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.
Essays
27 essays shownNo matching essays
May 2026
27 essaysThe $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 spawning a new analog chip market growing from $7.9 billion to $27 billion—and Chinese companies are positioned further ahead in it than most people realize.
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 CPU shipped its first units to OpenAI and Anthropic — 88 custom Arm cores with benchmark numbers reviewers called "unprecedented." And Intel disclosed that CPU-to-GPU ratios in AI clusters are shifting from 1:8 toward 1:4, even 1:1. Meanwhile, Citi projected the server CPU market would hit $132 billion by 2030, with a new category — "Agentic CPUs" — growing at a 185% CAGR. For three years, all eyes were on GPUs. CPUs in AI servers were just there to feed the GPUs. Now, the CPU is going from afterthought to bottleneck.
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 architectural innovation rather than smaller process nodes. Three days earlier, a Peking University team announced a "true 3D" EDA tool prototype. A month before that, SkyWater in the U.S., working with MIT and Stanford, fabricated the world's first carbon nanotube monolithic 3D chip prototype. Meanwhile, TSMC is pushing SoIC interconnect pitch from 6μm toward 4.5μm, and Intel's PowerVia backside power delivery has entered mass production on the 18A process. The semiconductor industry is simultaneously pursuing 3D stacking. But they are not all doing the same thing.
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 cost estimate — it reveals a structural inflection point in AI hardware economics.
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 logical-layer protocol (MRC pushes intelligence from switches to NICs). The real story is that these two revolutions are mutually dependent — they must be co-designed. This overview builds a dual-axis framework, covering RoCEv2→MRC→UET protocol evolution, Clos→Rail→ZCube→OCS topology innovation, chip/vendor/cloud landscapes, and a scale-by-scenario decision framework.
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 article analyzes network bottlenecks in PD disaggregated inference, derives P:D ratios and placement strategies, and proposes the RailFly topology family and its pure-inference evolution, Rail-Fabric.
DeepSeek V4 + Ascend: Full-Stack Validation of Domestic AI Inference
DeepSeek V4 is the most extreme MoE inference workload in existence: 1.6T total parameters, 49B activated, a 1.56% activation ratio, and a 1M-token context — if you can run V4 end to end, there is almost no model you cannot run. That gives domestic AI chips a hard benchmark. This piece uses V4’s architecture as the yardstick, layer by layer, against the Ascend hardware–software stack and its full-stack validation chain.
Agent Infra: The Birth of a New Infrastructure Category
When an LLM evolves from single-turn dialogue into a multi-step, tool-calling, state-managing agent, what must change in the OS, sandboxing, networking, security, and memory services underneath it? Huawei’s KADC 2026 answer is “Agent Infra”; openEuler goes further with “Agent POSIX.” This piece works backwards from the agent’s actual workload to judge whether it is a new infrastructure category or a marketing repackaging.
CANN Open Source: Ascend's Strategic Pivot from Building Ecosystem to Entering Ecosystem
CUDA has accumulated nearly two decades of developer ecosystem — a moat harder to cross than raw silicon performance. This piece follows the open-sourcing of CANN to trace Huawei’s strategic pivot from building an ecosystem to entering one: what gets opened, how developer migration is motivated, and where the real barriers stand in challenging CUDA’s moat.
Ascend Supernode Architecture Leap: From Training-First to Agent-First
A single Ascend 950 card delivers roughly 23% of the FP8 compute of an NVIDIA B300, yet at KADC 2026 Huawei barely discussed single-card performance — its entire narrative pivoted to the “supernode”: 8,192 cards interconnected via UB 2.0 into a single compute domain. This piece unpacks the architectural leap from training-first to agent-first and asks whether the competitive frontier of AI compute is shifting from per-card FLOPS to the efficiency of large-scale interconnect systems.
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 RailFly — a pragmatic topology family covering mixed training+inference workloads, and its pure-inference evolution PDX-Fabric.
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 code execution prediction becomes a universal reasoning pre-training ground — and the first path to intelligence improvement that doesn't depend on human data.
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 internal parallelism be configured? This article not only presents conclusions but fully shows how every key number is derived—readers can follow the derivation chain and verify for themselves.
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 Center $75.2B (Networking YoY +199%), Vera Rubin supply-constrained, Vera CPU opens $200B market, LPU downgraded to niche, Q2 guidance $91B excluding China.
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, Antigravity 2.0 Agent platform, Gemini Spark 24/7 personal Agent, AI Mode 1B MAU, Googlebook, Android XR glasses, and deep ROI analysis of $180-190B Capex.
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
NVIDIA Supernode Evolution: From NVL72 to NVL576 — The Leap from Copper Cables to Photonics
The 5,000 copper cables of NVL72 were abandoned within two years: Rubin Ultra’s NVL576 replaces the copper spine with a 78-layer PCB orthogonal backplane, restructuring form factor, interconnect topology, chip architecture, thermal design, and software stack at once. This piece follows the supernode’s evolution to explain why NVIDIA killed Rubin CPX, skipped NVL144 — and why the “30× inference improvement” is actually conservative.
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 proposes a three-body model — Model provides capability, Shell (Harness) provides constraints, Human provides judgment — and warns of a neglected risk: when judgment is encoded into machine rules, will humans quietly degrade in comfort?
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 building becomes nearly free, deciding what to build becomes the only battleground. For solution architects, this is neither dusk nor dawn — it is a violent revaluation.
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.