The AI Weekly · Week of August 24, 2026
Agentic inference reshapes hardware, economics, and the gap between benchmarks and reality
- Agentic Inference Becomes the Battleground as CUDA Moat Faces Huawei and AMD ParityTwenty-six days. That is how long it took the SGLang team to close a 100x performance gap on DeepSeek V4 running on AMD's MI355X, a chip that at launch could only run FP8 inference and little else.
- The Benchmark-Reality Gap Widens as Cheap Open Models Top Leaderboards but Falter on Real TasksGLM-5.3 posted a perfect score—100% pass rate across 28 practical tasks—for $0.28. GPT-5.5 cost $1.43 to run the same suite.
- OpenAI Pauses Frontier RL Over Cyber Risk While Cutting PricesArticle — "OpenAI Pauses Frontier RL Over Cyber Risk While Racing to Own the Lowest Price on Every Task": Article — "OpenAI Pauses Frontier RL Over Cyber Risk While Racing to Own the Lowest Price on Every Task": Every sa…
- Anthropic's Revenue Rockets to $65B as Its Best Model Loses Users to Cheaper RivalsArticle — "Anthropic's Revenue Rockets to $65B Annualized as Its Best Model Loses Users to Cheaper Rivals": Two numbers describe the same company in July, and they point in opposite directions. Anthropic's annualized rev…
- China's Embodied AI Moves From WRC Demos to Pre-Orders and Real RevenueTwo numbers were on display at WRC 2026, and most readers will have filed them under the same heading. UBTech booked over 2 billion yuan in 2025 revenue. Sunway New Materials logged 210 million yuan for its Qiyuan robots…
Threads we are following
- Whether AMD's Helios rack-scale ramp can clear its SerDes/retimer bottleneck and translate serving-stack parity into system-level parity
- Whether routing hardens into a standard procurement layer or gets exposed as another untested scaffold masking model weaknesses
- Whether Astra-tier restraint survives once a competitor ships something that makes a pause commercially expensive
- Whether Opus 5 overtakes Fable 5 in Ramp's next monthly billing breakdown, and what that does to Anthropic's average prices ahead of its IPO
- Whether Qiyuan's 210 million yuan in advance receipts converts into recognized revenue after the promised September delivery.
One issue a week. Free.
sift. reads hundreds of sources so you don't have to, and sends one issue you can actually finish. No daily noise. Unsubscribe anytime.