Newsletter comparison
SemiAnalysis vs Vik's Newsletter
Short answer: Use SemiAnalysis for the full AI-infrastructure stack and its economics. Use Vik's Newsletter for concentrated work on the component bottleneck most likely to break a forecast. Reading both makes sense for anyone building or underwriting capacity.[1][2][3][4][5][6]
An independent guide. Scottie isn't affiliated with, endorsed by, or sponsored by SemiAnalysis and Vik's Newsletter. Each publication owns its name, writing, and subscription terms.
A brief from both newsletters · July 31, 2026
See both newsletters in one brief
See what made the brief, what didn't, and the original links behind every included story.
In this brief: Scottie kept the selected stories separate because they added different value.
Reader priorities
An AI-infrastructure reader wants to understand which hardware limits will determine data-center performance, availability, and total cost.
These are illustrative priorities, not a customer’s data.
What made the brief
- SemiAnalysis2 read · 2 included
- Vik's Newsletter1 read · 1 included
Scottie
Scottie example brief
July 31, 2026 · Executive brief
01 / The rundown
- High Bandwidth Memory scaling limits and three-fold capacity bottlenecks are forcing chipmakers to pursue new memory architectures for future AI clusters.
- AMD is making software strides with major customer deployments, but internal GPU cluster shortages risk slowing its ROCm development velocity.
- Data center operators are adopting modular prefabricated builds to bypass electrician shortages, reducing construction timelines by 36 percent.
02 / The briefing
01 / main
High Bandwidth Memory faces physical limits as scaling challenges mount
The takeaway: Stacking DRAM chips to build HBM creates thermal issues, lower yields, and severe wafer inefficiency. Each HBM die requires three times the wafer capacity of standard DRAM due to silicon vias, creating a memory supply crunch lasting two to three years.
Concrete details
- HBM bit-density is one-third that of standalone DRAM, consuming three times more silicon wafers.
- Memory manufacturers currently enjoy margins near 90% amid an ongoing two-to-three-year supply crunch.
Why it matters for this reader: For your focus on hardware performance and total cost, HBM wafer inefficiencies and 90% supplier margins directly drive up long-term data-center memory expenses.
Original sourcesVik's Newsletter
02 / main
AMD software gains ground while rack supply and cluster shortages linger
The takeaway: AMD is winning major deployments with Anthropic and Microsoft, but infrastructure constraints hold back progress. Helios rack production faces backplane reliability challenges, while internal developers lack stable GPU clusters for automated software testing.
Concrete details
- Up to 85% of the Helios rack backplane requires retiming, needing over 550 Broadcom retimers per rack.
- Anthropic plans to deploy 2 gigawatts of AMD accelerators using agentic software workflows.
Why it matters for this reader: Understanding AMD's rack integration delays and software cluster limits helps you evaluate whether non-Nvidia hardware options can lower your total cost of ownership.
Original sourcesSemiAnalysis
03 / main
Modular construction speeds data center deployment amid trade labor shortages
The takeaway: Data center builders are adopting off-site modular prefabrication to bypass trade labor bottlenecks. Factory-built concrete panels, pre-wired electrical rooms, and steel frames allow operators to erect structures faster while lowering capital expenditures.
Concrete details
- Modular construction compresses facility build windows by 36% and reduces capital costs by 8% per megawatt.
- Electricians represent 30% to 40% of construction man-hours, with widespread shortages projected for 2027.
Why it matters for this reader: If data-center availability and time-to-market dictate your compute deployment strategy, modular builds offer a faster path to expanding online capacity.
Original sourcesSemiAnalysis
Action items
- Evaluate modular data center vendors to reduce facility deployment timelines ahead of predicted 2027 trade labor shortages.
- Audit internal cluster allocations to ensure software development and automated testing teams receive stable GPU access.
See every issue behind this brief
High Bandwidth Memory limits and the future of AI hardware scaling
- HBM is the Ugly Baby: What Now?Vik's Newsletter · Included
AMD's software progress and the CUDA moat
- Can AMD break the CUDA Moat? AMD Advancing AI 2026SemiAnalysis · Included
Modular construction techniques in large data centers
- The Wild Wild West Of LEGO DatacentersSemiAnalysis · Included
The useful difference
The difference that matters
SemiAnalysis assembles the broad compute platform and economic picture. Vik's Newsletter zooms into bottlenecks such as power delivery, high-bandwidth memory, and optical components.[1][2][3][4][5][6]
What each is best for
SemiAnalysis
SemiAnalysis supplies the platform-level comparison: accelerators, systems, inference economics, and competitive hardware roadmaps.[1][2][3]
Vik's Newsletter
Vik's Newsletter supplies component-level scrutiny of power, memory, and optical technologies that can cap the platform's real performance.[4][5][6]
When it’s worth reading both
A system forecast is only as good as its tightest component assumption. Scottie can connect the broad model to the specific bottleneck and keep conflicting estimates visible.[1][2][3][4][5][6]