A research and investment information platform tailored for financial professionalsCrafted with care
Expectation gaps are productivity.
—— Jingbei Moonlight
📢 Reports indicate that Google may begin selling Tensor Processing Units (TPUs) to external data centers. Currently, NVIDIA sells approximately 8 million chips annually; can TPUs achieve sales of 500,000 to 1 million units per year? The answer may be affirmative. Each sale of 500,000 TPUs is expected to increase Google’s cloud revenue by about 10% in 2027, boost earnings per share (EPS) by 3%, and drive up valuation multiples. This report also elaborates on the key impacts of this event on the semiconductor industry.
What happened?
🗣️ Reports show that Google is in talks with metaverse platform companies (META) and other cloud service customers, allowing them to use Google’s TPUs in their own data centers. Notably, there are indications that META plans to invest billions of dollars in 2027 to deploy TPUs in its data centers, while also renting TPUs from Google Cloud next year. It is particularly important to note that discussions between META and Google focus on using TPUs for model training rather than supporting existing inference workloads.
🔧 The backdrop of this development is that Google has developed software called the “TPU Command Center,” aimed at lowering the barrier for customers to use TPUs—customers can interact with the TPU server cluster through the PyTorch framework and the command center software. Additionally, Anthropic has previously reached an agreement with Google Cloud to gain access to approximately 1 million TPUs, laying the groundwork for this collaboration.
💰 Each sale of 500,000 branded TPUs could increase EPS by 3% in 2027.
🤝 The recent agreement between Anthropic and Google for the use of 1 million TPUs, along with current discussions with META regarding procurement, marks a significant breakthrough in Google’s strategy to deploy and sell TPUs externally after about 12 years of innovation in this field. It is estimated that Broadcom produced approximately 1.8 million TPUs for Google in 2025, most of which were used internally; Morgan Stanley’s semiconductor team predicts that TPU production will continue to expand, potentially reaching 3 million units by 2027.
📈 Reports suggest that positive feedback from external markets may further boost Google’s TPU procurement in 2027, a metric worth monitoring. Sensitivity analysis indicates that the scale of Google’s external sales of TPUs through a first-party (1P) model will significantly impact performance: each sale of approximately 500,000 chips is expected to bring in $13 billion (about 11%) in growth for Google’s cloud revenue in 2027 (higher than market expectations) and increase EPS by 3% (approximately $0.37).
✨ The accelerated growth of Google’s cloud business and its expansion in the TPU market may also drive up its valuation multiples (similar trends have been observed in recent months). Morgan Stanley analysts Joe Moore and the NVIDIA team predict that NVIDIA will ship about 8 million graphics processing units (GPUs) in 2027. Considering two factors—one is capacity supply capability, and the other is that Google TPU continues to overcome historical pain points of developer familiarity and limited programming flexibility—the goal of achieving annual sales of 500,000 to 1 million TPUs is reasonable.
❓ Three unknown factors to watch regarding Google’s TPU external strategy
📌 TPU business model: Google has expressed a willingness to flexibly meet customer needs by providing various TPU acquisition methods (including Google Cloud rentals, licensing, chip sales, and equipment rentals). This will affect the revenue recognition of related transactions, potentially impacting Google’s profit margins and capital expenditures. Although the above calculations assume that Google sells chips directly to META for use in its data centers, specific pricing models, contract terms, and additional services remain unclear.
💰 TPU pricing strategy: There is a significant difference in gross margins across the chip industry, with NVIDIA around 75% and AMD around 50%, which significantly impacts the pricing of Google TPUs. In sensitivity analysis, we assume that Google’s TPU gross margin is at the lower end of industry peers because Google needs to first validate product competitiveness and gain initial customer acceptance. Additionally, the calculations are based on single-chip pricing, but if Google sells rack-level products to META, the pricing logic may change.
🛠️ TPU application scenarios: Although META is considering using TPUs for model training, the training process can be broken down into multiple stages, and it is currently unclear how dependent META is on TPUs for training its Llama model. If TPUs become a core component of the Llama model training infrastructure, it will be significant—this means that two of the five leading large language models (LLMs) will be trained based on TPUs.
🤔 What does this mean for META?—Achieving a more cost-effective computing solution
💸 While META has been investing in building cutting-edge large language models (LLMs) and optimizing its core business, it has also been focused on controlling the growth of capital expenditures (capex) and operating expenses (opex). The high cost-performance ratio of TPUs (outstanding performance per dollar) is a key driver for META’s consideration of adopting this chip. Given the excellent cost-effectiveness of TPUs, deploying them in 2027 is expected to help META offset some capital expenditure pressures. The ultimate impact will depend on two factors: one is the scale of META’s procurement (the internet industry team currently predicts that META will procure over 1 million NVIDIA chips in 2026), and the other is the efficiency advantage of TPUs compared to other chips.
🔧 Impact on the semiconductor industry
✅ The expansion of Google’s TPU applications will benefit Broadcom (AVGO), but whether it can be widely adopted by other customers remains uncertain. Currently, Broadcom has disclosed five dedicated processor (XPU) customers, which we speculate to be Google, META, ByteDance, Anthropic, and OpenAI. If META confirms the adoption of TPUs, it will mean that three out of five customers (potentially four, including OpenAI) are using this product.
🏆 From an industry competition perspective, TPUs are becoming another commercial competitor to NVIDIA in the dedicated processor (XPU) market, but this does not equate to Broadcom’s dedicated integrated circuit (ASIC) strategy gaining industry recognition.
⚠️ If META’s scale of TPU application exceeds previous expectations, it may delay the rollout of its self-developed dedicated integrated circuit (MTIA) planned for the second half of 2026, thereby reducing the potential growth space for EPS. This could be a positive signal for NVIDIA: given that ASIC development must contend with NVIDIA’s annual product iteration pressure, customers may prefer to choose the more mature TPU ecosystem—especially since product development involves not only the chip itself but also rack-level connectivity solutions, where NVIDIA leads in commercialization.
📌 In the short term, although news of META adopting TPUs may raise market concerns about NVIDIA, the overall demand in the AI industry continues to rise, which will not significantly change our view on NVIDIA. Broadcom is expected to achieve growth advantages next year (noting that some growth comes from lower-margin rack business), but current valuation multiples are already at a high level. We are optimistic about both companies, as the TPU supply chain is clearly accelerating, while NVIDIA’s supply chain also maintains a growth trend.
🤝 In the large language model (LLM) space, NVIDIA has gained more market share from Google this year—Google’s spending on NVIDIA is about $20 billion this year, while spending on TPUs is about several billion dollars. This ratio may adjust next year, but we do not expect the market to exhibit a “winner-takes-all” pattern. Currently, even if the Gemini model surpasses GPT-5 to become the industry-leading model, competitors will not slow down their investments—technology industry leaders have repeatedly emphasized the importance of this competition, and the laws of industry scaling remain effective, ensuring that market competition will remain fierce. However, if competitors’ market shares are impacted, their subsequent financing and ongoing investment capabilities will also be affected.
🤔 What about AMD’s situation?
📌 We believe AMD’s overall situation remains largely unchanged. In the short term, one of its core customers, META, seeking alternative solutions may affect market confidence in the upcoming launch of AMD’s MI450 chip. However, META’s strategic goal has not changed: to diversify its computing infrastructure by investing in self-developed ASICs, collaborating with Broadcom to develop external ASICs, and procuring GPUs from NVIDIA and AMD, now adding TPUs as an option.
💹 In the long term, return on investment (ROI) will be a key decision factor; AMD needs to demonstrate the ROI advantages of its products to gain a significant share in this market. Previously, a notable advantage of AMD over ASICs was adaptability to all cloud platforms, but as TPUs gradually become a commercial alternative, this advantage has diminished. We have high expectations for AMD’s MI450 chip performance, but without clear ROI validation data, market confidence is unlikely to significantly improve.
For complete content, more research summaries, and to receive hard-hitting insights in advance, scan the code to join:
If you found the article insightful, please click the “Read” button in the upper right corner and share it with your friends. After reading, feel free to give it a thumbs up.
And Read
Disclaimer: The views expressed in this article and any articles from this public account are for discussion purposes only and do not constitute any investment advice.