Alphabet has addressed worries about the potential threat of artificial intelligence to its Google tech empire, primarily leveraging its proprietary silicon chips. The company’s custom-built tensor processing units (TPUs) act as the driving force behind its Gemini chatbot, which has solidified Google’s competitive edge over adversaries like OpenAI’s ChatGPT over the past year. These TPUs are also a crucial element in Google's rapidly expanding cloud computing sector, where clients, such as the emerging AI company Anthropic, rent access to these chips. In some instances, customers can now purchase TPUs for their own data centers. Recently, Google has also embarked on an AI computing collaboration with investment giant Blackstone, focusing on the TPU technology.
Strong demand is fueling Google’s compute business, leading Wall Street to predict a remarkable 64% rise in Google Cloud revenue this year, translating to approximately $96 billion, as reported by FactSet. Analysts forecast substantial growth continuing through 2027, with projections suggesting increases exceeding 50%. With the surge in demand for AI computational power, Google's TPUs are increasingly regarded as a viable alternative to Nvidia’s dominant graphics processing units (GPUs). This positions Alphabet as a significant player in the AI infrastructure landscape, despite Google Cloud's revenue still trailing that of Amazon Web Services and Microsoft Azure. This dual position boosts both Google's internal AI initiatives and attracts external clientele, contributing to Jim Cramer’s positive outlook on the stock.
Brad Gastwirth, Circular Technology’s global head of market research, stated that Google is likely the most underrecognized contender against Nvidia. The appeal of TPUs, in essence, lies in the simple principle of maximizing computing power per dollar spent, which has become increasingly crucial for businesses striving to implement AI at scale.
In terms of AI computing, two main stages are crucial: the initial training phase, where AI models are taught using vast datasets to recognize patterns, and the inference phase, in which trained models make predictions based on new input. Training is computationally intensive, making it one of the more costly aspects of developing AI systems, while inference is less demanding but occurs constantly once a model is live, often resulting in cumulative costs that can surpass training expenses. Here, TPUs excel, as they are specialized for machine learning tasks, optimizing both training and inference efficiency.
TPUs are categorized as application-specific integrated circuits (ASICs), akin to a custom-fitted suit tailored for precise applications rather than general requirements. Co-designed by Google and Broadcom, TPUs exhibit notable energy efficiency, consuming 20% to 40% less than Nvidia processors, as noted by William Blair analyst Ralph Schackart. This efficiency allows Google to provide compute capacity at approximately 20% to 30% lower prices, thereby attracting AI startups to its cloud services.
Despite these advantages, Google’s AI computing ambitions face stiff competition, necessitating continuous innovation to remain at the forefront. Additionally, various industry-wide risks regarding component availability, from memory chips to essential materials, pose potential constraints on growth. Rising memory costs have also impacted major tech stocks recently. Moreover, concerns have emerged about the departure of skilled AI researchers from Google to competitors like OpenAI and Anthropic. Although these personnel focused on AI model development and not specifically on TPUs, seamless integration of advanced AI systems with optimized hardware is crucial for success.
Alphabet’s shares have seen a 16% decline from their early May peak amidst a general slump in the tech sector. However, year-to-date, Alphabet stocks remain about 8% higher, outpacing rivals such as Microsoft, Amazon, and Meta Platforms. Nvidia continues to dominate with its GPUs, recognized for their versatility, initially designed for enhanced 3D graphics but now utilized widely across AI applications. The company advantages include a well-established CUDA software system, vital for developers, and its presence across all major cloud platforms, enhancing its attractiveness to AI developers.
On the downside, GPUs come with high costs, significant power consumption, and limited availability due to high demand. Despite this, analysts maintain that Nvidia's substantial market share remains well-protected in the near term. Nevertheless, as AI adoption accelerates, the market is transitioning from initially training-centric approaches toward inference-centered models by late 2026.
The market saw its first wave of AI computing driven by the training phase, particularly following ChatGPT's release in late 2022. While still reliant on Nvidia chips, ChatGPT has also diversified its computational resources in recent times. Presently, the push towards inference is intensifying as consumer and enterprise adoption of AI models widens. Consequently, there is a growing emphasis on computational costs and return on investment, prompting hyperscalers to explore homegrown ASICs and alternatives, sometimes referred to as AI accelerators.
Central processing units (CPUs) have recently gained traction after initially being overlooked amid the AI computing boom. Although they fulfill different roles compared to GPUs and TPUs, CPUs are essential for managing general tasks and orchestrating AI systems. Their resurgence is driven by the increasing need for agentic AI systems that can manage diverse tasks, such as web browsing and data processing.
Many tech giants, including Nvidia and Google, have produced their own CPUs, while traditional CPU manufacturers like AMD and Intel remain noteworthy players. The rising demand for AI computing has resulted in a strained supply landscape, prompting companies with necessary resources to develop specialized chips to meet computational needs. Amazon has created custom chips like the Graviton CPU and Trainium AI accelerator, while Microsoft has also rolled out its proprietary silicon, Maia, for cloud efficiency. Meta Platforms, meanwhile, is designing MTIA processors for AI applications across its platforms, including Instagram and Facebook. OpenAI is slated to release its first proprietary chip in collaboration with Broadcom later this year.
Google has been at the forefront of TPU development since its inception in 2013. Faced with estimations of unprecedented demand for computation, Google’s leadership realized early on that existing infrastructure would not suffice. Initial TPU deployment began in 2015, and the chip quickly became pivotal within various Google services. The first-generation TPUs were initially projected to be limited in number, yet demand quickly escalated, leading to the production of over 100,000 units for diverse applications, including search algorithms, advertising, and AI initiatives such as AlphaGo.
The chip’s evolution has witnessed remarkable advancements, with the latest eighth-generation TPUs introduced at the Google Cloud Next conference, optimizing performance for both training and inference with two dedicated variants: TPU 8t for model training and TPU 8i for inference. The new designs significantly enhance speed and efficiency for demanding AI workloads, enabling the construction of expansive training clusters capable of extraordinary computational tasks.
As Google streamlines AI operations through TPUs, its cloud services are becoming more versatile, allowing for improved pricing strategies and profitability. The company reports substantial cost reductions in operational units driven by TPU efficiencies. The attractiveness of Google’s offering has caught the attention of leading AI research labs, including Anthropic, which plans to utilize Google’s TPUs extensively, and Meta Platforms, which has committed to a multi-billion-dollar agreement with Alphabet. Beyond the usual tech companies, TPUs are emerging as infrastructure solutions across various industries, including finance and energy. Recent reports highlight that Citadel Securities employs Google’s TPUs for advanced financial modeling, and all 17 U.S. Department of Energy national laboratories are utilizing an AI framework developed by Google, powered by its TPUs.
Alphabet's cloud backlog reflects strong growth momentum, nearly doubling to $472 billion driven by heightened enterprise AI adoption and TPU hardware sales expansions. This trend marks a significant shift in how customers engage with TPUs, representing an additional revenue stream for Google Cloud. Analysts predict that Google’s revenue from TPU infrastructure could reach approximately $3 billion by 2026, potentially soaring to $25 billion by 2027. The commercialization of TPUs remains underappreciated in current market valuations, suggesting significant upside potential.
Moreover, Google is positioned to benefit regardless of how clients access TPUs. By controlling its proprietary chip designs, the company maintains advantageous profit margins and maintains an economically viable model that can adapt to market demands effectively. The new joint venture with Blackstone, which involves $5 billion of initial investment aimed at building a robust cloud infrastructure, further validates Google’s commitment to advancing TPU capabilities while minimizing capital expenditure. This collaboration reflects a strategic, capital-efficient means for Google to accelerate its TPU momentum in the competitive AI landscape. As Alphabet continues to carve a path in the AI sector, its TPU development remains a focal point of its ongoing strategy.
