Alibaba has just made a bold bet that controlling the entire lifecycle of AI tokens, the basic computational units powering every AI model, is the key to dominance in the AI era. On June 8th, the Chinese tech giant merged its Tongyi Large Model Division with the Future Life Laboratory to create the Token Foundry, a new division personally led by Group CEO Wu Yongming. The move signals that the real competition in AI is no longer just about building the best models, but about controlling the infrastructure and ecosystem that makes those models possible.

This restructuring reveals a fundamental shift in how tech giants are competing. For years, the focus was on model capability, raw computing power, and who could train the largest language models. But as open-source models have caught up faster than expected, and token prices have plummeted, companies are realizing that owning the entire supply chain, from silicon to software to distribution, is the only way to maintain a lasting advantage.

Why Is Token Control Becoming the New Battleground?

Every AI model, whether it is ChatGPT, Claude, or Alibaba's Qwen, generates and consumes tokens at massive scale. These tokens require GPU (graphics processing unit) computing power to create. Historically, NVIDIA has collected what amounts to a toll on every token generated worldwide, because nearly all AI companies rely on NVIDIA's hardware. But the real moat, according to industry analysis, is not the hardware itself, but the developer ecosystem that locks companies into a platform.

NVIDIA released CUDA, a software platform that lets researchers and developers harness computing power, back in 2007. For nearly two decades, CUDA has been the de facto standard for AI development. This lock-in effect means that even if competitors build better chips, developers remain reluctant to switch because rewriting code for a new platform is expensive and time-consuming. NVIDIA's strategy has been to keep the most powerful model capabilities behind proprietary APIs, creating a closed loop where the generation and consumption of tokens happen entirely within NVIDIA's ecosystem.

However, the competitive landscape is shifting. Anthropic, the AI safety company behind Claude, has taken a different approach. Instead of trying to own the entire stack, Anthropic is penetrating the application layer through developer tools and protocols. The company developed the MCP (Model Context Protocol), which has become the de facto standard for the agent ecosystem, with over 97 million monthly SDK (software development kit) downloads. OpenAI, Google, and Microsoft have all connected to it. Claude Code, Anthropic's AI programming tool, has the fastest adoption growth rate among all enterprise AI coding tools.

How Are Tech Giants Restructuring to Win the Token Wars?

• Alibaba's Vertical Integration: Alibaba's Token Foundry controls chips through Pingtouge, models through Qwen and its video generation tools, and distribution through Bailian MaaS. The company has assembled the complete puzzle from silicon to applications, aiming to own every step of the token supply chain.
• NVIDIA's Ecosystem Expansion: NVIDIA is extending beyond hardware into software and inference platforms, recognizing that the real competitive advantage lies in developer stickiness and the CUDA ecosystem rather than chip superiority alone.
• OpenAI's Platform Strategy: OpenAI is developing its own chips with Broadcom and has run models on Cerebras chips, while simultaneously building proprietary toolchains and protocols to make switching costs prohibitively high for developers.
• Anthropic's Toolchain Approach: Anthropic is focusing on the protocol and toolchain layer, where replacement costs are extremely high, rather than competing purely on model capability.

Alibaba's new Token Foundry includes several world-leading models. Qwen 3.7 Max, released recently, was designed for agents by nature and supports an ultra-long context of 1 million tokens, allowing AI to run continuously for dozens of hours and adjust thousands of tools by itself. Happy Horse, Alibaba's video generation model, topped the global video generation list in April. Happy Oyster, positioned as a "real-time world engine," generates images while users are speaking, functioning like a real-time interactive world simulator.

The boundaries between hardware, software, and applications are blurring rapidly. Companies that once focused on a single layer are now expanding into adjacent layers. This cross-over intensification means that the winner in the AI era will not be the company with the best single component, but the company that can make switching costs prohibitively high across multiple layers of the stack.

What Does This Mean for NVIDIA's Dominance?

NVIDIA's position remains formidable, but it is not invincible. The company's Blackwell GPU architecture is still processing a large number of orders, and the next-generation Rubin architecture is already on the agenda. This business currently has the highest certainty in the semiconductor industry. However, every major tech company is now developing its own chips. Google has TPU (tensor processing unit), and Amazon has Trainium. The proliferation of custom silicon means that NVIDIA can no longer rely on hardware superiority alone.

The real test for NVIDIA will be whether the CUDA ecosystem can maintain its dominance as open-source models catch up and token prices continue to fall. If developers can achieve comparable results with cheaper alternatives, the switching costs that have protected NVIDIA may erode faster than expected. The window for pure model-capability premium will not last long, and companies like OpenAI, Anthropic, and Alibaba are racing to build irreplaceable toolchains and protocols before that window closes.

What Should Companies Do to Avoid Becoming Dependent on NVIDIA?

For companies outside the AI giant ecosystem, the stakes are particularly high. Jensen Huang, NVIDIA's CEO, is a disciplined strategist who will closely examine the competitiveness of manufacturing industries, including semiconductors. If companies hope to avoid becoming dependent on NVIDIA's AI platform, they must preserve overwhelming technological advantages of their own. This requires continuous innovation and the willingness to sharpen capabilities relentlessly, just as Huang has pushed both himself and NVIDIA beyond their perceived limits over more than three decades.

The battle for control of the token supply chain has just entered its main event. The companies that win will not be those with the best single technology, but those that can build ecosystems so sticky and valuable that switching becomes economically irrational. Alibaba's Token Foundry, NVIDIA's CUDA ecosystem, Anthropic's MCP protocol, and OpenAI's proprietary toolchain are all bets on this same fundamental principle: in the AI era, control of the supply chain is control of the future.

" }