In June 2017, eight researchers at Google published a paper called “Attention Is All You Need”. The title was a quiet joke about the mechanism the paper introduced. Attention. The idea that a neural network could weigh which parts of an input mattered most for predicting what came next. The paper was 15 pages, dense with math, and barely registered outside the machine learning community.
Every large language model in commercial use today is built on top of that paper. ChatGPT. Claude. Gemini. Llama. Mistral. Grok. All of them. The transformer architecture proposed by Ashish Vaswani and seven colleagues at Google Brain is the foundation of the modern AI industry. Google invented the technology, published it openly, and then watched competitors use that invention to threaten Google's core business.
For most of the next decade, that story looked like it was going to end badly.
It is not ending badly.
The Fall That Should Have Happened
When ChatGPT launched on November 30, 2022, the question on every desk at Wall Street and every all-hands at Mountain View was the same. Was this the moment that broke Google's search business? OpenAI's chatbot reached 100 million users in two months, the fastest consumer adoption of any technology in history. Microsoft (NASDAQ: MSFT) followed with a $10 billion investment in OpenAI, then bolted GPT-4 into Bing. Sundar Pichai called a “code red.” Larry Page and Sergey Brin came back to the Googleplex.
Alphabet's (NASDAQ: GOOGL) initial response, the rushed launch of Bard in February 2023, produced a factual error in its first demo video and erased $100 billion in market value in a single day. For most of the next 18 months, the market priced Alphabet like a defender, not a builder. Bear cases multiplied. Search query volume was about to collapse. Advertising revenue was about to migrate to chat. Microsoft was finally going to win consumer tech.
That story is now visibly wrong.
Alphabet stock is up roughly 130% in the past 12 months. On May 10, 2026, the company briefly passed Nvidia (NASDAQ: NVDA) as the world's most valuable company in after-hours trading, closing the week with a market cap of $4.8 trillion against Nvidia's $5.2 trillion. Google Cloud revenue grew 63% year over year in Q1 2026, with operating margins expanding to 33%. Enterprise deals over $1 billion in 2025 surpassed the previous three years combined.
The financial reversal is not the true story, however. In October 2024, Google CEO Sundar Pichai has announced a restructuring to accelerate the company’s AI initiatives, consolidating the Gemini app team and putting all AI under Google DeepMind, led by its Co-founder and Nobel Laureate Demis Hassabis. We're now seeing the results in the past two events: Google Cloud Next ’26 in April and Google I/O ’26 in May.
Completing the AI Stack
There is a way to understand who is winning the AI race that survives quarterly mood swings and analyst pile-ons. It is a simple test. There are eight categories every major AI company is competing in. Today, Google is the only company that occupies all eight.
The eight layers, in plain English:
AI Server Chips. The physical processors that train and run AI systems. Training builds the model, which requires massive parallelized computation that can run for weeks. Inference runs the model for users in real time, requiring low-latency responses at massive scale. Different workloads, different optimization, increasingly different silicon. The company that designs its own chips sets its own cost curve and does not wait for Nvidia's roadmap.
Data (Model Training). What models learn from. Without unique, proprietary data, two models with similar architectures converge on similar capabilities. The company with the most instrumented, highest-signal data at training time builds a compounding advantage that competitors cannot replicate by adding compute alone.
Cloud Infrastructure. The rented infrastructure that hosts the models. It is where the chips physically live. It is the highway system that moves data between them, and between models and the users or systems querying them. Cloud is also the surface on which enterprise software is increasingly delivered.
Frontier AI Model. The trained AI system that defines the capability ceiling of an era. The smartest model, the most capable model, the model that sets the benchmark every competitor is measured against. These are the things consumers see when they say "ChatGPT" or "Claude" or "Gemini."
Enterprise AI Apps. AI systems deployed inside companies to augment or replace human workflows. Code generation, security monitoring, document analysis, financial modeling, procurement, customer service. The enterprise layer is where the majority of AI revenue flows today, and where governance matters most.
Consumer AI Apps. AI products that reach individual users at scale. Search. The Gemini app. ChatGPT. Claude. The consumer layer creates the distribution moat. A company without consumer AI apps at scale has no user base to convert to agent users when the agent layer matures.
Wearable AI Devices. Hardware worn on or near the body that delivers AI assistance in the physical world. Smart glasses. Earbuds with real-time translation and search. Devices that put the AI interface in the peripheral vision and ambient audio of daily life, not behind a screen. The wearable layer is where the next consumer interface will be built.
Physical AI (Robotics). AI systems that operate in the physical world through robotic form factors. Industrial automation. Autonomous vehicles. Humanoid robots. The physical AI layer is the final frontier, where software intelligence meets physical action in the world.
Now go down the list and apply it to each major AI company.
OpenAI has the data, the frontier model, the enterprise apps, and the consumer app. It does not have its own chips. It does not have its own cloud at scale. It has no wearables and no robotics play. Its relationship with Microsoft governs almost every layer it does not own.
Anthropic has the frontier model, a strong enterprise app in Claude for Work, and a consumer app in Claude.ai. It has none of the other five layers. No chips, no cloud, no data moat beyond training runs, no wearables, no physical AI.
Microsoft has chips (custom silicon through its OpenAI partnership), data, cloud, a frontier model (via OpenAI), enterprise apps through Copilot and Azure, and consumer apps. It has no wearables and no physical AI strategy at scale.
Meta has chips (MTIA), data (the most behaviorally instrumented social graph in the world), a frontier model in Llama, consumer apps reaching billions of users, and a wearable AI position through Ray-Ban Meta glasses. It does not have a cloud infrastructure business or meaningful enterprise AI apps.
Amazon has chips (Trainium and Inferentia), data, cloud (AWS, the largest in the world), a frontier model through its investment in Anthropic and Nova model family, enterprise apps, and consumer apps through Alexa and its retail ecosystem. It has no wearables. It does have a physical AI play through its logistics and warehouse robotics programs.
The SpaceX/xAI/Tesla grouping in the chart is the most interesting competitor entry. Data, yes: Tesla's fleet generates more real-world physical training data than any other source on earth. Frontier model, yes: Grok is credible and improving. Consumer apps, yes, through Grok. Physical AI, yes: Tesla's Optimus humanoid robot and the autonomous vehicle program represent the most advanced physical AI deployment outside of specialized industrial applications. No chips of their own, no cloud, no wearables, no enterprise apps. A formidable physical AI position built on a narrow stack.
Apple has consumer apps and wearables. Two layers. The most profitable two-layer position in the history of technology, built on hardware margins and an install base of two billion devices. But two layers in an eight-layer race. Which is why Apple chose Google as a preferred cloud provider and reportedly tapped Gemini to power a rebuilt Siri. The company that owns the device relationship is renting the intelligence.
Google has all eight. The chips (TPUs, now in their eighth generation, with dual architectures for training and inference). The data (Search, YouTube, Maps, Workspace, Android, the most heavily instrumented web crawl in the world). The cloud (Google Cloud, plus its own internal infrastructure spanning more than one million TPUs). The frontier model (the Gemini 3.5 family, plus Gemini Omni). The enterprise apps (Gemini Enterprise Agent Platform, Workspace Intelligence, Google Cloud's AI suite). The consumer apps (Search, Gmail, YouTube, Photos, Maps, Drive, Docs, Chrome, Android, the Gemini app with 900 million monthly active users). The wearable AI devices (Android XR audio glasses launching fall 2026, display glasses to follow). And physical AI, through its investment in and relationship with robotics programs that run on its cloud and model infrastructure.
No other company is even close.
The Compounding Logic
The reason this matters is that the eight categories are not independent. They compound.
Better chips lower the cost of training. Cheaper training enables better models. Better models improve the apps. More apps generate more usage data. More usage data feeds the next round of training. Apps run on top of the cloud, which runs on top of the chips. Agents run on top of the apps. Agent-to-agent commerce runs on top of the protocols. Every layer feeds the layer above it. Every layer also creates leverage for the layer below it.
The company that controls the whole loop benefits at every step. The company that controls only some of the layers has to pay rent to whoever controls the rest.
At Cloud Next, Sundar disclosed that 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall. Google's security operations agents now triage tens of thousands of unstructured threat reports each month, reducing threat mitigation time by more than 90%. The company that builds the platform also runs on the platform. Production benchmarks are first-party. The gap between vendor demo and enterprise reality is shorter than at any prior moment in cloud history.
That is what compounding looks like in practice. Each loop sharpens the next.
What Cloud Next and I/O Made Visible
Take the two keynotes together. They are not separate product launches. They are coordinated declarations of a single thesis: the agentic enterprise is operational, and Google is its infrastructure.
Cloud Next ’26 in Las Vegas was the enterprise half. Google unveiled two new eighth-generation custom AI chips. TPU 8t for training, which scales to 9,600 chips in a single superpod with two petabytes of shared memory and three times the processing power of the previous Ironwood generation. TPU 8i for inference, which delivers 80% better performance per dollar than the prior generation. Plain English translation: training a frontier model is faster and cheaper, and running millions of concurrent agents in production is dramatically cheaper. Google also launched the Gemini Enterprise Agent Platform, a unified environment to build, scale, govern, and optimize AI agents. The Agentic Data Cloud, which lets Google's agents query data inside Amazon Web Services or Microsoft Azure without moving the data to Google. Workspace Intelligence, which embeds agent capabilities across the productivity suite that already serves three billion users.
Cloud Next disclosed the operating reality of the agentic enterprise. 75% of Google Cloud customers now use AI products. 330 customers each processed over one trillion tokens in the last 12 months. Google's first-party models process more than 16 billion tokens per minute via direct API use.
Google I/O ’26 in Silicon Valley was the consumer and developer half, and it doubled the dose. Google now processes 3.2 quadrillion tokens per month across all its products, a 7x increase from one year ago. (A token is the basic unit of AI processing, roughly one word.) Model APIs are processing 19 billion tokens per minute. 375 Google Cloud customers each processed more than one trillion tokens in the last twelve months, up from 330 only a month earlier. 8.5 million developers are building with Google's models monthly. Alphabet plans to spend between $180 billion and $190 billion on capital expenditures this year. In 2022, that figure was $31 billion. Six years, six times the spending. For context, $190 billion is greater than the annual GDP of Hungary or Greece. One company is spending that on AI infrastructure in 12 months.
That is the scaffolding. Now look at what Google built on top of it.
The Gemini 3.5 Family Comes for the Model Layer
The model layer is where the visible scoreboard of the AI race plays out. For most of the last three years, that scoreboard read as a two-horse race. OpenAI's GPT-4, then GPT-5. Anthropic's Claude. Gemini, persistently a half-step behind on benchmarks that mattered most for enterprise adoption.
That changed at I/O.
Gemini 3.5 Flash, the first model in the 3.5 family, launched as the new default across the Gemini app and AI Mode in Search globally. The benchmark gap closed in both directions. 3.5 Flash beats Gemini 3.1 Pro on nearly every dimension. It scores 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval (a benchmark of real-world economically valuable tasks), and 83.6% on MCP Atlas. It also beats comparable frontier models from OpenAI and Anthropic on output speed by four times. The phrase “frontier intelligence at Flash speed” is no longer marketing. It is a measurable position.
Pricing reinforces the position. Gemini 3.5 Flash delivers frontier-level capabilities at less than half the price of comparable frontier models. Google's own math: if a company processing one trillion tokens per day shifts 80% of its workloads from other frontier models to 3.5 Flash, the savings exceed $1 billion annually. That is not a feature update. It is a structural pricing move designed to make leaving Google's platform economically irrational.
Gemini 3.5 Pro is coming next month. The pattern is now visible. Ship Flash to billions immediately. Refine internally on Pro. Release Pro as the quiet quality upgrade six to eight weeks later. The cadence is faster than any competitor can match.
Behind the model release sits the more consequential announcement: Gemini Omni. This is not another LLM. It is a new model family designed to generate samples in any output modality from any input. Text in, video out. Image in, text out. Multimodal generation, multimodal editing, character and voice consistency preserved across scenes. Google describes the leap as moving from “predicting text to simulating reality.” The first model in the family, Gemini Omni Flash, launched at I/O and is available immediately in the Gemini app, Google Flow, and YouTube Shorts.
The implication for the AI race is straightforward. The smartest model is no longer the only competition. The most general model is. Whoever solves multimodal generation first, at frontier quality, at consumer scale, captures applications that text-only models cannot. That is the layer Google just claimed.
Antigravity Is Google's ClaudeCode
Every era of computing creates a layer that defines who builds on top of it. In the 1990s, that layer was the operating system. In the 2000s, it was the browser. In the 2010s, it was the mobile app store. In the 2020s, the layer is the agent runtime. The environment where AI agents are designed, orchestrated, supervised, and run.
OpenAI is building toward this layer with Codex. Anthropic is building toward this layer with Claude Code. Google is building toward this layer with Antigravity.
Antigravity is Google's agent-first development platform. The Antigravity 2.0 release announced at I/O turned it into a standalone desktop application, a central home where developers and increasingly non-developers orchestrate cohorts of autonomous AI agents. It runs an optimized version of Gemini 3.5 Flash that operates twelve times faster than other frontier models, not four times. It powers Gemini Spark. It powers the agentic capabilities now appearing in Search. It is being adopted internally at Google faster than any tool in the company's history.
The internal numbers tell the story. In March 2026, Google was processing half a trillion tokens per day across its AI developer tools. By I/O, that figure was more than three trillion tokens per day, doubling every few weeks. 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% last fall. One complex code migration that would have taken six months with engineers alone was completed in roughly six weeks with agents and engineers working together.
Antigravity is not competing with Codex and Claude Code on developer experience alone. It is competing on distribution. Codex lives inside OpenAI's product surface. Claude Code is a standalone command-line tool. Antigravity ships into Workspace, Search, Chrome, Android, the Gemini app, and the Gemini Enterprise Agent Platform on day one. A developer who builds an agent in Antigravity has it deployable across a three-billion-user surface area immediately. That is not a feature advantage. That is a structural advantage that compounds with every release.
The agent runtime is the operating system of the next decade of computing. The company that owns it earns rent on every transaction that flows through it. Google just shipped its candidate.
Gemini Spark: The First Agentic Consumer App at Hyperscale
If Antigravity is the runtime, Gemini Spark is the consumer-facing proof that the runtime works. Spark is Google's 24/7 personal AI agent, and it is the single most important announcement from I/O.
Spark runs on dedicated virtual machines inside Google Cloud. It does not require a user's laptop to be open. It is powered by Gemini 3.5 and the Antigravity harness. It works inside Gmail, Docs, Slides, Chrome, and more than 30 third-party services connected through the Model Context Protocol, including Canva, OpenTable, and Instacart. It can draft emails, manage calendars, compile auto-updating documents, place reservations, complete shopping orders. It launches on the new $100 per month Google AI Ultra tier for U.S. subscribers next week.
Three structural facts make Spark the iPhone moment for consumer agents.
First, Spark sits inside Google's existing distribution. The Gemini app already has 900 million monthly active users, more than double a year ago. Daily requests have grown over seven times in the same period. Spark inherits that user base on day one. No other company has 900 million monthly active users to convert into agent users.
Second, Spark runs on Google's own infrastructure, using Google's own chips and Google's own model. The unit economics work because Google captures margin at every layer. OpenAI cannot ship a comparable always-on consumer agent without renting compute from Microsoft. Anthropic cannot ship one without renting from Google or Amazon. The infrastructure economics belong to Google.
Third, Spark connects to the rest of the world through standards Google co-authored. It uses MCP for tool connections. It uses AP2 for payments. It uses UCP for commerce. When Spark places an order through OpenTable or buys something through Instacart, the transaction runs on protocols Google helped design.
A personal AI agent that lives inside your productivity stack, connects to your enterprise tools, runs 24/7 on infrastructure built and owned by one company, and transacts through standards that company helped write, is not visiting your workflow. It is residing in it.
Search Becomes Agentic
Search was supposed to be the casualty of the AI revolution. ChatGPT was framed, in 2022 and 2023, as the inevitable replacement for the ten blue links. Every analyst note for two years assumed Search query volume would collapse. The bear case on Alphabet was, fundamentally, a bear case on Search.
The bear case is now visibly wrong.
AI Overviews in Search reaches 2.5 billion monthly active users. AI Mode, the deeper conversational search experience, surpassed one billion monthly active users in its first year, with queries more than doubling every quarter since launch. Last quarter, total Search queries hit an all-time high. Users who adopt AI features in Search do not search less. They search more.
I/O made Search the most aggressively agentified surface at Google. Four announcements matter.
First, the Search box itself. Google described it as the biggest upgrade to the Search box in over 25 years. It now dynamically expands to give users space to describe what they need. It anticipates intent with AI-powered suggestions. It accepts inputs in any modality, text, images, files, videos, Chrome tabs. This is the front door to the agentic internet, redesigned for the first time since 1998.
Second, information agents. Personalized AI agents that run in the background 24/7, monitoring the web for what each user cares about. Apartment listings that meet specific criteria. Pro athlete sneaker drops. Real-time finance, shopping, and sports updates. A user describes what they want once. The agent monitors continuously and notifies when criteria are met. Information agents launch this summer for Google AI Pro and Ultra subscribers.
Third, agentic booking. Search now handles increasingly complex multi-criteria requests across local experiences and services. A user can ask Google to find a private karaoke room for six on a Friday night that serves food late. Search returns availability, pricing, and direct booking links. For categories like home repair, beauty, and pet care, Search will call businesses on the user's behalf. The consumer agent layer is now fused into the world's most-used search engine.
Fourth, generative UI in Search. Powered by Gemini 3.5 Flash and Antigravity, Search now builds custom experiences on the fly. Interactive visuals. Dynamic layouts. Persistent dashboards and trackers that users return to and update. Google demoed a custom fitness tracker built in response to a single search query, tapping live data, maps, and weather. Search is becoming a runtime for personalized mini-apps, generated by AI, on demand. Free, for everyone, this summer.
The Search transformation answers the question that hung over Alphabet for four years. What happens to Search in the age of AI? The answer is that Search becomes the layer where AI agents live, work, and act on the user's behalf. Not a casualty. The interface for the entire agentic era.
Universal Cart and the Agentic Commerce Stack
Google handles more than one billion shopping queries per day. The Shopping Graph, Google's product catalog, contains over sixty billion product listings, the most comprehensive in the world. Until I/O, these were assets. After I/O, they are the foundation of a new commerce stack that may be the single most underestimated announcement of the keynote.
The new front door is Universal Cart. An intelligent shopping cart that works across merchants, across services, and across product categories. Users add items while browsing Search, chatting with Gemini, watching YouTube, or reading Gmail. The cart works in the background, finding deals and price drops, tracking price history, sending back-in-stock alerts. It uses intelligent reasoning to anticipate needs. If a user assembles components for a custom PC from multiple retailers, the cart flags incompatibilities and suggests alternatives. It applies loyalty programs and merchant offers automatically through Google Wallet integration.
The retail launch partners are not small. Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, and the Shopify ecosystem, including Fenty and Steve Madden, are live for UCP-powered checkout in the U.S. Critically, the brand remains the merchant of record. Google is not becoming the retailer. Google is becoming the operating system on top of which retailers transact.
Below the cart sits the protocol layer. The Universal Commerce Protocol (UCP) is the open standard, co-developed with Amazon, Shopify, Walmart, Meta, Microsoft, Salesforce, and India's Flipkart, that lets AI agents move across the entire shopping journey, product discovery, pricing, inventory, checkout, shipment tracking, without breaking context. UCP is expanding from retail into hotel booking and local food delivery, and from the U.S. into Canada, Australia, and the U.K. in the coming months. The intent is clear. UCP is being built as the horizontal commerce layer for the agentic economy, not just shopping.
Below UCP sits the payment layer. The Agent Payments Protocol (AP2), updated this week to v0.2.0, is the infrastructure that lets AI agents make purchases on a user's behalf. It is backed by American Express, Mastercard, Visa, PayPal, and more than sixty other payment networks. AP2 uses cryptographically signed digital contracts called Mandates that create a tamper-proof audit trail for every transaction. Users set limits on what an agent can spend, on which brands, on which products. The agent transacts only when those limits are met. Returns, disputes, and refunds reference the same digital record. AP2 launches first with Gemini Spark in the coming months.
The supporting protocols complete the stack. MCP (Model Context Protocol) lets agents connect to tools and data sources where work actually lives, with more than 14,000 servers already deployed. A2A (Agent-to-Agent) lets one AI agent ask another to do work for it, launched with more than 50 partners including Atlassian, Box, PayPal, and Workday. AGUI (Agent-User Interface) lets a long-running agent show users what it is doing, ask for approval at sensitive steps, and accept course corrections. X42 is a separate payment protocol from Coinbase, adopted by Cloudflare, that lets agents buy compute or data autonomously.
McKinsey estimates the agentic commerce market could reach $5 trillion by 2030. The companies that will share that market are the companies whose products are reachable through UCP and whose payments clear through AP2. Both are protocols Google co-authored. Six standards in total, with one company in every room.
Amazon, until I/O, was the company most analysts assumed would own AI-mediated commerce because Amazon owns the destination. Google's move is to make the destination obsolete. The cart lives wherever the user is. The protocol runs every transaction. The payment clears through Google's own rails. Universal Cart is, structurally, Amazon's worst nightmare.
The Rest of the Sweep
The remaining announcements from I/O matter less for what each piece does and more for what they signal in aggregate.
Daily Brief, an out-of-the-box morning digest agent that synthesizes inbox, calendar, and tasks into prioritized next steps. Google Pics, a new image creation tool built on the Nano Banana model that treats every image element as an editable object. Ask YouTube, which jumps users directly to the part of a video most relevant to their question. Docs Live, voice-powered document creation rolling out across Gmail, Docs, and Keep. Intelligent eyewear, with audio glasses launching first in fall 2026 followed by display glasses, both Gemini-native. Gemini for Science, connecting Antigravity to over thirty major life science databases. Project Genie expanded globally to AI Ultra subscribers, including a new capability that uses Street View to build navigable worlds anchored in reality. Stitch updates for real-time design guidance.
A restructured AI subscription stack: a new $100 per month Google AI Ultra plan aimed at developers and knowledge workers, and a price reduction on the top Ultra tier from $250 to $200 per month. Both tiers include Gemini Spark in the U.S. SynthID adoption by OpenAI, Kakao, and ElevenLabs, the first time a major AI safety standard has crossed competitive lines, with more than one hundred billion images and videos and sixty thousand years of audio already watermarked.
Read together, these are not isolated launches. They are the simultaneous activation of every surface in the Google ecosystem with agents, models, and protocols designed to work together. No other company can ship at this surface area in a single keynote.
What Anthropic and Apple Are Telling You
Two corporate decisions made in the last six months tell you more about Google's structural position than any quarterly earnings report.
The first is the Anthropic-Google relationship. Anthropic recently committed approximately $200 billion to Google Cloud over five years for roughly five gigawatts of compute capacity. Google has invested up to $40 billion directly in Anthropic. Anthropic competes with Google's Gemini model. Customers pick Claude over Gemini all the time. But when an enterprise chooses Claude, the actual computing happens on Google's chips, inside Google's cloud, under a $200 billion contract. Google gets paid on the infrastructure. Google benefits through its equity stake. Anthropic's growth spends Google's money back into Google's services.
One analyst at Wall Street firm Deepwater Asset Management put it directly. If customers prefer Claude, Google still wins on the infrastructure, on the cloud revenue, and through its investment stake.
That is not a partnership. It is a deliberate setup where every possible outcome flows back to the same company. No other AI company has built that kind of structural hedge. OpenAI does not own the cloud it runs on. Anthropic does not have its own chips. Microsoft does not own its frontier model outright. Google built every layer of the stack first, then designed financial relationships where its competitors' growth funds its own infrastructure.
The second decision is Apple's. Apple (NASDAQ: AAPL) has chosen Google as a preferred cloud provider and reportedly tapped Gemini to power a rebuilt Siri. The most disciplined platform owner in technology, a company whose entire brand is built on owning the experience end to end, selected another company's AI to live inside its flagship product. When Tim Cook's company picks your AI, the rest of the C-suite is paying attention.
Putting Things In Perspective
It seems like Google is now hitting all cylinders.
However, if we've learned anything about competing in the AI era it is that things move fast and can change quickly. Indeed, there are a few risk factors to consider for Google.
The first is concentration risk. The roughly $462 billion Google Cloud backlog has questions hanging over it. If Anthropic's $200 billion commitment represents more than 40% of Google's contracted cloud revenue, that is concentration on a scale Oracle (NYSE: ORCL) experienced last year, when a 360% backlog jump tied largely to OpenAI cost the stock roughly half its value over five months once analysts priced in the dependency. Microsoft faces the same scrutiny today around OpenAI. Whether Google can demonstrate diversified billion-dollar deals across many customers will determine whether the rally extends or stalls.
The second is the capex bet. Spending $190 billion in a single year on AI infrastructure assumes demand keeps growing at current rates and competitors do not invent a fundamentally cheaper architecture. Both assumptions are reasonable. Neither is certain. If inference costs collapse faster than capex amortizes, the operating leverage that looks attractive today gets pressured tomorrow. Gemini 3.5 Pro lands next month, and if it does not extend the lead 3.5 Flash just established, the story tightens.
The third is regulatory. Google is already the target of multiple antitrust actions in the United States and the European Union. A company that owns every layer of the AI stack is, by definition, in a position to be accused of using dominance in one layer against competitors in another. Expect those cases to multiply. Expect them to be slow. Expect them to shape how aggressively Google can monetize the stack, particularly the commerce protocols UCP and AP2.
None of these risks are hypothetical. All three are knowable. And all three together do not unwind the structural advantage. They constrain its monetization.
What This Looks Like From a Distance
Step back from the keynote demos and the analyst notes, and the pattern is straightforward.
Eight researchers at Google invented the transformer in 2017 and published it openly. Google then proceeded to do what large incumbents almost always do with their own breakthroughs. It underestimated the speed at which others would commercialize the work. ChatGPT exposed that lag. The next two years were, internally, an existential mobilization. Sundar cancelled vacations. Larry Page returned to active engagement. The DeepMind merger with Google Brain consolidated research and product. Custom silicon accelerated through three generations in two years. Gemini went from a delayed launch to a daily-updated model family. The Gemini app went from a defensive product to the fastest-growing consumer AI service. Antigravity made every developer at Google twice as productive. Spark made the agent personal. Search made the bear case disappear. Universal Cart made commerce a Google stack.
The result, nine years after “Attention Is All You Need,” is not a better version of the company that existed in 2017. It is a different company entirely. One that owns the chips, the cloud, the data, the model, the apps, the agents, the protocols, the consumer distribution, the enterprise relationships, the developer environment, and increasingly the rails through which AI agents will transact with each other.
The race for the smartest AI model will continue. So will the race for the most consumer users. The race for enterprise software will keep being fought. All of those races are now being run on tracks that one company built, owns, and is renting back to the competition.
Google is not the dark horse in the AI race. Google might just become the company everyone else is running on top of.
A Note On Governance
There is a governance question buried under all of this, and it deserves one paragraph because it explains why the stack matters more than the demos.
The agentic enterprise is no longer a strategy slide. It is the operating reality of every Fortune 500 company that adopts AI from a major vendor between now and the EU AI Act enforcement deadline in August 2026. AI agents are already making thousands of decisions per minute on platforms most boards have never inventoried. Most companies have arrived at full-stack dependency on a single vendor without any single committee approving it, because each step looked like a product decision, not a platform decision. The governance frameworks built for a static enterprise, annual reviews, quarterly board meetings, charters written for human accountability on human time, were never designed for systems that act at machine speed. Continuous governance is the only model that fits the moment.
