Alibaba Group is building its own artificial intelligence chips. It’s also pouring resources into agentic AI — the kind of AI that doesn’t just answer questions but takes actions, makes decisions, and executes multi-step tasks with minimal human oversight. These two moves, taken together, signal that China’s largest e-commerce and cloud computing company is mounting a serious challenge to the American incumbents that have dominated the AI hardware and software stack for years.
The ambition is enormous. And the timing isn’t accidental.
According to TechRepublic, Alibaba has been accelerating development of its in-house AI processors while simultaneously advancing its agentic AI capabilities through its cloud division. The company’s Hanguang 800 chip, first introduced in 2019, was an early signal of intent. But the latest efforts go much further. Alibaba’s DAMO Academy — its global research arm — has been working on next-generation chips designed specifically for the inference workloads that agentic AI systems demand. These aren’t training chips meant to compete head-to-head with Nvidia’s H100 or Blackwell architectures. They’re purpose-built silicon aimed at running AI models efficiently at scale, particularly within Alibaba’s own cloud infrastructure.
That distinction matters more than it might seem at first glance. Training large language models requires brute computational force — the kind Nvidia has monopolized with its GPU platforms. But inference, the process of actually running a trained model to produce outputs in real time, is where the long-term economics of AI will be won or lost. Every time a customer asks an AI agent to book a flight, reconcile an invoice, or analyze a contract, that’s an inference workload. And as agentic AI proliferates, inference demand will dwarf training demand by orders of magnitude.
Alibaba knows this. So does every other major cloud provider.
The company’s cloud division, Alibaba Cloud, is already the largest cloud computing platform in Asia. It serves millions of businesses across China and Southeast Asia, and it has been aggressively expanding its AI services portfolio. The integration of custom AI chips into this infrastructure would reduce Alibaba’s dependence on external suppliers — most critically, on Nvidia and AMD, both of which face U.S. export restrictions that limit the performance of chips they can sell to Chinese companies.
Those export controls, tightened repeatedly by the Biden administration and maintained under President Trump, have become the defining constraint on China’s AI ambitions. Washington has restricted the sale of advanced AI chips to China, citing national security concerns. Nvidia’s A100 and H100 processors were among the first to be curtailed. Subsequent rounds of restrictions closed loopholes that allowed modified chips to slip through. The result: Chinese companies can’t easily buy the world’s best AI hardware. They have to build their own.
This is exactly what Alibaba is doing. And it isn’t alone. Huawei has developed its Ascend series of AI processors. Baidu has its Kunlun chips. ByteDance and Tencent have explored custom silicon as well. But Alibaba’s effort stands out because of the company’s sheer scale in cloud computing and its aggressive push into agentic AI, which creates a natural demand-side pull for custom inference hardware.
Agentic AI represents a fundamental shift in how artificial intelligence is deployed. Traditional AI systems — chatbots, recommendation engines, image classifiers — respond to individual prompts. They generate an output and stop. Agentic AI systems, by contrast, can pursue goals over extended interactions. They can break complex tasks into subtasks, use external tools, query databases, interact with APIs, and adjust their strategies based on intermediate results. Think of the difference between asking a chatbot to summarize a document and asking an AI agent to research a topic, draft a report, fact-check it against multiple sources, format it according to company guidelines, and email it to a distribution list.
The commercial implications are staggering.
Alibaba’s Qwen family of large language models, developed by the company’s cloud intelligence unit, has been gaining traction as a foundation for agentic applications. The Qwen models are open-weight, meaning external developers can build on top of them — a strategic choice that mirrors Meta’s approach with its Llama models and contrasts sharply with OpenAI’s closed model. By making Qwen freely available, Alibaba is cultivating a developer community that builds agentic applications running on Alibaba Cloud, consuming Alibaba’s inference compute, and potentially running on Alibaba’s custom chips.
It’s a vertically integrated play. Hardware, models, cloud platform, developer tools, end-user applications — all connected.
Recent developments have only intensified the competitive dynamics. In May 2025, Alibaba reported strong quarterly earnings driven in part by surging demand for its AI cloud services. The company disclosed that AI-related revenue within Alibaba Cloud grew at triple-digit rates year over year, though it didn’t break out exact figures. CEO Eddie Wu told analysts that AI inference workloads were growing faster than any other category on the platform, validating the company’s bet on custom inference silicon.
Meanwhile, the broader agentic AI market is heating up globally. OpenAI, Google DeepMind, Microsoft, Amazon Web Services, and Anthropic are all racing to build and deploy AI agents. Microsoft has embedded agentic capabilities into its Copilot products across Office, Dynamics, and Azure. Salesforce launched Agentforce. Google introduced agent-building tools in Vertex AI. The consensus among industry leaders is that agentic AI will be the primary interface through which businesses interact with AI systems within the next two to three years.
But there’s a catch. Agentic AI systems are computationally expensive. Each agent interaction involves multiple model calls, tool invocations, and reasoning steps. A single agentic task might require ten or twenty times the inference compute of a simple chatbot exchange. This is why custom inference chips matter so much — and why Alibaba’s dual investment in chips and agentic AI is strategically coherent in a way that isolated bets on either one wouldn’t be.
The geopolitical dimension can’t be separated from the commercial one. U.S. restrictions on chip exports to China have created a bifurcated global AI supply chain. American companies dominate the high end of AI training hardware. Chinese companies, locked out of that market, are investing heavily in domestic alternatives. The risk for Washington is that these restrictions, while slowing China’s progress in the short term, are accelerating the development of a fully independent Chinese AI hardware stack in the medium term. Alibaba’s chip efforts are a case study in this dynamic.
There are real technical challenges, of course. Designing competitive AI chips is extraordinarily difficult. Nvidia’s dominance rests not just on its silicon but on CUDA, its proprietary software platform that makes it easy for developers to write GPU-accelerated code. Any challenger — Chinese or otherwise — must offer not just competitive hardware but a software development environment that engineers actually want to use. Alibaba has been building out its own AI compiler and runtime tools, but closing the gap with CUDA is a multi-year project with uncertain outcomes.
Fabrication is another bottleneck. The most advanced AI chips are manufactured by TSMC in Taiwan using extreme ultraviolet lithography at process nodes of 5 nanometers and below. Chinese foundries, led by SMIC, have made progress but still trail TSMC by at least one or two generations. This means Alibaba’s custom chips may not match the raw performance of Nvidia’s latest offerings — but they don’t necessarily need to. If the chips are good enough for inference workloads and cheap enough to deploy at scale within Alibaba’s own data centers, the economic math can still work.
Good enough, deployed at massive scale, often beats best-in-class deployed in limited quantities. That’s a lesson the Chinese technology sector has internalized across multiple industries, from smartphones to electric vehicles to solar panels.
Alibaba’s approach also reflects a broader trend among hyperscale cloud providers worldwide. Amazon has its Trainium and Inferentia chips. Google has its Tensor Processing Units. Microsoft is developing its Maia AI accelerator. The logic is the same everywhere: vertical integration reduces costs, improves performance for specific workloads, and decreases dependence on a single supplier. For Chinese companies, the additional motivation of circumventing export controls makes the case even more compelling.
The agentic AI piece of Alibaba’s strategy deserves closer examination. The company has been building what it calls “AI agents for enterprise” — pre-built agentic systems designed for specific business functions like customer service, supply chain management, financial analysis, and software development. These agents run on Alibaba Cloud, use Qwen models as their reasoning backbone, and are designed to integrate with the enterprise software tools that Chinese businesses already use.
This is a direct analog to what Microsoft is doing with Copilot agents in the Western market. The difference is that Alibaba controls more of the stack. It builds the models, designs the chips, operates the cloud, and in many cases provides the enterprise applications (through its DingTalk collaboration platform and other business tools) that the agents interact with. That level of vertical integration gives Alibaba optimization advantages that a more modular competitor would struggle to match.
Not everyone is convinced the strategy will work. Skeptics point to Alibaba’s mixed track record in cloud computing outside China, its ongoing regulatory challenges with the Chinese government, and the sheer difficulty of building world-class AI chips without access to the best fabrication technology. There’s also the question of whether Alibaba can attract and retain the engineering talent needed to execute on so many fronts simultaneously. The company has undergone significant organizational restructuring in recent years, including a brief flirtation with splitting into six independent business units before partially reversing course.
But the direction of travel is clear. Alibaba is building toward a future where it controls the full AI stack — from silicon to models to agents to applications — within its own infrastructure. Whether it fully achieves that vision or not, the attempt itself is reshaping competitive dynamics in cloud computing and AI hardware across Asia and beyond.
For American chipmakers, the implications are sobering. Every custom AI chip that Alibaba deploys internally is a chip it didn’t buy from Nvidia. Multiply that by Huawei, Baidu, Tencent, and the dozens of smaller Chinese AI companies pursuing similar strategies, and the aggregate demand destruction for U.S. chip exports becomes significant. Nvidia CEO Jensen Huang has warned repeatedly that overly aggressive export controls risk permanently ceding the Chinese market to domestic competitors. Alibaba’s chip program is evidence that this risk is materializing.
The agentic AI race, meanwhile, is still in its early innings. No company — Chinese or American — has yet built AI agents that work reliably enough for fully autonomous enterprise deployment. Current agents still hallucinate, lose track of multi-step plans, and occasionally take actions their users didn’t intend. The technology is advancing rapidly, but the gap between demo-ready and production-ready remains substantial.
Alibaba is betting that closing this gap will require tight integration between hardware and software — that the companies best positioned to build reliable AI agents are the ones that also control the inference infrastructure those agents run on. It’s a compelling thesis. Whether it proves correct will depend on execution, and execution at this scale is never guaranteed.
What is guaranteed: the competition between U.S. and Chinese technology companies over AI’s future architecture is intensifying, not abating. Alibaba’s simultaneous push into custom chips and agentic AI isn’t just a product strategy. It’s an industrial policy response to geopolitical reality, wrapped in a commercial offering. And it’s one that the rest of the industry — in both hemispheres — will have to reckon with.
WebProNews is an iEntry Publication