Be part of our day by day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Be taught Extra
Google Cloud unveiled its seventh-generation Tensor Processing Unit (TPU) referred to as Ironwood on Wednesday, a customized AI accelerator that the corporate claims delivers greater than 24 instances the computing energy of the world’s quickest supercomputer when deployed at scale.
The brand new chip, introduced at Google Cloud Subsequent ’25, represents a big pivot in Google’s decade-long AI chip growth technique. Whereas earlier generations of TPUs had been designed primarily for each coaching and inference workloads, Ironwood is the primary purpose-built particularly for inference — the method of deploying skilled AI fashions to make predictions or generate responses.
“Ironwood is built to support this next phase of generative AI and its tremendous computational and communication requirements,” stated Amin Vahdat, Google’s Vice President and Common Supervisor of ML, Methods, and Cloud AI, in a digital press convention forward of the occasion. “This is what we call the ‘age of inference’ where AI agents will proactively retrieve and generate data to collaboratively deliver insights and answers, not just data.”
Shattering computational obstacles: Inside Ironwood’s 42.5 exaflops of AI muscle
The technical specs of Ironwood are putting. When scaled to 9,216 chips per pod, Ironwood delivers 42.5 exaflops of computing energy — dwarfing El Capitan‘s 1.7 exaflops, currently the world’s quickest supercomputer. Every particular person Ironwood chip delivers peak compute of 4,614 teraflops.
Ironwood additionally options important reminiscence and bandwidth enhancements. Every chip comes with 192GB of Excessive Bandwidth Reminiscence (HBM), six instances greater than Trillium, Google’s previous-generation TPU introduced final 12 months. Reminiscence bandwidth reaches 7.2 terabits per second per chip, a 4.5x enchancment over Trillium.
Maybe most significantly in an period of power-constrained information facilities, Ironwood delivers twice the efficiency per watt in comparison with Trillium, and is sort of 30 instances extra energy environment friendly than Google’s first Cloud TPU from 2018.
“At a time when available power is one of the constraints for delivering AI capabilities, we deliver significantly more capacity per watt for customer workloads,” Vahdat defined.
From mannequin constructing to ‘thinking machines’: Why Google’s inference focus issues now
The emphasis on inference slightly than coaching represents a big inflection level within the AI timeline. For years, the {industry} has been fixated on constructing more and more large basis fashions, with firms competing totally on parameter dimension and coaching capabilities. Google’s pivot to inference optimization suggests we’re getting into a brand new part the place deployment effectivity and reasoning capabilities take middle stage.
This transition is sensible. Coaching occurs as soon as, however inference operations happen billions of instances day by day as customers work together with AI techniques. The economics of AI are more and more tied to inference prices, particularly as fashions develop extra advanced and computationally intensive.
In the course of the press convention, Vahdat revealed that Google has noticed a 10x year-over-year enhance in demand for AI compute over the previous eight years — a staggering issue of 100 million general. No quantity of Moore’s Regulation development might fulfill this progress curve with out specialised architectures like Ironwood.
What’s notably notable is the concentrate on “thinking models” that carry out advanced reasoning duties slightly than easy sample recognition. This implies Google sees the way forward for AI not simply in bigger fashions, however in fashions that may break down issues, purpose by means of a number of steps, and basically simulate human-like thought processes.
Gemini’s pondering engine: How Google’s next-gen fashions leverage superior {hardware}
Google is positioning Ironwood as the inspiration for its most superior AI fashions, together with Gemini 2.5, which the corporate describes as having “thinking capabilities natively built in.”
On the convention, Google additionally introduced Gemini 2.5 Flash, a less expensive model of its flagship mannequin that “adjusts the depth of reasoning based on a prompt’s complexity.” Whereas Gemini 2.5 Professional is designed for advanced use circumstances like drug discovery and monetary modeling, Gemini 2.5 Flash is positioned for on a regular basis purposes the place responsiveness is crucial.
The corporate additionally demonstrated its full suite of generative media fashions, together with text-to-image, text-to-video, and a newly introduced text-to-music functionality referred to as Lyria. An illustration confirmed how these instruments may very well be used collectively to create an entire promotional video for a live performance.
Past silicon: Google’s complete infrastructure technique consists of community and software program
Ironwood is only one a part of Google’s broader AI infrastructure technique. The corporate additionally introduced Cloud WAN, a managed wide-area community service that provides companies entry to Google’s planet-scale non-public community infrastructure.
“Cloud WAN is a fully managed, viable and secure enterprise networking backbone that provides up to 40% improved network performance, while also reducing total cost of ownership by that same 40%,” Vahdat stated.
Google can also be increasing its software program choices for AI workloads, together with Pathways, its machine studying runtime developed by Google DeepMind. Pathways on Google Cloud permits clients to scale out mannequin serving throughout tons of of TPUs.
AI economics: How Google’s $12 billion cloud enterprise plans to win the effectivity conflict
These {hardware} and software program bulletins come at an important time for Google Cloud, which reported $12 billion in This fall 2024 income, up 30% 12 months over 12 months, in its newest earnings report.
The economics of AI deployment are more and more turning into a differentiating issue within the cloud wars. Google faces intense competitors from Microsoft Azure, which has leveraged its OpenAI partnership right into a formidable market place, and Amazon Net Providers, which continues to develop its Trainium and Inferentia chip choices.
What separates Google’s strategy is its vertical integration. Whereas rivals have partnerships with chip producers or acquired startups, Google has been growing TPUs in-house for over a decade. This provides the corporate unparalleled management over its AI stack, from silicon to software program to providers.
By bringing this expertise to enterprise clients, Google is betting that its hard-won expertise constructing chips for Search, Gmail, and YouTube will translate into aggressive benefits within the enterprise market. The technique is evident: provide the identical infrastructure that powers Google’s personal AI, at scale, to anybody keen to pay for it.
The multi-agent ecosystem: Google’s audacious plan for AI techniques that work collectively
Past {hardware}, Google outlined a imaginative and prescient for AI centered round multi-agent techniques. The corporate introduced an Agent Growth Equipment (ADK) that permits builders to construct techniques the place a number of AI brokers can work collectively.
Maybe most importantly, Google introduced an “agent-to-agent interoperability protocol” (A2A) that permits AI brokers constructed on completely different frameworks and by completely different distributors to speak with one another.
“2025 will be a transition year where generative AI shifts from answering single questions to solving complex problems through agented systems,” Vahdat predicted.
Google is partnering with greater than 50 {industry} leaders, together with Salesforce, ServiceNow, and SAP, to advance this interoperability commonplace.
Enterprise actuality verify: What Ironwood’s energy and effectivity imply to your AI technique
For enterprises deploying AI, these bulletins might considerably cut back the fee and complexity of operating subtle AI fashions. Ironwood’s improved effectivity might make operating superior reasoning fashions extra economical, whereas the agent interoperability protocol might assist companies keep away from vendor lock-in.
The actual-world impression of those developments shouldn’t be underestimated. Many organizations have been reluctant to deploy superior AI fashions resulting from prohibitive infrastructure prices and power consumption. If Google can ship on its performance-per-watt guarantees, we might see a brand new wave of AI adoption in industries which have to this point remained on the sidelines.
The multi-agent strategy is equally important for enterprises overwhelmed by the complexity of deploying AI throughout completely different techniques and distributors. By standardizing how AI techniques talk, Google is trying to interrupt down the silos which have restricted AI’s enterprise impression.
In the course of the press convention, Google emphasised that over 400 buyer tales could be shared at Subsequent ’25, showcasing actual enterprise impression from its AI improvements.
The silicon arms race: Will Google’s customized chips and open requirements reshape AI’s future?
As AI continues to advance, the infrastructure powering it’s going to turn out to be more and more crucial. Google’s investments in specialised {hardware} like Ironwood, mixed with its agent interoperability initiatives, recommend the corporate is positioning itself for a future the place AI turns into extra distributed, extra advanced, and extra deeply built-in into enterprise operations.
“Leading thinking models like Gemini 2.5 and the Nobel Prize winning AlphaFold all run on TPUs today,” Vahdat famous. “With Ironwood we can’t wait to see what AI breakthroughs are sparked by our own developers and Google Cloud customers when it becomes available later this year.”
The strategic implications lengthen past Google’s personal enterprise. By pushing for open requirements in agent communication whereas sustaining proprietary benefits in {hardware}, Google is trying a fragile balancing act. The corporate needs the broader ecosystem to flourish (with Google infrastructure beneath), whereas nonetheless sustaining aggressive differentiation.
How shortly opponents reply to Google’s {hardware} developments and whether or not the {industry} coalesces across the proposed agent interoperability requirements can be key elements to look at within the months forward. If historical past is any information, we will anticipate Microsoft and Amazon to counter with their very own inference optimization methods, probably organising a three-way race to construct essentially the most environment friendly AI infrastructure stack.
