From ‘catch up’ to ‘catch us’: How Google quietly took the lead in enterprise AI

Only a yr in the past, the narrative round Google and enterprise AI felt caught. Regardless of inventing core applied sciences just like the Transformer, the tech big appeared perpetually on the again foot, overshadowed by OpenAI‘s viral success, Anthropic‘s coding prowess and Microsoft‘s aggressive enterprise push.

However witness the scene at Google Cloud Subsequent 2025 in Las Vegas final week: A assured Google, armed with benchmark-topping fashions, formidable infrastructure and a cohesive enterprise technique, declaring a surprising turnaround. In a closed-door analyst assembly with senior Google executives, one analyst summed it up. This feels just like the second, he mentioned, when Google went from “catch up, to catch us.” 

This sentiment that Google has not solely caught up with however even surged forward of OpenAI and Microsoft within the enterprise AI race prevailed all through the occasion. And it’s extra than simply Google’s advertising and marketing spin. Proof suggests Google has leveraged the previous yr for intense, centered execution, translating its technological property right into a performant, built-in platform that’s quickly profitable over enterprise decision-makers. From boasting the world’s strongest AI fashions operating on hyper-efficient customized silicon, to a burgeoning ecosystem of AI brokers designed for real-world enterprise issues, Google is making a compelling case that it was by no means truly misplaced – however that its stumbles masked a interval of deep, foundational improvement. 

Now, with its built-in stack firing on all cylinders, Google seems positioned to guide the following section of the enterprise AI revolution. And in my interviews with a number of Google executives at Subsequent, they mentioned Google wields benefits in infrastructure and mannequin integration that rivals like OpenAI, Microsoft or AWS will battle to duplicate.

The shadow of doubt: acknowledging the current previous

It’s inconceivable to understand the present momentum with out acknowledging the current previous. Google was the birthplace of the Transformer structure, which sparked the trendy revolution in giant language fashions (LLMs). Google additionally began investing in specialised AI {hardware} (TPUs), which are actually driving industry-leading effectivity, a decade in the past. And but, two and a half years in the past, it inexplicably discovered itself enjoying protection. 

OpenAI’s ChatGPT captured the general public creativeness and enterprise curiosity at breathtaking velocity and have become the fastest-growing app in historical past. Rivals like Anthropic carved out niches in areas like coding.

Google’s personal public steps typically appeared tentative or flawed. The notorious Bard demo fumbles in 2023 and the later controversy over its picture generator producing traditionally inaccurate depictions fed a story of an organization probably hampered by inner forms or overcorrection on alignment. It felt like Google was misplaced: The AI stumbles appeared to suit a sample, first proven by Google’s preliminary slowness within the cloud competitors, the place it remained a distant third in market share behind Amazon and Microsoft. Google Cloud CTO Will Grannis acknowledged the early questions on whether or not Google Cloud would stand behind in the long term. “Is it even a real thing?,” he recalled individuals asking him. The query lingered: Might Google translate its plain analysis brilliance and infrastructure scale into enterprise AI dominance?

The pivot: a aware resolution to guide

Behind the scenes, nevertheless, a shift was underway, catalyzed by a aware resolution on the highest ranges to reclaim management. Mat Velloso, VP of product for Google DeepMind’s AI Developer Platform, described sensing a pivotal second upon becoming a member of Google in Feb. 2024, after leaving Microsoft. “When I came to Google, I spoke with Sundar [Pichai], I spoke with several leaders here, and I felt like that was the moment where they were deciding, okay, this [generative AI] is a thing the industry clearly cares about. Let’s make it happen,” Velloso shared in an interview with VentureBeat throughout Subsequent final week.

This renewed push wasn’t hampered by a feared “brain drain” that some outsiders felt was depleting Google. In reality, the corporate quietly doubled down on execution in early 2024 – a yr marked by aggressive hiring, inner unification and buyer traction. Whereas rivals made splashy hires, Google retained its core AI management, together with DeepMind CEO Demis Hassabis and Google Cloud CEO Thomas Kurian, offering stability and deep experience.

Furthermore, expertise started flowing in direction of Google’s centered mission. Logan Kilpatrick, as an example, returned to Google from OpenAI, drawn by the chance to construct foundational AI throughout the firm, creating it. He joined Velloso in what he described as a “zero to one experience,” tasked with constructing developer traction for Gemini from the bottom up. “It was like the team was me on day one… we actually have no users on this platform, we have no revenue. No one is interested in Gemini at this moment,” Kilpatrick recalled of the start line. Folks acquainted with the inner dynamics additionally credit score leaders like Josh Woodward, who helped begin AI Studio and now leads the Gemini App and Labs. Extra just lately, Noam Shazeer, a key co-author of the unique “Attention Is All You Need” Transformer paper throughout his first tenure at Google, returned to the corporate in late 2024 as a technical co-lead for the essential Gemini undertaking

This concerted effort, combining these hires, analysis breakthroughs, refinements to its database expertise and a sharpened enterprise focus general, started yielding outcomes. These cumulative advances, mixed with what CTO Will Grannis termed “hundreds of fine-grain” platform parts, set the stage for the bulletins at Subsequent ’25, and cemented Google’s comeback narrative.

Pillar 1: Gemini 2.5 and the period of pondering fashions

It’s true {that a} main enterprise mantra has develop into “it’s not just about the model.” In any case, the efficiency hole between main fashions has narrowed dramatically, and tech insiders acknowledge that true intelligence is coming from expertise packaged across the mannequin, not simply the mannequin itself – for instance, agentic applied sciences that permit a mannequin to make use of instruments and discover the online round it.

Regardless of this, to own the demonstrably best-performing LLM is a vital feat – and a strong validator, an indication that the model-owning firm has issues like superior analysis and probably the most environment friendly underlying expertise structure. With the discharge of Gemini 2.5 Professional simply weeks earlier than Subsequent ’25, Google definitively seized that mantle. It rapidly topped the unbiased Chatbot Enviornment leaderboard, considerably outperforming even OpenAI’s newest GPT-4o variant, and aced notoriously tough reasoning benchmarks like Humanity’s Final Examination. As Pichai said within the keynote, “It’s our most intelligent AI model ever. And it is the best model in the world.” The mannequin had pushed an 80 % enhance in Gemini utilization inside a month, he Tweeted individually. 

For the primary time, Google’s Gemini demand was on hearth. As I detailed beforehand, except for Gemini 2.5 Professional’s uncooked intelligence, what impressed me is its demonstrable reasoning. Google has engineered a “thinking” functionality, permitting the mannequin to carry out multi-step reasoning, planning, and even self-reflection earlier than finalizing a response. The structured, coherent chain-of-thought (CoT) – utilizing numbered steps and sub-bullets – avoids the rambling or opaque nature of outputs from different fashions from DeepSeek or OpenAI. For technical groups evaluating outputs for crucial duties, this transparency permits validation, correction, and redirection with unprecedented confidence.

However extra importantly for enterprise customers, Gemini 2.5 Professional additionally dramatically closed the hole in coding, which is without doubt one of the largest software areas for generative AI. In an interview with VentureBeat, CTO Fiona Tan, the CTO of main retailer Wayfair, mentioned that after preliminary checks, the corporate discovered it “stepped up quite a bit” and was now “pretty comparable” to Anthropic’s Claude 3.7 Sonnet, beforehand the popular selection for a lot of builders. 

Google additionally added a large 1 million token context window to the mannequin, enabling reasoning throughout total codebases or prolonged documentation, far exceeding the capabilities of the fashions of OpenAI or Anthropic. (OpenAI responded this week with fashions that includes equally giant context home windows, although benchmarks counsel Gemini 2.5 Professional retains an edge in general reasoning). This benefit permits for advanced, multi-file software program engineering duties.

Complementing Professional is Gemini 2.5 Flash, introduced at Subsequent ’25 and launched simply yesterday. Additionally, a “thinking” mannequin, Flash is optimized for low latency and cost-efficiency. You’ll be able to management how a lot the mannequin causes and steadiness efficiency together with your finances. This tiered method additional displays the “intelligence per dollar” technique championed by Google executives.

Velloso confirmed a chart revealing that throughout the intelligence spectrum, Google fashions provide one of the best worth. “If we had this conversation one year ago… I would have nothing to show,” Velloso admitted, highlighting the speedy turnaround. “And now, like, across the board, we are, if you’re looking for whatever model, whatever size, like, if you’re not Google, you’re losing money.” These charts have been up to date to account for OpenAI’s newest mannequin releases this week. See beneath:

For any given value, Google’s fashions provide extra intelligence than 90 % of the time. Supply: Pierre Bongrand.

Wayfair’s Tan mentioned she additionally noticed promising latency enhancements with 2.5 Professional: “Gemini 2.5 came back faster,” making it viable for “more customer-facing sort of capabilities,” she mentioned, one thing she mentioned hasn’t been the case earlier than with different fashions. Gemini may develop into the primary mannequin Wayfair makes use of for these buyer interactions, she mentioned.

The Gemini household’s capabilities lengthen to multimodality, integrating seamlessly with Google’s different main fashions like Imagen 3 (picture technology), Veo 2 (video technology), Chirp 3 (audio), and the newly introduced Lyria (text-to-music), all accessible by way of Google’s platform for Enterprise customers, Vertex. Google is the one firm that provides its personal generative media fashions throughout all modalities on its platform. Microsoft, AWS and OpenAI should associate with different firms to do that.

Pillar 2: Infrastructure prowess – the engine below the hood

The flexibility to quickly iterate and effectively serve these highly effective fashions stems from Google’s arguably unparalleled infrastructure, honed over many years of operating planet-scale providers. Central to that is the Tensor Processing Unit (TPU).

At Subsequent ’25, Google unveiled Ironwood, its seventh-generation TPU, explicitly designed for the calls for of inference and “thinking models.” The size is immense, tailor-made for demanding AI workloads: Ironwood pods pack over 9,000 liquid-cooled chips, delivering a claimed 42.5 exaflops of compute energy. Google’s VP of ML Methods Amin Vahdat mentioned on stage at Subsequent that that is “more than 24 times” the compute energy of the world’s present #1 supercomputer. 

Google said that Ironwood presents 2x perf/watt relative to Trillium, the earlier technology of TPU. That is vital since enterprise clients more and more say vitality prices and availability constrain large-scale AI deployments.

Google Cloud CTO Will Grannis emphasised the consistency of this progress. 12 months over yr, Google is making 10x, 8x, 9x, 10x enhancements in its processors, he informed VentureBeat in an interview, creating what he referred to as a “hyper Moore’s law” for AI accelerators. He mentioned clients are shopping for Google’s roadmap, not simply its expertise. 

Google’s place fueled this sustained TPU funding. It must effectively energy large providers like Search, YouTube, and Gmail for greater than 2 billion customers. This necessitated creating customized, optimized {hardware} lengthy earlier than the present generative AI growth. Whereas Meta operates at an analogous shopper scale, different rivals lacked this particular inner driver for decade-long, vertically built-in AI {hardware} improvement.

Now these TPU investments are paying off as a result of they’re driving the effectivity not just for its personal apps, however in addition they permit Google to supply Gemini to different customers at a greater intelligence per greenback, every little thing equal.

Why can’t Google’s rivals purchase environment friendly processors from Nvidia, you ask? It’s true that Nvidia’s GPU processors dominate the method pre-training of LLMs. However market demand has pushed up the value of those GPUs, and Nvidia takes a wholesome lower for itself as revenue. This passes vital prices alongside to customers of its chips. And in addition, whereas pre-training has dominated the utilization of AI chips thus far, that is altering now that enterprises are literally deploying these purposes. That is the place ” inference” is available in, and right here TPUs are thought-about extra environment friendly than GPUs for workloads at scale. 

Once you ask Google executives the place their principal expertise benefit in AI comes from, they normally fall again to the TPU as crucial. Mark Lohmeyer, the VP who runs Google’s computing infrastructure, was unequivocal: TPUs are “certainly a highly differentiated part of what we do… OpenAI, they don’t have those capabilities.”

Considerably, Google presents TPUs not in isolation, however as a part of the broader, extra advanced enterprise AI structure. For technical insiders, it’s understood that top-tier efficiency hinges on integrating more and more specialised expertise breakthroughs. Many updates have been detailed at Subsequent. Vahdat described this as a “supercomputing system,” integrating {hardware} (TPUs, the newest Nvidia GPUs like Blackwell and upcoming Vera Rubin, superior storage like Hyperdisk Exapools, Wherever Cache, and Fast Storage) with a unified software program stack. This software program contains Cluster Director for managing accelerators, Pathways (Gemini’s distributed runtime, now out there to clients), and bringing optimizations like vLLM to TPUs, permitting simpler workload migration for these beforehand on Nvidia/PyTorch stacks. This built-in system, Vahdat argued, is why Gemini 2.0 Flash achieves 24 instances greater intelligence per greenback, in comparison with GPT-4o.

Google can also be extending its bodily infrastructure attain. Cloud WAN makes Google’s low-latency 2-million-mile non-public fiber community out there to enterprises, promising as much as 40% sooner efficiency and 40% decrease complete value of possession (TCO) in comparison with customer-managed networks. 

Moreover, Google Distributed Cloud (GDC) permits Gemini and Nvidia {hardware} (by way of a Dell partnership) to run in sovereign, on-premises, and even air-gapped environments – a functionality Nvidia CEO Jensen Huang lauded as “utterly gigantic” for bringing state-of-the-art AI to regulated industries and nations. At Subsequent, Huang referred to as Google’s infrastructure one of the best on the planet: “No company is better at every single layer of computing than Google and Google Cloud,” he mentioned.

Pillar 3: The built-in full stack – connecting the dots

Google’s strategic benefit grows when contemplating how these fashions and infrastructure parts are woven right into a cohesive platform. Not like rivals, which regularly depend on partnerships to bridge gaps, Google controls practically each layer, enabling tighter integration and sooner innovation cycles.

So why does this integration matter, if a competitor like Microsoft can merely associate with OpenAI to match infrastructure breadth with LLM mannequin prowess? The Googlers I talked with mentioned it makes an enormous distinction, and so they got here up with anecdotes to again it up.

Take the numerous enchancment of Google’s enterprise database BigQuery. The database now presents a data graph that permits LLMs to go looking over knowledge way more effectively, and it now boasts greater than 5 instances the shoppers of rivals like Snowflake and Databricks, VentureBeat reported yesterday. Yasmeen Ahmad, Head of Product for Information Analytics at Google Cloud, mentioned the huge enhancements have been solely attainable as a result of Google’s knowledge groups have been working carefully with the DeepMind crew. They labored via use instances that have been laborious to unravel, and this led to the database offering 50 % extra accuracy based mostly on frequent queries, a minimum of in accordance with Google’s inner testing, in attending to the appropriate knowledge than the closest rivals, Ahmad informed VentureBeat in an interview. Ahmad mentioned this kind of deep integration throughout the stack is how Google has “leapfrogged” the {industry}.

This inner cohesion contrasts sharply with the “frenemies” dynamic at Microsoft. Whereas Microsoft companions with OpenAI to distribute its fashions on the Azure cloud, Microsoft can also be constructing its personal fashions. Mat Velloso, the Google govt who now leads the AI developer program, left Microsoft after getting annoyed attempting to align Home windows Copilot plans with OpenAI’s mannequin choices. “How do you share your product plans with another company that’s actually competing with you… The whole thing is a contradiction,” he recalled. “Here I sit side by side with the people who are building the models.”

This integration speaks to what Google leaders see as their core benefit: its distinctive means to attach deep experience throughout the total spectrum, from foundational analysis and mannequin constructing to “planet-scale” software deployment and infrastructure design. 

Vertex AI serves because the central nervous system for Google’s enterprise AI efforts. And the combination goes past simply Google’s personal choices. Vertex’s Mannequin Backyard presents over 200 curated fashions, together with Google’s, Meta’s Llama 4, and quite a few open-source choices. Vertex gives instruments for tuning, analysis (together with AI-powered Evals, which Grannis highlighted as a key accelerator), deployment, and monitoring. Its grounding capabilities leverage inner AI-ready databases alongside compatibility with exterior vector databases. Add to that Google’s new choices to floor fashions with Google Search, the world’s greatest search engine.

Integration extends to Google Workspace. New options introduced at Subsequent ’25, like “Help Me Analyze” in Sheets (sure, Sheets now has an “=AI” components), Audio Overviews in Docs and Workspace Flows, additional embed Gemini’s capabilities into each day workflows, creating a robust suggestions loop for Google to make use of to enhance the expertise. 

Whereas driving its built-in stack, Google additionally champions openness the place it serves the ecosystem. Having pushed Kubernetes adoption, it’s now selling JAX for AI frameworks and now open protocols for agent communication (A2A) alongside assist for current requirements (MCP). Google can also be providing a whole bunch of connectors to exterior platforms from inside Agentspace, which is Google’s new unified interface for workers to seek out and use brokers. This hub idea is compelling. The keynote demonstration of Agentspace (beginning at 51:40) illustrates this. Google presents customers pre-built brokers, or workers or builders can construct their very own utilizing no-code AI capabilities. Or they will pull in brokers from the surface by way of A2A connectors. It integrates into the Chrome browser for seamless entry.

Pillar 4: Concentrate on enterprise worth and the agent ecosystem

Maybe probably the most vital shift is Google’s sharpened concentrate on fixing concrete enterprise issues, significantly via the lens of AI brokers. Thomas Kurian, Google Cloud CEO, outlined three causes clients select Google: the AI-optimized platform, the open multi-cloud method permitting connection to current IT, and the enterprise-ready concentrate on safety, sovereignty, and compliance.

Brokers are key to this technique. Except for AgentSpace, this additionally contains:

Constructing Blocks: The open-source Agent Growth Package (ADK), introduced at Subsequent, has already seen vital curiosity from builders. The ADK simplifies creating multi-agent methods, whereas the proposed Agent2Agent (A2A) protocol goals to make sure interoperability, permitting brokers constructed with totally different instruments (Gemini ADK, LangGraph, CrewAI, and so forth.) to collaborate. Google’s Grannis mentioned that A2A anticipates the size and safety challenges of a future with probably a whole bunch of hundreds of interacting brokers.

This A2A protocol is de facto vital. In a background interview with VentureBeat this week, the CISO of a significant US retailer, who requested anonymity due to the sensitivity round safety points. However they mentioned the A2A protocol was useful as a result of the retailer is searching for an answer to differentiate between actual individuals and bots who’re utilizing brokers to purchase merchandise. This retailer needs to keep away from promoting to scalper bots, and with A2A, it’s simpler to barter with brokers to confirm their proprietor identities.

Goal-built Brokers: Google showcased professional brokers built-in into Agentspace (like NotebookLM, Concept Era, Deep Analysis) and highlighted 5 key classes gaining traction: Buyer Brokers (powering instruments like Reddit Solutions, Verizon’s assist assistant, Wendy’s drive-thru), Artistic Brokers (utilized by WPP, Brandtech, Sphere), Information Brokers (driving insights at Mattel, Spotify, Bayer), Coding Brokers (Gemini Code Help), and Safety Brokers (built-in into the brand new Google Unified Safety platform). 

This complete agent technique seems to be resonating. Conversations with executives at three different giant enterprises this previous week, additionally talking anonymously on account of aggressive sensitivities, echoed this enthusiasm for Google’s agent technique. Google Cloud COO Francis DeSouza confirmed in an interview: “Every conversation includes AI. Specifically, every conversation includes agents.” 

Kevin Laughridge, an govt at Deloitte, a giant person of Google’s AI merchandise, and a distributor of them to different firms, described the agent market as a “land grab” the place Google’s early strikes with protocols and its built-in platform provide vital benefits. “Whoever is getting out first and getting the most agents that actually deliver value – is who is going to win in this race,” Laughridge mentioned in an interview. He mentioned Google’s progress was “astonishing,” noting that customized brokers Deloitte constructed only a yr in the past may now be replicated “out of the box” utilizing Agentspace. Deloitte itself is constructing 100 brokers on the platform, focusing on mid-office features like finance, danger and engineering, he mentioned.

The client proof factors are mounting. At Subsequent, Google cited “500 plus customers in production” with generative AI, up from simply “dozens of prototypes” a yr in the past. If Microsoft was perceived as approach forward a yr in the past, that doesn’t appear so clearly the case anymore. Given the PR warfare from all sides, it’s tough to say who is de facto profitable proper now definitively. Metrics differ. Google’s 500 quantity isn’t instantly corresponding to the 400 case research Microsoft promotes (and Microsoft, in response, informed VentureBeat at press time that it plans to replace this public rely to 600 shortly, underscoring the extraordinary advertising and marketing). And if Google’s distribution of AI via its apps is important, Microsoft’s Copilot distribution via its 365 providing is equally spectacular. Each are actually hitting thousands and thousands of builders via APIs.

[Editor’s note: Understanding how enterprises are navigating this ‘agent land grab,’ and successfully deploying these complex AI solutions, will be central to the discussions at VentureBeat’s Transform event this June 24-25 in San Francisco.]

However examples abound of Google’s traction:

• Wendy’s: Deployed an AI drive-thru system to hundreds of places in only one yr, bettering worker expertise and order accuracy. Google Cloud CTO Will Grannis famous that the AI system is able to understanding slang and filtering out background noise, considerably decreasing the stress of stay buyer interactions. That frees up employees to concentrate on meals prep and high quality — a shift Grannis referred to as “a great example of AI streamlining real-world operations.”
• Salesforce: Introduced a significant enlargement, enabling its platform to run on Google Cloud for the primary time (past AWS), citing Google’s means to assist them “innovate and optimize.”
• Honeywell & Intuit: Firms beforehand strongly related to Microsoft and AWS, respectively, now partnering with Google Cloud on AI initiatives.
• Main Banks (Deutsche Financial institution, Wells Fargo): Leveraging brokers and Gemini for analysis, evaluation, and modernizing customer support.
• Retailers (Walmart, Mercado Libre, Lowe’s): Utilizing search, brokers, and knowledge platforms.

This enterprise traction fuels Google Cloud’s general development, which has outpaced AWS and Azure for the final three quarters. Google Cloud reached a $44 billion annualized run charge in 2024, up from simply $5 billion in 2018.

Navigating the aggressive waters

Google’s ascent doesn’t imply rivals are standing nonetheless. OpenAI’s speedy releases this week of GPT-4.1 (centered on coding and lengthy context) and the o-series (multimodal reasoning, software use) reveal OpenAI’s continued innovation. Furthermore, OpenAI’s new picture technology characteristic replace in GPT-4o fueled large development over simply the final month, serving to ChatGPT attain 800 million customers. Microsoft continues to leverage its huge enterprise footprint and OpenAI partnership, whereas Anthropic stays a robust contender, significantly in coding and safety-conscious purposes.

Nonetheless, it’s indeniable that Google’s narrative has improved remarkably. Only a yr in the past, Google was considered as a stodgy, halting, blundering competitor that maybe was about to blow its likelihood at main  AI in any respect. As an alternative, its distinctive, built-in stack and company steadfastness has revealed one thing else: Google possesses world-class capabilities throughout your entire spectrum – from chip design (TPUs) and world infrastructure to foundational mannequin analysis (DeepMind), software improvement (Workspace, Search, YouTube), and enterprise cloud providers (Vertex AI, BigQuery, Agentspace). “We’re the only hyperscaler that’s in the foundational model conversation,” deSouza said flatly. This end-to-end possession permits for optimizations (like “intelligence per dollar”) and integration depth that partnership-reliant fashions battle to match. Rivals usually must sew collectively disparate items, probably creating friction or limiting innovation velocity.

Google’s second is now

Whereas the AI race stays dynamic, Google has assembled all these items on the exact second the market calls for them. As Deloitte’s Laughridge put it, Google hit a degree the place its capabilities aligned completely “where the market demanded it.” For those who have been ready for Google to show itself in enterprise AI, you could have missed the second — it already has. The corporate that invented lots of the core applied sciences powering this revolution seems to have lastly caught up – and greater than that, it’s now setting the tempo that rivals must match.

Within the video beneath, recorded proper after Subsequent, AI professional Sam Witteveen and I break down the present panorama and rising tendencies, and why Google’s AI ecosystem feels so robust: