(Illustration: AWS re:Invent 2025 Keynote with Peter DeSantis and Dave Brown. Image source: AWS.)

Peter DeSantis, SVP of Utility Computing at AWS, and Dave Brown took the stage at re:Invent 2025 to deliver a keynote that went deep into the infrastructure fundamentals powering the AI era. Rather than chasing the latest AI hype, they made a compelling case that the core cloud attributes we have relied on for two decades, security, availability, elasticity, agility, and cost, matter more than ever. The announcements ranged from Graviton5 with 192 cores in a single package and 5x more L3 cache, to Lambda Managed Instances that bridge the EC2-Lambda divide, to S3 Vectors hitting GA with sub-100ms queries over 2 billion vectors. On the AI acceleration front, Trainium3 UltraServer delivers 5x output tokens per megawatt, and PyTorch native support means porting GPU code to Trainium is literally a one-line change.

✳️ tl;dr

One theme “Infrastructure Fundamentals for the AI Era” runs throughout, with six sections:

• Core Cloud Attributes: Security, availability, elasticity, agility, and cost remain the foundation. These attributes guided every AWS decision for 20 years and are even more critical in the AI era.
• Nitro and Graviton Evolution: From custom silicon (Nitro) eliminating virtualization jitter to Graviton5 delivering 192 cores with 5x L3 cache. M9g instances offer up to 25% better performance than M8g. Guest: Payam Mirrashidi (Apple) on Swift + Graviton achieving 40% performance gains.
• Serverless Expansion: Lambda Managed Instances bridges the gap between EC2 performance and Lambda simplicity. Your Lambda functions run on EC2 instances you choose, while Lambda manages provisioning, patching, and scaling.
• Inference and Bedrock Architecture: Project Mantle inference engine powers Bedrock with service tiers (priority, standard, flexible), per-customer queue fairness, Journal for fault tolerance, and confidential computing.
• Vector Search and S3 Vectors: Nova Multimodal Embeddings unifies text, image, video, audio into shared vector space. S3 Vectors GA achieves sub-100ms queries on 2 billion vectors. 250K+ vector indexes created in 4 months. Guest: Jae Lee (TwelveLabs) on video intelligence.
• Trainium3 and AI Acceleration: Trainium3 UltraServer with 144 chips, 360 PetaFLOPS, 20TB HBM. 5x output tokens per megawatt. NKI GA and Neuron Explorer for performance profiling. PyTorch native support. Guest: Dean Leitersdorf (Decart) on real-time visual intelligence.

✳️ Live Experience

(Caption: This session used to be held on Monday evenings, originally known as Monday Night Live, and I’d often miss it due to dinner or meeting conflicts — only to catch up later. This year it clashed with other commitments again, so here are some stand-in photos of the Amazon EC2 UltraServer displayed on the CEO Keynote stage and the cute, mischievous Kiro who loves playing hide-and-seek — giving you a glimpse of the physical and protocol world that truly exists behind the cloud’s abstraction.)

View this post on Instagram

Read More

(Illustration: AWS re:Invent 2025 Keynote with Dr. Swami Sivasubramanian. Image source: AWS.)

Dr. Swami Sivasubramanian, AWS VP of Agentic AI, delivered the developer-focused keynote at re:Invent 2025, showcasing how AI agents are transforming the way we build software. From Blue Origin using agentic AI to design lunar infrastructure 75% faster, to Vercel powering 11 million customers with self-driving infrastructure, this keynote demonstrated that the age of autonomous AI agents is here. The announcement of Nova Act with 90% reliability for enterprise workflows marks a turning point for production-ready automation.

✳️ tl;dr

One theme “Agentic AI” runs throughout, with five sections:

• Understanding Agents: Agents sense environments, convert objectives into executable steps, and continuously learn; unlike chatbots that give advice, agents investigate and initiate solutions.
• Building Agents: Strands Agent SDK (5M+ downloads) and Amazon Bedrock AgentCore provide model-driven development with Identity, Policy, Evals, and Episodic Memory capabilities.
• Customizing Models: Reinforcement Fine-Tuning in Amazon Bedrock delivers 66% accuracy gains; Nova Forge enables custom frontier models by mixing proprietary data during mid-training.
• Ensuring Trust: Amazon Nova Act achieves 90% reliability through neuro-symbolic AI combining automated reasoning with LLMs; Cedar policy language provides deterministic controls.
• Reimagining Work: Kiro autonomous agent, AWS Security Agent, AWS DevOps Agent, plus Amazon Connect with 8 new agentic capabilities and Nova Sonic integration.

✳️ Live Experience

(Caption: A few nights before the Dr. Swami Keynote, on the second floor of The Venetian in Las Vegas, we luckily caught Dr. Swami in the wild. We introduced ourselves as AWS Heroes from Taiwan and got an enthusiastic handshake from this super approachable leader — absolutely thrilled.)

View this post on Instagram

Read More

(Illustration: AWS re:Invent 2025 Keynote with CEO Matt Garman. Image source: AWS.)

AWS CEO Matt Garman opened the 14th annual re:Invent to over 60,000 attendees and nearly 2 million online viewers, framing AWS as a $132 billion business accelerating at 20% year over year. Under the theme “Freedom to Invent,” Garman walked through five interconnected areas: AI infrastructure (Trainium3 GA, Trainium4 preview, P6e-GB300, AI Factories), models and intelligence (Nova 2 family with Lite/Pro/Sonic/Omni, Nova Forge open training models), agents at scale (AgentCore Policy and Evaluations), developer transformation (AWS Transform Custom, Kiro IDE, three frontier agents), and a rapid-fire round of 25 core service launches across compute, Lambda, storage, EMR, security, and databases.

✳️ tl;dr

One theme “Freedom to Invent” runs throughout, with five sections:

• AI Infrastructure: AWS leads GPU reliability, launches P6e-GB300 with Nvidia, announces Trainium3 GA (first 3nm AI chip in AWS Cloud), previews Trainium4 (6x FP4 compute), and introduces AWS AI Factories for dedicated on-premises AI infrastructure.
• Models & Intelligence: Amazon Bedrock surpasses 100K customers with 50+ trillion-token accounts (by comparison, when I compiled my 2025 Year in Review, feeding an entire year of Heptabase journal notes only came to 200k tokens); new open-weights models from Mistral, Google, MiniMax, Nvidia; Nova 2 family (Lite, Pro, Sonic, Omni) delivers frontier-level intelligence at optimized cost; Nova Forge introduces open training models for custom Novellas.
• Agents at Scale: Amazon Bedrock AgentCore adds Policy (Cedar-based deterministic controls) and Evaluations (continuous quality inspection with 13 pre-built evaluators); guest speakers from Sony, Adobe, and Writer showcase enterprise adoption.
• Developer Transformation: AWS Transform Custom supports any code modernization; Kiro becomes Amazon’s official AI IDE; three frontier agents launched: Kiro autonomous agent, AWS Security Agent, AWS DevOps Agent.
• 25 Core Service Launches: Rapid-fire announcements including new EC2 instance families (X, C8a, C8ine, M8azn, Mac), Lambda durable functions, S3 50TB max object size, S3 Vectors GA, EMR serverless storage, GuardDuty for ECS, Security Hub GA, RDS 256TB capacity, and database savings plans.

✳️ Live Experience

(Caption: On the day of the CEO Keynote, I woke up early in Las Vegas, lined up to get in, and for the first time challenged myself to take notes live on site using only a reMarkable Paper Pro for handwritten notes — quickly jotting down keywords, rapidly categorizing and capturing the structure, while simultaneously mapping out how I could help brief teammates if they were pulled into other tasks on short notice. Grateful for my experience leading camps as a kid and the brain-intensive training from a disciplined morning routine.)

View this post on Instagram

(Caption: After the CEO Keynote ended, we had only an hour and a half to regroup, move locations, grab food, and brief each other before heading straight into the re:Invent Podcast recording studio. Every year we challenge ourselves with different tasks that push the bar even higher.)

View this post on Instagram

Read More

(Caption: Ricoh GR IV, the latest chapter of the street photography god machine.)

Oh oh oh! The “Ginkgo Viewing Set” that I’ve been waiting for two and a half months has finally leisurely left the warehouse! These two and a half months were too long, so long that I felt I needed Ginkgo to supplement my brain power.

• Ricoh released GR IV on 2025-08-20,
• I rushed to B&H to pre-order one (and related accessories) as soon as possible on 2025-08-21,
• Waiting and hoping, waiting and waiting, I received two Backorder Status updates in between.
• In mid-October, I couldn’t help but ask customer service if there was an estimated delivery date, because I really wanted to take the GR IV on my upcoming business trips and street sweeping.
• Luckily, I received the Order Shipped notification on 2025-10-31,
• But encountering Double 11 and African Swine Fever, the entire customs was jammed,
• It wasn’t until 2025-11-09 that I received the import duty and tax notification. Fortunately, I received the box immediately after payment, and I can happily unbox it.

My last Ricoh was the R3 from 2005 (it even lacked a G!), and in a blink of an eye, it’s been twenty years. This time, I even bought the camera from B&H. Is it true love or destiny? Am I really destined with these two letters? Although the Ginkgo camera (plus batteries, flash, and other accessories) was stuck in customs for quite a few days, and for the first time I didn’t receive any EZ WAY notification (maybe because I manually filled out the paper appointment letter first?), I finally got it before Thanksgiving, giving me time to learn how to set it up (play with it?!) before my business trip.

Today, GR IV has finally arrived. Used some other bricks to exchange for these bricks in front of me. Tomorrow I’ll tell the kids this is also a type of Minecraft?

Okay, okay, no more chatting, let’s hurry up and unbox this year’s “Unbearable Lightness of Being Focused”. (At least compared to last year’s “Unbearable Lightness of Being Focused”, it is really light!)

Read More

(Illustration: Screenshot of AWS Health Dashboard at 2025-10-20 12:51 PDT. Image source: Ernest.)

✳️ tl;dr

• The following content is from an official AWS report ¹, segmented and highlighted by AWS Community Hero Ernest ² from the perspective of a developer and technical manager, aiming to stay close to the facts and conduct reasoning and extended learning based on these facts.
• Through studying this report, we hope that both parties (AWS and us as AWS customers) can accumulate experience and continue to improve together, whether in the cloud or on-premises.
• Unless otherwise specified, all times below are in Pacific Daylight Time (PDT) from AWS Seattle headquarters on the West Coast.
• This note will begin with a knowledge graph, followed by a breakdown of the original official report content, divided into four sections: Amazon DynamoDB, Amazon EC2, Network Load Balancer (NLB), Other AWS Services
• If you have the budget to adjust your architecture for cross-region high availability but don’t have enough time for major architectural changes, it is recommended to take a look at AWS services with “global” in their name. For example, “Amazon DynamoDB Global Tables” from the same DynamoDB family was almost unaffected during this incident.

• We wanted to provide you with some additional information about the service disruption that occurred
  • in the N. Virginia (us-east-1) Region ³
  • on October 19 and 20, 2025.
  • While the event started at 11:48 PM PDT on October 19 (Taipei Timezone UTC+8, 2025-10-20 14:48)
  • and ended at 2:20 PM PDT on October 20 (Taipei Timezone UTC+8, 2025-10-21 05:20),
  • there were three distinct periods of impact to customer applications.
    • First, between 11:48 PM on October 19 and 2:40 AM on October 20, Amazon DynamoDB experienced increased API error rates in the N. Virginia (us-east-1) Region.
    • Second, between 5:30 AM and 2:09 PM on October 20, Network Load Balancer (NLB) experienced increased connection errors for some load balancers in the N. Virginia (us-east-1) Region.
      • This was caused by health check failures in the NLB fleet, which resulted in increased connection errors on some NLBs.
    • Third, between 2:25 AM and 10:36 AM on October 20, new EC2 instance launches failed and, while instance launches began to succeed from 10:37 AM, some newly launched instances experienced connectivity issues which were resolved by 1:50 PM.

Read More

(Hand-drawn by Ernest Chiang. You might also be interested in his Ernest PKM workflow.)

1️⃣ Introduction: Making Zettelkasten Learnable and Replicable

In 2017, German scholar Sönke Ahrens published a book that transformed the knowledge management field: How to Take Smart Notes. ¹

This book accomplished something important: systematizing, proceduralizing, and making actionable Niklas Luhmann's Zettelkasten method.

Before Ahrens, Zettelkasten was more like a “legend”—we knew Luhmann wrote 70 books using this method, but weren’t quite clear how ordinary people could replicate this system. Luhmann’s own 1981 paper “Communicating with Slip Boxes” was more philosophical reflection than operational manual. (I personally prefer reflection, but many friends have been asking about methods, so I compiled this note.)

“Writing is not what happens after thinking.
Writing is the medium of thinking."

— Sönke Ahrens, How to Take Smart Notes (2017)

Read More

(Hand-drawn by Ernest Chiang. You might also be interested in his Ernest PKM workflow.)

1️⃣ Introduction: A Sociologist and His Thinking Machine

Niklas Luhmann (1927–1998) was a German sociologist renowned for his systems theory. During his academic career, he achieved astonishing productivity: 70 books and over 400 scholarly articles.

But even more remarkable, he attributed all of this to what seemed like a simple tool: the Zettelkasten (slip box).

• This was not an ordinary note-taking system.
  • Luhmann began building this system in the 1950s, eventually accumulating over 90,000 index cards.
• He called this system his communication partner
  • An external brain capable of dialoguing with him, facilitating thinking, and even “surprising” himself. ¹
  • Doesn’t this sound like a manual, century-old version of an AI Agent or knowledge assistant?!

In his famous 1981 paper “Communicating with Slip Boxes” (Kommunikation mit Zettelkästen), Luhmann described in detail how this system worked. But interestingly, much of the modern understanding of the Zettelkasten method is actually mixed with interpretations and adaptations by later scholars.

This note explores the original method that Luhmann himself actually used.

“Without writing, one cannot think;
at least not in a sophisticated, connectable manner.”

— Niklas Luhmann, Kommunikation mit Zettelkästen (1981)¹

Read More

(Illustration: KKR HQ locates at 30 hudson yards in New York. Image source: Photo by Illya Goloborodko.)

✳️ tl;dr

• News tracking 2024-10-30 ¹ → 2025-07-30 ²
• KKR is no stranger to the tech industry, having acquired an 80% stake in Philips’ semiconductor division in 2006, which was later renamed NXP Semiconductors. This was a client Ernest once served, leaving a particularly deep impression.

• 【2024-10-30】KKR and Energy Capital Partners announced a $50 billion strategic partnership focused on accelerating data center, power generation, and transmission infrastructure development to support global AI and cloud computing expansion ¹
• The partnership combines over 8GW of existing data center pipeline and 100GW of operating and development-ready power generation capacity, with KKR owning over 100 data center facilities worldwide ¹
• ECP invests in clean energy asset base, owning and operating over 83GW of power generation capacity in the U.S. market, spanning five asset classes including power generation, renewables, and storage ¹
• The partnership aims to collaborate with utilities, power producers, and data center developers to rapidly and responsibly develop large data center campuses for hyperscalers ¹
• Context: BlackRock launched a $30 billion AI infrastructure fund in the same month, backed by Microsoft and Nvidia, demonstrating capital’s rush into AI infrastructure ¹

• 【2025-07-30】First investment lands: 190MW hyperscale data center campus in Bosque County, Texas, marking 9 months from strategic announcement to first project announcement ²
• Innovative co-location model: Data center adjacent to Calpine’s Thad Hill Energy Center natural gas power plant, representing the first such dedicated power agreement with a hyperscaler ²
• Constructed through a joint venture between CyrusOne and ECP, expected to be operational in Q4 2026, with total investment approaching $4 billion, initial IT capacity of 144MW, spanning over 700,000 square feet ²
• Behind-the-Meter model: Calpine provides 190MW of dedicated power, which can be redirected to support system reliability and local demand during ERCOT grid emergencies ²

• AI-driven power demand: Goldman Sachs predicts global data center power demand will grow 165% by 2030 compared to 2023, with AI data center power density surging from traditional 5-10kW/rack to 50-200kW/rack ³⁴
• Grid bottleneck severity: New data center grid connection delays have reached 5 years, while Behind-the-Meter natural gas generation can be deployed within 18-24 months, becoming a pragmatic choice ⁵
• Texas regulatory environment: SB6 new legislation (signed June 2025) requires large loads over 75MW to bear grid costs, accept emergency curtailment, and install ERCOT-controlled “kill switches” ⁶⁷
• Private equity continues to double down: 2024 data center investment reached $108 billion (triple that of 2023), with KKR's 2021 acquisition of CyrusOne for $15 billion laying the foundation for this partnership ⁸⁹

Read More

(Illustration: Aligned Data Centers. Image source: Aligned Data Centers.)

✳️ tl;dr

• AIP, MGX, and GIP formed a consortium to acquire Aligned Data Centers for $40 billion ¹²
• This marks AIP’s (AI Infrastructure Partnership) first investment since its establishment

• AIP was founded by BlackRock, Microsoft, NVIDIA, and MGX in September 2024 ³
• AIP targets mobilizing $30 billion in equity capital, with potential to reach $100 billion including debt financing
• Aligned owns 50 campuses with over 5 GW of operational and planned capacity across key digital hubs in the US and Latin America

• The hyperscale data center market is projected to reach $167.3 billion in 2025
• Growing at a 23.58% CAGR to $602.4 billion by 2030 ⁴
• Global data center electricity consumption is expected to double from 415 TWh in 2024 to 945 TWh in 2030, accounting for 3% of global electricity usage ⁵
• Aligned holds over 50 patented cooling technologies, including air, liquid, and hybrid cooling systems designed specifically for high-density AI workloads ¹⁶

• NVIDIA’s latest GB200 chip requires power density up to 120 kilowatts per rack, making liquid cooling essential for racks above 20 kilowatts ⁷⁸
• Kuwait Investment Authority and Singapore’s Temasek serve as anchor investors in AIP, demonstrating sovereign wealth funds’ long-term commitment to AI infrastructure ²³
• Notably, Macquarie Asset Management first invested in Aligned in 2018, expanding it from 2 facilities to 50 campuses over 7 years, achieving an exit valuation of approximately $40 billion ⁹¹⁰

• Major tech companies are expected to invest $400 billion in AI infrastructure in 2025, with OpenAI’s Stargate initiative alone reaching $500 billion ²¹¹
• US data center power demand is projected to double from 35 GW in 2024 to 78 GW by 2035, with average hourly electricity consumption tripling ¹²

• For you as a manager: Macquarie entered investments (in multiple data center companies) in 2018 and exited 7 years later. As we approach 2026, have you activated your radar to identify targets for the next 3-5 years?

Read More

(Photo by Ray Hennessy on Unsplash)

✳️ tl;dr

• Meta completed nearly $300 billion in financing through SPV structure
• Building the Hyperion data center in Louisiana, setting a record for the largest private equity transaction in history ¹
• Hyperion data center covers 4 million square feet,
• When fully operational, it will consume 5 gigawatts of electricity, equivalent to the power consumption of 4 million American households

• Meta retains only 20% equity yet maintains full operational control,
• This “control without consolidation” accounting technique keeps $270 billion in debt off the balance sheet ²
• Equity accounts for only 8.5% of total financing ($2.5 billion/$29.5 billion)
• Insurance companies invest heavily in such projects through private credit,
• But face asset-liability mismatch risks and may be forced to liquidate investments during economic downturns ³
• Historical lesson: In the 1990s, telecom companies laid 80 million miles of fiber optic cables,
• Four years after the bubble burst, 85%-95% remained unused, earning the nickname “dark fiber” ⁴
• Meta, Amazon, Google, Microsoft committed to a record $320 billion in capital expenditure this year, mostly for AI infrastructure, yet Meta’s 10-K admits: “there can be no assurance that the usage of AI will enhance our products or services” ⁵

• Power infrastructure becomes one of the bottlenecks
• Grid Strategies estimates data centers will need an additional 60 gigawatts of electricity by 2030, equivalent to Italy’s national peak demand ⁶
• Cooling technology is also crucial: from air cooling to direct liquid cooling to immersion cooling, affecting long-term operating costs ⁷

• Morgan Stanley serves as the exclusive underwriter, while also providing financing advisory for multiple similar projects ¹
• Bonds are issued in 144A format private placement, with a spread of 225 basis points above Treasury bonds ¹

• What can technology managers do? (1) Quantify the actual ROI timeline for AI investments (2) Assess power supply chain risks (3) Maintain appropriate financial leverage ratios

^{8 9}

Read More