Vik’s Research Library — 2026

73 posts published in 2026, through November 13.

AI compute architectures, high-bandwidth memory, optical networking, semiconductor lasers, power delivery, advanced packaging, and industry analysis.

Welcome to the 2026 research library.

During 2026, Vik’s Newsletter has focused increasingly on the engineering challenges behind AI infrastructure and the semiconductor technologies required to solve them.

From the growing importance of CPUs in agentic AI to custom HBM base dies, next-generation optical interconnects, and 800V power delivery, the central question has remained the same: What are the physical and architectural constraints driving technological change, and which companies are positioned to benefit?

I approach these subjects from an engineer’s perspective. The goal is to understand how something works, why particular architectures are chosen, what trade-offs they introduce, and how those decisions affect the semiconductor industry.

This library organizes my writing by subject rather than publication date. It includes technical deep dives, architecture comparisons, industry analysis, conference reports, and shorter commentary.

Some articles establish the fundamentals. Others develop specific theses about where the technology is going and how competitive positions may change.

If you’re new to a subject, I recommend following the articles in sequence.

1. AI Compute Architectures and Inference

How AI workloads shape the architectures of CPUs, GPUs, custom accelerators, and the systems connecting them.

A Close Look at SRAM for Inference in the Age of HBM Supremacy — When SRAM can outperform HBM for inference, and what the trade-offs look like.
The CPU Bottleneck in Agentic AI and Why Server CPUs Matter More Than Ever — How agent orchestration changes CPU requirements and the CPU-to-GPU balance.
The AI Datacenter CPU Yellow Pages — A reference guide to server CPUs used in AI infrastructure.
GTC 2026 Preview: NVIDIA’s SRAM-Decode Hardware — The implications of SRAM-based decoding architectures for inference.
Beyond GTC: Compute, LPX, and the Untold Story of SpecDec — A deeper examination of compute architecture and speculative decoding.
TurboQuant: Inner Workings and Implications — Understanding compression techniques and their implications for AI hardware.
Tokenmaxxing and the Token Value Chain — Following the economics of tokens through the AI infrastructure stack.
Cerebras IPO and the Four Bottlenecks in Its Custom-Everything Architecture — Wafer-scale computing, architectural trade-offs, and the challenges of commercial scaling.
Inside SambaNova’s Inference Architecture — Reconfigurable dataflow computing and a three-tier memory hierarchy.
How Many CPUs Do Personal AI Assistants Need? — Modeling the CPU and memory requirements of personal AI agents at scale.

2. High Bandwidth Memory, DRAM, NAND, and Storage

Memory bandwidth, capacity, manufacturing, and data movement are central to AI system performance. This collection examines how those constraints reshape the memory hierarchy.

Context Memory Storage Systems, Agentic AI Tokenomics, and Flash vs. DRAM — The role of memory pooling and storage in agentic AI workloads.
Why the DRAM Structural Floor in the AI Era Makes This Time Different — Understanding how AI demand could change the traditional memory cycle.
What AI Inference Actually Demands From a NAND SSD — NAND architecture, KV-cache storage, and the characteristics of inference workloads.
HBM is the Ugly Baby: What Now? — How memory architecture innovation could address HBM’s cost and capacity constraints.
Hot Chips 2026: Tuning into Memory Vibes — Technical and industry observations from Micron, Samsung, and SK Hynix.
How Custom Base Die Drives HBM Bandwidth, and Who Stands to Benefit — Why custom base dies could shift differentiation and bargaining power in HBM.

3. Copper Interconnects, Networking, and Scale-Up

How do thousands of accelerators communicate? The answer depends on reach, signaling speed, power, latency, cost, and increasingly difficult physical constraints.

Why Optical Circuit Switching is Arista Networks’ Long-Term Problem — The architectural implications of optical circuit switching for traditional networking.
XPO vs. CPO: Speed, Power, and Modularity in Next-Gen AI Networking — Comparing optical integration approaches and their trade-offs.
How Rack Power Density Is Opening a New Market for Active Copper — Why higher rack density can create demand for active electrical cables.
Kyber Midplane Design, Making NVL72x2 Work, and All Roads Lead to CPO — NVIDIA’s scale-up networking challenges viewed through manufacturing and signal integrity.
Astera Labs: From the Tray to the Rack, and What Comes After — PCIe switches, retimers, rack connectivity, and emerging interconnect standards.
Astera Labs Is a Switch Company in Connectivity Clothing — Examining the company’s evolution from retimers toward switching.
OCP APAC: Copper vs. Optics Contradictions — Competing views on where copper ends and optical networking begins.
The PAM4 vs. PAM6 Dilemma for 448G Copper Interconnects — The signal integrity and modulation trade-offs involved in extending copper to higher speeds.

4. Photonics, Lasers, and Optical Networking

From the physics of semiconductor lasers to co-packaged optics, microLEDs, and optical component suppliers.

Semiconductor Lasers and Co-Packaged Optics

Lumentum: Laser Demand, OCS, CPO and Optical Scale-Up — Lumentum’s position across the growing optical networking market.
Grating Coupled COUPE for CPO, and Himax’s Fragile FAU Moat — TSMC’s optical coupling technology and its implications for packaging suppliers.
MicroLEDs vs. Lasers: The Linewidth Tradeoff — How spectral linewidth and dispersion determine optical link performance.
Lasers for CPO/NPO, Part 1: The InP DFB Laser — Semiconductor laser physics and the four limitations on high-power operation.
Lasers for CPO/NPO, Part 2: Lumentum’s Technology and Moat — High-power laser design and the sources of competitive differentiation.
How Coherent Can Earn Investor Confidence — Evaluating laser performance claims, qualification, and technical evidence.
Lasers for CPO/NPO, Part 3: Master Oscillator, Power Amplifier — Comparing MOPA and single-cavity approaches to ultra-high-power lasers.

Emerging Optical Technologies and Industry Developments

A MicroLED Field Trip to Avicena — Firsthand observations of microLED interconnect technology, manufacturing, and performance.
ECOC 2026: Important Product Announcements Around Optical Scale-Up — MicroLED testing, VCSELs, lasers, and emerging optical solutions.
OCP 2026: MicroVCSELs Are Gaining Attention, but MicroLEDs Still Have a Case — Where different microemitter technologies fit in short-reach networking.
What a Marvell-Lumentum Deal Means for DSPs, Lasers, and Modules — Vertical integration, optical DSPs, laser supply, and module economics.

5. AI Datacenter Power, Cooling, and Physical Infrastructure

The limitations of electricity delivery, thermal management, and datacenter construction are increasingly shaping AI hardware architectures.

AI Datacenters Drink More Water Than You Think — Understanding water consumption, cooling, and environmental trade-offs.
Power Delivery As The Next Physics Wall In AI Datacenters — The physical argument for higher voltages and the resulting power-conversion architectures.
800VDC: Mapping the Power Semi Space for Sidecar Deployments — Near-term architectures, semiconductor opportunities, and SiC/GaN/silicon trade-offs.
Why AI Data Center Constraints Make Efficiency More Valuable — Why the output of existing facilities matters when approvals, power, and physical space become constrained.

6. Semiconductor Manufacturing, Foundries, and Advanced Packaging

Understanding the manufacturing constraints behind architectural choices, and the companies capable of delivering them.

High Stakes Advanced Node Poker: TSMC Goes All-in, Intel Folds — Competitive strategy and manufacturing commitments at advanced process nodes.
Huawei’s Tau Scaling Is Really a Hybrid Bonding Bet — Advanced packaging and hybrid bonding as alternatives to conventional scaling.
Battle for Advanced Packaging: TSMC Monopoly, Intel Challenge, and Amkor Arms-Dealing — Competitive positions in the advanced packaging supply chain.
Qualcomm’s High Bandwidth Compute and the Packaging Problem It Moved — The architectural advantages and manufacturing challenges of three-dimensional memory integration.

7. Conference Research and Field Reports

Technical developments and firsthand observations from semiconductor conferences, with an emphasis on what the presentations imply beyond the announcements themselves.

2025 International Electron Devices Meeting: Detailed Conference Report — Device technology and manufacturing developments from IEDM, published in January 2026.
Computex 2026 Debrief — Key developments and observations from Computex.
2026 OCP APAC: Full Conference Report — Copper, optics, XPO, and power delivery architecture developments.
TEF and OCP Global Summit 2026: Game Plan — The technical questions and conference sessions selected for investigation.

Additional conference-derived research appears in the relevant sections above, including Hot Chips memory coverage, ECOC optical announcements, and OCP microVCSEL developments.

8. Industry Strategy, Market Dynamics, and Company Analysis

Connecting technology developments to commercial outcomes, industry structure, and investment expectations.

The AI Trade Burned Down: Let Round 2 Begin — What an AI-sector market correction reveals about expectations, financing, and infrastructure demand.
The AI Buildout Will Continue — The implications of AI development, safety concerns, and demand for compute.
Intel: Why Didn’t We See This Coming? — A retrospective on Intel’s earnings, missed warning signals, and the lessons for analyzing semiconductor companies.

9. This Week in Chips — News and Commentary

In addition to the longer research articles, I published This Week in Chips (TWiC) roundups covering company announcements, technology developments, earnings, and shorter technical observations. These articles are all free, but I discontinued writing these midway through the year, as I felt it was becoming more noise than signal.

These are organized from the beginning of 2026 onward.

10. Publication Updates and Corrections

These posts document developments in the publication and corrections to earlier articles.

You can also browse the complete publication chronologically through the Vik’s Newsletter archive.

Get Full Access to Vik’s Newsletter

The purpose of this newsletter is to help you understand semiconductor technologies at a deeper level, not simply follow the latest announcements.

A paid subscription unlocks the complete collection of paid technical articles and industry analysis, including the full archive and all future paid research.

Whether you’re an engineer, investor, researcher, or industry professional, the goal is to give you the technical understanding to evaluate emerging technologies and the companies building them.

Thank you for supporting independent, engineering-first semiconductor research.

Engineer first. Analyst later.