Daily Alpha · X
· Midday Alpha · by Buzzberg Research
Nvidia's lower Vera Rubin HBM4 bandwidth spec drew analysis from @svtrivo and @damnang2, while Gavin Baker highlighted how little Fortune 500 customers spend with Anthropic.
Themes on this desk
Vera Rubin HBM4
@svtrivo notes Vera Rubin's datasheet HBM4 bandwidth fell to 19.2 TB/s from 22 TB/s, and @damnang2 says AMD's MI455X now leads per-GPU HBM bandwidth by 21%.
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Gavin Baker flags Anthropic's under-0.1% share of large Fortune 500 customers' expenses
Baker quotes the S-1: Anthropic's revenue from Fortune 500 customers paying $100K+ a year averaged under 0.1% of those customers' total expenses in the September 2026 quarter, and over half the Fortune 500 is in that group.
Most surprising line in the S-1:
Baker calls the figure "absolutely wild" given about $75 billion in revenue.
Source →Alibaba, Tencent and ByteDance are reportedly negotiating memory LTAs with SK hynix and Samsung
Citing Maeil News, the post says Chinese cloud providers including Alibaba, Tencent and ByteDance are negotiating memory long-term agreements with SK hynix and Samsung, with some customers reportedly pushing for 10-year contracts into 2035.
Oh look… more memory durability evidence for $MU and co.
The post reads the deals as more memory durability evidence for Micron and peers.
Source →Meta-AMD paper shows MI350X averaging 101.2% of B200's speed across 10 models
Citing a Meta and AMD paper, the post says Meta trains production recommendation models on tens of thousands of MI350X GPUs, which averaged 101.2% of B200's speed across 10 models and were faster on 4 of them by up to 1.3 times. It says MI455X now leads per-GPU HBM bandwidth by 21% after Nvidia lowered Vera Rubin's HBM4 spec.
🚨🚨 How well can $AMD's Helios be used for training, and how much of $NVDA's ecosystem can it break down?
Meta used AI agents to port Nvidia-tuned kernels, succeeding more than 80% of the time and cutting development time to about one-tenth, per the post.
Source →CenterPoint says about 2 GW of studied ERCOT load withdrew its investment
Per the post, CenterPoint management said on its FQ3 2026 call that approximately 2 gigawatts of studied load, whose customers were finalizing details during the state's audit process, decided to withdraw their investment.
AI TRADE NEGATIVE SIGNAL • CNP (CenterPoint Energy, Inc.), FQ3 2026 call, reported 10/27/26 • Speaker: unknown, management • Topic:
The post labels it a negative AI trade signal on ERCOT Batch Zero studied load.
Source →Nvidia's datasheet cuts Vera Rubin per-GPU HBM4 bandwidth to 19.2 TB/s from 22 TB/s
Per-GPU HBM4 bandwidth went from 22 TB/s to 19.2 TB/s and NVL72 from 1,580 TB/s to 1,400 TB/s, with capacity at 288GB and NVFP4 compute unchanged. The post reads the lower pin target as widening the pool of stacks that pass qualification, which it labels industry interpretation, not an Nvidia statement.
NVIDIA quietly cut Vera Rubin's HBM4 bandwidth by about 13%.
The fastest-binning supplier loses some of its speed premium and a chaser like Micron gets an easier bar, per the post.
Source →