Monday, September 7

Monday, September 7, 2026

Everything we published on this day.

28 stories — the front page that day →

The Week Ahead

Nvidia’s Jensen Huang called GPT-6 Astra AGI as Anthropic contracted 14.8 GW of compute, while the Fed’s September decision hinges on…

The Quality Check #1

The Quality Check #1 reviews recent developments at Arista, Cummins, Veeva, ServiceNow and AppLovin following earnings season.

@convequity

Convequity on X

Kirin 9050 Pro is the first Tau Scaling chip. The bigger story is Ascend. Huawei just put LogicFolding in the Mate XT 2. First commercial SoC on the Tau (τ) law He Tingbo laid out in May / updated in July. Not a surprise if you read that paper. Still a real jump: DUV + SAQP + logic stacking took density from 155 → 238 MTr/mm² on the same SMIC N+3 node. +55% in one gen. Power went down, not up. That’s N3-class equivalent density. Headline number matches Intel 18A HD (238) and sits next to Samsung SF2 (~231). TSMC N2 HD is still ahead (~313). Caveat: 238 is Huawei’s LogicFolding equivalent, not a planar N3 shrink. Same device node as 9030 Pro. Some analysts put the industry-formula read closer to ~175. Mobile scoreboard vs last-gen trifold / 9030 Pro: • Device +42% • CPU ST +24% / MT +52% • P-core 3.1 GHz (was 2.75) • CPU / GPU / NPU iso-perf power −41% / −58% / −66% • GPU render +142% • On-device 30B-total-param MoE Multi-core still trails $QCOM’s latest. The gap is no longer multi-generation. Next Tau Kirin can catch, maybe pass, a leading-edge mobile SoC. The packaging tell is sharper than the phone benches. 1.4–1.5μm hybrid-bond pitch. ~50M vertical interconnects. Western HVM hybrid bonding is still ~6μm (TSMC SoIC) to ~9μm (Intel Foveros Direct). If 1.4μm holds, that is SOTA vs Western leading-edge bonding — not the Geekbench print. Phone catch-up is the demo. AI silicon is the payoff. Apply the same fold to Ascend and the process-density gap vs $NVDA peers compresses from the lithography side over the next few quarters. Remaining bottlenecks: domestic DUV tool volume, HBM, packaging, software. I wouldn’t assume they outproduce ASML on DUV. I also wouldn’t fade the direction. Mature DUV + stacking + almost-unlimited power is enough to put China in the token-factory fight vs the US. Tables below. 🧵
@convequity

Convequity on X

GPT-6 Astra: scaling still works. Algo is the larger remaining gap. Astra shows both scaling laws and algorithms are still progressing. It is reportedly distilled from a 10T+ teacher named Bel. Maybe with $NVDA Vera Rubin and the next compute generation coming online, 10T+ becomes the new standard and performance keeps compounding. That means continued — even accelerating — demand for AI infra. The more interesting signal is algo. IMO we are still just touching the surface of how algorithms can improve. There is still a large gap between SOTA perf/watt and the human brain. GPT-6’s recurrent depth is the part that stuck with me. CoT / test-time reasoning lets the model write intermediate steps out as tokens, then produce a better final answer than a direct reply. Recurrent depth is different. It lets the model think longer and harder in hidden space before any word is emitted. Not just write the thinking out. Think harder on each next bit of the word. I am perplexed this arrived so late in GPT-6. Stable Diffusion introduced and spread this kind of latent iterative method in image generation ~5 years ago. The main reason it came so late: hidden thinking is harder to safety-check and monitor than explicit CoT. That constraint is real. It is also more human-like. People do not narrate every step. Plenty of global-max algorithms are still unimplemented for similar reasons. Models only get stronger from here.
UBER 2026-09-07

A quick update on Deep Dives

MBI Deep Dives will update prior company analyses through year-end, beginning with Uber, while deferring new work on Micron and S&P Global.

Celsius: The Pepsi Put

Celsius holds 20% of the U.S. energy-drink market and 40% of sugar-free sales, but its $1.65 billion Alani acquisition complicates…