*The content is solely for presenting diverse market perspectives and research viewpoints and does not imply endorsement of the views or conclusions in the article by this official account.

"We are aggressively expanding AWS."

At the Morgan Stanley TMT Conference, NVIDIA CEO Jensen Huang's seemingly offhand remark may be one of the most significant signals in the cloud computing industry in recent months.

The surface meaning of this statement is straightforward: as the core supplier of global AI computing power, NVIDIA observed that OpenAI is significantly increasing its resource investment in Amazon Web Services (AWS). However, the deeper implications extend far beyond this. It reveals an ongoing structural shift: AI computing power demand is moving from 'single-cloud platforms' to 'multi-cloud platforms.'

Over the past two years, the market has widely believed that Microsoft Azure has secured the cloud computing dividends of the AI era. However, if this trend continues, the growth structure of the global cloud computing market may undergo a profoundly underestimated transformation. The biggest potential beneficiary could very well be Amazon Web Services (AWS), once considered a laggard in the AI wave.

From 'Azure Dominance' to 'Multi-Cloud Computing Power': The First Crack in AI Infrastructure

Over the past two years, the narrative around AI infrastructure has been nearly monopolized by an 'Iron Triangle': OpenAI + Azure + NVIDIA.

After Microsoft invested over $13 billion in OpenAI, the two companies formed a deeply exclusive partnership. OpenAI's training and inference computing power almost entirely ran on Microsoft Azure's dedicated supercomputers. This deep integration made Azure the biggest beneficiary of AI cloud computing dividends over the past two years. Between 2024 and 2025, Microsoft disclosed multiple times in its earnings reports that AI services contributed over 7 percentage points of additional growth to Azure, prompting capital markets to grant Microsoft a significant valuation premium.

However, this structure has a natural physical and commercial bottleneck: a single cloud platform can hardly sustain the exponential growth in computing power demand throughout the AI era.

As model scales continue to expand, training costs and inference call volumes are growing non-linearly. Industry estimates suggest that the single-training cost of a GPT-4-level model has reached hundreds of millions of dollars, while inference request volumes for large models are expanding at a 10-fold annual rate. Under these circumstances, computing power supply quickly becomes a bottleneck. Whether it's power supply, data center space, or delivery cycles for high-end GPUs, a single cloud provider cannot infinitely scale up in a short period.

More importantly, for leading AI companies like OpenAI, placing their fate entirely in the hands of a single cloud provider means significant 'vendor lock-in' risk. If Azure experiences service disruptions or gains absolute dominance in pricing power, OpenAI's business continuity will be threatened.

This is why more and more AI companies are actively adopting a Multi-cloud Strategy. From Huang's statement, it appears that OpenAI has already begun expanding some of its computing power needs to AWS. This marks the first time the AI cloud market has seen a 'diversion' at the infrastructure level. This is not just a tactical adjustment by OpenAI but a collective correction by the entire industry against 'single-cloud dependency.' When AI giants start seeking second or even third computing power sources, AWS, with the most mature infrastructure and the largest capacity reserves, naturally becomes the top choice.

AWS's AI Flywheel: OpenAI, Anthropic, and the Inference Economy

If Azure won the first phase of the AI era through exclusivity in 'training computing power,' AWS may be betting on the 'Inference Economy' in the second phase.

Two key variables are reshaping AWS's growth logic.

The first variable is OpenAI's computing power expansion. As model scales continue to grow and application scenarios expand from consumer-facing chatbots to enterprise B-end applications, AI companies often adopt cross-cloud deployments. On the one hand, this is to secure more GPU resources to meet peak demand; on the other hand, it is to reduce platform dependency risks and maintain bargaining power. For AWS, this means a new possibility: some of OpenAI's GPU clusters are now entering AWS. Even if it's just handling inference traffic, it represents a significant revenue increment (incremental revenue) for AWS.

The other, more critical variable is Anthropic.

Unlike the complex relationship between OpenAI and Microsoft, Anthropic has established a deep and clear strategic partnership with Amazon from the outset. Since 2023, Amazon has invested over $8 billion in Anthropic and plans to invest further. The infrastructure for Claude series models runs extensively on AWS and is made available to enterprise customers through the Amazon Bedrock platform.

The underlying business logic is clear: AI training is a one-time big deal, but AI inference is the long-term cash flow.

Every model call means more than just GPU computing. It involves API calls, database read/write operations, object storage, network bandwidth transmission, and security services. In other words, the more inference occurs, the more stable the cloud revenue becomes, and the richer the variety of cloud services involved.

If Claude, OpenAI, and other models running on AWS continue to grow, AWS will not only sell GPU computing power but also offer a full suite of cloud infrastructure services. This 'all-in-one' revenue structure has higher stickiness and profit margins than simply renting out computing power. This is why many investors are beginning to re-examine a question: AWS's growth cycle may not be over yet. In the training phase, the market focuses on who has the most H100s; but in the inference phase, the market focuses on whose ecosystem can best support large-scale, low-latency, high-concurrency application deployments. In this regard, AWS's decade-long enterprise service capabilities form its new moat.

Agentic AI: The 'Next Phase' That Truly Amplifies Cloud Computing Demand

Even greater changes may come from Agentic AI.

Over the past two years, most AI products have remained in the 'question-answering tool' stage. Users input a question, and the model returns a text response. This model's cloud resource consumption is relatively singular, primarily focused on GPU inference.

However, with the rise of AI Agents, models are shifting from 'answering questions' to 'executing tasks.' A typical AI Agent task may involve: multi-round model inference, autonomous planning, database access, third-party API calls, backend service execution, and long-term memory storage.

In other words, AI no longer just consumes GPUs. It simultaneously drives CPU, storage, database, network, and the entire cloud service ecosystem.

For example, an AI Agent for automated financial analysis needs to read reports stored in S3, query historical data from a database, use a model for analysis, and finally write the results to DynamoDB and send notifications via SNS. In this process, GPU computing may only account for a portion of the cost, with significant cloud resource consumption occurring in data flow and logical execution.

What does this mean for cloud providers? It means AI is not just a computing power business but a platform business.

Over the past few years, the market has been discussing whether AI will compress cloud providers' profits because GPUs are expensive, and some inference may occur at the edge. However, if Agentic AI becomes mainstream, the situation may be the opposite. Because each AI Agent is essentially a software system running 24/7, it continuously consumes cloud resources and generates cloud revenue.

This is precisely AWS's strongest area. AWS offers the richest variety of cloud services (over 200), ranging from serverless computing to managed databases, from message queues to identity authentication. As AI evolves from a 'dialog box' into an 'operating system,' the value of cloud platforms like AWS that provide comprehensive infrastructure will be re-amplified. The more complex Agentic AI becomes, the deeper its reliance on cloud-native services, and the wider AWS's moat grows.

Conclusion: From Computing Power Arms Race to Ecosystem Marathon

Over the past two years, the narrative around AI infrastructure has almost entirely revolved around one question: Who can secure the most GPUs?

But Huang's remark hints at a new phase. AI computing power is no longer just a GPU competition; it's becoming a cloud platform competition.

As AI moves from training to inference and from tools to Agents, the cloud computing market's true growth cycle may just be beginning. This is no longer a simple hardware arms race but a comprehensive competition involving stability, ecosystem, cost control, and multi-cloud architecture capabilities.

For investors, a shift in perspective is crucial. We should no longer focus solely on NVIDIA's graphics card shipments but on how cloud providers can convert AI traffic into lasting platform revenue. In this new race, AWS is likely to return to the center stage, leveraging its deep infrastructure accumulation and the dividends of its multi-cloud strategy. The AI wave hasn't receded; it's simply adopted a more pragmatic and sustainable approach to drive the next round of cloud computing growth.