Alphabet has reportedly ordered 3 million Tensor Processing Units (TPUs) from Intel for delivery through 2028 to diversify its AI chip supply chain. This move shifts Google away from its exclusive reliance on Taiwan Semiconductor Manufacturing (TSMC) as the massive demand for artificial intelligence hardware threatens to outpace current manufacturing capacity.
Why is Alphabet shifting its chip manufacturing to Intel?
The decision to partner with Intel stems from the need to avoid production bottlenecks. While Google has collaborated with Intel on specialized processors for years, those chips were historically manufactured by TSMC. According to reporting from The Motley Fool, Google’s recent confidence in Intel follows months of testing Intel’s specific chip packaging technology to ensure it meets Google’s rigid engineering standards.
By adding Intel as a primary manufacturer, Alphabet gains a vital second source for its application-specific integrated circuits (ASICs). These chips are purpose-built for the specific matrix and vector-based mathematics required to run large-scale AI models. Relying on a single provider like TSMC creates a single point of failure in an era where AI compute demand is growing exponentially.
How do Google’s new TPU architectures improve AI performance?
Google is moving away from “all-purpose” AI chips in favor of specialized hardware. At the recent Cloud Next conference, Google introduced two distinct architectures: the TPU 8t and the TPU 8i.
Amin Vahdat, Google’s senior VP and chief technologist for AI and infrastructure, stated that the community benefits from chips specialized for individual tasks. The distinction is clear:
- TPU 8t: Dedicated specifically to training workloads, where models learn from massive datasets.
- TPU 8i: Designed for inference, which is the process of a trained model providing answers or predictions to users.
Vahdat noted that this specialization allows the company to run its most demanding AI workloads two to four times faster. Additionally, the company reports these specialized chips operate at a 30% lower cost than previous-generation TPUs.
Comparing TPU Architectures
| Feature | TPU 8t | TPU 8i |
|---|---|---|
| Primary Use | Model Training | Model Inference |
| Optimization Goal | Computational Throughput | Serving Latency/Efficiency |
What happens next for Google’s AI business model?
Alphabet is transitioning its TPU strategy from internal use to a commercial product. While Google historically used these chips only for its own services, executives recently announced they will sell TPUs to a select group of external customers. This shift is intended to significantly expand the company’s total addressable market.
This commercial expansion is already reflected in the company’s financial metrics. Google’s backlog has nearly doubled year over year, reaching a reported $460 billion. This surge suggests that enterprise demand for specialized AI hardware is outpacing the availability of general-purpose chips like those produced by Nvidia.
Frequently Asked Questions
Why is Google using Intel instead of just TSMC?
Google is using Intel to diversify its supply chain. By not relying solely on TSMC, Google can avoid production delays and capacity shortages caused by the global AI boom.
What is the difference between training and inference?
Training is the process of teaching an AI model using massive amounts of data. Inference is the process of the AI actually working—such as when you ask a chatbot a question and it generates a response.
How much has Google’s AI backlog grown?
According to reported data, Google’s backlog has nearly doubled year over year to approximately $460 billion.
What do you think about Google’s move to Intel? Will this help them catch up in the AI race? Let us know in the comments below or subscribe to our newsletter for more deep dives into the semiconductor industry.
