Self-developed AI chips (TPU) and the world's largest AI computing infrastructure and cloud services
Deep coupling of self-developed TPUs with TensorFlow/JAX, backed by the world's largest AI clusters and data center network, builds an integrated hardware-software ecosystem moat
Company Overview
Google was founded in 1998 and reorganized into Alphabet Inc. in 2015 (listed on NASDAQ in 2004, ticker symbols GOOGL/GOOG), headquartered in Mountain View, California, USA. In the AI industry chain, Google is both a self-developer of core computing hardware (TPU series chips) and the world's largest operator of AI infrastructure (renting out AI computing power through Google Cloud), while also being deeply involved in the development of AI foundation models (Gemini). Its AI hardware business directly serves internal use cases such as Search, Ads, YouTube, and Waymo, and also provides cloud-native AI computing services externally.
TPU (Tensor Processing Unit) is a specialized ASIC chip designed by Google for deep neural networks, and has now iterated to the v5p version. This chip is not sold separately but is embedded in Google data centers, providing computing power to external customers on a pay-as-you-go basis through Google Cloud (Cloud TPU service). In 2024, Google announced the launch of its self-developed data center network chip Axelion to further remove cluster communication bottlenecks.
| Product Version | Purpose/Positioning | Key Performance Characteristics | Deployment Scale Reference |
|---|---|---|---|
| TPU v1 | Inference only | 8-bit integer arithmetic | Internal Search/Street View |
| TPU v2 | Training + Inference | 45 TFLOPS per chip (bf16) | 4x4=16-chip Pod |
| TPU v3 | Training + Inference | Liquid-cooled, 420 TFLOPS per chip (bf16) | 4x4=64-chip Pod |
| TPU v4 | Large-scale training | 2.7x performance improvement, optical interconnect | 4096-chip supercomputer |
| TPU v5e | Medium-scale training/inference | 2x performance per watt compared to v4 | More flexible cluster scaling |
| TPU v5p | Top-tier training | 2x FLOPs of v4, doubled memory bandwidth | 8960-chip Pod |
Google Cloud provides a complete AI development platform (Vertex AI, Colab), an AI supercomputer (AI Hypercomputer), and TPU/GPU-based rental services, serving as the primary channel for external customers to use Google AI hardware. In 2023, Google Cloud revenue grew 25% year-over-year, with accelerating contributions from AI and generative AI.
Gemini series models (Ultra/Pro/Nano) partially rely on TPU for training and inference, and the Nano version can run on client devices (Pixel phones, Chromebooks). Although this does not directly generate large-scale hardware revenue, it strengthens client-side demand for TPU.
Technical Moat
Moat 1: Ultimate Software-Hardware Integration
TPUs are deeply integrated with Google's self-developed ML frameworks, TensorFlow and JAX, with optimizations targeting the TPU microarchitecture spanning from the compiler to the runtime. In addition, Google operates the world's largest private AI training cluster (built on TPUs), accumulating massive distributed training expertise and network topology optimization patents (such as embedded High-Bandwidth Memory and optical interconnect architectures). External competitors find it difficult to replicate this full-stack synergy from chip design to compute scheduling.
Moat 2: Large-Scale Ecosystem and Data Flywheel
Google's internal AI applications (Search, YouTube, Advertising, Translate) generate PB-scale inference requests daily, providing real-world, rigorous scenario validation for TPU iteration. Meanwhile, Google Cloud attracts developers through TPU Pod rental, creating a flywheel effect of "more users → more optimizations → stronger chips → lower costs." As of 2024, over 60% of AI unicorns use Google Cloud's TPU or GPU services.
Moat 3: Data Center Network Innovation
The self-developed Axelion Ethernet switch released in 2024 elevated TPU cluster communication efficiency to an industry-leading level, reducing reliance on NVIDIA InfiniBand. Combined with Google's proprietary Jupiter network (the world's largest software-defined network), it forms end-to-end infrastructure competitiveness.
| Dimension | Data |
|---|---|
| Global AI Data Center Accelerator Market Share (2024) | ~10% (revenue basis, including in-cloud self-use) |
| Global Public Cloud AI Compute Market Share | ~35% (Top 3, competing with AWS and Azure) |
| Key Competitors | NVIDIA (GPU + NVLink + InfiniBand), AMD (Instinct), Intel (Gaudi), Amazon (Trainium/Inferentia) |
| Downstream Customers | Enterprise in-house AI teams, Google Cloud enterprise customers (e.g., Salesforce, AppLovin), AI research institutions |
| Competitive Landscape | TPU excels in high cost-performance training and deep software-hardware binding, but is constrained by ecosystem compatibility in the general-purpose GPU market |
Financial and Growth
| Metric | Data (FY 2023, as of 2023.12.31) |
|---|---|
| Group Total Revenue | $307 billion |
| Google Cloud Revenue | $33 billion (25% YoY increase) |
| AI-related capital expenditure (est.) | Approximately $32 billion (including TPU R&D and data center construction) |
| Group Gross Margin | 55% |
| Group Net Margin | 24% |
| Key Growth Drivers | AI-driven cloud services accelerate growth; large-scale deployment of TPU v5p reduces customer training costs; Gemini ecosystem attracts developers to switch; in-house network chips improve cluster economies of scale |
Key Risks:
- Reliance on Internal Demand: TPUs currently primarily serve Google's own businesses and cloud customers, with low penetration in the external standalone chip market (e.g., enterprise-owned data centers). If cloud growth decelerates, it could undermine economies of scale.
- Nvidia GPU Dominance: The CUDA ecosystem remains unassailable, with most mainstream AI frameworks and open-source models prioritizing GPUs. TPUs must rely on Google Cloud integration to drive adoption.
- High R&D and Manufacturing Costs for Custom Chips: Tape-out costs for 7nm and more advanced process nodes, along with rising data center power consumption, could lead to declining ROI if demand falls short of expectations.
- Potential Antitrust Risks: Bundling TPUs with Google Cloud may invite regulatory scrutiny—for example, the EU's DMA requires interoperability between cloud services.
Core Investment Thesis / Industry Value Summary:
Google AI positions its custom TPU chips and Axelion networking at the core of the compute and communications hardware layer, leveraging a tightly integrated software-hardware ecosystem to lock in internal use cases and cloud customers—forming a closed-loop flywheel from training to inference. Despite competition from Nvidia, TPUs have demonstrated cost advantages in specific domains such as large-scale Transformer training and high-throughput inference. Within the AI industry chain, Google stands as one of the few vertically integrated players spanning chip design, cloud computing, and frontier models (Gemini). Its hardware evolution roadmap plays a critical role in reducing AI compute costs and driving the democratization of computing power.