Read Damnang’s Substack with Scottie
Keep Damnang’s Substack. Lose the inbox pileup.
Scottie reads the newsletters you choose, keeps the useful differences, and attaches every original link. Damnang’s Substack stays in your reading life without becoming another email to manage.
An independent guide. Scottie isn't affiliated with, endorsed by, or sponsored by Damnang’s Substack. The publication owns its name, writing, and subscription terms.
From Scottie · July 31, 2026
See Damnang’s Substack in a Scottie brief
Scottie read Damnang’s Substack, SemiAnalysis, and Vik's Newsletter for a reader with the priorities shown below. Start with the rundown, open the full brief, or check every issue behind it.
3sources
6issues read
6stories included
3 of 3 Damnang’s Substack issues included
What shaped this brief
Reader priorities
A semiconductor or infrastructure investor tracking memory, data-center power, and the capital cycle behind AI compute.
These are illustrative priorities, not a customer’s data.
Sources in this brief
- Damnang’s SubstackThe publication this guide is about
- SemiAnalysisAdds compute architecture, chip economics, and AI-infrastructure cost analysis.
- Vik's NewsletterAdds specialist coverage of power delivery, memory, optics, and data-center hardware.
Scottie
Scottie example brief
July 31, 2026 · Executive brief
01 / The rundown
- Memory market anxiety driven by overcapacity fears appears overblown as high bandwidth memory consumes three times the wafer capacity of ordinary DRAM.
- Data center operators are pivoting to modular, factory-built designs to bypass critical electrician shortages and accelerate AI compute deployments.
- AMD is challenging Nvidia's AI accelerator market share with its 2nm MI455X silicon, despite facing rack production and software testing bottlenecks.
Read the complete brief 6 stories · 3 action items
02 / The briefing
01 / main
AI capex fear overstates memory supply risks ahead of updates
The takeaway: Market anxiety over AI capital spending assumes 70% of SK Hynix peak earnings will vanish within four quarters. However, that assumption ignores how HBM production consumes significantly more wafer area per bit, dampening potential 2028 DRAM oversupply.
Concrete details
- SK Hynix stock priced in a drop from 150,000 won peak quarterly EPS down to 180,000 won annual EPS.
- Shipping 1GB of 16-high HBM4 requires four times the wafer area of commodity DRAM.
Why it matters for this reader: For your memory cycle thesis, this selloff pricing misjudges how HBM4's fourfold wafer area penalty prevents the severe DRAM bit glut markets fear.
Original sourcesDamnang’s Substack
02 / main
CXMT's IPO pop triggers memory panic despite severe cost headwinds
The takeaway: CXMT surged 472% on its A-share debut, sparking fears it will break global DRAM pricing. Yet CXMT carries 30% higher production costs and 40% larger die sizes than incumbent makers, limiting its ability to undercut prices.
Concrete details
- CXMT opened at 49.50 yuan against its 8.66 yuan IPO price, reaching a 3.31 trillion yuan valuation.
- CXMT processes 13% of global DRAM wafer starts but produces only 6% of effective memory capacity.
Why it matters for this reader: Tracking the memory capital cycle shows CXMT's technical gap and low yield mean its IPO proceeds will not yield net capacity risks until 2028.
Original sourcesDamnang’s Substack
03 / main
Tech leaders detail physical trade-offs as HBM architecture scales up
The takeaway: High bandwidth memory provides critical bandwidth for AI, but stacking DRAM dies reduces bit density to one-third of standalone chips while creating thermal and yield bottlenecks. Future iterations will require custom base dies and integrated cooling.
Concrete details
- HBM4 stacks sixteen 4GB dies for 64GB capacity and 2TB per second total bandwidth.
- HBM consumes 22% of global DRAM wafer capacity while producing just 9% of usable memory capacity.
Why it matters for this reader: As an investor following AI compute infrastructure, HBM's shifting design transforms memory from fungible commodities into high-margin custom silicon.
Original sourcesVik's Newsletter
04 / main
Korean industry updates reveal suppliers pushing for custom LTA contracts
The takeaway: Direct conversations with Korean industry leaders show that while AI memory is transitioning toward custom products and long-term contracts, suppliers still face intense pressure to satisfy customer architectural demands across HBM4 and logic integrations.
Concrete details
- Multiyear long-term agreements are increasing earnings visibility and supplier bargaining power in AI workloads.
- HBM4 and HBM4E transitions to 12-high and 16-high stacks face physical manufacturing bottlenecks.
Why it matters for this reader: Navigating the semiconductor capital cycle requires assessing whether memory suppliers are gaining permanent pricing power or remaining tied to customer capex preferences.
Original sourcesDamnang’s Substack
05 / main
Modular construction accelerates data centers to bypass skilled labor shortages
The takeaway: Operators are adopting prefabricated modular structures to defeat trade labor shortages, specifically among electricians. Moving structural and electrical work offsite compresses build timelines dramatically while cutting capital expenditure per megawatt.
Concrete details
- Modular construction compresses data center build schedules by 36% and reduces capital expenditure per megawatt by 8%.
- Electricians represent 30% to 40% of total construction man-hours on data center sites.
Why it matters for this reader: For an infrastructure investor monitoring AI compute buildouts, modular construction directly resolves localized labor constraints that threaten facility delivery schedules.
Original sourcesSemiAnalysis
06 / main
AMD targets Nvidia's market share with 2nm silicon despite software bottlenecks
The takeaway: AMD is making strides against Nvidia by deploying its 2nm MI455X silicon and open-source software stack. However, internal GPU cluster shortages for testing and rack assembly integration issues still delay its competitive rollout.
Concrete details
- The MI455X features 3,470mm2 of logic silicon using TSMC 2nm tiles on a 5.5x reticle CoWoS-L package.
- AMD's Helios rack backplane requires over 550 Broadcom ethernet retimers due to SerDes design constraints.
Why it matters for this reader: Evaluating AI compute market dynamics requires balancing AMD's hardware progress against internal software testing constraints that stall customer deployments.
Original sourcesSemiAnalysis
Action items
- Model HBM wafer penalties into 2028 DRAM supply forecasts to adjust for CXMT capacity expansion delays.
- Evaluate data center infrastructure portfolios for modular design adoption to mitigate forecasted electrician shortages.
- Track AMD Helios rack production milestones to evaluate whether software testing bottlenecks delay AI accelerator adoption.
See every issue behind this brief
Market Concerns Over AI Capex and Memory
- What Would Stop This Selloff?Damnang’s Substack · Included
Industry Perspectives on High Bandwidth Memory
- HBM is the Ugly Baby: What Now?Vik's Newsletter · Included
CXMT's IPO Surge and Its Impact on Memory Competitors
- After CXMT’s Listing, Do the Memory Three Have Further to Fall?Damnang’s Substack · Included
Modular Construction Solutions for Datacenters
- The Wild Wild West Of LEGO DatacentersSemiAnalysis · Included
Korean Expert Insights on Memory Sector Trends
- I Met with Semiconductor Experts in KoreaDamnang’s Substack · Included
AMD's AI Software Push to Challenge CUDA
- Can AMD break the CUDA Moat? AMD Advancing AI 2026SemiAnalysis · Included
About Damnang’s Substack
Should you add Damnang’s Substack to Scottie?
Follow the physical infrastructure and market forces that constrain AI growth.[1][2][3]
Who it’s for
A semiconductor or infrastructure investor tracking memory, data-center power, and the capital cycle behind AI compute.[1][2]
What you’ll find in it
The issues linked below include “AI Datacenter Power Investment Map: From 765kV to 0.65V”, “What Would Stop This Selloff?”, and “After CXMT’s Listing, Do the Memory Three Have Further to Fall?”. Open them to judge the publication in its own words.
How to read Damnang’s Substack with Scottie
Add its public feed for publicly available issues. If you subscribe to paid issues, forward those emails to your private @scottie.fyi address. Scottie does not need access to your inbox.
Scottie read 3 Damnang’s Substack items and included 3 for the brief. The sources covered different things, so Scottie kept their stories separate instead of forcing a connection.
Read three issues from Damnang’s Substack
Read the publication in its own words. Scottie keeps these original links attached; it does not replace the writing.
- AI Datacenter Power Investment Map: From 765kV to 0.65VPublic issue[2]
- What Would Stop This Selloff?Public issue[3]
- After CXMT’s Listing, Do the Memory Three Have Further to Fall?Mon, 27 Jul 2026[4]