Start with a sharp question
A useful essay begins with a decision, bottleneck or contradiction, not a trend label.
02
Writing
79 essays · 9 themes · Source-backed technical analysis
I write to understand
the systems I build.
Every essay from the original archive now lives here. Each starts with a technical question, follows the evidence and identifies the constraint that actually matters.
IssueSystemConstraint
002ModelsAttention
019MemoryAllocation
022InterconnectBandwidth
043ComputeArchitecture
055LithographyThroughput
064FoundryCapacity
Selected essaysSix systems.
Six systems.
Six hard constraints.
A focused selection across model architecture, memory, interconnects, CPU design, foundry capacity and lithography. Every essay now opens inside this portfolio.
002Model architecture
The title was wrong. That is what made it historic.
Why removing recurrence made parallel training practical and changed the direction of modern AI.
10 min019AI infrastructureThe AI Memory Tax
How AI servers are repricing DRAM and NAND for smartphones, PCs and the old semiconductor cycle.
12 min022AI infrastructureWhen AI Runs Out of Copper
Why optical I/O and co-packaged photonics are becoming a critical AI systems bottleneck.
12 min043Compute architectureThe Open ISA Talent Bet
What Rivos reveals about elite CPU teams, software, memory architecture and hyperscaler control.
12 min064Semiconductor systemsTSMC's Capex Became the AI Industrial Plan
How foundry spending became the physical buildout of the AI compute economy.
15 min055Semiconductor systemsThe EUV Throughput Bottleneck
Why resolution is only half the problem when AI demand requires more photons, uptime and wafers.
14 minComplete archiveAll 79 essays.
All 79 essays.
The complete record.
The shortlist above is where a recruiter should start. This is the complete record, preserved inside the new portfolio and searchable by subject.
Showing 79 of 79 essays
079Apple Silicon16 min read→078Compute systems16 min read→077Foundries16 min read→076Memory16 min read→075Optical and RF16 min read→074Memory16 min read→073Compute systems16 min read→072Compute systems16 min read→071Apple Silicon15 min read→070Optical and RF16 min read→069Policy16 min read→068Policy16 min read→067Equipment15 min read→066Foundries15 min read→065Packaging15 min read→064Foundries15 min read→063Foundries15 min read→062Foundries15 min read→061Packaging15 min read→060Compute systems14 min read→059Equipment14 min read→058Foundries14 min read→057Optical and RF14 min read→056Foundries14 min read→055Equipment14 min read→054Optical and RF14 min read→053Policy14 min read→052Apple Silicon14 min read→051Equipment14 min read→050Memory14 min read→049Memory14 min read→048Packaging14 min read→047Apple Silicon13 min read→046Packaging12 min read→045Optical and RF12 min read→044Compute systems12 min read→043Compute systems12 min read→042Policy12 min read→041Compute systems12 min read→040Packaging12 min read→039Packaging12 min read→038Packaging12 min read→037Equipment12 min read→036Compute systems12 min read→035Foundries12 min read→034Optical and RF12 min read→033Compute systems12 min read→032Compute systems12 min read→031Memory12 min read→030Compute systems12 min read→029Foundries12 min read→028Optical and RF12 min read→027Compute systems12 min read→026Foundries12 min read→025Compute systems12 min read→024Optical and RF12 min read→023Compute systems12 min read→022Optical and RF12 min read→021Compute systems12 min read→020Compute systems12 min read→019Memory12 min read→018Packaging12 min read→017Foundries12 min read→016Optical and RF10 min read→015Compute systems12 min read→014Compute systems12 min read→013Equipment12 min read→012Memory12 min read→011Strategy10 min read→010Strategy12 min read→009Strategy13 min read→008Strategy11 min read→007Strategy10 min read→006Strategy10 min read→005Strategy12 min read→004Strategy10 min read→003Strategy14 min read→002Strategy10 min read→001Strategy10 min read→
Apple M2 Was the Costly Transition Chip. M5 Shows Apple's Real AI Silicon Direction.
M2 looked underwhelming because the CPU uplift was modest and the die grew. Looking back from M5, M2 was the costly transition chip that widened the platform so M3, M4 and M5 could turn Apple Silicon into a local AI engine.
Ampere's Cloud-Native Pitch Wasn't Fluff. The IPO Story Became a SoftBank AI Infrastructure Bet.
Ampere's cloud-native CPU thesis was real: many small cores, no SMT, predictable performance and low cost per core. But by 2026, hyperscalers built their own Arm CPUs and SoftBank acquired Ampere for $6.5B. Cloud-native won; Ampere's merchant moat narrowed.
Meteor Lake Proved Intel's Chiplet Future. It Did Not Prove Process Leadership.
Meteor Lake did not deliver a clean 2x Intel 4 product shrink. Its real lesson was chiplets: Intel 4 compute, external foundry tiles, Foveros, and node choice by function. The real comeback test is Panther Lake and Clearwater Forest on 18A.
Marvell's Tanzanite Bet Was Early. CXL Is Now Becoming the Memory Pooling Layer for AI Infrastructure.
Marvell's Tanzanite acquisition was an early CXL memory-pooling bet. In 2026, AI inference, KV cache, and memory utilization make CXL a serious composable infrastructure layer, but not an HBM replacement. HBM feeds the GPU. CXL helps feed the system.
Ayar Labs Was Early on Optical I/O. Now It Has to Prove CPO Can Ship at AI Scale.
Ayar Labs saw the AI data movement bottleneck early. Nvidia, TSMC, Alchip, Wiwynn, Lightmatter, AMD, and Marvell now validate optical I/O, but CPO still has to prove production reliability at AI scale. The bottleneck has moved from physics to production.
Samsung's Semiconductor Crisis Was Real. The AI Memory Boom Gave It a Second Chance.
Samsung's 2022 semiconductor crisis was real, but AI memory changed the story. HBM4 and server memory are driving a comeback, while Foundry still has to rebuild trust against TSMC. Memory comeback is real. Foundry is still a trust rebuild.
Nvidia's Empire Was Not Eroded. But Intel's Network and Edge Playbook Still Matters.
Nvidia's AI empire expanded, but Intel's Network and Edge playbook still matters. Edge AI, vRAN, OpenVINO, IPDK, P4, Open Edge Platform, BlueField-4, Aerial, Vera Rubin, and the Nvidia-Intel partnership define the 2026 platform-vs-abstraction reality.
Nvidia's Hopper Supplier Cut Was a Warning: AI Hardware Design Wins Are Not Moats
A 2022 Hopper supplier cut showed how dangerous Nvidia customer concentration can be. In 2026, the risk is bigger as Nvidia controls rack-scale AI systems, power, cooling, networking, and platform architecture across Blackwell GB200 NVL72 and Vera Rubin.
Apple's iPhone Chip Split Became the New Normal. Then AI Changed the Rules Again.
Apple's iPhone chip split was not a one-off. Moore's Law economics forced Pro-first silicon, then Apple Intelligence raised the baseline chip floor across the lineup. The modern iPhone strategy is a silicon ladder shaped by A-series tier, GPU, memory, thermals, camera pipeline, N1 wireless, and C-series modem subsystems.
GlobalFoundries Fotonix Was Early. Now Silicon Photonics Is Becoming AI Infrastructure.
GlobalFoundries Fotonix was early to the optical I/O future. In 2026, Nvidia, Lightmatter, Marvell, AMD, TSMC, and GF are pushing silicon photonics deeper into AI infrastructure through co-packaged optics, photonic interposers, AMF, and the New York Advanced Packaging and Photonics Center.
Nvidia Was Hacked in 2022. The Real Lesson Is AI Infrastructure Security.
The Nvidia hack was not just a corporate breach. It showed that AI infrastructure security depends on identities, firmware, drivers, code signing, simulation tools, repositories, supplier portals, and the confidential-computing layer around advanced GPUs.
I, Semiconductor: Regionalization Became the New Cost of AI Sovereignty
Semiconductors are not deglobalizing. They are becoming regionally duplicated around trusted blocs through the US CHIPS Act, the EU Chips Act, Rapidus, TSMC Arizona, and the SEMI capex cycle. The cost of AI sovereignty is higher capex, duplicated fabs, and dependence on chokepoint suppliers.
Veeco's Transformation: The Small Toolmaker Under AI's Biggest Hardware Bottlenecks
Veeco does not make GPUs or HBM, but its tools sit underneath AI hardware bottlenecks: EUV mask blanks, laser annealing, advanced packaging, GaN power, photonics, and data storage. The Axcelis merger turns this niche turnaround into a US wafer fabrication equipment scale story.
Intel's Turnaround Is Now an 18A and 14A Execution Test
Intel's 2022 turnaround story was built on culture change, EUV, foundry ambition, and aggressive process roadmaps. In 2026, 18A is the technical proof point and 14A is the economic test that will decide whether Intel Foundry can become real.
The Bonding Step Became the AI Packaging Bottleneck
AI packaging is not only about CoWoS and HBM. Thermocompression bonding, fluxless TCB and hybrid bonding are becoming strategic because advanced AI packages are too valuable to lose at the bond. ASMPT, Besi, K&S, Intel Foveros Direct and TSMC SoIC are quietly rewriting what counts as a leading edge step.
TSMC's Capex Became the AI Industrial Plan
TSMC's 2026 capex is not just foundry spending. It is the industrial buildout behind AI compute across 2nm, 3nm, CoWoS, SoIC, advanced packaging and HPC capacity.
The Semiconductor Cycle Split in Two
The 2026 semiconductor market is approaching $1 trillion, but AI has split the cycle. Logic, HBM, EUV, CoWoS and hyperscaler silicon are booming while traditional electronics recover unevenly.
TSMC Became the Allocation Layer of AI Compute
TSMC is not just a wafer manufacturer. In the AI era, it allocates scarce advanced-node wafers, CoWoS capacity, SoIC, HBM integration, yield-learning windows and packaging slots across Apple, Nvidia, AMD, Qualcomm, Broadcom, MediaTek and Intel.
The Package Became the Computer
Advanced packaging moved from a back-end detail to the main AI scaling layer. CoWoS, HBM, SoIC, chiplets, hybrid bonding and die-to-die IO now matter as much as transistor scaling.
AWS Turned the CPU Into Cloud Infrastructure
Graviton3 was not just an Arm CPU. It showed AWS using Nitro, chiplets, custom storage, packaging and server design to turn the CPU into a fleet-level cloud economics weapon.
The EUV Materials War Behind AI Chips
ASML gets the attention, but EUV scaling also depends on photoresist, coaters, developers, dry resist, MOR, etch and defect control. The hidden battle is turning scarce EUV photons into high-yield wafers.
Intel Bet the Farm. Now It Has to Prove the Foundry
Intel's 2021 turnaround bet was necessary, but 2026 is the proof phase. 18A is in production, foundry losses are real, expansion is being disciplined, and Intel must earn external foundry trust.
Qualcomm's Modem-to-Antenna Moat
Qualcomm's ultraBAW filters were never just a component story. They were part of a broader modem-to-antenna strategy across RFFE, 5G, Wi-Fi 7, automotive, IoT and future wireless systems.
The Node Race Became a System Race
TSMC's 3nm delay was an early sign that node transitions were getting harder. In 2026, the foundry race is about more than 2nm: EUV, backside power, packaging, HBM, yield, customer trust, and AI-scale capacity.
The EUV Throughput Bottleneck
EUV source power was a bottleneck in 2021. By 2026, ASML improved throughput, High-NA readiness, and source power, but AI chip scaling still depends on more photons, uptime, and lower cost per wafer.
The Foundry That Quit Moore's Law Found a Second Life
GlobalFoundries did not become the next TSMC. But RF SOI, FD-SOI, silicon photonics, secure supply, automotive, defense, and AI data-center interconnects made its specialty foundry strategy far more important by 2026.
China's Silicon Carbide Bet: From Subsidy Boom to Overcapacity War
China built serious silicon carbide capacity, but SiC is not won by capex alone. The hard test is yield, reliability, automotive qualification, 8-inch manufacturing, and profitability through overcapacity.
The A15 Die Shot That Predicted Apple Silicon's Future
Apple's A15 looked weak if judged only by CPU gains. But the die shot showed Apple moving silicon budget into cache, GPU, neural compute, ISP, media, and system-level efficiency.
The Reliability Layer Behind AI Hardware
As AI processors, silicon photonics, silicon carbide devices, and advanced packages become more expensive, burn-in and reliability screening become critical infrastructure.
The Memory Oligopoly Became AI's Toll Booth
In 2021, Micron said it would not add DRAM or NAND wafers. In 2026, AI turned memory supply discipline into a bottleneck across HBM, server DRAM, NAND, and enterprise SSDs.
China's NAND Breakout Was Delayed, Not Defeated
YMTC did not create the NAND apocalypse on schedule. Export controls delayed it, AI demand absorbed it, and domestic equipment constraints reshaped it. But China's NAND breakout is still reshaping the global memory market.
The Boring Back-End of AI
The AI boom is not only about GPUs, HBM, EUV and CoWoS. Kulicke & Soffa shows why semiconductor back-end assembly, bonding, packaging and interconnects matter more than most people realize.
Apple Silicon After the CPU Era
In 2021, Apple's A15 looked like a warning sign. In 2026, the verdict is clearer: CPU gains slowed, but Apple moved the battleground to GPU, Neural Engine, memory, thermals, and the whole chip.
The Package Escaped the Wafer
How Applied Materials' 600mm panel-level packaging bet aged into the AI mega-package race, as CoWoS expands, HBM counts rise, glass substrates emerge, and the package becomes the scaling surface.
The Dead-Zone Layer
How the iPhone satellite rumor evolved into a real battle over low-bandwidth dead-zone connectivity across Apple, Amazon, Globalstar, Starlink, T-Mobile, carriers, 3GPP NTN, and FCC SCS.
The Vehicle Compute Backbone
Why NXP's 2021 Samsung-rumor story aged into a cleaner thesis about software-defined vehicles, zonal architecture, secure connectivity, edge AI, and the embedded layers that make intelligent machines work.
The Open ISA Talent Bet
Why Rivos shows that high-performance RISC-V is not about free cores, but elite CPU teams, full-stack software, memory architecture, and hyperscaler control.
The Architecture Gatekeeper
Why Arm China shows that in semiconductor IP, losing control over the distribution channel can become almost as serious as losing control over the technology itself.
The Custom AI Hardware Trap
Why Tesla Dojo showed that custom AI hardware is not won by impressive silicon alone, but by memory balance, software, interconnect, packaging, cooling, power delivery, and economics.
The Package Became the Computer
How Tesla Dojo showed that AI scaling is not just about faster chips, but package-scale integration, bandwidth, SRAM locality, power delivery, cooling, software, and rack-level systems.
The Back-End Bottleneck
Why Kulicke & Soffa's wire-bonder cycle is becoming a broader story about AI packaging, memory interconnect, thermo-compression bonding, hybrid bonding, and power semiconductor assembly.
The Wafer-Scale Training Bet
How Tesla's Dojo showed that AI scaling is not just about faster chips, but packaging, interconnect, memory, power delivery, cooling, and software.
The Semiconductor Substitution Ladder
Why China's chip-tool localization is real in cleaning, etch, deposition, CMP, and mature nodes, but still hits the hardest walls in lithography, photoresist, metrology, and advanced-node process integration.
The Power Efficiency Layer
Why Power Integrations is no longer just an ESG semiconductor story, but a hidden power-conversion layer for GaN, EVs, grids, battery storage, industrial systems, and 800 VDC AI factories.
The Foundry Trust Test
Why Intel's foundry comeback will not be proven by logos, evaluations, or packaging wins, but by committed external wafer volume.
The Android Flagship Breakout
How MediaTek moved from cheap-chip stigma to TSMC-node flagship, on-device AI, Smart Edge, automotive, and cloud AI ASIC ambitions.
The Bubble That Became Infrastructure
Why Nvidia's 2021 overvaluation story aged half right: crypto, gaming, and stock-split speculation faded, but the AI factory platform became real.
The AI Chip Software Wall
Why Graphcore's IPU shows that great silicon is not enough to beat Nvidia: AI hardware is a full-stack war of CUDA, kernels, frameworks, benchmarks, cloud access, trust, and capital.
The Fab That Outlived 3D XPoint
How Micron's failed 3D XPoint memory fab became Texas Instruments' 300mm analog and embedded semiconductor supply-chain weapon.
The Wafer-Scale Latency Bet
Why Cerebras is betting that the fastest path for inference and scientific AI is not more distributed complexity, but fewer chip boundaries.
The GAA Credibility Test
Why Samsung Foundry's 2nm comeback is about trust, not transistor marketing: from 3GAE delays to SF2, HBM base dies, advanced packaging, and the race to become TSMC's credible alternative.
The AI-Native Network
How Qualcomm's 5G infrastructure bet evolved into the AI-native network: private 5G, edge AI, Open RAN acceleration, positioning, AI-RAN, and 6G.
The Networked AI Bet
Why Tenstorrent is attacking Nvidia with open software, RISC-V IP, Ethernet-native scale-out, and switch-light AI systems rather than simply building a bigger GPU.
The Foundry Toll Road
Why TSMC's pricing power got stronger in the AI era: customers are buying performance-per-watt, CoWoS capacity, HBM integration, yield, scale, and trust.
The Clinic-on-Wrist Reality Check
Why Rockley's clinic-on-wrist vision was directionally right but early, and why non-invasive glucose and continuous biomarkers remain one of the hardest problems in wearable health.
The Other Leading Edge
Why GlobalFoundries still matters in an AI world without chasing 2nm: silicon photonics, RF-SOI, FD-SOI, GaN, eMRAM, and specialty manufacturing. TSMC dominates the logic frontier. GF is trying to own selected specialty frontiers.
The Inference Efficiency War
How Qualcomm is turning Cloud AI 100's edge-inference DNA into AI200 and AI250 rack-scale infrastructure for cost-per-token AI serving. Training gets headlines. Inference gets the bill.
When AI Runs Out of Copper
Why AI factories are hitting data-movement limits, and why optical I/O, co-packaged optics, Intel photonics, Nvidia networking, Broadcom switches, and TSMC COUPE matter. The race to replace copper with light has begun.
The Custom Silicon Flywheel
How Google's Argos VCU became the blueprint for hyperscaler custom silicon, from video acceleration to Axion CPUs, Ironwood TPUs, Meta MSVP, AWS Trainium, and Microsoft Maia. The cloud is becoming a portfolio of domain-specific machines.
Nvidia's Earnings Quality Test
From crypto GPUs in 2021 to AI capex in 2026, the real question is whether Nvidia's explosive data-center revenue turns into durable utilization and customer cash flow. The numbers are real. The question is who ultimately pays for the AI buildout.
The AI Memory Tax
How AI servers are repricing DRAM and NAND, squeezing smartphones, PCs, and low-end consumer electronics, and splitting the old semiconductor cycle. AI infrastructure is booming while consumer electronics absorbs the cost.
The Boring Back-End Boom
Why mature nodes, wirebonding, power modules, OSATs, and semiconductor back-end equipment are becoming strategic again. AI made the front-end famous. The back-end is where every chip becomes a product.
The Density Illusion
Why Apple's A14 exposed the gap between theoretical transistor density and real chip density, and why AI-era scaling now depends on SRAM, memory bandwidth, packaging, power, and architecture. Moore's Law did not die. It became a system problem.
The Modem-to-Antenna War
How Apple is taking apart Qualcomm's modem-RF moat, from the iPhone 12 mmWave antenna decision to the Apple C1 modem. Qualcomm still has the best complete engine. Apple is trying to own the car.
Nvidia Didn't Buy Arm. It Built the AI Factory Anyway.
How Nvidia failed to buy Arm but still built the CPU-GPU-DPU-networking-software stack behind the modern AI factory. The acquisition failed. The strategy did not.
MediaTek and the Fragmented Compute War
How MediaTek moved from Huawei-substitution thesis to a broader role in flagship phones, edge AI, automotive platforms, and custom data-center ASICs. In 2020, the question was who supplies Huawei. In 2026, the question is who supplies compute.
The Dry Resist War
Why Lam's Aether dry resist matters for EUV, High-NA lithography, DRAM, HBM, and AI-era chipmaking. EUV exposed the wafer. The resist has to do the rest.
The AI Memory Wall
Why DRAM, HBM, advanced packaging, and semicap became central to the AI infrastructure race. The AI boom is usually described as a GPU shortage. That is only half true. The deeper bottleneck is memory.
The Prompt Is Not Infrastructure
Prompts are instructions, not controls. As AI agents gain access to tools, private data, and external actions, trustworthy AI will require permissions, sandboxes, logs, and real infrastructure.
The Interface Coup
AI will not merely replace search. It may replace the interface layer. When assistants become the front door, apps become callable infrastructure and distribution moves from visibility to permission.
The Memory Moat
The next AI moat may not be the best model. It may be memory: the accumulated context that makes an assistant feel like it knows you. When AI remembers who you are becoming, switching tools becomes harder.
The Model Monoculture
What happens when millions of people think, write, code, research, and decide through the same few AI models? The danger is not only wrong answers. It is correlated cognition.
Liability Laundering
AI agents will not just answer questions. They will act across systems, tools, data, and customers. The hidden danger is liability laundering: humans remain responsible while practical control moves elsewhere.
The Free Intelligence Trap
Free AI is not just the next ad-supported product. Search monetized what people wanted. Social monetized who they were. AI may monetize the moment before they decide.
The Cheap Code Era
AI will automate coding before it can own software. The cheap code era is here, but the expensive part of software is moving from syntax to systems, judgment, verification, security, and accountability.
Yann LeCun's Billion-Dollar Bet: Intelligence Beyond the Next Token
LeCun's argument is not that LLMs are useless. It is that real intelligence may require world models that predict consequences, plan actions, and understand reality beyond text.
The Boiling Frog Economy
How AI quietly breaks the ladder before it breaks the job market. The danger is not that some jobs disappear. It is that the first rung disappears first.
The title was wrong. That's what made it historic.
Everyone remembers 'Attention Is All You Need' as the attention paper. The real innovation was something later summaries under-emphasized, and it explains everything that came after.
There is no Alexander without Aristotle
On intelligence as infrastructure, why the education system was never designed for you, and what it took to understand the world on my own terms.
Editorial methodA repeatable way
A repeatable way
to think in public.
The writing is useful evidence of how I frame an unfamiliar system, test a claim and turn research into an explanation someone else can use.
Build an evidence ledger
Separate sourced facts, reasonable inference and open uncertainty before writing the conclusion.
Explain the system
Connect the mechanisms so the reader leaves with a model that transfers to the next problem.
Read nextFrom written model
From written model
to working map.
The interactive atlases turn the same research habit into navigable systems. Follow a field from its foundations and papers through compute, applications and open questions.
Explore the atlases