Google has spent years building its own AI chips for internal use and renting access to them through its cloud platform. Now, for the first time, it is selling those chips directly to outside customers who will install them in their own data centers.
The announcement, made during Alphabet’s first-quarter 2026 earnings call on April 29, marks a fundamental shift in how Google plans to compete in the AI hardware market. More importantly, it puts the company more squarely in the business that has made Nvidia one of the most valuable companies in the world.
What Google Announced
Alphabet CEO Sundar Pichai confirmed on the earnings call that Google will begin selling its custom Tensor Processing Units (TPUs) to a select group of customers, who will install the chips directly in their own data centers. This is a departure from Google’s previous model, where companies could only access TPUs by renting compute capacity from Google’s own cloud infrastructure.
CFO Anat Ashkenazi said that some TPU sales revenue will be recorded this year but a more significant financial impact on Alphabet’s balance sheet is not expected until 2027. She also noted that revenue from TPU hardware sales will vary by quarter depending on when shipments are made.
Pichai acknowledged that directly selling hardware chips is a departure from the typical cloud model, which carries a higher long-term return on investment because Google can continue charging customers for access to the chips throughout their lifespan. The company is clearly trading some of that recurring revenue for the opportunity to grow its total addressable market.
Who the Customers Are
Pichai pointed to AI research labs, capital markets firms, and high-performance computing operators as the primary customer groups driving this demand, specifically citing capital markets clients that want TPUs running directly inside their own data centers.
Meta has reportedly been in discussions with Google to spend billions of dollars integrating TPUs into its own data centers starting in 2027, while also planning to rent TPU capacity from Google Cloud starting next year. A deal of that scale, if it goes through, would represent one of the most consequential chip agreements between two major tech companies in years.
Google has also signed a large TPU capacity agreement with Anthropic, the AI safety and research company, with chips expected to begin coming online in 2027.
The Chips Behind the Push
Google unveiled its eighth-generation TPUs at Google Cloud Next, with two distinct variants – the TPU 8t, built for AI model training; and the TPU 8i, designed for inference workloads. These are the chips that will form the foundation of the new direct-sales strategy.
Pichai said that external chip sales would help secure funding needed for next-generation semiconductor research and enable economies of scale, a clear signal that Google sees hardware sales not just as a revenue line, but as a way to fund the next wave of chip development.
What This Means for Nvidia
Nvidia currently controls the overwhelming majority of the AI chip market, with its data center GPUs serving as the default infrastructure for training and running AI models across most of the industry. Google’s move does not challenge that dominance overnight, but it does something meaningful – it gives large enterprises and AI labs a credible alternative that they can physically own and operate.
Executives inside Google Cloud have suggested that expanding TPU adoption could help the company capture as much as 10% of Nvidia’s annual revenue.
Google Cloud revenue reached $20 billion in the first quarter of 2026, a 63% increase year-over-year, with CFO Ashkenazi confirming that the largest contributor to that growth was AI solutions driven by strong demand for Google’s models and infrastructure.
Goldman Sachs raised its 2027 estimate for the custom AI silicon market to $240 billion after the TPU 8t and 8i announcement, up from $195 billion, which analysts described as a $45 billion shift in capital expenditure moving away from Nvidia architectures.
Why the Timing Matters
The AI compute market is entering a phase where inference – running AI models in real time – is becoming just as critical as training them. Companies running inference at scale are increasingly price-sensitive and looking for hardware that offers better cost-per-computation ratios. Google’s chips, now available for direct ownership rather than cloud rental, give those companies more control over their infrastructure costs.
Alphabet has positioned itself as both an AI infrastructure provider and a hyperscaler customer, sitting on a cloud backlog of $460 billion, a number that underscores how much enterprise demand for AI infrastructure is building up and how much of it Google now wants to capture through direct hardware deals and not just cloud subscriptions.
The competitive map in AI infrastructure has been relatively stable for the past two years, with Nvidia at the top and everyone else renting its chips or building alternatives that haven’t reached meaningful scale. As such, Google’s decision to sell its own chips to paying customers, starting now, changes the shape of that map.
