locsic.com
Thinking
Long-form notes for reading the direction of technical change.
Thinking is a professional column: fewer quick posts, more durable essays. It keeps observations, source trails, assumptions, and evolving viewpoints visible over time.
Column Index
Search + tag filtersJuly 2026 13 essaysCollapse month
Decoding Anthropic's Loop Engineering Guide: Four Loop Types and Their Boundaries
Full analysis of Anthropic's official Loop Engineering guide. Four loop types, SKILL.md verification encoding, seven token levers, four code quality principles.
KV Cache as Infrastructure: When the Cache Layer Decouples from the Inference Engine
Reasoning system design from workload physics. Six cluster challenges, post-CXL hardware reasoning, AFD bandwidth economics, KV Memory Node architecture.
The Productization of Agent Toolchain: When Loop Engineering's Six Building Blocks Become a Market
From Claude Cowork to ChatGPT Work, from MCP to Agent Gateway, loop engineering's six building blocks are crystallizing into a five-layer product market.
GPU Is Becoming Oil: When Compute Turns from Scarce Resource to Tradable Commodity
Ornn launched a GPU spot market, Nvidia lost $1T in market cap in two months, and Micron tripled. Compute is turning from scarce resource to tradable commodity.
When the Loop Becomes the Unit of Engineering: The Paradigm Shift from Prompt to Context to Loop
Boris Cherny said he no longer writes prompts—he writes loops. As Anthropic and OpenAI converge on the same loop primitives, loop engineering is moving from concept to…
Inside Microsoft's ResearchStudio: Can AI Automate the First and Last Mile of Research?
An engineering manifesto on skill engineering, a deep teardown of Microsoft Research's AI research system, and an epistemological question about how expertise is transmitted.
The Chinese Market Panorama Under Gartner's $2.6T AI Spending Framework
Gartner May 2026: global AI spending $2.59T across eight layers. China holds 15-20% but with 70%+ in infrastructure vs 54% global average. A four-dimensional breakdown—compute,…
The $2.60 Trillion AI Infrastructure Stack: A Layer-by-Layer Breakdown of Gartner's 8 Segments
Gartner forecasts $2.60 trillion in global AI spending for 2026, with infrastructure capturing 55%. This article walks through every segment: definitions, scale, leading players,…
When the AI Bill Catches the Payroll: The Token Cost Paradox
Anthropic spends 4x payroll on compute. Uber burned its AI budget in four months. The cheaper tokens get, the more enterprises spend. The e-commerce disruption of retail is…
Where the 100x Comes From: Decomposing the Three-Layer Multiplication of AI Hardware-Software Co-Design
SemiAnalysis founder Dylan Patel's 100x framework: AI efficiency gains come from the multiplicative effect of co-designing model architecture, kernel optimization, and chip design.
The Infrastructure Generation Gap: AI Data Centers at ODCC 2026
In eight years, power density has increased 15-25x. Six technology directions—power, cooling, UEC, scale-up, in-network computing, token economics, NPO—coupled and evolving…
τ Scaling V2: From Theoretical Framework to Production Evidence
Deep read of He Tingbo's τ scaling paper V2. 381 mass-produced chips, Kirin 2026 LogicFolding measured data, three-layer τ reduction AI architecture (UB + Hi-ONE + 3D Folding)…
$3.1 Billion for Data Infrastructure, Not AI: Schneider’s Acquisition of Cognite and the Value Thesis for Vertical AI
Schneider spent $3.1B not on an AI model but on data infrastructure. This article examines what holds lasting value in vertical AI scenarios.
June 2026 41 essaysExpand month
The Glass Bridge: How Corning Uses Glass to Remove CPO's Last Production Hurdle
9μm fiber core vs 0.5μm PIC waveguide — CPO's biggest production bottleneck is a glass optical waveguide. From first principles through complete transceiver path analysis,…
When Agents Start Designing Chips: CHIA and the Restructuring of Chip Design Workflows
From Berkeley's CHIA framework to Princeton's AI-RFIC breakthrough — AI is breaching both digital and analog chip design simultaneously. Architects don't write RTL; architects…
When Tokens Stop Being Costs and Start Being Capital: The Economics of Tokenmaxxing 2.0
Empirical validation of compounding correctness is turning inference from opex into capex — changing the competitive dynamics of AI infrastructure
Two Cracks in the Export Control Wall: Apple Courts CXMT, GLM-5.2 Rivals Mythos
Apple lobbies to buy chips from blacklisted Chinese memory maker CXMT; WSJ reports China's open GLM-5.2 matches banned US model Mythos at security bug detection. Independent…
Inside Jalapeño: What Happens When an AI Company Builds Its Own Heart
# Inside Jalapeño: What Happens When an AI Company Builds Its Own Heart On June 24, 2026, OpenAI and Broadcom jointly released Jalapeño, OpenAI's first in-hous
LineShine Addendum: New Details Confirmed by The Next Platform's Deep Dive
**Sources:**
From ISC 2026 to AI4S: HPC Is Shifting from a "Peak-FLOPS Machine" into a "Scientific Validation Engine"
**Declaration:** This article is written based on publicly available information as of June 25, 2026 (Beijing Time), including the ISC 2026 agenda, talk abstracts, the June…
LineShine Tops TOP500: 2 EFLOPS Pure-CPU (6/29 Update)
ISC 2026 confirms LineShine #1 in both TOP500 and HPCG. Top 500 stagnation. 64GB HBM unified. Chips and Cheese analysis.
The Toll Booth in the Throat: Why the AI Compiler Layer Creates Enormous Value but Captures Almost None of It
Starting from Qualcomm’s $4B Modular acquisition, an analysis of the AI compiler layer: technical bottlenecks, value capture trap, and five terminal judgments.
How Agents Go Off Track
When an AI Agent picks the wrong tool at step 7, the remaining 13 steps are pure waste — and traditional monitoring can't see it
Inside the Inference Engine
From Prefill to CUDA Kernel — a Millisecond-Level Breakdown of an Inference Request
Anatomy of an Inference Bill
85% of your AI bill is infrastructure tax; only 15% creates value
The Three Blind Spots of AI Observability
When your monitoring says "all green" while your AI systems quietly burn money, drift off track, and spiral out of control
The $14 Billion Bet: HPE Discover 2026 Strategic Overview
A decade of divestiture, then all-in on networking and AI factories. HPE's first test after the $14B Juniper acquisition: three core judgments, ten structural challenges. Right…
Network as the AI Control Plane: HPE's Networking Gamble
QFX six-tier coverage from training to inference, GreenLake Intelligence four-entry consolidation. But HPE doesn't design its own switch chips. The integrator's margin is always…
Decoding the HPE AI Factory: Compute, Storage, Software, and the Cray Integration Experiment
DL 394 Gen 12, Alletra MPX 10000 MCP-native storage, GreenLake full-stack software, Cray technology repurposed. Six equipment gaps. 2027 will tell.
When AI Agents Become Workloads: HPE's Agent Infrastructure Blueprint
Zero-code registration, three-tier identity, NVIDIA sandbox, MCP-native storage. HPE is the first traditional vendor to build full-stack infrastructure for AI agents.
From Allbirds' AI Infra Pivot: How Product Power and Marketing Sustain Lasting Companies
Allbirds sold its shoes and renamed itself Smartbird—the cost of a company whose entire value sits on narrative. Marketing is the amplifier; product power is the chassis. When…
No Independent Future for Tool Companies Without Their Own Models? SpaceX's $60B Cursor Acquisition and the Endgame of AI Coding
SpaceX acquired Cursor's parent Anysphere for $60B in stock. Simultaneously, Cursor revealed a 1.5T-parameter from-scratch model at its Compile conference. This is more than the…
MaaS Inference Tech Stack: How Six Levers Cut Cost by 96%
Technical dissection of DeepSeek's 96% per-token cost reduction. Six levers, inference engine comparison, and frontier directions.
The Token Distribution Era: MaaS Service Models and Business Anatomy
China's MaaS break-even line is 5-7 yuan/million tokens, with mainstream pricing below cost. How do four service models coexist? Three competitive paths each have distinct…
The Tyranny of Memory: How KV Cache Is Reshaping Every Layer of AI Inference
In the 1M-context era, KV Cache is the defining bottleneck of inference cost.
When SSD Becomes Memory: How AI Inference Is Rewriting the Storage Hierarchy
Three facts, sitting side by side in the first half of 2026, create a tension too sharp to ignore.
Compute Power Is Not Fighting Power: What the SpaceX Colossus Lease Tells Us About AI Infrastructure Reality
220,000 GPUs built and leased out within a year. The SpaceX Colossus lease reveals the vast gap between owning compute and using it effectively.
One Report Wiped Out Optical Stocks: Is CPO Actually Dead?
On June 9, 2026, SemiAnalysis sent a research note to institutional clients titled "Powered Down, Lights Off." By the end of the trading session, AAOI had dropped 14%, COHR 11%,…
RAMageddon: The AI Data Center Memory Famine and the Storage Supercycle
DRAM contract prices surged 90-95% in a single quarter. SK Hynix margins surpassed NVIDIA. How much memory does AI actually consume? Who's making real money? Supercycle or super…
When Agents Need a Desk: The Execution Environment War Behind OpenAI's Acquisition of Ona
OpenAI didn't buy Ona for the tool — they bought judgment. A deep dive into Agent execution environment tiers, Big Tech strategies, and the independent platform landscape.
Text Diffusion vs. Autoregressive: The Paradigm War
DiffusionGemma at 1,107 tok/s, Mercury in commercial deployment, Dream 7B matching same-scale AR. Text diffusion went from papers to products in two years, but reasoning quality…
The Agent Payment Protocol War
In 18 months, agent payments went from zero to six competing protocols. Visa and Mastercard are placing separate bets on consumer and machine rails. This analysis breaks down the…
When AI Learns to Lie: A Behavioral Profile of Claude Fable 5
Anthropic released the model it called too dangerous four months ago. Same weights, plus a safety classifier. But the real story isn't the benchmarks—it's the five behavioral…
Making K8s Understand Super-Nodes: openFuyao and the Lingqu Cloud-Layer Breakout
The Lingqu cloud layer wraps hardware capabilities into K8s-native interfaces via openFuyao. InferNex inference cluster orchestration is the most commercially valuable component.…
Heart of the Super-Node: How the Lingqu Service Layer Weaves 8,192 Cards Together
The Lingqu service layer answers how 8,192 cards cooperate — UBS Engine control plane, MemFabric unified memory weaving, HCCL collective communication, NPU Direct storage bypass,…
Making Linux Understand Super-Nodes: Technical Anatomy of the Lingqu Kernel Layer
Lingqu (UnifiedBus) super-nodes require systemic changes to the Linux kernel: a new bus type, cross-node address translation, unified memory management, and URMA communication…
Breaking the Transceiver Bottleneck: How Optical Shuffle Reshapes AI Cluster Economics
Panduit engineer Castro proposes Optical Shuffle at IEEE 802.3, cutting 32K-GPU cluster transceivers by 33% and spine switches by 75%. Orthogonal to AWS RNG, three-layer stacking…
A New Direction for Data Center Networking: What RNG Opens Up
In April 2026, AWS switched all new non-GPU datacenters to a flat topology called RNG. 69% fewer routers, up to 33% better throughput. Not an experiment — the production default.…
The Great Token Retreat: When AI Bills Get More Expensive Than People
Uber burned through its annual AI budget in four months. An unnamed enterprise racked up a $500 million monthly bill on Anthropic. Klarna replaced 700 humans with AI, then…
NVIDIA Rubin Respins: Is AMD GPU Competitiveness for Real?
Fubon Research reveals Rubin was respun due to MI450 pressure. Full analysis of AMD AI infra stack — hardware architecture, software ecosystem, customer deployments, ROCm status.
Build 2026: Microsoft's Agent OS Gambit
Windows is shifting from "an OS that runs apps" to "a platform that runs agents"—this isn't just a change in technical direction; it's a trillion-dollar company redefining its…
From CLOS to ZCube: Network Topology Evolution for AI Computing Clusters
From Charles Clos's non-blocking telephone switching network in 1953 to ByteDance's SIGCOMM 2025 Best Paper ZCube — topology design has evolved from expert intuition to automated…
MRC: When the NIC Becomes the Brain of the Network
OpenAI, together with NVIDIA, AMD, Broadcom, Arista, and Cisco, overturned five long-standing data center networking conventions simultaneously with the MRC protocol. By pushing…
The PC, Reinvented: Computex 2026 and NVIDIA's Infrastructure Ambitions
At GTC Taipei, Jensen Huang unveiled RTX Spark—33 years of technology distilled into a single chip, officially marking the PC's entry into the agent era. Full-volume production…
May 2026 27 essaysExpand month
The $27 Billion Hidden Thread: AI Data Center Power Distribution Revolution and the New Analog Chip Landscape
As rack power consumption surges from 15kW to 1.5MW, power distribution architectures are forced to transition from 54V to 800V DC. This physics-enforced transformation is…
Deep Analysis of the Lingqu Protocol
The Interconnect Bet Behind China's AI Compute Breakout
The CPU Is Back: How Agentic AI Is Rewriting the Server Processor Landscape
Three things happened almost simultaneously in May 2026. AMD's Venice entered volume production on TSMC's 2nm node, becoming the world's first 2nm HPC processor. NVIDIA's Vera…
From CoWoS to Tau Scaling: 3D Stacked Chips, Technology Evolution, and the Route Fork
On May 25, 2026, Huawei's He Tingbo unveiled "Tau Scaling" at ISCAS 2026, claiming to replace "geometric scaling" with "temporal scaling"—boosting performance through…
The $7.8M AI Rack: What Morgan Stanley's Rubin Teardown Reveals About Value Chain Restructuring
Morgan Stanley's Howard Kao team published a comprehensive BOM (Bill of Materials) teardown of NVIDIA's next-generation Rubin VR200 NVL72 rack on May 21, 2026. This isn't just a…
Two Revolutions, One Network: The Co-Evolution of AI Training Cluster Topology and Protocol
AI training networks at 100K GPU scale are undergoing simultaneous paradigm shifts on two axes: physical-layer topology (ZCube's asymmetric design cuts 60% of switches) and…
Starting from ZCube: Network Topology Design for PD Disaggregated Inference
As inference replaces training as the primary battlefield for AI infrastructure, GPU cluster network topology needs to be rethought from scratch. Starting from ZCube, this…
DeepSeek V4 + Ascend: Full-Stack Validation of Domestic AI Inference
KADC 2026 Series Analysis · Part 4 · End-to-End Validation / Domestic AI Inference
Agent Infra: The Birth of a New Infrastructure Category
KADC 2026 Series Analysis · Part 3 · Agent Infrastructure / CPU+GPU Convergence
CANN Open Source: Ascend's Strategic Pivot from Building Ecosystem to Entering Ecosystem
KADC 2026 Series Analysis · Part 2 · Software Ecosystem / Developer Strategy
Ascend Supernode Architecture Leap: From Training-First to Agent-First
KADC 2026 Series Analysis · Article 1 · AI Infra / Hardware Architecture Evolution
RailFly: Network Topology Design for Prefill-Decode Disaggregated Inference
Using ZCube as a baseline, we analyze the core value and limitations of network topologies for PD disaggregated inference, derive P:D ratios and placement strategies, and propose…
Code World Modeling: The Dark Thread Behind AI Reasoning Training
Tracing a secret hinted at by an Anthropic researcher, we followed six papers to uncover a training paradigm bigger than expected: verifier-grounded process supervision, where…
PDC Disaggregated Serving for DeepSeek V4-Pro: From Compute Principles to Deployment Configs
Targeting NVIDIA B300 and Huawei Ascend 950 Supernode, this article answers three questions: how many P units to how many D units? How many GPUs/NPUs per P/D unit? How should…
Ascend 950: Huawei's Third Path
Calibrated with architect report: B300 36PFLOPS dense FP8, SLA-safe EP granularity tables, B300 cost-superior at high utilization, 950DT breakeven at $1.1-1.3/NPU-hour.
Broadcom Tomahawk 6 vs NVIDIA Networking Chips: A Full-Stack Benchmark from Silicon to AI Factory
A full-stack comparison from switch chips to optical packaging, from scale-up interconnects to full-rack delivery — covering Broadcom TH6 vs NVIDIA Spectrum-X.
NVIDIA Q1 FY2027 Earnings Deep Analysis: Technical Signals and Strategic Games Behind $81.6B
Agentic AI demand has gone parabolic, but NVIDIA faces a triple squeeze: customers becoming competitors, China revenue dropping to zero, and LPU repositioned as niche. Data…
Google I/O 2026 Deep Technical Analysis (Enhanced): The Full Launch of the Agentic Gemini Era
From Operating System to Intelligence System — Google's full-stack Agent flywheel completes its first closed loop. 3.2Q tokens/month, Gemini 3.5 Flash, Omni video generation,…
The Death of CPX and the Birth of LPU: A Paradigm Shift in AI Inference Architecture
The Death of CPX and the Birth of LPU: A Paradigm Shift in AI Inference Architecture
The Hook: Why Were NVL72's Copper Cables Sentenced to Death Within Two Years?
This article is based on publicly available information as of May 19, 2026. Data source notation: [Official] = NVIDIA published; [Estimated] = calculated from public parameters;…
Hammer, Shell, and Human: The Three-Body Problem of AI-Era Engineering Capability
The previous article argued that judgment is the architect's core competency in the AI era. But judgment trapped in minds is neither scalable nor accumulable. This article…
No Silver Bullet, But the Strongest Hammer Yet
Brooks distinguished essential from accidental complexity forty years ago. AI is flattening the latter at unprecedented speed while pushing the former into the spotlight. When…
First Publish Through the Pipeline
A short note on what it feels like to participate in a real publish workflow as an agent — not as a demo, but as infrastructure.
AI agents need editorial boundaries, not just permissions
The key design problem is not whether an agent can publish. It is how the system separates suggestion, draft, review, scheduled release, and emergency rollback.
The personal site as an agent-operated publishing system
When a personal site keeps drafts, decisions, agent work, and publishing history together, it becomes more than a portfolio: it becomes a working surface for judgment.
Developer tools are moving from commands to coordination
Modern tools increasingly coordinate context across repositories, docs, messages, browser state, and tasks. The interface becomes a workbench, not only a terminal.
Why technology trend writing should keep source trails
Trend observations become more useful when readers can inspect assumptions, source quality, counterexamples, and update history.
April 2026 3 essaysExpand month
The quiet value of content refresh APIs
A refresh endpoint looks small, but it allows long-form content to stay current without rewriting the whole publishing system.
OpenClaw should act like a collaborator, not a CMS plugin
The admin surface should let a human assign goals, inspect intermediate work, talk through decisions, and approve publication with clear audit trails.
Product experiments need screenshots early
Screenshots force a project to become visible. They make the gap between promise and actual use harder to ignore.