "intelligence is the one utility with no demand ceiling" — Stratechery published a joint interview with Altman and AWS CEO Matt Garman, timed with the deepening AWS-OpenAI partnership: OpenAI models and Codex on Bedrock, plus Bedrock Managed Agents powered by the OpenAI agent harness (limited preview). Altman's core points: agent tooling is "just way too hard to do," and tightly integrating harness and model will enable products that can't reliably be built today at all. On demand: intelligence is the only utility where he'd say "just give me more, as long as the price is low enough" — with a promise to keep cutting prices dramatically. Direct relevance for AI application engineers: managed harness-model integration (the opinionated counterpart to AWS's open AgentCore lane) is becoming a cloud battleground — add a "managed integration" column to your multi-model vs. self-built harness architecture comparison. Stratechery

  • Simon Willison: Claude Code Opus 5 Auto Mode can be broken out of its sandbox — A security writeup published August 27 (citing Johann Rehberger's research): Auto Mode has paths around its sandbox via prompt injection. A direct alarm bell for anyone wiring coding agents into production: treat agent permission models and injection defenses as a pre-deploy audit item — don't assume vendor defaults are safe. Simon Willison

  • No other significant new statements in the window — Amodei's "The Adolescence of Technology" (19,000 words) and his Dwarkesh podcast appearance both date to earlier this month — outside 24h with no new increment, so we're not backfilling. Karpathy, LeCun, and Jensen Huang (whose earnings-call comments belong to the Nvidia item above) had no major independent statements.

📝 Fresh Blog Posts (24h)

  • Cloudflare · How we saved 100TB of memory by optimizing 1.1.1.1's DNS cache — Published Aug 27. An engineering postmortem on how cache and data-structure trade-offs reclaimed 100TB across the world's busiest public DNS resolvers. Worth a close read if you run infrastructure. Post

  • GitHub Blog · OpenClaw went viral. Meet the maintainers building and securing it — Published Aug 27. An interview with the maintainers of the viral open-source agent project OpenClaw on supply-chain security and community governance after hypergrowth — useful for anyone maintaining or contributing to OSS. Post

  • Simon Willison · Qwen3.8-Flash-Next — Published late Aug 26 (UTC), at the edge of the window. Hands-on notes on Alibaba's newest Flash-tier model, with Simon's usual capability-boundary assessment. Post

  • Vercel · Cursor in the AI SDK harness layer / Run Claude Managed Agents with Chat SDK — Two posts on Aug 27-28: Cursor's editor capabilities land in Vercel's AI SDK harness layer, and Chat SDK can now run Claude Managed Agents. Landing in the same week as AWS's Bedrock Managed Agents — the "harness layer" is becoming the industry's platform flashpoint. AI app architects should take note. Post

⭐ GitHub Trending

  • tt-a1i/archify (JavaScript · ⭐23,038 · +4,260 today) — An agent skill for beautiful, verifiable architecture/workflow/sequence/data-flow/lifecycle diagrams: self-contained HTML with motion and crisp export. #1 on the daily overall board. Solves the specific pain of "make my agent draw engineering diagrams" — a new breakout in doc automation, complementary to the Claude Code/OpenClaw agent ecosystem. repo

  • calesthio/OpenMontage (Python · ⭐52,331 · +1,284 today) — Billed as the world's first open-source agentic video production system: 12 production pipelines, 100+ tools, 700+ agent skill and production-knowledge files. Highly relevant if you work in AI video — an open-source reference for agentic video workflow design worth dissecting pipeline by pipeline. repo

  • DietrichGebert/ponytail (JavaScript · ⭐114,000 · +1,610 today) — "Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote." Its virality is counter-cyclical: while the industry piles on agent capabilities, this project argues the agent's first job is cutting scope. repo

  • bilawalsidhu/gods-eye-view (JavaScript · ⭐7,882 · +1,984 today) — A spy satellite simulator in your browser, except the data is real: live open-source spatial intelligence on a photorealistic 3D globe. A phenomenon-grade demo of remote-sensing/OSINT data visualization — #2 by daily gain. repo

  • freestylefly/awesome-gpt-image-2 (JavaScript · ⭐22,970 · +2,093 today) — "Prompt as Code": an industrial-grade GPT-Image2 prompt engine and template library, reverse-engineered from 530+ cases, 20+ industrial templates, distilled into Skills. From the Chinese community — a systematic treatment of image-generation prompt engineering, relevant to visual creation toolchains. repo

  • K-Dense-AI/scientific-agent-skills (Python · ⭐35,290 · +494 today) — Turn any AI agent into an "AI Scientist": 163 ready-to-use validated skills, claiming 175,000+ scientist users. A flagship of the agent-skills ecosystem spreading from coding into the scientific domain. repo

References