Cloud Platforms Regain AI 「Pricing Power」

NextFin News -- In the artificial intelligence (AI) era, who will be the first to cash in on the dividends? And who will ultimately seize the dominant position in the market?

Since ChatGPT set off this new wave of AI at the end of 2022, those have remained among the questions investors and the market care about most.

So far, hardware vendors that 「sell shovels,」 such as Nvidia, have gotten there first—becoming the foundation and core of 「AI economics,」 and building the business model with the largest scale of growth and the richest profits.

But higher up the AI stack—including the cloud platform layer, the model layer, and the application layer—the picture has long been far less clear.

Large models have consistently been at the center of both technical and market attention, and the application layer has been widely expected to be where 「AI lands」 and enters all kinds of real-life scenarios. Yet judging from some recent moves, cloud platforms—now clawing back their 「pricing power」—may be the side that can truly hold the initiative.

No More 「Model Porters」

In the past, the industry debated whether cloud platforms might be replaced by model vendors—or reduced to mere 「pipes.」

Under the 「replacement」 view, if large models themselves become capable of reaching both enterprise and consumer markets, the infrastructure value of cloud platforms would be diluted. Moreover, AI could also upend the cloud industry』s server-time billing model. If the industry shifts toward 「outcome-based」 pricing and charges by tokens, companies would no longer be content to pay for 「compute time,」 and would instead prefer to buy 「intelligence」 directly. This logic is similar to the narrative during this year』s 「SaaS stock crash,」 which claimed that the 「per-seat pricing model」 was coming to an end.

So-called 「pipelining」 in the telecom market mainly refers to operators spending vast sums to build networks, only to be reduced to mere 「pipes」 for users, while the fat profits are siphoned off by higher-layer applications. By analogy in the AI space, this becomes cloud platforms that have invested heavily, yet are very likely to be squeezed by upstream model vendors and application providers when it comes to profit sharing—ultimately relegated to low-margin, low-visibility compute 「porters.」

At the outset of this AI wave, large models did steal the spotlight, and cloud platforms without a flagship model partnership did face some skepticism in the market. But as the AI industry has evolved to this point, not only has no 「replacement」 occurred; cloud platforms have also had no intention of settling for being 「model porters.」 Instead, leveraging their own accumulated strengths and new operating models, they have quietly gained the upper hand in the profit split.

At present, cloud providers are actively expanding their model lineups and pushing MaaS platforms at full speed, integrating foundation models from different vendors into standardized APIs.

In late April this year, Microsoft Azure and OpenAI first ended their seven-year exclusive partnership agreement, allowing OpenAI to make its full product suite available to all cloud providers. Amazon AWS then announced that it would bring OpenAI』s latest models, the Codex coding assistant, and Managed Agents into its Bedrock platform. This was seen as a pivotal counterstrike by AWS against Azure AI Foundry and Google Vertex AI, and it also underscored just how fierce AI cloud competition has become—as well as the determination of the giants to pull more models into their own ecosystems.

What may be even more disruptive for model vendors, however, is the cloud platform 「model supermarket」 approach.

Multi-model subscription services in the 「model supermarket」 vein—such as AWS』s Amazon Bedrock Marketplace, Alibaba Cloud』s Bailian platform Token Plan, and ByteDance』s Volcano Engine Ark Coding Plan—allow users to pay a monthly fee and call multiple leading models at once, at a price not much different from subscription bundles for top-tier foundation models.

At its core, a 「model supermarket」 lowers the barrier to entry, but it also makes the underlying models appear highly homogeneous—and easily interchangeable—to developers. At this stage in AI』s development, large models have indeed run into issues such as increasingly convergent capabilities, a narrowing technology gap, and a long-standing difficulty in turning a profit. Once large models are 「downgraded」 from being product-facing offerings to becoming upstream technology supply, cloud platforms can more easily capture value by controlling the entry point—thereby reclaiming pricing power and taking the lead in the market.

On the other hand, cloud platforms can use this to strengthen their ecosystems: even if they sell APIs at low prices, they can still drive higher-value services such as cloud storage and data platforms through ecosystem bundling, ultimately maximizing their own returns.

Cloud Platforms Fit in the Agent Era

This year』s biggest buzz has been AI Agents. From the 「lobster craze」 to the wide array of Agent products launched by major vendors, the Agent era has truly begun.

Cloud providers have also found an opportunity to come into their own in this new wave.

If we want Agents to achieve real commercial deployment—especially at the enterprise level, and then be generalized into industry-specific scenarios—the first thing that must be solved is security and stable operations. Next, Agents/Agent teams also need to be easier to manage, aligning with enterprise workflows and business-scenario requirements.

For now, whether it』s model vendors building Agents or Agent apps that go directly to market, they still fall short in these areas. Moreover, for enterprises and industries, when the foundation isn』t solid, locking into a single model or application also carries risks.

For cloud platforms, however, security, stability, and enterprise services are exactly where their strengths lie.

From some real-world cases, cloud platforms can not only swap underlying models freely within a unified security framework, but also provide enterprise users with infrastructure-grade security guardrails and a governance foundation to protect core assets such as operations and data. In addition, they can deeply embed Agents into an enterprise』s existing identity, access-control, and audit/governance systems, making them easier for enterprises to operate and manage.

At present, cloud providers』 Agent push has entered an accelerated phase.

At the 「2026 Alibaba Cloud Summit」 on May 20, Alibaba Cloud launched the official website for its 「Qwen Cloud」 product, offering APIs for more than 150 mainstream models. Reportedly, the 「users」 of this model are not 「humans」 but Agents. By packaging the core capabilities of model services into Skills and CLI tools, it enables Agents to call models on their own and build AI applications. Based on what has been revealed at the summit so far, 「Qwen Cloud」 also redesigned its underlying architecture for this purpose. In the end, this means Agents can invoke models without getting bogged down in low-level adaptation, while models must comply with the security framework, identity system, and audit rules defined by 「Qwen Cloud.」 For enterprises, this translates into using Agents with greater confidence and flexibility.

AWS has also been working to position Bedrock as an AI Agent marketplace. When integrating OpenAI models, it laid out detailed requirements—for example, that relevant models on Bedrock must 「inherit the enterprise control mechanisms customers depend on,」 among other stipulations. Bedrock Managed Agents also emphasizes that each agent has an independent identity, every action is logged, agents run within the customer』s AWS environment, and all inference is completed on Bedrock. All of this makes it feasible for enterprises to connect to and use Agents more easily, while further binding this market to AWS』s own cloud ecosystem.

In addition, Azure AI Foundry introduced an 「agent identity」 feature this year, giving each Agent an independent identity within the service and enabling organizations to track behavior through permission controls, auditing, and logging. Google Cloud is also playing to its long-standing strength in 「security.」 Vertex AI—now upgraded into the 「Gemini Enterprise Agent Platform」—recently added security features such as an entry point for enforcing security policies and mechanisms to detect anomalous agent behavior, and on May 14 it officially launched security protection for Vertex AI.

Cloud platforms stepping in may prove to be the key to rolling out Agents at scale and making their business models work. In fact, from a market perspective, what enterprises care about most when procuring AI Agents may not be 「which model is best,」 but rather 「which model fits us best—and can run within our systems in the safest, most controllable, and most efficient way.」

2026 is only the 「first year」 of Agents being deployed at scale; the industrial and commercial race—and the reshuffling of the competitive landscape—has only just begun. So far, cloud platforms not only seem poised to reclaim their 「pricing power」; by bringing model companies under their own security governance frameworks, they may further become the dominant players in the market, and even the ones who set the rules of the game.

(Author | Hu Jiameng, Editor | Yang Lin)