Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Monday, October 5, 2026

Back to today

From X

@convequity

Convequity on X

High-bandwidth memory generally costs about 5x more per GB than conventional DRAM. As AI systems support longer context windows and more complex reasoning, memory demand is spreading into conventional DRAM and NAND. The mechanism is a tiered memory system. HBM holds active model weights and the context needed now. External DRAM holds a broader pool of context that the model can retrieve, while NAND-based SSDs store less frequently used context and model checkpoints. This keeps more context accessible at a lower cost, with a latency trade-off. Memory producers are operating at approximately 80% gross margins and 60%+ EBIT margins. The industry's through-cycle average gross margin has been 35-40%, and past peaks rarely exceeded 55-60%. Supply explains much of this. Samsung, SK Hynix and Micron invested heavily in capacity during the 2021-2022 boom. That capacity came online in 2023 and 2024, when smartphone, PC and non-AI server demand was weak. All three cut production, recorded operating losses and set conservative capex plans. Samsung cut DRAM output by approximately 15-20% in 2023. The capacity response as of mid-2026: 1. Samsung has targeted a roughly 50% increase in HBM production capacity this year. 2. SK Hynix has accelerated its HBM4 ramp through the M15X fab, while reportedly redirecting some planned HBM4 capacity toward DDR5. 3. $MU has accelerated DRAM and HBM output in Taiwan. Meaningful wafer starts from its new US and Singapore facilities are not expected until 2027 and beyond. These additions are weighted toward HBM and high-end server DRAM, not broad conventional memory capacity. They appear insufficient to close the gap in the near term, so the unusually high margins look likely to persist for several more quarters.
@convequity 3 clicks

Convequity on X

$NET's AI hosting is limited by the physics of its edge footprint and by the role of its products. Carrier-hotel sites built for fast network traffic are a different proposition from purpose-built AI datacenters. Cloudflare’s own engineering disclosures acknowledge power constraints, and management distinguishes its strategy from the hyperscalers’ GPU spending race. Our assessment starts with a two-H100-per-server edge configuration. H100 PCIe and NVL-class cards operate around 350–400 watts per GPU, versus up to 700 watts for H100 SXM. Hyperscalers commonly pack eight SXM GPUs into a machine and connect those machines into dense clusters. The distinction is power, cooling, memory bandwidth and GPU interconnects — not simply whether a model can run. The existing edge footprint cannot acquire hyperscaler-class cooling and power density through server upgrades. Dense, liquid-cooled Blackwell and Vera Rubin systems require a different infrastructure commitment. We see little chance that NET’s existing edge footprint becomes a competitive home for those dense clusters. For sustained, large-model inference, our assessment is that sparse edge GPUs cost more per token than dense AI clusters. Geographic proximity saves network transit time, but it does not guarantee a faster first token or a cheaper response. Greater compute capacity, faster interconnects and efficient batching can outweigh the edge’s network advantage. We expect heavy training, frontier-scale hosting and the most demanding next-generation inference to remain concentrated with hyperscalers. Workers AI’s progress with larger open models is real. Hosting those models does not establish an economic advantage over dense clusters. AI Gateway also lets customers route requests to external providers, matching the model and infrastructure to the task. Cloudflare can still win substantially. More agents mean more requests to secure, more Workers executions, more orchestration and more demand for fine-grained access controls. Those workloads build on infrastructure and capabilities Cloudflare already has. On the SaaS Survival Scorecard, we score $NET at 2.5 for Model-Layer Leverage. That reflects its substantial work building an agent harness across inference, orchestration and execution, while stopping short of model post-training and reinforcement learning.
@convequity 2 clicks

Convequity on X

Anthropic's $518bn of planned compute obligations is about the same size as $AMZN's entire stack of debt, leases and purchase commitments. Convequity's ABBX barometer adds up borrowings, leases, purchase obligations and leases not yet started for each hyperscaler. On that all-in measure, AMZN and Anthropic are roughly level. 1. Against valuation, the number looks normal. At the $2 trillion valuation being debated, Anthropic's obligations are about a quarter of its value. That is close to $GOOGL's ratio and inside the range of the four big hyperscalers. 2. Against revenue, it is large but not extreme. Anthropic told investors its annualized run rate was above $65bn at the end of July. The obligations are about eight times that, spread over roughly a decade. A run rate is the latest month multiplied by twelve, not full-year booked revenue, and the hyperscalers fund their stacks from large, established cash flows. 3. Reuters reports that about 80% of the obligations are non-cancelable or payable regardless of usage, while many of Anthropic's largest customers are not on long-term contracts. I think the size is about right. It fits a $2 trillion company, and at the July pace Anthropic can grow into it. The risk is in the terms: most of the cost is locked in for years, while much of the revenue can leave at short notice. The number to watch now is how much of Anthropic's revenue is under long-term contract.

Investor Letters

All letters

British American Tobacco: Harvest

British American Tobacco’s U.S. reset revives combustibles and rebuilds its nicotine-pouch portfolio after legacy-brand writedowns and Velo…

BTI $120B
Invariant 5 clicks

Practical Parametric Yield

Cerebras’s WSE-3 faces parametric-yield constraints that could restrict supply, making power and clock-speed tradeoffs central to its next…

CBRS $39B
Irrational Analysis 3 clicks

Nike, Part Two: The Deep Dive

Nike’s second Latticework program examines the company and the global athletic footwear and apparel industry for professional investors.

NKE $51B
Latticework by MOI Global 1 click

Nike, Part One: The Story

Nike grew from a $1,000 shoe-distribution venture in 1964 into the largest sports brand, using athlete endorsements, product innovation and…

NKE $51B
Latticework by MOI Global 2 clicks

Publications we follow

Every publication on this site, hand-picked.

109 publications

Newsletters 92

YouTube 4

Blogs 12

Seeking Alpha 45 of 216

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

74 accounts
Nominate an account

Investor letters we follow

189 firms

Sectors

Ordered by how often our editor reaches for them.