Google I/O is the company's yearly developer conference. This year, on 19–20 May 2026 in Mountain View, Google used it to announce that almost every product it makes — Search, Gmail, Android, Chrome, the Gemini app, even glasses — will be driven by AI models that can act, not just answer. The umbrella name is “agentic Gemini.” The headline pieces: a new multimodal model called Gemini Omni; a cheaper, faster default model called Gemini 3.5 Flash; a personal AI agent called Spark that reads your inbox and takes action; a developer tool called Antigravity 2.0 that competes head-on with Cursor and GitHub Copilot; a redesigned Search that hands queries to AI agents; and Intelligent Eyewear smart glasses built with Samsung, Warby Parker and Gentle Monster. For a DAX40 CIO, the question is no longer whether Gemini shows up in the stack. It is on what terms.

Sundar Pichai walked onto the Shoreline Amphitheatre stage under a deliberately understated banner — three words, no hardware on a plinth — and told ten thousand developers and several thousand more on the live stream that Google was “firmly in our agentic Gemini era.” Within ninety minutes he and DeepMind chief Demis Hassabis had unveiled six product lines that, taken together, attempt to make Gemini the default surface for almost every kind of computer interaction Google touches. The model news came first. Gemini Omni, a multimodal family that ingests image, audio, video and text and generates the same, was positioned as DeepMind's answer to a year of catch-up with OpenAI's Sora and the conversational voice modes from Anthropic and Meta. Gemini Omni Flash, the cheaper sibling, started rolling out to the Gemini app and to Google Flow the same day for AI Plus, Pro and Ultra subscribers, with API access following. Underneath sits Gemini 3.5 Flash — a smaller, agent-tuned workhorse priced at $1.50 per million input tokens and $9 per million output tokens, roughly 40 percent below Gemini 3.1 Pro, with a one-million-token context window. Google's published benchmarks show it edging Pro on Terminal-Bench 2.1 (76.2 percent versus 70.3) while running, the company claims, four times faster than competing frontier models. From Wednesday it became the default in both the Gemini app and AI Mode in Search. Then came the agents. Gemini Spark, demonstrated by product VP Josh Woodward, is a 24/7 personal agent that lives on a dedicated VM inside Google Cloud, integrates with Workspace, and — the detail that drew the loudest applause — has its own email address, so you can simply forward it a task. Spark can parse credit-card statements, surface hidden subscriptions, watch school emails for deadlines and draft both a Google Doc and the email that launches the project around it. It is US-only beta at launch, gated to the new $100-a-month AI Ultra tier. The developer story was Antigravity 2.0, a year-old experiment that Google has now expanded into a full agent-first platform: redesigned desktop IDE, a Go-based CLI, an SDK for building custom agents, native voice control, and a Manager Surface that lets you fan out five subagents in parallel — one refactoring auth, another writing tests, a third checking a library question — and review the artifacts on return. Models supported include Gemini 3 Pro, Gemini 3.5 Flash, Anthropic's Claude Sonnet 4.5 and OpenAI's GPT-OSS. That is not accidental: by undercutting Cursor on price and out-flanking it on multi-agent orchestration, Google is trying to retake the developer surface it has been losing for two years. The big swing, though, was Search. AI Mode, once a Labs toggle, is now the default for many queries, has crossed one billion monthly users and will soon spawn “information agents” that run in the background. And, to bookend the morning, Samsung's mobile chief joined Pichai to unveil Intelligent Eyewear — audio glasses this autumn, display glasses to follow — built with Warby Parker and Gentle Monster. Not by accident, Google did not pause to explain how any of this would land in the European Union.

Strip away the keynote choreography and a clear architecture emerges. At the bottom sits TPU 8i, Google's newest in-house silicon, which Pichai briefly framed as the cost lever that lets Gemini 3.5 Flash be priced where it is. One tier up sits the model family: Gemini 3 Pro for hard reasoning, Gemini 3.5 Flash for default and agentic work, Gemini Omni for multimodal generation, and small on-device variants for Android and the new glasses. Above the models, Google has consolidated what used to be a sprawl of agent runtimes into a single “agentic harness”: the same plumbing that powers Spark, ships inside Antigravity 2.0, and is offered to enterprises as the renamed Gemini Enterprise Agent Platform — formerly Vertex AI. Above that sit the surfaces — Search, Gmail, Docs, Chrome, the Gemini app, Android, Android XR and now Intelligent Eyewear — each rewired to call the harness rather than a stateless model endpoint. The shift is less about any single launch than about Google finally collapsing its product lines onto one stack. For the past two years the criticism from Ben Thompson and others was that Google had the strongest models but no compelling harness — leaving Anthropic and OpenAI to capture the margin in the agentic layer. Thomas Kurian's pre-keynote Stratechery interview was a tacit acknowledgement. Antigravity 2.0 and the rebranded Enterprise Agent Platform are Google's attempt to close that gap in a single quarter. The numbers are worth pausing on. One billion AI Mode users is, in population terms, larger than the European Union and the United States combined, and roughly twice the monthly active user base of the entire iOS App Store paid-app ecosystem at its peak. Each AI Mode query is, on Google's own data, three times the length of a classic query and far more expensive to serve — which is precisely why the company needed a Flash-tier model cheap enough to be the default. Gemini 3.5 Flash at $1.50/$9 per million tokens makes a back-of-envelope agent call cost a few cents; at that price, embedding it inside Search, Workspace and Android is defensible at advertising margins. The catch — and the reason the room politely declined to dwell on it — is geographical. Google Cloud's own documentation confirms that Gemini 3.x models are not yet available in EU regions on the Enterprise Agent Platform. Frankfurt, Belgium and the Netherlands customers must stay on Gemini 2.5 Pro and 2.0 Flash in europe-west4 for GDPR-compliant deployments, with no announced timeline for 3.x parity. For a DAX40 CIO designing an agent strategy, the Wednesday keynote read very differently than for a US peer: the new toys are visible through the shop window, but the EU regional shelves still hold last year's stock. The same applies to Spark (US-only beta) and to AI Mode's full agentic features, which Google declined to commit to a European rollout date. Meanwhile Brussels is moving in the opposite direction. On 27 January 2026 the European Commission opened two parallel DMA specification proceedings against Google: one requiring Android interoperability for third-party AI assistants on the hardware features Gemini uses, and one ordering Google to share anonymised Search ranking, query, click and view data with rival search engines and AI chatbot providers. The binding decision is due by 27 July 2026 — six weeks after Spark's planned US ramp. The “agentic Gemini era” will arrive in Europe wrapped in a thicker layer of regulatory paperwork than anywhere else Google ships.

Exa, founded in 2021 as Metaphor by Harvard freshmen Will Bryk and Jeff Wang, runs its own web crawler and a neural search index built on embeddings rather than the keyword inverted index that powers Google and Bing. Instead of matching strings, Exa converts queries and documents into high-dimensional vectors and returns pages whose meaning sits closest in that mathematical space. The pitch is narrow and specific: most search engines were built for humans typing a few words into a box; Exa was built for software agents firing thousands of structured queries per task. On May 20, 2026, the company closed a $250 million Series C led by Andreessen Horowitz at a $2.2 billion valuation, up from $700 million eight months earlier. Existing backers Lightspeed, Benchmark, Y Combinator and Nvidia's NVentures followed on.

Will Bryk, 27, published the announcement himself from Exa's office in San Francisco's Mission District, in a post co-signed by a16z general partner Martin Casado. The tone was less press release than manifesto. “We're organizing the world's knowledge, but this time for AI,” Bryk wrote, in a line designed to sit alongside Larry Page's original 1998 framing. His central claim is empirical rather than rhetorical: AI agents will issue more web searches than humans this year, and within a handful of years they will issue roughly a thousand times more. “As trillions of agents come online over the coming years, search needs will grow thousands of times beyond the total search volume of Google,” he argued. The Series C is sized for that scenario. Of the $250 million, the bulk is earmarked for GPU capacity to train Exa's next-generation embedding models and to scale the crawl and serving stack to hundreds of thousands of queries per second. The backstory matters because it explains why a16z paid this price. Bryk and Wang met as Harvard freshmen, dropped into Y Combinator's Summer 2021 batch as Metaphor Systems, and spent two years building a semantic search demo that almost nobody used. The pivot came when OpenAI shipped the function-calling API and developers started wiring LLMs to external tools. Suddenly the customers Metaphor could not find as a consumer product showed up as developers needing a search API that returned clean, structured passages instead of ad-laden SERP HTML. The company rebranded to Exa in January 2024, raised a $17 million Series A from Lightspeed and Nvidia that July, then an $85 million Series B led by Benchmark in September 2025 at a $700 million valuation. Revenue tracked the agent boom: roughly $0.9 million ARR in September 2024, $10 million a year later, $12 million by January 2026, on a customer list that now includes Cursor, Cognition, HubSpot, OpenRouter and Monday.com, plus the search backends inside Perplexity and OpenAI's Deep Research mode. The a16z thesis, articulated by Casado, is that search is being rebuilt in the same way databases were rebuilt for the cloud era: not as a feature on top of an existing stack, but as new infrastructure with different physics. Human search optimises for one query, ten blue links, a click-through, and an ad auction. Agent search optimises for hundreds of parallel queries per task, structured passages instead of pages, sub-second latency at scale, and a per-query payment model with no advertiser in the loop. Microsoft chose the opposite path: it retired the Bing Search API in August 2025 to push customers toward Copilot. Google still gates its index behind a Custom Search product that was never designed for agent traffic. Into that vacuum Bryk is pouring $250 million.

Exa's technical bet is that the right primitive for agent search is the embedding, not the inverted index. A traditional search engine tokenises a query, looks up the terms in a posted-list structure, and ranks pages with a learned model on top. Exa runs the query through a transformer encoder, converts it to a dense vector of several hundred dimensions, and finds nearest neighbours among the billions of page vectors already stored in its custom vector database. The advantage is semantic recall: the query “companies hiring Rust engineers in Berlin” returns the right job postings even when the page text never contains those exact tokens. The cost is computational. Naive nearest-neighbour search across a web-scale corpus is prohibitive, so Exa stacks several optimisations: Matryoshka embeddings that let the same vector be truncated to a shorter form for a coarse first pass, hierarchical clustering that lets a query skip whole regions of the index, binary compression that turns 32-bit floats into single bits with controlled accuracy loss, and assembly-level SIMD on the hot path. The serving layer is written in Rust. According to Bryk's own technical talks, the stack delivers semantic search latencies competitive with keyword engines while running on a fraction of the hardware footprint a naive implementation would demand. The competitive set divides into three groups. First, own-index neural players: Exa, Brave Search, You.com and the newer entrant Parallel. Independent benchmarks published in early 2026 placed Brave narrowly ahead of Exa on result quality for agent workloads, with Tavily close behind, though latencies vary by a factor of twenty across the field. Second, LLM-optimised wrappers on top of Google or Bing: Tavily, Serper, Linkup and Firecrawl, which add ranking, snippeting and citation formatting tuned for LLM consumption. Third, the incumbents: Google's Custom Search, the retired Bing API, and Perplexity's own newly opened search endpoint, all designed for very different shapes of demand. The second-order story is web access itself. Cloudflare in 2025 launched Pay Per Crawl, a pricing layer that lets publishers charge bots per request, and began serving fabricated content to scrapers that ignore robots.txt. If agent traffic genuinely grows a thousandfold, the economics of crawling shift from a roughly free externality to a measurable cost. An own-index player like Exa pays that cost once on the crawl side and amortises it across queries; a wrapper that proxies live Google requests inherits whatever Google chooses to charge. That asymmetry is part of what a16z is buying.

For most of the GPT era, the AI industry assumed public opinion was somewhere between curious and indifferent. That assumption has now collapsed. A Gallup release on May 13, 2026 found 71% of Americans oppose having an AI datacenter built within ten miles of their home — a stronger NIMBY response than nuclear power plants. Alberto Romero of The Algorithmic Bridge compiled the full picture across Gallup, Pew, Marquette, NBC, Washington Post and Change Research and concluded the trend is unanimous: every major pollster, every demographic, every party. The hostility is now flowing into ballots, county zoning meetings and state-house bills. For DAX40 firms placing AI workloads, for hyperscalers planning German and Irish capacity, and for Brussels watching the EU AI Act take effect, the political ground under generative AI has shifted in under twelve months.

Picture Alberto Romero at his desk in Madrid on the morning of May 20, 2026, a second monitor stacked with PDF crosstabs from six pollsters. He had been tracking individual surveys for months, but the Gallup release a week earlier — telephone interviews with 1,000 adults conducted March 2–18, 2026 — was the moment the pattern became impossible to dismiss. By the end of the day he had published “How America Turned Against AI According to the Poll Data,” a 6,000-word compilation that has since become the most-shared sentiment piece of the spring. “Across every pollster — Gallup, Pew, Change Research, the Washington Post, Marquette, Morning Consult, Heatmap, NBC, Politico, the University of Maryland — the direction is the same,” Romero wrote. “Americans have soured on AI datacenters, on AI itself, and on the people building it.” The Gallup headline number is the one the industry now cannot avoid. Seventy-one percent of US adults would oppose construction of a large AI datacenter within ten miles of their home; 48% are strongly opposed. For comparison, only 53% would oppose a nuclear power plant in the same radius — meaning AI hyperscaler campuses are now, in the American imagination, scarier than fission. Half of opponents cite resource consumption; 18% specifically water, 18% electricity, 16% pollution and noise. The Marquette Law School survey, fielded April 8–16, 2026, reached the same conclusion from a different angle: majorities of Republicans, Democrats, independents, every age cohort and every income band believe the costs of datacenters outweigh their benefits. Pew, in a June 2025 baseline that has only worsened, found half of US adults more concerned than excited about AI, against just 10% more excited — up from 37% concerned in 2021. The pivot from polling to policy is already visible on the ground. Loudoun County, Virginia — the densest datacenter cluster on Earth — voted in March 2025 to end by-right zoning, forcing every new application through public hearings. Maine became the first state to legislate a statewide cap, banning facilities over 20 megawatts until November 2027. Good Jobs First counts 69 US jurisdictions with active moratoria as of April 2026, most lasting six to twelve months. Georgia's one-year ban passed the Senate floor before adjournment killed it; Virginia's HB 1515 was carried into 2027. The parallel in Europe is no longer theoretical: a Groß-Gerau council formally denied Vantage a 174 MW campus outside Frankfurt on March 31, 2026, ruling disadvantages outweighed benefits — the first major German hyperscaler permit rejection of the AI era. A proposed 700 MW build near Freyenstein in Brandenburg has triggered organised resident protests since February. The historical analogue Romero reaches for is not Facebook 2018 or Cambridge Analytica. It is fracking circa 2013, when a technology with elite consensus support collided with local-government zoning fights and never recovered its social licence. The difference: fracking took a decade to flip public opinion. AI did it in eighteen months.

The longitudinal arc is the part that should alarm anyone modelling AI capex through 2028. In 2023, public AI sentiment in the US was net-positive: most Pew tracking polls showed Americans roughly split between excitement and concern, with knowledge-worker cohorts strongly enthusiastic. By June 2025, Pew's ratio had shifted to 50% concerned versus 10% excited — a five-to-one negative skew. By spring 2026, the Gallup datacenter numbers translated diffuse anxiety into a hard veto on physical infrastructure. Partisan breakdowns are the most strategically important detail because they kill the “regulatory wedge” thesis the industry has leaned on. The Marquette April 2026 survey found majorities opposed across Republicans, Democrats and independents alike. Pew on AI regulation trust shows a mirror image: 54% of Republicans trust the US to regulate AI well versus just 36% of Democrats, but both numbers are below the threshold needed to insulate hyperscalers from bipartisan legislative pressure. When Bernie Sanders and AOC introduced an AI datacenter moratorium bill in May 2026, Y Combinator's Garry Tan responded that “the people who say they want American jobs are trying to block the biggest job creation engine since the interstate highway system” — but the bill nonetheless attracted Republican co-sponsorship interest, a configuration that would have been unimaginable two years ago. The trust-in-AI-companies layer underneath is even worse for incumbents. Gallup, Marquette and Change Research all triangulate to the same finding: roughly two-thirds of Americans do not trust the people running OpenAI, Google, Microsoft, Anthropic or Meta to act in the public interest on AI. The April 2026 attack on Sam Altman's San Francisco residence — covered by Fortune as an inflection point — was less an isolated incident than the predictable downstream of a sentiment curve that has been visible in the polls for nine months. Alex Rafter, a Gartner cloud-infrastructure analyst, captured the industry's bind to the Washington Times: “Microsoft and its peers are caught in an uncomfortable place. Their AI services are in more demand than ever, but the physical infrastructure required to deliver them is becoming politically toxic.” For European observers, two cross-references matter. First, Pew's 24-country comparison shows Americans are now the most AI-anxious population in the developed world — meaningfully more so than Germans, who polled closer to the global median. That asymmetry will not last. Second, the European backlash is already being imported via the same vectors: AlgorithmWatch documented in March 2026 that Germany's grid is being pushed to its limits by hyperscaler buildout, and Frankfurt's rejected Vantage permit cited explicitly the same resource arguments — power draw and water — that drive the US Gallup numbers. The Bavarian permit fights of 2025 over Microsoft and AWS sites near Munich foreshadowed it. The historical comparison that fits best is not Facebook 2018 — that was a discrete scandal cycle — but US public sentiment on fracking between 2010 and 2015, where a frontier technology with elite economic backing lost its social licence one county hearing at a time, and never got it back even as the federal climate calculation shifted in its favour.

For three years, generative image models have been creative toys with one cursed flaw: they could not spell. Logos came out smeared, infographic labels read like Klingon, banner headlines garbled in the second word. That single defect kept image AI out of serious enterprise pipelines, where a misspelled product name on a campaign visual is a legal review event, not a meme. OpenAI's new GPT Image 2, rolled out broadly across the ChatGPT and API surface in late April and pushed into Microsoft Azure AI Foundry on May 19–20, takes a different approach: it runs a reasoning loop before any pixel is drawn. The model plans the image, can search the web for references, generates candidates, then verifies its own output against the prompt. The headline result — near-perfect text rendering and dramatically better instruction following — moves image generation out of the meme economy and into the brand book.

Oliver Korzen, the German analyst behind the This Week in AI newsletter, opened his laptop on Wednesday morning and typed a single prompt into the new GPT Image 2 endpoint: “Build me a complete brand kit for a fictional Berlin fintech called Schwarzwald Capital — wordmark logo, three-color palette with hex codes, and a LinkedIn banner showing the logo plus the tagline ‘Patient capital for impatient founders.' Photorealistic office background, no garbled text.” Forty-two seconds later the model returned a coherent kit. The logo was a clean serif wordmark. The palette swatches carried legible hex values. The banner tagline rendered correctly, including the apostrophe. “This is the first time I have ever shown an AI-generated brand asset to a client without manually retouching the text,” Korzen wrote in the piece that landed in inboxes Thursday morning. He ran three more enterprise-shaped tests in front of his audience: relighting a product shot of a coffee machine from harsh daylight to soft morning studio (no reshoot, no Photoshop), mass-removing a crowd of tourists from a brand photo of the Brandenburg Gate, and converting a messy hand-drawn process diagram into a presentation-ready infographic with correctly spelled labels. The demo lands inside a deliberate enterprise push by OpenAI. Sam Altman teased the launch on X with a one-liner — “we've got something interesting in the works” — at 3 a.m. on April 21, and the model dropped hours later as gpt-image-2 in the API and Codex. DALL-E 3 was retired on May 12, and on May 19 Microsoft made the model generally available in Azure AI Foundry, explicitly pitching the “full thinking mode capabilities — web search, up to 8 consistent images per prompt, and self-verification” at compliance-conscious buyers. Adobe, in a move that would have been unthinkable two years ago, now exposes gpt-image-2 as a partner model inside Firefly Boards. The technical pivot is the reasoning step. Previous diffusion models were essentially pattern-matchers — given a noisy tensor and a text embedding, they iteratively denoised toward something that looked like the prompt. They had no internal representation of meaning. GPT Image 2, built natively on the GPT-4o-class architecture, plans the scene as a structured intermediate before the first pixel exists. OpenAI claims 99% text-rendering accuracy in English and above 90% in Chinese, Japanese, Korean, Hindi, Bengali, and Arabic — a leap that resembles, in scale, the GPT-2 to GPT-3 jump on multi-step reasoning benchmarks five years ago. For DAX40 marketing teams who have spent two years budgeting around “AI for ideation only, designer for final asset,” that ratio is now upside-down. The bottleneck is no longer the model. It is the brand-governance workflow wrapped around it. Korzen's closing line in the briefing was less giddy than the prompt-bros on X: “The interesting question is not whether your agency uses this. It is whether your in-house team can move fast enough to make the agency optional.”

Four workflows from Korzen's walkthrough map almost directly onto line items in a DAX40 marketing budget. The first is brand-kit generation: one prompt produces a logo, palette, and primary social asset that holds together visually. For a Henkel detergent sub-brand launching in a new market, that compresses a process which previously meant two weeks with a creative agency to under an hour of prompt engineering plus legal review. Henkel was already a public reference customer for Adobe Firefly's Custom Models programme; the question now is whether the on-brand fine-tuning Adobe sells as a moat survives a model that can hold brand consistency from a single reference image plus a prompt. The second workflow is product-photo relighting. GPT Image 2 takes a product shot and re-renders it under a different lighting setup — golden hour, soft studio, hard noon — while preserving the object's geometry and surface. For BMW's configurator photography, where a single trim variant historically requires a physical reshoot or hours of CGI work, the implication is direct: the variant matrix collapses. A pricing model that charges roughly $0.21 per high-quality 1024×1024 image makes generating 200 lighting variants of one model cost less than catered lunch for the photo studio. The third is mass background cleanup. Korzen's Brandenburg Gate test removed approximately forty tourists from a crowded plaza shot, leaving a coherent paved surface and intact architectural detail. For Lufthansa marketing — long dependent on either stock photography or expensive empty-airport shoots at 4 a.m. — clean editorial imagery becomes a prompt away. The legal team will have opinions; the budget owner will have a smile. The fourth, and the most strategically important, is presentation-ready infographics. Until now, no image model could be trusted with a chart label, an arrow, or a numeric callout. GPT Image 2's reasoning loop verifies labels before output. McKinsey-style decks built end-to-end inside ChatGPT — long the dream of every consulting analyst — are now a single-prompt request away from a rough draft. The Information Architects of an Accenture or a Bain will not lose their jobs to this. The intern who spent Friday night cleaning up the labels on slide 47 might. The pricing structure deserves attention. OpenAI is charging $8 per million image input tokens, $30 per million image output tokens, with cached image input at $2 per million — yielding roughly $0.006 per low-quality output, $0.053 at medium, and $0.211 at high. Azure mirrors the same tier. Adobe Firefly's enterprise contract runs roughly $0.02–$0.10 per image with a ~$1,000/month minimum. The per-unit economics now favor OpenAI on the high end and Firefly on bulk low-fidelity production — which is exactly why Adobe has chosen to host gpt-image-2 inside Firefly Boards rather than fight it. The CMO at a DAX40 firm renewing a six-figure Firefly contract this quarter will note the choice.