At its annual Cloud Next gathering in Las Vegas, Google signaled a strategic maturation in the AI race — not merely competing on model intelligence, but asserting dominance across the entire technological stack, from custom silicon to enterprise applications. The unveiling of Ironwood TPUs, purpose-built for inference at planetary scale, reflects a broader philosophical shift: the industry's obsession with training ever-larger models is giving way to the harder, quieter work of deploying them efficiently. By proposing open interoperability protocols alongside proprietary infrastructure, Google
Google Unveils Ironwood TPUs and Agent Ecosystem to Challenge Microsoft, Amazon in AI
machines that can effectively talk to each other
Mark
Why does it matter that Ironwood was built specifically for inference rather than training? Aren't both important?
Mimi
Training is the expensive, one-time event. Inference is what happens every single day, millions of times, when a customer actually uses the model. Google is saying: we've optimized for the real cost center. That's where the energy bill lives.
Mark
And this Agent2Agent protocol—is that just Google trying to lock in customers by making them dependent on Google's platform?
Mimi
The opposite, actually. By proposing an open standard that lets agents from different vendors talk to each other, Google is saying: build with us, but don't be trapped by us. It's a bet that interoperability is more valuable than lock-in.
Mark
Fifty partners signed on. Is that a lot?
Mimi
For a protocol that was just announced? Yes. And these aren't small companies—Salesforce, ServiceNow, SAP. These are the systems enterprises already depend on. If agents can talk across those platforms, that changes how work gets done.
Mark
The Wendy's drive-through handling 60,000 orders a day—is that replacing people?
Mimi
It's handling the orders. People are still involved in fulfillment. But it's a real example of where AI is moving from "interesting experiment" to "running actual business operations."
Mark
What's the risk for Google here?
Mimi
Execution. They're announcing a lot of pieces that have to work together seamlessly. And they're betting that enterprises care about interoperability more than they care about staying within a single vendor's ecosystem. That's not guaranteed.
Il Polso
- Google arrived at Cloud Next not as a company catching up, but as one attempting to redefine the terms of the AI competition entirely.
- Ironwood TPUs — delivering 42 exaflops across pods of 9,000+ chips — represent a direct challenge to Microsoft and Amazon, promising 24 times the power of the world's leading supercomputer with twice the energy efficiency.
- The Agent2Agent open protocol threatens to dissolve vendor lock-in across the industry, with 50+ partners already committed — a move that could reshape how enterprises build and connect AI systems.
- Google is threading AI into the mundane fabric of daily enterprise work: Workspace tools that surface insights without formulas, no-code agent builders, and real-world deployments like Wendy's 60,000 daily AI-handled orders.
- The full-stack strategy — chips, networking, models, agent frameworks, and applications under one roof — is landing as a coherent answer to what critics have called the fragmented AI offerings of its rivals.
At its annual Cloud Next gathering in Las Vegas, Google signaled a strategic maturation in the AI race — not merely competing on model intelligence, but asserting dominance across the entire technological stack, from custom silicon to enterprise applications. The unveiling of Ironwood TPUs, purpose-built for inference at planetary scale, reflects a broader philosophical shift: the industry's obsession with training ever-larger models is giving way to the harder, quieter work of deploying them efficiently. By proposing open interoperability protocols alongside proprietary infrastructure, Google is wagering that the next era of enterprise AI belongs not to the most powerful single machine, but to whoever becomes the connective tissue between many.
Google used its Cloud Next '25 conference in Las Vegas to deliver a pointed message: it is no longer chasing competitors in artificial intelligence, but attempting to set the terms of the race itself. The company's strategy centers on controlling every layer of the AI stack, from custom hardware to the applications enterprises rely on daily.
At the hardware level, Google introduced Ironwood, its seventh-generation Tensor Processing Unit. Each pod houses more than 9,000 chips and produces 42 exaflops of computing power — 24 times more than El Capitan, currently the world's most powerful supercomputer. Crucially, Ironwood was designed not for training models but for running them at scale, achieving twice the power efficiency of its predecessors. As Google's VP of ML systems Amin Vahdat put it, the focus has shifted from what goes into a model to what the model can do once trained. Google also opened its 2-million-mile fiber network — the backbone of YouTube and Gmail — to enterprise customers through Cloud WAN, promising 40 percent performance gains and equivalent cost reductions.
On the software side, Google introduced Gemini 2.5 Flash, a cost-efficient model with multi-step reasoning capabilities that adjusts its thinking depth to the complexity of each task. Its counterpart, Gemini 2.5 Pro, targets demanding domains like drug discovery and financial modeling. Google also expanded its generative media suite, unifying image, video, audio, and music generation tools on a single platform.
Perhaps the most consequential announcement was Agent2Agent, an open protocol enabling AI agents from different vendors to communicate with one another. With over 50 partners already signed on — including Salesforce, SAP, and ServiceNow — the protocol signals Google's ambition to become the interoperability layer of the enterprise AI world rather than a closed ecosystem. Developers can now build multi-agent systems in under 100 lines of code, while non-technical users gained a no-code Agent Designer tool.
Real-world deployments illustrated the stakes: Wendy's AI system processes 60,000 drive-through orders daily, The Home Depot's agent guides customers through home projects, and WPP uses Google's tools to produce marketing campaigns at scale. As Microsoft deepens its OpenAI integration and Amazon builds out Anthropic-powered offerings, Google is betting that the decisive advantage in enterprise AI will belong to whoever can most efficiently orchestrate many specialized agents working in concert — and whoever builds the open roads between them.
Google walked into its annual Cloud Next conference in Las Vegas with a message: the company is no longer playing catch-up in artificial intelligence. Instead, it's laying down a full-stack platform designed to outmaneuver Microsoft and Amazon by controlling everything from the silicon up to the applications that enterprises actually use.
The centerpiece is Ironwood, Google's seventh-generation Tensor Processing Unit. Each pod contains more than 9,000 chips and delivers 42 exaflops of computing power—a figure the company claims is 24 times more powerful than El Capitan, the world's leading supercomputer. But the real innovation isn't raw speed. Ironwood was built for a specific job: running trained AI models at scale, not training them from scratch. The chips achieve twice the power efficiency of their predecessors, which matters enormously when you're talking about the energy costs of deploying generative AI across an enterprise. "It's no longer about the data put into the model, but what the model can do with data after it's been trained," said Amin Vahdat, Google's vice president and general manager of ML systems and cloud AI.
Google is also opening its internal infrastructure to customers. Cloud WAN gives enterprises access to Google's 2-million-mile fiber network—the same backbone that powers YouTube and Gmail. The company claims this improves network performance by 40 percent while cutting total cost of ownership by the same margin. It's an unusual move for a hyperscaler, essentially commodifying what was once a competitive advantage.
On the software side, Google introduced Gemini 2.5 Flash, a cost-effective variant of its flagship AI model that includes what the company calls "thinking capabilities." Unlike traditional language models that generate answers directly, these systems break down complex problems through multi-step reasoning and self-reflection. Gemini 2.5 Pro, launched two weeks earlier, targets high-complexity work like drug discovery and financial modeling. The Flash version adjusts its reasoning depth based on how hard the problem is, balancing performance against cost. Google also expanded its generative media tools—Imagen for images, Veo for video, Chirp for audio, and a new text-to-music model called Lyria—all designed to work together on a single platform.
But perhaps the most forward-looking announcement concerns what Google calls a "multi-agent ecosystem." The company introduced an Agent Development Kit that lets developers build multi-agent systems in fewer than 100 lines of code. More significantly, Google proposed Agent2Agent, an open protocol that would allow AI agents from different vendors to communicate with each other. More than 50 partners have already signed on, including Salesforce, ServiceNow, and SAP. For non-technical users, Google enhanced its Agent Space platform with Agent Gallery and Agent Designer, a no-code interface for building custom agents. In a demonstration, the company showed how a banking account manager could analyze client portfolios, forecast cash flow problems, and draft client communications—all without writing code.
Google is also weaving AI throughout its Workspace productivity suite. New features in Sheets automatically identify data insights without requiring formulas or pivot tables. Docs now offers Audio Overviews that convert documents into human-sounding audio. The company highlighted five categories where specialized agents are gaining traction: customer service, creative work, data analysis, coding, and security. Wendy's AI drive-through system now handles 60,000 orders daily. The Home Depot's "Magic Apron" agent guides customers through home improvement projects. WPP uses Google's AI to conceptualize and produce marketing campaigns at scale.
These announcements arrive as competition in cloud AI intensifies. Microsoft has deeply integrated OpenAI's technology into Azure. Amazon has been building out Anthropic-powered offerings and its own specialized chips. Google's strategy appears designed to differentiate itself by offering the full stack—from custom silicon to global networking to models to agent frameworks to application integration. Thomas Kurian, CEO of Google Cloud, emphasized the company's commitment to "delivering world-class infrastructure, models, platforms, and agents; offering an open, multi-cloud platform that provides flexibility and choice; and building for interoperability."
What's significant about Google's approach is its maturity. The industry has spent years obsessing over training ever-larger models. Google is now betting that the real competitive advantage lies in deploying those models efficiently at scale and, more importantly, in orchestrating multiple specialized AI agents working in concert. By proposing open protocols like Agent2Agent, Google is positioning itself not as a walled garden but as the connective tissue in a heterogeneous AI ecosystem. The company is betting that the future of enterprise AI isn't just about smarter machines, but about machines that can effectively talk to each other.
Citazioni salienti
It's no longer about the data put into the model, but what the model can do with data after it's been trained.— Amin Vahdat, VP and GM of ML systems and cloud AI at Google
2025 will be a transition year where generative AI shifts from answering single questions to solving complex problems through agented systems.— Amin Vahdat
Vuoi la storia completa?
Leggi l'originale su VentureBeat ↗
Technology