Google orders 3 million chips from Intel: Supply chain restructuring under the AI computing power famine

From Capacity Bottleneck to Strategic Substitution: A Turning Point in the AI Chip Supply Chain

In 2025, AI model parameters have already broken through the trillion-level threshold, with the computing power required for training and inference growing exponentially. This technological revolution is reshaping the power structure of the semiconductor industry. The latest news shows that Google has ordered at least 3 million TPUs (Tensor Processing Units) from Intel, with delivery expected between 2027 and 2028. At the same time, NVIDIA is also actively testing Intel's manufacturing process, seeking alternative capacity for its next-generation GPU architecture, "Feynman." Behind these moves lies a massive gap between global AI computing demand and TSMC's production capacity.

As the world's most advanced semiconductor foundry, TSMC's 3nm and 5nm capacity has long been occupied by AI chip orders. Reports indicate that its capacity is fully booked until 2028, and even the yet-to-be-completed Arizona plant has been reserved entirely. This supply-demand imbalance is driving a key trend: tech giants are beginning to proactively build "dual-source" or even "multi-source" chip supply systems.

Intel Foundry's Comeback: From Follower to Backup Option

For years, Intel lagged behind TSMC in advanced process technology, and its foundry business was once considered a marginal option. But the capacity pressure from AI has given Intel some breathing room. Google's TPUs were primarily produced by TSMC; now, shifting some orders to Intel means that Intel's 18A process has at least met Google's reliability requirements. NVIDIA's test contract is even more critical: if the Feynman architecture (integrating four GPU chips) can be successfully manufactured by Intel, it would formally establish Intel as the only large-scale advanced foundry alternative to TSMC.

Intel recently announced an investment of $3.3 billion to build a new substrate packaging plant in India, in addition to its expansion plans in the United States and Europe, demonstrating its determination. However, pricing and yield rates remain challenges. The report did not disclose the contract amount, but considering that Google plans to produce more than 6 million TPUs between 2027 and 2028, with Intel taking on at least half of that volume, it will be enough to reshape its foundry business revenue structure.

Techno-Geopolitics: A Distributed Future for Chip Supply Chains

This shift is not purely commercial. TSMC is concentrated in Taiwan, while the United States and Europe are vigorously promoting domestic chip manufacturing. As the only company in the US with advanced process capabilities, Intel naturally benefits from the national security narrative. Google and NVIDIA's shift toward Intel objectively reduces single-supplier risk. In the future, we may see more of a "design in the US, manufacture in multiple locations" model—TSMC, Intel, and Samsung will form a tripolar landscape, but Intel, leveraging its geographic advantages and political support, could capture 20–30% of high-end orders.

AI Demand Will Not Slow Down: Long-Term Challenges for Computing InfrastructureThe scaling of AI is accelerating. Model training clusters at companies like OpenAI, Google, and Meta already require tens of thousands of GPUs, and in the future, they will need millions. Nvidia's Feynman architecture was specifically designed to handle ultra-large models. Even if Intel and TSMC expand production simultaneously, industry-wide capacity may still remain tight until 2028. This, in turn, will spur architectural innovations (such as more efficient chip interconnects) and exploration of non-silicon computing (optical computing, in-memory computing).

Conclusion: The Chip War Enters a New Phase

Google's order to Intel is ostensibly about capacity diversification, but in reality, it is a strategic move to gain autonomous control over the AI industry chain. When computing power becomes the new oil, controlling manufacturing nodes equals controlling the energy switch. Nvidia's testing further shows that this GPU giant is also unwilling to put all its eggs in one basket. Over the next two years, we will see more such "dual sourcing" agreements, and Intel is likely to achieve a true revival through this. For the entire industry, this is a signal: the AI boom is reshaping semiconductor geography, and the ultimate beneficiaries of this transformation are those companies that can simultaneously master technology, capital, and geopolitics.