AI云端狂想曲：亚马逊云科技的算力突围、Agent重构与卓越运营

On the third anniversary of OpenAI's ChatGPT release, the AI ​​war has entered a new round of fierce competition.

Google's major release of Gemini 3 in late November, along with news of large orders for TPU cluster training of Gemini and Meta , triggered a "Code Red" alert for OpenAI and sent chills down the spines of GPU giants like Nvidia.

As model giants begin a new round of competition, AI cloud providers are facing both intense speed races and pressure to innovate in the rapidly expanding computing power market . I was curious about how

. I was curious about how Amazon Web Services (AWS), the world's largest cloud provider, would strategize its next move in this cloud war.

So, we flew to Las Vegas again.

At this year's conference, we discovered that this cloud giant had practically turned Las Vegas into an AI utopia.

Even the trash cans were AI-powered!

Wow, amazing! (The recognition was) incredibly accurate.

And at this year's re:Invent, we found that not only were we doing something big on-site, but AWS was also doing something big. With the

world's largest and most widely deployed AI cloud infrastructure, we were once again in a state of...

well, let's just say we were impressed. The freedom to transform concepts into tangible impact during the Renaissance era is happening at an unprecedented pace.

What we're showcasing today proves that "magic" isn't a myth; they're changing the game.

They're selling shovels and sacks during a gold rush — there really is gold out there . Amazon Web Services' annual re:Invent conference is undoubtedly

. Amazon Web Services' annual re:Invent conference is undoubtedly a bellwether for the AI ​​industry.

Over 60,000 people traveled from around the world to attend, and millions joined online.

While many of the showcased technologies and case studies are B2B enterprise-level innovations , this demonstrates how companies are "voting with their feet" and how effective AI truly is.

They've built such powerful computing power and infrastructure that AI can be directly utilized . Having these readily available,

. Having these readily available, Lego-like components to build the solutions you need is fantastic.

I plan to expand further.

We're developing these capabilities for AI, and in this video, we'll take you to Las Vegas.

We'll explore Amazon Web Services ' cloud-based symphony from various angles, including computing power models , agents, and enterprise applications.

We'll discuss its vision for future AI development trends , how to fight this cloud-based battle , and we'll also take you through this year's exhibition area to experience firsthand how companies across various industries are using AI to make products more creative , management more efficient , interactions smoother , and truly help businesses achieve the "last mile" of AI implementation . By the way, I had a blast at the event, so

. By the way, I had a blast at the event, so be sure to watch until the end!

After several days of intensive announcements, I feel like Amazon Web Services released hundreds of updates.

This video certainly can't cover everything , but I think there are four key points that have been very significant. The updates

are the pioneering "open training model" platform and Agent for computing power models , and the signal is very clear : to accelerate the investment in AI , and helping enterprises "achieve practical implementation" is the core task.

First, let's look at the upgrades in computing power support from Amazon Web Services (AWS), which the market is very concerned about . The conclusion is that by the end of 2025,

. The conclusion is that by the end of 2025, this cloud giant's strategy is not only to continue to deeply integrate with NVIDIA GPUs for better adaptation, but also to significantly improve the performance of its self-developed chips.

There are two key points here.

The first key point is that, as in previous years, AWS emphasizes its cooperation with NVIDIA and aims to be the "best habitat" for GPUs.

Simply put, AWS does not just deploy NVIDIA's most advanced GPUs , but will also perform deep optimization together.

Such optimization can enable GPU computing power to achieve extremely high reliability and availability.

For example, the CEO Matt Garman introduced the P6e-GB300 instance , a supernode server built on the NVIDIA GB300-NVL72.

Amazon Web Services (AWS) aims to become a superior choice for running NVIDIA graphics cards.

While this sounds easy , achieving it requires deep optimization at every level of hardware and software – something only AWS can do.

In fact, there are no shortcuts.

In early November 2025, AWS and OpenAI announced a multi-year strategic partnership worth $38 billion.

OpenAI uses Amazon EC2 UltraServers computing clusters powered by NVIDIA GPUs, enabling extremely low-latency communication between interconnected systems , allowing OpenAI to efficiently run AI tasks.

However, what has drawn more market attention is the second key point : AWS's self-developed chips , similar to Google's TPUs.

These ASIC-specific chips pose a threat to NVIDIA's current monopoly.

But for cloud providers like AWS, this is a battle for computing power.

CEO Matt Garman revealed in his keynote that Amazon provides various model API calls.

The vast majority of users on the Bedrock platform use Amazon Web Services' self-developed Trainium chip for inference, including Anthropic's Claude model inference, which also uses Trainium chips.

At this year's conference, the Trainium chip series received an upgrade : the third-generation AI accelerator, Trainium 3, built on a 3nm process.

Its energy efficiency is 40% higher than the previous generation, and its computing power is doubled.

Trainium 3 brings high cost-effectiveness to the industry for large-scale AI training and inference. While

Trainium 2 has already achieved remarkable results in the past year , the Trainium 3 system has made a huge leap forward, with its computing power increased by 4.4 times, memory bandwidth increased by 3.9 times , and, most importantly, the AI ​​processing capacity per megawatt of power... The token count increased fivefold.

There was also a special surprise: I brought an UltraServers rack to the stage today , and Matt even gave a sneak peek at the Trainium 4: FP4 performance improved sixfold, memory bandwidth increased fourfold, and HBM capacity increased twofold. This

performance enhancement, coupled with reduced energy consumption, further demonstrates Amazon Web Services ' determination and commitment to developing its own internal ASIC chips.

This signal is quite important because Google's ASIC accelerator chip, TPU , received a lot of discussion recently . Amazon Web Services' actions

. Amazon Web Services' actions show that tech giants are all increasing their investment in ASIC chips.

The ASIC chip market may see even more positive development in the future. For those unfamiliar with ASIC chips, let me briefly explain : ASIC chips are integrated circuits tailored for specific functions or application scenarios . Compared to NVIDIA's general-purpose GPUs,

. Compared to NVIDIA's general-purpose GPUs, it achieves higher efficiency , lower power consumption, and better performance for specific tasks.

Simply put, NVIDIA's GPU chips are highly versatile and can handle almost any task , while ASIC chips require a certain adaptation time —currently several months.

However, once adapted, they can achieve more efficient training and inference.

There are many ways to optimize performance , and Amazon Trainium puts the tools in the hands of developers to help them achieve superior model and application performance.

Amazon Trainium also significantly reduces the operating costs of workloads . It must be said that

. It must be said that the Trainium series has developed very rapidly in the last two years , a testament to the reaping of the rewards of the previous $8 billion investment in Anthropic. At re:Invent in 2024, Amazon Web Services and Anthropic announced Project Rainier, involving 30 data centers and costing $1.1 billion. The gigawatt (GW) power grid utilizes 500,000 Trainium 2 chips

. This large-scale cluster adaptation and collaboration

. This large-scale cluster adaptation and collaboration has enabled Trainium to be adapted and optimized not only for inference but also for training large models, providing high-performance and low-cost computing power services to users like Anthropic's Claude model . It seems the $8 billion investment was worthwhile.

. It seems the $8 billion investment was worthwhile.

However, the investors we interviewed believe that ASICs and GPUs are not actually opposed, but rather complementary at this stage.

If you look at these cloud computing vendors who are aggressively developing their own chips and investing in ASICs, they are competing.

Ultimately, who hosts the best Nvidia GPUs? Therefore, it's difficult to

Nvidia GPUs? Therefore, it's difficult to frame it as an ASIC versus GPU scenario.

There are many workloads on GPUs , and we offer you the best GPUs, the best support, the most reliable infrastructure, and the best security.

Data ... Whether it's storing data or related supporting components , if you want certain workloads, running them on our ASICs is very effective cost-effective and we welcome you to use our ASICs.

I think our ASICs are more like a dynamic situation.

So, to some extent, each company's ASICs serve as a supplement to Nvidia, making them more effective in certain workloads and helping them better manage their margins and cost structures.

On another level, I think their self-developed ASICs can be customized as a differentiating factor to compete with other cloud providers, whether it's competing with the ones mentioned earlier or with Neocloud CoreWeave Nebius etc. Besides their self-developed chips, another highlight this year is Amazon Web Services' self-developed model combination.

The Nova model update has greatly expanded and upgraded Amazon Web Services' model ecosystem.

Amazon Web Services' second rhapsody is its model layer update , and this year the key word for the model layer is expansion.

Today, there are tens of thousands of customers using Amazon Nova , including marketing giants like Dentsu, Infosys, and Blue Origin. Tech leaders like Robinhood and innovative startups like NinjaTech AI are all involved.

Today, Amazon Nova takes it to the next level as we officially release the new generation, Amazon Nova 2.

We know that Amazon Web Services' model family is called Nova, with four updated models: Nova 2 Lite, Nova 2 Pro , Nova 2 Sonic, and Nova 2 Omni.

All have seen significant upgrades in the past year.

Interestingly, Amazon Web Services' self-developed model strategy doesn't compete on performance with the top-tier models , but rather focuses on "cost-effectiveness" while maintaining parity with leading models.

Leveraging the support of the entire cloud ecosystem, it provides customers with more choices.

Understanding this logic helps us grasp the underlying logic behind Amazon Web Services' model releases this year.

Let's look at the upgrades to these four Nova models. Nova 2 Lite is a fast and economical inference model designed for everyday workloads.

It can process text, images, and video input and generate text output. In other words, Nova 2 Lite's capabilities are well-suited for tasks such as processing various documents, extracting key information from videos, generating code, providing accurate fact-based answers , and automating multi-step agent workflows.

Nova 2 Pro is Amazon Web Services' most intelligent inference model, capable of processing text, image, video, and voice input and generating text output.

It's suitable for highly complex tasks requiring high accuracy, such as agentic coding , long-term planning, and complex problem solving.

This model can also be used as a "teacher model," transferring its capabilities to a smaller, more efficient "student model" through knowledge distillation for specific vertical domains and application scenarios . In other words,

. In other words, Nova 2 Pro is suitable for multi-document analysis, video inference, complex instruction execution , solving high-order mathematical problems , and performing agent and software engineering tasks.

The third model release is Nova 2 Sonic , Amazon Web Services' end-to-end speech model, which deeply integrates text and speech understanding and generation, achieving a real-time, human-like conversational AI experience.

This model supports more languages ​​and expressive timbres, boasts higher recognition accuracy , and provides a context window of up to 1 million tokens, supporting long-term interaction and rapid switching between voice and text.

The model can process tasks asynchronously; while users continue natural conversations or even change topics, the system can still perform operations in the background, such as booking tickets.

Furthermore, Nova 2 Sonic can also integrate with Amazon Web Services. Connect

integrates third-party voice service providers and conversational AI frameworks to support interaction in customer service AI assistants and interactive voice experiences.

The fourth released model is called Nova 2 Omni , a unified multimodal inference and generation model that can handle text, images , video, and voice input , and can simultaneously generate text and images.

Nova 2 Omni can process text of up to 750,000 words.

Hours of audio and video, along with hundreds of pages of documents, can be analyzed simultaneously, including complete product catalogs, user reviews, brand guidelines, and video material libraries.

This reduces the cost and complexity of connecting multiple professional models . Such a "full-modal" model

. Such a "full-modal" model is extremely useful for enterprises. As

a result, we see many companies, including Cisco and Siemens, using the Nova 2 model to build a variety of innovative applications, from agent threat detection to video understanding and voice AI assistants.

Another highlight of this re:Invent conference was the launch of four external models on Amazon Bedrock.

Two of these are the large models MiniMax and Kimi from China, and the other two are Gemma from Google and Nemotron from NVIDIA.

This makes Amazon Bedrock's rich model ecosystem very friendly to developers and enterprises.

In addition to the models themselves , there are two other highlights : the launch of Nova Forge , the first innovative service for building cutting-edge AI models , and Nova Act Nova, a highly reliable AI agent service for UI workflows. The initial motivation behind Nova Forge was that when embracing AI, it's crucial for companies to integrate their private domain and proprietary knowledge.

However, in the past two years, companies have often faced three unsatisfactory choices : First, limited fine-tuning of closed-source models only allows for a superficial infusion of company expertise.

Second, continuous training of open-source weighted models without original training data can lead to "degradation" in basic capabilities such as instruction following.

Third, building models from scratch involves huge costs and time.

What companies truly need is a solution that can both acquire the capabilities of cutting-edge models and deeply integrate their own expertise. Nova Forge is the solution found by Amazon Web Services.

Nova Forge is a so-called "open training model" concept.

By combining a company's proprietary data with Nova's model capabilities, it helps companies create custom Nova optimized variants , also known as "custom models" or Novellas.

Currently, many companies and organizations, including travel websites Booking.com, Reddit, and Sony, have begun using Nova Forge to build custom models that better suit their needs.

Another significant service is Nova Act , which is used to build and deploy highly reliable AI agents in browsers to automate various operations.

This service is powered by a customized version of Nova. 2. The Lite model provides computing power support and is a fast track for building and managing large-scale browser automation agent clusters.

Amazon Web Services' Nova Act achieved 90% execution reliability in early customer workflows.

Some customer case studies include startup Sola Systems , which integrated Nova Act into its platform, automating hundreds of thousands of workflow tasks for customers each month , covering tasks such as payment reconciliation, freight coordination, and medical record updates.

1. Password uses Nova Act to help users access login information with less manual operation ; a simple prompt can automatically complete login steps on hundreds of different websites.

Hertz used Nova Act to automate end-to-end testing on a car rental platform , increasing software delivery speed by 5 times.

This platform processes millions of dollars in bookings daily; testing processes that previously took weeks can now be completed in hours. Amazon Leo

eliminated quality testing bottlenecks with Nova Act before the launch of its satellite internet service.

Test scenarios were written in natural language and automatically executed and adapted across thousands of web and mobile test cases, significantly reducing engineer time investment . The release of the Act

. The release of the Act also demonstrates Amazon Web Services' emphasis on the Agent field and the ecosystem support built around Agents is a key focus of this year's re:Invent.

I firmly believe that AI Agents will become one of the most transformative breakthroughs of this era , and Amazon Web Services is the best platform for building and running these Agents.

Let's take a look at how Amazon Web Services aims to reshape the collaborative boundaries of software development teams around Agents . Agents are expected to explode in 2025,

. Agents are expected to explode in 2025, especially as enterprises rush to deploy them.

However, the challenge we see lies in integrating Agents into internal production and management, requiring rapid, large-scale deployment, remembering past interactions and learning , identifying and accessing all Agent tools and controls, mastering Agent tools used to execute complex workflows, and finally observing and debugging problems. This process is quite complex. So, how can Amazon Web Services help customers build

and deploy secure, production-grade Agents at scale?

Amazon Web Services' solution is called Amazon Bedrock AgentCore. This is an Agent platform designed specifically for securely building and deploying Agents at scale . It includes a series of key components

. It includes a series of key components providing the complete set of services needed to run production-grade Agents at scale . In other words, customers can use Amazon Bedrock AgentCore to build Agents or combine it with any open-source Agent building framework . AgentCore

. AgentCore has three key new features.

First, Policy in AgentCore allows enterprises to create granular policies simply by describing rules in natural language.

Policies can be defined for agents, such as accessible tools and data, executable operations, and applicable conditions.

This essentially sets clear boundaries for agent operations , solving the most challenging permission issues.

This is considered a solution to the "last mile" of agent building within enterprises . Second, AgentCore Evaluation

. Second, AgentCore Evaluation monitors agent performance in real-world scenarios through built-in evaluation tools . Evaluation dimensions include accuracy and helpfulness.

. Evaluation dimensions include accuracy and helpfulness.

Custom evaluation tools can also be added to meet specific business needs.

Third, Episodic Functionality in AgentCore Memory automatically saves key events and states during interactions, helping agents learn from past experiences and improve decision-making.

Now that building agents is becoming increasingly convenient , a major question arises: how to make them more efficient?

Existing large models possess extensive intelligence; they can handle complex tool calls, multi-step inference, and unexpected scenarios.

However, their efficiency can still be improved.

This efficiency concerns not only cost but also latency— how fast can your agents respond?

Regarding scalability, its ability to handle peak traffic; regarding agility, its ability to iterate and optimize rapidly— these are all real challenges that must be faced when deploying AI on a large scale . We can see that

. We can see that the core goal of Amazon Web Services' (AWS) deployment of these functions is to accelerate the process of agents from idea to large-scale production . In addition, the three newly released agents—

. In addition, the three newly released agents— Kiro Autonomous Agent, which can be developed and programmed autonomously , like a skilled programmer completing development tasks independently under guidance— fundamentally change the way software is developed.

It can operate in parallel with the development process, maintain context understanding , and automate a series of tasks from feature delivery and defect classification to improving code coverage.

Amazon Security Agent acts as a virtual security engineer, serving as a security consultant in application design, code review, and penetration testing.

Amazon DevOps Agent acts as a virtual operations expert, assisting teams in solving and preventing operational failures and continuously improving system reliability and performance.

To summarize, over the several days of re:Invent, hundreds of updates were announced.

We simply cannot explain them all , but the most important one is what we just mentioned: building AI. The comprehensive upgrade of the four pillars of Agents

building AI. The comprehensive upgrade of the four pillars of Agents —AI chips, model ecosystem, pioneering open training model platform , and Agent development tools —and the accelerated investment in the entire ecosystem clearly convey a key message: although AI is changing the way applications are developed , the fundamental attributes of cloud computing —security, availability, elasticity, cost, and agility—

are more important than ever.

Google is clearly promoting its own Gemini, Microsoft, and, relatively speaking, OpenAI's models.

Amazon in my opinion, is the most balanced among these companies . Although it also has its own Nova series models

. Although it also has its own Nova series models , it doesn't feel like it 's making its own models a primary focus. Instead

, it provides its focus, I think, more on the Bedrock level.

For example, as a multi-model platform, it integrates other supporting elements, whether it's security or data... Storage, or

data... Storage, or

building a complete platform, feels somewhat like the e-commerce model , aggregating a large amount of supply and demand and acting as a platform. Having discussed these core updates, let's look at some easier-to-understand customer case studies and our on-site experiences to see how Amazon Web Services (AWS) customers are embracing AI strategies. While the re:Invent conference in Las Vegas features many very technical

aspects, I think one of the conference's strengths is its use of customer case studies to vividly and engagingly explain how cloud technologies empower products, making it easy for

the audience to understand.

This year, several keynote presentations particularly impressed me with collaborations.

First, Adobe, the parent company of Photoshop, is closely watched by designers, who are most directly impacted by generative AI .

Adobe CEO Shantanu Narayen stated that to accelerate its AI transformation, Adobe is leveraging AWS infrastructure to deploy AI in its core products, such as the Adobe One-Stop Platform. Firefly

includes Adobe's generative AI models , supporting text-to-image, text-to-video, and generative fill capabilities.

These models are trained using Amazon EC2 P5 and P6 compute instances powered by NVIDIA GPUs.

Adobe Firefly models have already generated over 29 billion creative assets, demonstrating the creative control AI gives creators.

Adobe has also integrated its AI assistant into Adobe Express , and its partnership with Amazon Web Services ensures these agents are not only highly efficient but also highly secure.

Another interesting case is media giant Condé Nast.

If you don't recognize the name, you'll surely know Vogue magazine.

Condé Nast is the parent company of Vogue , and this century-old publishing giant is partnering with Amazon to transform into a modern digital media company . Condé Nast's Chief Product Officer and Chief Technology Officer,

. Condé Nast's Chief Product Officer and Chief Technology Officer, Sanjay... On stage, Bhakta stated that

Sanjay... On stage, Bhakta stated that a key challenge in the transformation from traditional media was the highly fragmented nature of its media brands and assets across different countries and languages.

The company addressed this by building a unified content system and data lake warehouse based on Amazon Web Services (AWS) , enabling cross-regional content distribution through AI translation.

This resulted in digital revenue accounting for 70% of total revenue , with *The New Yorker* boasting over 1 million paid subscribers . Condé Nast also leverages AWS

. Condé Nast also leverages AWS to collaborate on large-scale live events, including the Met Gala —yes , the one where everyone wears dazzling gowns and is known as the "Fashion Oscars."

Through its partnership with AWS, Condé Nast utilizes real-time analysis of events and live streams to enhance its business.

The engagement of fashion events like the Gala is 54% higher than the Grammys, nearly 400% higher than the Golden Globes , and even 522% higher than the Super Bowl.

For media giants like Condé Nast, AI transformation is crucial.

As the CTO said in the final slide, print media defined us, digital media expanded us , and data will revolutionize us. We

also had two "sky-high and sea-low" case studies that impressed me deeply.

Let's start by talking about how AI can "go to the sky."

At this year's re:Invent conference, Bezos's commercial space company , Blue Origin, absolutely blew the show , treating the "away" venue like its own home ground, releasing a series of major announcements.

First, Blue Origin unveiled its first spacecraft capable of autonomous vertical launch and landing.

Meanwhile, its New... The Glenn orbital rocket successfully launched and landed, with the next goal being a return to the moon.

The company's collaboration with Amazon Web Services has integrated 2,700 agents into its business processes, resulting in over 3.5 million interactions for 70% of its employees in the past month.

Blue Origin also released a "lunar vacuum cleaner," an AI -powered device that converts lunar dust into energy.

Blue Origin has boldly utilized Amazon Web Services' agent functionality...

Introduced into aerospace system design, from refining requirements and material communication to higher-level system architecture and physical simulation, Blue Origin leverages its internal BlueGPT platform to utilize multiple agents to support R&D , increasing overall delivery speed by 75%.

With Agentic AI , we believe the future is here, just not yet accessible to everyone.

We haven't applied AI to all product development, but in the future, when millions live and work in space, we will be able to autonomously design entire rockets using AI.

I believe we can launch one hundred rockets by one person instead of one hundred people launching one.

Finally, AI will be used to "enter the sea."

In the final keynote speech, Amazon CTO Werner Vogels shared insights from the Ocean Cleanup Organization (Ocean Cleanup). Cleanup

leverages AI technology to optimize plastic detection models, predict waste movement trajectories, maximize cleanup efficiency , and ensure operations are carried out in the world's most critical areas.

There were numerous such customer case studies at the event.

Beyond the keynote speeches, we heard many inspiring stories about how B2B companies are using AI to improve efficiency , and at the exhibition area, we could personally experience the future integration of products and AI.

This year, my experience at the conference focused primarily on the sports event category , and Amazon Web Services dedicated a huge space specifically to this industry, showcasing everything from small games like dart throwing to visual recognition using edge cameras . From very small scenarios like

. From very small scenarios like real-time score calculation in the cloud to real-time billiards angle and power recognition to guide players , to real-time cloud golf coaching , and e-sports live streaming with real-time win rate analysis, the collaboration between F1 racing and Amazon Web Services has certainly broadened my knowledge considerably.

This year's Apple movie *F1: Racing* brought the behind-the-scenes mechanisms of this sport to the public's attention.

We know that each car is equipped with 300 sensors, generating over 1.1 million data points per second , making F1 a truly data-driven sport.

But what's impressive is that at such high speeds , in those few seconds before a car is about to enter the pit lane, the real-time data transmission and analysis of cloud services must be timely and efficient. This

and efficient. This accuracy allows teams to achieve efficient and precise responses and support in fierce competition.

This technology is truly amazing.

By the way, I also experienced changing tires in F1 firsthand.

My first attempt was rather disastrous , but I believe that with diligent practice and a time of 2 seconds , I could work part-time changing tires in F1!

Also, the collaboration between Amazon Web Services and the NBA is very interesting.

Previously, basketball game analysis was all handwritten , making data analysis extremely difficult.

Now, AI can help the NBA track and share statistical data to measure player performance that was previously impossible to quantify directly, including using Amazon Web Services' machine learning to track shot difficulty, defensive ratings, and gravity.

As a casual fan, this was the first time I 'd heard of this... The concept of "gravity"

of this... The concept of "gravity" is a tactical concept referring to a player's ability to attract the attention of the opposing defense.

When an offensive player poses a significant threat, he forces defenders to focus more on closing in on or helping to defend him, thus altering the entire defensive positioning and decision-making.

Before the advent of AI, data like "gravity" was difficult to record and analyze , and viewers were not particularly aware of it.

Now, with the support of AI, the analysis of gravity has become one of the most crucial data indicators in the NBA , changing the way viewers watch games by focusing solely on the "ball-handler." By

the way, let's test your knowledge: who is the most "gravitaminable" NBA star?

The answer is Stephen Curry. Curry, did you guess?

So, there's a saying in the NBA, "Curry has gravity."

Curry probably never imagined that someone would become his fan because of Amazon Web Services' re:Invent conference . Actually, after

. Actually, after spending so much time at re:Invent in Las Vegas , listening to so many presentations, and seeing so many announcements, my feeling is that I'm even more excited about the development of AI in 2026.

Along the way, my colleagues and I were talking about the past three years of AI development. The

first year was filled with excitement, everyone thought this model was amazing, that model was powerful.

Then the second year brought a huge sense of disappointment , realizing that this wasn't very effective and that had many bugs.

Now, in the third year, everyone is truly putting their full effort into thinking about how to "implement" AI across various ecosystems, directions, and technology stacks, innovating and developing products and services that solve problems. This is actually the third stage of technological explosion.

And at this year's re:Invent, I truly felt this third stage.

Like the words in the opening animation of the CEO's keynote speech at this year's conference : if the question isn't "Is this possible?"

but simply "Why not?"...

And Amazon's CTO Werner... Vogels

delivered a very moving speech.

This was his fourteenth and final keynote address at re:Invent . He told the developers in attendance that

. He told the developers in attendance that most of what we build will ultimately go unseen.

The only reason we can do it so well is our professional pride in operational excellence.

Amazon customers click a button and their packages arrive, but who thinks about the people behind the scenes and the supply chain ? Nobody sees all this work.

? Nobody sees all this work.

Customers will never tell you how well your data engineers have done; only you understand the effort involved.

I think it's important that everyone be proud of their work.

We consistently provide stable system services, clear and smooth deployments, and undetectable rollbacks.

Most of what we do goes unseen.

The only reason we can do all this well is our professional pride in operational excellence.

This is the quality of top builders ; they are meticulous in their work, even if nobody sees it.

Every time I come to re:Invent, I'm deeply moved.

Many of the people here are actually B2B. Enterprise server-side

engineers developers , and so-called "builders" do very basic, very infrastructure -level work.

Unlike 2C end users who can readily appreciate the fancy features , amazing results, and immediate feedback of the product model , these "unseen values" are the key to solving problems in the path of technological development.

Therefore, on the third anniversary of ChatGPT's launch, Amazon Web Services has spent the past three years pursuing its strengths with a very "long-term" approach —what Werner calls "operational excellence" —rather than striving for single-point dominance in any particular area.

Instead, it's about achieving ultimate optimization to create a full-stack ecosystem of computing power platform, model, data management, and agent tools, truly helping AI to be implemented.

This makes me even more excited for 2026.

Alright, that's all for this episode.

Your comments, shares, and likes are the best motivation for "Silicon Valley 101" to create in-depth technology and business content.

See you in the next video!