Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Sunday, October 25, 2026

Back to today

From X

@convequity

Convequity on X

$NVDA: custom ASICs eat into the GPU share Our scenarios include custom ASICs rising from roughly 20% of the overall AI compute mix toward 40–50%. ABBX records 21.2GW of planned capacity for OpenAI and 13.8GW for Anthropic. Within Anthropic’s plan, ABBX allocates 5GW to AWS Trainium and 5GW to Google/Broadcom TPUs, putting these custom-chip platforms at roughly 72% of its planned capacity. That is a substantial shift in hardware demand if the deployments proceed, although the plans span different timelines. With purchasing concentrated among a few large customers, their decisions to expand on custom silicon could reduce NVIDIA’s share of spending even as the overall AI infrastructure market continues to grow.
@convequity

Convequity on X

$SITM supplies timing devices used in optical modules in NVIDIA-based AI data centers. In an AI data center, a clock exists to keep parts synchronized. Training and inference run across thousands of chips in parallel, and each step depends on data arriving at the right moment. When timing slips, GPUs wait, networks stall, and expensive compute goes unused. The move from 800G to 1.6T optical modules shows how this becomes demand for timing parts. Bandwidth does not simply double. The link is split into more lanes running at higher baud rates. Each lane has a tighter jitter budget, and it often adds clock domains for the DSPs, SerDes groups and laser drivers. So each module carries more timing components, and each component has to meet a higher performance requirement. That pushes average selling prices up alongside unit volumes. The same pattern appears in the switches inside that cluster. A high-radix switch multiplies SerDes lanes as bandwidth scales. When the switch increases radix and speed at the same time, it drives more clock domains across the fabric, the ports, and the control logic. Timing content in the switch therefore multiplies rather than rising in step with the size of the cluster. This is why we see continued hypergrowth for SiTime. In NVIDIA-based data centers, SiTime's parts are discrete devices, separate from the GPU and switch ASICs, allowing timing performance to improve independently as system requirements change, and demand for them rises in several places in the rack at once.

Investor Letters

All letters

Tower Semiconductor Deep Dive

Tower Semiconductor’s low-loss photonics chips halve lasers in 1.6T optical modules, supporting a $680 million annualized photonics run…

TSEM $19B
CV's Substack 10 clicks

Deep Dive: $ASML

ASML’s €230 million EUV machines underpin advanced AI chip production, with demand booked through 2028 after nearly $9 billion and two…

ASML $631B
Rebound Capital 8 clicks

Margin Call

Petrobras’s Morpho-1 well revived hopes that Brazil’s Equatorial Margin can become a major new oil frontier after hydrocarbons were…

PBR $83B
Doomberg 10 clicks

Our Views on CoreWeave

CoreWeave's financial statements point to a business model familiar to the authors after a closer review of the AI infrastructure company.

CRWV $30B
WHAT ARE WE DOING? “Contrarians At The Gate” 12 clicks
DLTR · long

Dollar Tree INC

Dollar Tree is long after divesting Family Dollar, separating its stronger banner from its weaker business.

DLTR $25B
Value Investors Club

Capital Wars

CoreWeave’s $35 billion to $39 billion 2026 spending plan relies on debt as unsecured yields reach 15%, making its GPU cloud expansion a…

CRWV $30B
Net Interest 12 clicks

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

YouTube 4

Blogs 15

Seeking Alpha 62 of 256

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

91 accounts
Nominate an account

Investor letters we follow

190 firms

Sectors

Ordered by how often our editor reaches for them.