Read Vik's Newsletter with Scottie
Keep Vik's Newsletter. Lose the inbox pileup.
Scottie reads the newsletters you choose, keeps the useful differences, and attaches every original link. Vik's Newsletter stays in your reading life without becoming another email to manage.
An independent guide. Scottie isn't affiliated with, endorsed by, or sponsored by Vik's Newsletter. The publication owns its name, writing, and subscription terms.
From Scottie · July 31, 2026
See Vik's Newsletter in a Scottie brief
Scottie read Vik's Newsletter, SemiAnalysis, and The Diligence Stack 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
1 of 1 Vik's Newsletter issues included
What shaped this brief
Reader priorities
A semiconductor or data-center specialist tracking power delivery, memory, optics, and the physical limits of AI infrastructure.
These are illustrative priorities, not a customer’s data.
Sources in this brief
- Vik's NewsletterThe publication this guide is about
- SemiAnalysisAdds full-system AI compute economics and architecture comparison.
- The Diligence StackAdds capacity, storage, supplier, and hyperscaler investment context.
Scottie
Scottie example brief
July 31, 2026 · Executive brief
01 / The rundown
- Severe utility delays are forcing hyperscalers to adopt behind-the-meter power generation and leased shells to bring compute online faster.
- Data center construction is shifting to modular offsite fabrication to compress schedules by seven months and bypass acute electrician shortages.
- Surging demand for non-chip AI components is pushing frontier rack power to 142 kilowatts and driving semiconductor selling prices upward.
Read the complete brief 6 stories · 3 action items
02 / The briefing
01 / main
Pat Gelsinger Calls HBM Lousy Memory as Tradeoffs Mount
The takeaway: High bandwidth memory is driving 90% margins for memory makers, but stacking DRAM creates severe thermal, yield, and capacity friction. Through-silicon vias cut HBM die bit-density to one-third of standard DRAM, requiring triple the wafer capacity to manufacture.
Concrete details
- HBM bit-density drops to one-third of standard DRAM, taking three times more silicon wafers per gigabyte.
- Memory suppliers are capturing near 90% margins while the broader AI memory shortage extends another two to three years.
Why it matters for this reader: As you monitor physical memory limits, HBM's wafer penalty and thermal walls necessitate new memory architectures before HBM5 limits scaling.
Original sourcesVik's Newsletter
02 / main
Behind-the-Meter Power Models Bypass Grid Delays for AI Sites
The takeaway: Grid interconnect queues are pushing hyperscalers toward behind-the-meter generation to get compute online faster. Lower power spikes from inference workloads and rack-level power smoothing make onsite power generation far easier to integrate without grid reliance.
Concrete details
- Interactive inference power spikes average 9%, compared to 37.5% maximum power swings seen during training workloads.
- Bloom Energy reported Q2 revenue of $1.065 billion, up 165.5% year-over-year as major hyperscalers adopt behind-the-meter power.
Why it matters for this reader: For your power delivery tracking, behind-the-meter generation offers a faster route to active megawatts while bypassing multi-year utility queue bottlenecks.
Original sourcesThe Diligence Stack
03 / main
Power Bottlenecks Force Hyperscalers Into Leased Data Center Shells
The takeaway: Local utility delays are worsening, with grid studies taking up to a year and substations requiring up to three years to build. To bridge capacity gaps, hyperscalers are leasing physical shells to maintain compute control while landlords deliver power.
Concrete details
- Data center grid bottlenecks are projected to constrain hyperscaler infrastructure deployments through at least 2030.
- The ratio of contracted backlog to cost-adjusted capital expenditures dropped from 44% in 2024 to 37% in 2026.
Why it matters for this reader: In your physical infrastructure analysis, leased shells let cloud providers bypass utility delays while preserving their high-margin compute and customer relationships.
Original sourcesThe Diligence Stack
04 / main
Modular Construction Speeds Data Center Builds Amid Labor Shortages
The takeaway: Datacenter builders are shifting work offsite to factory floors to bypass severe trade labor shortages. Prefabricated wall panels, pre-wired mechanical rooms, and steel frames compress construction timelines significantly while cutting overall capital expenditures.
Concrete details
- Modular methods shorten datacenter construction schedules by roughly 36%, saving seven to nine months while lowering capex 8%.
- An electrician shortage starting in 2027 threatens traditional builds, where electrical work accounts for 30% to 40% of man-hours.
Why it matters for this reader: Tracking data center deployment limits requires watching modular adoption, which helps offset critical trade shortages to bring capacity online faster.
Original sourcesSemiAnalysis
05 / main
Infrastructure Surrounding Accelerators Scales Faster Than Early Models Predicted
The takeaway: System-level demand around AI chips is outpacing original forecasts, pulling up semiconductor average selling prices and non-chip hardware requirements. Server DRAM, enterprise NAND, advanced packaging, and rack power levels have all surged well past initial expectations.
Concrete details
- Blended semiconductor average selling prices are projected to rise from $0.75 today to over $2.10 by 2028.
- Frontier rack power density hit 120 to 142 kilowatts, reaching 2030 projections four years ahead of schedule.
Why it matters for this reader: For your physical infrastructure modeling, scaling non-chip components like high-density power and packaging is driving the real capital intensity in AI setups.
Original sourcesThe Diligence Stack
06 / main
AMD Targets Nvidia With 2nm Silicon and Expanded Software Stack
The takeaway: AMD is advancing its hardware and software to challenge Nvidia, landing a 2GW deployment deal with Anthropic. However, rack manufacturing challenges and a shortage of internal GPU clusters for automated software testing continue to limit execution speed.
Concrete details
- The MI455X uses 2nm silicon across 3,470 square millimeters of logic with hybrid bonding in a 5.5x reticle package.
- Helios racks require over 550 ethernet retimers per rack due to backplane signal degradation and re-timing needs.
Why it matters for this reader: When assessing competitive AI hardware, watching AMD's 2nm packaging innovations against its backplane and software testing bottlenecks reveals key physical limits.
Original sourcesSemiAnalysis
Action items
- Evaluate behind-the-meter fuel cell deployments as a near-term bridge to bypass three-year utility substation construction delays.
- Audit high-density rack thermal and power specifications as frontier installations scale beyond 120 kilowatts per cabinet.
- Model memory wafer capacity requirements accounting for HBM's three-to-one production penalty compared to standard server DRAM.
See every issue behind this brief
Behind-the-Meter Power Models for AI Infrastructure
- The Behind-the-Meter AI BuildoutThe Diligence Stack · Included
Hyperscaler Demand and Local Utility Bottlenecks
- The Hyperscaler Capacity Partner HierarchyThe Diligence Stack · Included
Pat Gelsinger Critique of High Bandwidth Memory
- HBM is the Ugly Baby: What Now?Vik's Newsletter · Included
Modular Construction in Data Center Builds
- The Wild Wild West Of LEGO DatacentersSemiAnalysis · Included
AMD Software Strategy Against Nvidia Accelerators
- Can AMD break the CUDA Moat? AMD Advancing AI 2026SemiAnalysis · Included
Accelerator Surrounding Component Scaling
- In 2023-2024 We Weren't Bullish EnoughThe Diligence Stack · Included
About Vik's Newsletter
Should you add Vik's Newsletter to Scottie?
Understand which component bottlenecks will shape AI data-center cost and performance.[1][2][3]
Who it’s for
A semiconductor or data-center specialist tracking power delivery, memory, optics, and the physical limits of AI infrastructure.[1][2]
What you’ll find in it
The issues linked below include “Power Delivery As The Next Physics Wall In AI Datacenters”, “HBM is the Ugly Baby: What Now?”, and “Lasers for CPO/NPO: Part 2 – Lumentum’s Technology and Moat”. Open them to judge the publication in its own words.
- Byline
- Vikram Sekar[1]
How to read Vik's Newsletter 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 1 Vik's Newsletter item and included 1 for the brief. The sources covered different things, so Scottie kept their stories separate instead of forcing a connection.
Read three issues from Vik's Newsletter
Read the publication in its own words. Scottie keeps these original links attached; it does not replace the writing.
- Power Delivery As The Next Physics Wall In AI DatacentersPublic issue[2]
- HBM is the Ugly Baby: What Now?Public issue[3]
- Lasers for CPO/NPO: Part 2 – Lumentum’s Technology and MoatWed, 22 Jul 2026[4]