Company Overview
Google was founded in 1998, headquartered in Mountain View, California, USA. In 2015, it was reorganized as a core subsidiary under Alphabet Inc. (ticker: GOOGL). In the AI industry chain, Google has secured a position in the "core computing and communication hardware" segment with its self-developed Tensor Processing Unit (TPU) and one of the world's largest AI clusters. The TPU is an Application-Specific Integrated Circuit (ASIC), deeply optimized for frameworks such as TensorFlow, and is a pioneering force among North American cloud giants in the self-developed-chip showdown against NVIDIA.
TPU is a custom ASIC designed by Google for machine learning workloads. The latest generation TPU v5p delivers 459 TFLOPS of compute at BF16 precision, with a massive Pod supporting up to 8960 chips (interconnected via Google's self-developed optical switches). TPUs are not sold directly to external customers; instead, they serve as the core compute units for Google Cloud AI services. This business generates revenue primarily through a "compute rental" model. In 2023, related cloud AI revenue (including TPU rental) was approximately $5 billion, accounting for about 15% of Google Cloud's total revenue ($33.1 billion).
Google Cloud provides AI infrastructure based on TPUs and GPUs (A100/H100), including Vertex AI, AI Platform, and others. In 2023, Google Cloud generated $33.1 billion in revenue, up 26% year-over-year, with losses narrowing significantly. AI-related services contributed approximately 15%–20% of total revenue and were the fastest-growing segment. Additionally, Google supplies computing power to internal businesses (Search, Ads, YouTube, DeepMind) through its own data centers. This portion does not directly contribute to revenue but substantially reduces machine learning inference costs.
| Product Line | Revenue Share | Core Customers | Gross Margin (Est.) |
|---|---|---|---|
| Google Cloud (incl. TPU services) | ~10% of Alphabet | Enterprises, DeepMind | ~20% |
| Ads & Search (internal compute) | ~80% of Alphabet | Internal | ~60% |
Note: The cost savings from internal TPU usage are difficult to quantify, but they are considered a core factor supporting the gross margins of Ads/Search.
Technical Moat
Moat 1: Deep Integration of TPU with Google's Software Stack
From its inception, the TPU has been deeply integrated with frameworks such as TensorFlow and JAX, efficiently mapping computation graphs onto TPU hardware through the XLA compiler. This "hardware-software-algorithm" closed loop enables Google to achieve higher efficiency than general-purpose GPUs in large-scale training. For example, the Gemini model family runs on TPU v5p clusters. Third parties must access TPUs through Google Cloud services, resulting in technical lock-in.
Moat 2: Self-Developed Optical Communication and Large-Scale Interconnection
TPU Pods adopt self-developed Optical Circuit Switching (OCS) technology to achieve high-bandwidth, low-latency interconnect between chips without traditional switches. This architecture supports real-time reconfiguration of nearly arbitrary topologies, substantially improving training efficiency. To date, Google has deployed a single logical cluster of over 100,000 TPUs, ranking among the largest in the world.
Moat 3: Vertically Integrated Computing Ecosystem
Google simultaneously owns top-tier AI models (Gemini, PaLM, Gemma) and supporting chips, enabling additional performance gains through joint "model-chip" optimization. For instance, Gemini Nano runs efficiently on mobile devices, while larger Gemini models run on TPUs. This closed loop makes it difficult for competitors to replicate simply.
| Dimension | Data |
|---|---|
| Global AI accelerator chip market share | ~5% (2023, IDC estimate) |
| Industry ranking | 2nd (after NVIDIA) |
| Main competitors | NVIDIA (GPU), AMD (MI300), Amazon (Trainium), Microsoft (Maia), Intel (Gaudi) |
| Downstream customers | DeepMind, Google Research, Snapchat, Cohere, and other cloud customers, as well as internal Search/Ads |
Landscape analysis: Amid the current surge in AI compute demand, NVIDIA GPUs capture over 80% of the market with their versatility. However, ASICs represented by Google's TPU deliver a 2–4x cost-performance advantage in inference and specific training scenarios. As cloud vendors continue their in-house chip development, the ASIC market share is expected to rise from 5% in 2023 to 15% by 2028 (per various institutional forecasts). Google TPU currently holds the leading position in the ASIC segment.
Financials and Growth
| Metric | Data |
|---|---|
| Parent company Alphabet revenue (2023) | US$307.4 billion |
| Parent company net profit | US$73.8 billion |
| Google Cloud revenue (including TPU services) | US$33.1 billion |
| Google Cloud AI-related revenue (estimate) | US$5–6 billion |
| Gross margin (Cloud business) | ~18% (turned positive in 2023) |
| Core growth thesis | Global enterprise AI cloud adoption wave + TPU v5p cost-performance improvement + Google model ecosystem pull |
Growth Drivers:
- Enterprise demand for AI training and inference is growing explosively, and Google Cloud AI revenue is expected to maintain 50%+ annual growth.
- TPU v5p delivers 30%~50% better cost-performance than NVIDIA H100 for training large models, attracting price-sensitive customers.
- DeepMind continues to release top-tier models (Gemini, Gemma), forming a positive flywheel of "models → chips → cloud services".
Key Risks:
- NVIDIA GPU Dominance: NVIDIA holds an absolute advantage in the CUDA ecosystem, product iteration speed, and brand recognition. Google's TPU lacks general-purpose versatility, making it difficult to attract mainstream developers and enterprises.
- Supply Chain Dependency: TPUs are primarily manufactured by TSMC. If geopolitical factors lead to capacity constraints, chip delivery could be impacted.
- Internal Competition and Organizational Risk: Conflicts may arise between Google's internal chip teams and external cloud customer needs (e.g., TPU lacks optimization for PyTorch), while competitive pressure from the Microsoft/OpenAI and Amazon alliance continues to intensify.
- Technology Path Risk: If the Transformer architecture is replaced by other paradigms, the current TPU optimized for matrix multiplication may face performance degradation.
Core Investment Logic / Industry Value Summary:
Google's TPU serves as a benchmark for self-developed AI chips among North American cloud giants, representing an important direction toward reducing dependence on NVIDIA and achieving vertical integration. Although its market share is unlikely to surpass GPUs in the short term, the TPU offers significant cost advantages in inference and internal training scenarios, thanks to its software ecosystem, optical interconnect, and joint model optimization. As the ASIC market share grows, Google is well-positioned to secure a differentiated role in the AI computing power arena.