Google Developing Chip Claimed to Cut AI Running Costs by 90 Percent

What Happened

Alphabet, the parent company of Google, is developing a new proprietary chip that the company says could reduce the cost of running artificial intelligence workloads by up to ten times compared to existing solutions, according to reports published this week. The development comes as Alphabet disclosed quarterly capital expenditure of approximately 45 billion dollars, a figure that rattled investors and sent the company's stock down roughly three percent following its most recent earnings release.

Background

Google has invested heavily in custom silicon for AI over the past decade, most notably through its Tensor Processing Unit, or TPU, line of chips designed to accelerate machine learning tasks. The company has used successive generations of TPUs to reduce reliance on third-party semiconductor suppliers, particularly Nvidia, whose graphics processing units currently dominate AI data center deployments across the industry.

The scale of Alphabet's infrastructure spending has drawn sustained scrutiny from investors. The 45 billion dollar quarterly capital expenditure figure represents a significant commitment to building out the compute capacity needed to train and serve large AI models. Alphabet's cloud division and its core search business both reported continued growth in the same period, though the magnitude of the spending outweighed those signals for some market participants.

The Chip in Development

Details about the new chip remain limited. Reports indicate the device is intended primarily to reduce inference costs, meaning the expense associated with running a trained AI model to generate responses or complete tasks, rather than the cost of training models from scratch. Inference cost is an increasingly critical metric for technology companies deploying AI at scale, as it directly affects the economics of consumer and enterprise AI products.

A tenfold reduction in inference cost, if achieved, would represent a substantial shift in the unit economics of AI deployment. Google has not provided a public timeline for the chip's release or commercial availability, and independent verification of the reported performance figures has not been confirmed.

Market Context

Alphabet's move to develop more efficient in-house silicon is consistent with a broader industry pattern. Amazon Web Services, Microsoft, and Meta have all announced or deployed custom AI chips designed to reduce dependence on Nvidia hardware and lower operational costs at their respective data centers.

The competitive pressure on inference costs has intensified in 2025 and into 2026, in part because of the emergence of lower-cost AI models from Chinese developers, including DeepSeek, whose models have demonstrated competitive performance at significantly reduced computational expense. The availability of these models in the United States has prompted debate within the technology industry about how to respond, including through hardware and software efficiency gains on the domestic side.

Alphabet's cloud revenue and search advertising performance continued to grow in the most recent quarter, providing the financial basis for sustained capital expenditure at current levels, according to the company's earnings disclosures.

What It Means in Practice

If the reported efficiency gains are validated in production environments, the chip could lower the cost of serving Google's own AI products, including the Gemini model family, as well as services offered through Google Cloud to enterprise customers. Reduced inference costs could also affect pricing strategies across the cloud AI market, where cost-per-query is a factor in customer procurement decisions.

Google has not specified which fabrication partner or process node will be used to manufacture the chip, nor has the company detailed how it compares to its existing TPU v5 and TPU v6 generations in terms of performance benchmarks.

Alphabet is expected to provide additional detail on its custom silicon roadmap at upcoming developer and cloud infrastructure events scheduled for later in 2026.