Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Wednesday, October 21, 2026

Back to today

From X

@convequity 3 clicks

Convequity on X

Inference repeatedly reads the same model weights, with far fewer writes than training. 1. In training, weights are constantly updated, so the memory sees heavy read and write traffic. 2. In inference, the dominant pattern is repeated reading of the same weights, and the key-value cache alongside them, with far fewer writes. That is the case for putting AI model weights on flash. Flash has traditionally been weaker than DRAM on write endurance and write speed. A read-heavy, write-light workload plays to its strengths and gets around exactly those weaknesses. Today AI chips sit beside HBM built from stacked DRAM, and it is the expensive part of the memory bill. High-Bandwidth Flash (HBF) is the product built on the read-heavy observation. It aims to deliver HBM-like read bandwidth at far higher capacity and lower cost by using NAND flash instead of DRAM. Two caveats. These are aims, not measured specifications. As of our early-August note, the first prototype from Sandisk ($SNDK), developed with SK Hynix, was expected later in the year. HBF is also more specialised than commodity NAND, where Chinese producers are expanding capacity aggressively.
@convequity

Convequity on X

A bit of surpsied to see OpenAI Dot's VM spec is so low: CPU: 9 visible logical CPUs; host reports AMD EPYC 9V74 at ~2.6 GHz RAM: 9.7 GiB total, ~8.7 GiB available; no swap Storage: 32 GiB workspace filesystem, ~30 GiB free Also uses Cloudflare $NET network as egress like Grok Bot does but the computer spec is so much lower than Grok Bot. Maybe OpenAI is just more resource constrained vs. SpaceX given a much bigger user base. The demand for traditional compute, networking and storage will grow crazy as more people start to adopt this cloud agent architecture.

Investor Letters

All letters

Heico Corporation

HEICO sells FAA-approved aircraft replacement parts at 30% to 50% discounts to OEMs, gaining customers as restrictive aftermarket contracts…

HEI $42B
D Invests 13 clicks

Copper Fox Metals: Research Flash

Copper Fox’s C$9.1 million Schaft Creek program advances metallurgical, tailings, water and road studies toward a pre-feasibility study,…

TECK $31B
Atrium Research - Equity Research 6 clicks

AppLovin Deep Dive

AppLovin rebuilt its ad model with AXON 2.0, lifting LTM revenue to about $7 billion and margins to 77% as Unity competition threatens ad…

APP $92BU $18B
Speedwell Memos 68 clicks

Semtech Analyst Day Review

Semtech raised Q3 guidance and targets $3.2 billion of fiscal 2030 revenue, excluding upside from its AI data-center optical and copper…

SMTC $13B
CV's Substack 14 clicks

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

YouTube 4

Blogs 15

Seeking Alpha 54 of 247

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

90 accounts
Nominate an account

Investor letters we follow

190 firms

Sectors

Ordered by how often our editor reaches for them.