Google Expands Custom TPU Program as Ironwood Chip Enters Wide Release

Google is moving its in-house semiconductor effort into a new phase, introducing the seventh generation of its Tensor Processing Unit and signaling that custom silicon has become a central pillar of the company’s artificial-intelligence infrastructure strategy.

The latest processor, called Ironwood, will be offered to Google Cloud customers “in the coming weeks,” according to the company. Designed for both training and serving large AI models, Ironwood is rated at more than four times the performance of the previous generation TPU v5e and is expected to support applications ranging from generative chat systems to real-time agents.

Years of development reach commercial scale

Google began designing TPUs more than a decade ago to accelerate internal machine-learning workloads, releasing the first version to outside developers in 2018. The technology has since matured into one of the few production-scale alternatives to Nvidia’s graphics processing units, which dominate the market for AI computation.

Industry analysts estimate that Google is the only hyperscale provider routinely deploying application-specific integrated circuits in volumes comparable to GPU installations. Stacy Rasgon of Bernstein, who tracks the semiconductor sector, notes that creating and manufacturing a proprietary accelerator requires multi-year commitments of capital, engineering talent and fabrication capacity—barriers that competitors are still working to overcome.

Anthropic and other customers line up

Demand for Ironwood emerged quickly. AI start-up Anthropic intends to operate as many as one million of the new chips to run future versions of its Claude model. The two companies last month expanded an existing partnership in a transaction valued in the tens of billions of dollars, an agreement expected to add well over one gigawatt of compute capacity to Google’s data-center footprint in 2026.

Google’s cloud division has also signed a six-year infrastructure pact with Meta Platforms and secured a first-time contract with OpenAI, diversifying a customer base that historically relied on Google’s internal product groups. While the Meta arrangement did not specify how many TPUs are involved, corporate filings indicate that Alphabet’s business backlog reached $155 billion at the end of the third quarter, reflecting large, multi-year cloud commitments.

Balancing GPUs and custom silicon

Although TPUs are central to Google’s AI roadmap, the company continues to purchase substantial quantities of Nvidia GPUs to meet surging demand. A spokesperson described the procurement approach as “choice and synergy,” emphasizing that both architectures will coexist in Google data centers for the foreseeable future.

This blended strategy is mirrored by cloud customers. Anthropic, for example, plans to divide workloads across TPUs, Amazon’s Trainium chips and Nvidia’s H100 GPUs, optimizing cost, performance and redundancy. Executives at the start-up said preparatory engineering allowed its models to move interchangeably among the three platforms, enabling faster ramp-up of new capacity.

Cost and energy efficiency emerge as priorities

Analysts highlight efficiency as a key competitive factor. James Sanders of Tech Insights says Google can tailor TPUs to anticipated workloads, squeezing more compute per watt than general-purpose hardware. With data-center construction outpacing available electrical power in many regions, energy consumption is becoming as critical as chip supply.

Imagem: Internet

Google is exploring alternative approaches to power as well. Earlier this week the company unveiled Project Suncatcher, an initiative to test solar-powered satellites equipped with TPUs. Two prototypes are scheduled for launch by early 2027, a step Google says could open the door to large-scale orbital computing while reducing pressure on terrestrial resources.

Independent assessments underscore the cost advantage. Research published by brokerage firm Mizuho found that TPUs deliver competitive performance at lower operating expense, a view echoed by Morgan Stanley, which projected that growing developer familiarity with the architecture could accelerate Google Cloud revenue. A September analysis by D.A. Davidson estimated that a hypothetical stand-alone unit combining TPUs and Google DeepMind might be worth about $900 billion.

Capital spending rises to meet demand

Alphabet is backing its chip strategy with record investment. The company recently lifted the upper end of its 2025 capital-expenditure forecast to $93 billion, up from $85 billion. Management told investors that further increases are likely next year as new data-center campuses and manufacturing commitments come online.

Shareholders have, so far, accepted the spending spree. Alphabet’s stock advanced 38 percent in the third quarter—its strongest quarterly gain in two decades—and is up an additional double-digit percentage in the current period.

Google is not alone in pushing custom silicon, but competitors remain behind on the deployment curve. Amazon Web Services introduced its first inference chip, Inferentia, in 2019 and followed with Trainium in 2022. Microsoft unveiled its initial AI accelerator, code-named Maia, late last year and is still in early rollout. Meanwhile, Nvidia’s position as the industry standard persists, a dynamic explored in depth by the MIT Technology Review.

Whether Google eventually offers TPUs as discrete hardware remains an open question. Some analysts argue that selling systems directly to large research laboratories could unlock additional revenue streams. For now, the company appears committed to its service model, betting that vertical integration of chips, software and cloud capacity will keep its platform attractive as the race for AI compute enters its next stage.

Crédito da imagem: Sopa Images | Lightrocket | Getty Images