Google's TPU Sales Are Growing Fast: A New Era in the Chip Race

Nvidia's long-running dominance and the supply bottlenecks in the AI industry have forced tech giants to invest in their own silicon solutions. Google, which stands apart from its rivals in this area with more than a decade of R&D history, is not only optimizing its own model training processes with its Tensor Processing Unit (TPU) series; it is turning directly into an independent commercial force in the chip market.

Speaking on stage at the Communacopia conference hosted by Goldman Sachs, Google Cloud CEO Thomas Kurian shared the scale the company has reached in the TPU-focused AI infrastructure race. According to the data Kurian shared, Google has grown to twice the size of its closest rival in the market for cloud-based custom AI processors.

The Rise of Custom Silicon and a Financial Leap

In its second-quarter financial results, Google for the first time included direct TPU sales and supply figures to third-party enterprise customers in its cloud revenue. The impact of this strategic move on the balance sheet was striking:

Google Cloud's second-quarter revenue jumped to $24.77 billion, a record 82% year-over-year increase. Clearly beating Wall Street analysts' consensus expectation of $22.46 billion, this performance proves that enterprise demand for AI chips has moved beyond a theoretical investment and turned into a huge engine generating real revenue. The cloud unit's backlog rose from $460 billion in the first quarter to $514 billion in the second quarter.

3 Core Revenue Models Growing the Google TPU Ecosystem

In the past, Google used the TPU architecture only in its own internal operations and closed Google Cloud infrastructure. Today it brings the hardware to market with a multi-layered commercialization strategy:

1. Cloud-based compute rental (Cloud TPU): Companies access TPU pods in Google Cloud data centers through virtual machines and clusters to train large language models (LLMs) and lower live inference costs. 2. Direct TPU sales to customer data centers: Google takes on the role of hardware supplier by installing TPU hardware racks directly in customers' own on-premise or private data centers. Anthropic, one of the pioneers of generative AI, is among the leading customers actively choosing this method to diversify its model training infrastructure and reduce dependency. 3. Strategic infrastructure partnerships (the Blackstone model): Google's huge partnership with alternative asset management giant Blackstone makes it possible to deploy TPU systems globally, integrated into independent mega data centers.

This three-pronged approach takes Google beyond being just a "cloud service provider" and positions it as a full-fledged hardware maker competing directly with Nvidia.

The Eighth-Generation Beast: TPU v8 Ironwood

Google's biggest trump card in this market is "Ironwood", the eighth-generation TPU family announced last May. Groundbreaking in compute capacity per watt and optical circuit switching (OCS) architecture compared to the previous TPU v5p and TPU v6e generations, Ironwood offers unprecedented energy efficiency, especially in the distributed training of trillion-parameter models.

At a time when the energy crisis and power consumption ceilings in data centers are slowing AI development, the Ironwood architecture promises customers the advantage of more FLOPs on a lower energy budget.

Morgan Stanley Report: The $100 Billion Mark Will Be Passed in 2028

Financial circles are closely watching Google's move. Brian Nowak, Morgan Stanley's senior technology analyst, revised his expectations for Google's TPU operations upward in a research note.

According to Nowak's analysis:

• Google Cloud is expected to generate $27 billion in revenue in 2027 from direct hardware sales alone, excluding pure cloud rental.
• The total power capacity of TPU systems Google delivers to customers will jump from 0.3 gigawatts (GW) in the second half of 2026 to 3.2 gigawatts in 2027 and 4.2 gigawatts in 2028.
• Total Google Cloud revenue linked to the TPU ecosystem is forecast to reach $84 billion in 2027 and $108 billion in 2028.

To support this massive expansion, Google moved its 2026 capital expenditure (CapEx) guidance to a range of $195 billion to $205 billion. Capital investments reaching $44.9 billion in the second quarter alone, up 100% year over year, show that the company has no intention of shifting down a gear in the infrastructure race.

Why Is Hardware Independence Critical?

The biggest bottleneck companies face in the AI ecosystem is the high pricing margins and delivery delays that come with depending on a single chipmaker. By opening its TPUs to third-party customers, Google offers developers and AI labs a far more competitive alternative.

As organizations like Anthropic adopt Google's TPU infrastructure, it breaks the monopoly in training large models while cementing Google Cloud as one of the most profitable infrastructure players of the AI revolution.