So, as the conference was all about AI, I decided to let an AI do a summary of all the google cloud next videos on youtube. For those that couldn’t go, this should be a reference of the conference and a summary of each session to allow you to pick and choose what to watch.

So here is the list, all completely done via AI.

☁️ Google Cloud Next ‘26: The Agentic Cloud

Conference: Google Cloud Next ‘26 Dates: April 22–24, 2026 Location: Las Vegas, NV Scale: 32,000+ attendees, 3 keynotes, 25 spotlights, 700+ breakout sessions, 490+ sponsors Theme: “The Agentic Era” — moving AI from pilot to production across the enterprise

All session recordings are available on the Google Cloud YouTube channel.

The Big Picture

Google Cloud Next ‘26 was the largest Next yet, and the message was unambiguous: the era of AI experimentation is over. Every major announcement centered on agents — building them, scaling them, governing them, and running them on purpose-built infrastructure. CEO Thomas Kurian laid out a blueprint for the “agentic enterprise” built on a unified stack: chips designed for models, models grounded in your data, agents built with those models, and the whole thing secured by the infrastructure. It’s the same stack Google uses for Search, YouTube, and Chrome, now offered as a platform for everyone else.

The conference had three keynotes (opening, developer, infrastructure), and the content broadly fell into five themes: the new agent platform, next-gen infrastructure, the agentic data cloud, developer tooling, and customer stories showing agents in production.

Keynotes

Opening Keynote: Building the Agentic Enterprise

📺 Watch · 1:39:02 📺 10-minute recap · 10:29 📺 Highlights reel · 0:56

Thomas Kurian opened by noting that nearly 75% of Google Cloud customers now use AI products in production, declaring the experimentation phase over. Sundar Pichai announced Google plans to invest $175–185 billion in CapEx this year — nearly 6x the $31 billion in 2022 — with over half of ML compute directed at the cloud business. The headline announcement was the Gemini Enterprise Agent Platform, described as “mission control for the agentic enterprise” — a unified system for building, scaling, governing, and optimizing agents. Specific capabilities include: the low-code Agent Studio for building agents with natural language, Agent Registry for indexing every agent and tool across an org, Agent Identity with unique cryptographic IDs and auditable authorization policies, Agent Gateway for enforcing IAM policies on agent traffic, and Gemini Enterprise Projects giving teams a persistent high-fidelity workspace with Deep Think for complex tasks. Gemini 3.1 Pro (in preview) was announced as the new reasoning model optimized for complex workflow orchestration, and support was added for Claude Opus 4.7 — including a NASA partnership using Gemini Enterprise Agents for Artemis II flight readiness. The Agentic Data Cloud introduced four innovations: Knowledge Catalog (universal context engine that auto-tags PDFs and images on ingest), a new Data Agent Kit, Lightning Engine for Apache Spark (2x price-performance over previous market leader), and the Cross-Cloud Lakehouse built on Apache Iceberg, enabling data to stay in any cloud while agents reason across it. On security, Francis DeSouza announced an AI-native agentic SOC and officially welcomed Wiz to Google Cloud; Wiz’s AI APP (AI Application Protection Platform) uses a team of AI agents to autonomously identify, investigate, and fix security risks using Wiz’s security graph. Customer spotlights included Walmart (Gemini Enterprise on Pixel Fold for store leaders), Home Depot (Magic Apron shopping agent), Virgin Voyages (Project Ruby crew assistant), and Citadel Securities (2–4x faster workloads on Ironwood TPU).

Developer Keynote

📺 Watch · 1:07:21 📺 14-minute recap · 14:01 📺 10-minute version · 10:00 📺 Under a minute · 0:48

Brad Calder hosted a hands-on, demo-heavy keynote built around planning a simulated Las Vegas marathon through agents — an intentionally practical framing for multi-agent orchestration, stateful sessions, memory, observability, and governance. The entire demo used ADK (Agent Development Kit), with every Google Cloud service now MCP-enabled by default, so agents can communicate with any GCP resource via the Model Context Protocol. The keynote showed: the Agent Designer (low/no-code UI for building and previewing agents before generating SDK code), Google Cloud MCP servers (connecting agents to Maps, Finance, and other APIs with security included by default), Agent Skills (modular YAML+markdown capability bundles that agents load at runtime rather than stuffing into a single giant prompt), Event Compaction in ADK (Gemini-powered context summarization to keep long-running agents from hitting token limits), Cloud Assist (the AI agent for designing, debugging, and optimizing GCP workloads — demo showed it diagnosing a broken agent via one-click investigation from a Monitoring alert), and A2UI (open standard from Google for agents to dynamically generate rich UI components rather than plain text). New developer capabilities included: Agents CLI (terminal-native entry point that works with Gemini CLI, Claude Code, Codex, and other coding agents), Dart support for Firebase Functions (GA), and AI Studio Build (vibe-coded apps deployed to Cloud Run with Firestore + auth, with design previews and “Tab Tab Tab” prompt autocomplete). The livestream following the keynote featured Harrison Chase (LangChain), Michele Catasta (Replit), and Logan Kilpatrick (AI Studio).

Infrastructure Keynote: The Future of AI Infrastructure

📺 Watch · 32:54

Amin Vahdat (VP, Google Infrastructure) delivered the infrastructure keynote, framing Google’s unique ability to build vertically integrated end-to-end — from energy and data centers through hardware, software, and models — as the competitive moat. The major announcement was the 8th-generation TPUs, introduced as two separate chips for the first time: TPU 8t (training) and TPU 8i (inference). TPU 8t targets trillion-parameter frontier model training, delivering 121 exaflops of native FP4 compute in a single 9,600-chip superpod — a 3x increase over Ironwood — with 2 petabytes of shared HBM (enough to hold the Library of Congress digital collection 100 times), 2x ICI scale-up bandwidth, and 4x scale-out DCN bandwidth. Combined with Managed Lustre (10 TB/s throughput) and TPU Direct for RDMA access to storage, training at up to 1 million chips in a single cluster is now possible via JAX + Pathways. TPU 8i targets the “latency wall” of agentic inference and reinforcement learning, using a new Boardfly topology for ICI networking that cuts latency for all-to-all communication (critical for MoE and chain-of-thought reasoning) by up to 50%, with a 1,152-chip pod delivering 10x the floating-point exaflops of its predecessor and 7x the HBM capacity. Google Cloud Axion (Arm-based CPU) was also highlighted for general-purpose workloads. Both 8t and 8i will be available to Google Cloud customers by end of 2026. Citadel Securities demonstrated workloads running 2–4x faster with Ironwood, cutting research cycles from weeks to hours.

Agents & Agent Platform

The dominant theme of the conference. Google introduced the Gemini Enterprise Agent Platform as a comprehensive system for the full agent lifecycle.

Platform & Architecture

📺 Build agents with Agent Platform · 5:31 — Demo showing the Agent Designer flow: design a low/no-code agent visually, preview its behavior, then export pre-populated ADK Python SDK code as a starting point. The demo connected the agent to Google Maps via a Google Cloud MCP server in just a few lines of code, then added Agent Skills (YAML metadata + markdown instructions) for mapping, GIS, and race planning — allowing the agent to selectively load capabilities rather than maintain a single monolithic system prompt. The agent used gen_random_route() and GeoJSON processing tools to find viable marathon routes in Las Vegas.

📺 Google Announces Gemini Enterprise Agent Platform · 5:38 — Sundar Pichai’s announcement clip from the opening keynote. Key stats: 75% of all new code at Google is now AI-generated (up from 50% in fall 2025); Google recently completed a complex code migration 6x faster using a system of planner, orchestrator, and coder agents; Google’s security operations center now auto-triages tens of thousands of threat reports monthly, reducing threat mitigation time by over 90%.

📺 From Prototype to Production: Building with Gemini Enterprise Agent Platform · 22:23 — Dave Elliott and Addy Osmani in conversation about the four pillars of the platform: Build (ADK supports Python, Go, TypeScript, and Java), Scale (serverless Agent Runtime with Sessions for stateful users and Memory for persistent context), Govern (Agent Identity + Agent Gateway with IAM policy enforcement on all agent traffic), and Optimize (Agent Observability with production-integrated evals). The session emphasized that going from prototype to production requires solving identity, governance, and memory simultaneously — not as afterthoughts.

📺 Build and share no-code agents · 4:58 — Demo of sharing the marathon planner agent to Gemini Enterprise via the Agent Registry — agents deployed to Agent Runtime are automatically discoverable. The Agent Designer in Gemini Enterprise lets any user create a no-code agent from a prompt with a document as context, and the resulting agent is automatically registered back in the Registry for other agents and apps to call.

📺 Build and share no-code agents (full session) · 21:30 — Extended walkthrough with Ines Envid (Sr. Director, PM, Gemini Enterprise Agent Platform) explaining how Agent Gateway was built on Google’s “One Network” fabric — the same standardized networking layer used across all GCP compute modalities. Agent Gateway gives agents network-level identity that is immutable and auditable, unlike broad-purpose service accounts.

Agent Development

📺 A closer look at Agents CLI · 19:23 — Shubham Saboo (AI Product Manager, Google Cloud) explains the three big developer launches. Agents CLI is a global install that works as the entry point for any coding agent (Gemini CLI, Claude Code, Codex, etc.) — install once and it gives any connected coding agent complete knowledge of ADK and the full Agent Platform. Skills are highlighted as a platform-wide standard (not Google-specific): modular capability bundles that agents load at runtime, replacing the old pattern of stuffing everything into a single system prompt. The CLI handles the full agent development lifecycle: scaffolding, local testing, deployment, and debugging.

📺 Creating multi-agent systems · 6:08 — Demo of a three-agent system: a Planner agent, an Evaluator subagent (uses a separate model with limited context to grade routes on both deterministic criteria like exact marathon distance — 26 miles 385 yards / 42.195km — and non-deterministic criteria like community impact), and a Simulator agent. The Planner invokes the Evaluator as a tool. A2UI (Agent-to-User Interface), an open standard from Google, lets agents dynamically build rich UI components at runtime — shown generating a map visualization with evaluation scores — eliminating custom dashboard code.

📺 Enhancing agents with memory · 5:04 — Shows how to add session management and memory to an ADK agent in under 20 lines of code: attach to Agent Platform’s Memory Bank (an enterprise-ready, fully managed memory service) so the agent automatically creates long-term memories from completed tasks. Also demonstrated: a data engineering agent that built a complete RAG pipeline from a prompt — chunking documents with Document AI, generating embeddings, and storing them in AlloyDB — using Lightning Engine for Apache Spark for high-throughput processing.

📺 Securing agents · 9:40 — Demo of Agent Gateway in action: each deployed agent gets a unique, immutable Agent Identity (unlike service accounts, which are like “a single all-access hotel key shared by the event crew”). Egress policies govern what tools, models, and other agents each agent can call. The demo added a read_only_finance: true policy to the Planner agent to prevent it from writing to the financial database — policy enforcement happens at the gateway level without code changes.

📺 Debugging agents at scale · 5:20 — End-to-end observability workflow: added Event Compaction to the Simulator agent (ADK feature where Gemini periodically summarizes the agent’s workflow to reduce context size). When the agent broke, a one-click button in Cloud Monitoring logs launched a Cloud Assist Investigation — Gemini Cloud Assist gathered logs, examined agent runtime infrastructure, identified the root cause (malformed payloads to the Gemini model API), pointed to the specific line of code, and proposed a fix in natural language.

📺 Automating Creativity: Building Gen Media Agents with ADK and MCP · 20:09 — Katie Nguyen (DevRel, Gen Media) built a Character Story Agent using ADK that orchestrates Veo (video generation), Imagen (image generation), Lyria (music), and Gemini Audio (narration) into a coherent multi-asset story. The agent maintains memory across asset generation steps to ensure visual and narrative consistency — Gemini tracks which character images, video clips, and music were generated and references them when creating new assets to keep the story cohesive.

Agents in the Workplace

📺 Agents at Work: YouTube TV Customer Experience · 4:57 — Live demo of YouTube TV’s voice support agent, currently in production serving 100% of users. The agent handled complex product logic (recommending the YouTube TV Sports Plan at $18/month less than the base plan for a sports-only viewer), seamlessly switched to Spanish mid-call for a second caller, and offered to text a signup link — all without a software engineer touching a configuration. Agent Studio’s Visual Builder lets the YouTube TV team make changes via natural language rather than code.

📺 Agents at Work: Workspace Intelligence · 3:45 — Demo of Gemini Enterprise in Google Chat as a “Command Center”: the Regional Operations Agent proactively surfaced urgent tasks, linked relevant files, and flagged a 4pm deadline — all without the user opening a single extra tab. A skill-tagged prompt (@regional_campaign_skill) triggered an agent that cross-referenced emails, chats, documents, and live HubSpot win/loss data to produce a fully branded Google Slides deck, with citations for every source it pulled.

📺 Agentic Data Cloud · 5:28 — Demo breaking three data barriers: (1) Dark data — the Knowledge Catalog automatically extracts entities (recipe, ingredient, allergen) from supplier PDFs and maps hidden connections across documents, catching a soy allergen hidden inside “ingredient base 204” that a simple GenAI search would miss; (2) Data silos — the Cross-Cloud Lakehouse joined a Google-hosted allergen schema with a customer loyalty list on AWS S3 Iceberg without migrating any data; (3) Manual code — the Data Agent Kit wrote the data pipeline and SQL from a prompt in natural language.

📺 From systems of intelligence to systems of action: Yasmeen Ahmad · 20:24 — Yasmeen Ahmad (Managing Director, Data Cloud, Google Cloud) framed the shift from “systems of intelligence” (dashboards, predictive models) to “systems of action” (agents that drive real business outcomes). Key insight: historically only 10–20% of data insights made it into production. With agentic systems, data is now active in the reasoning loop — real-time, and acting through MCP tools into ledgers, operational systems, and marketing platforms. The session covered how customers need machine-readable, structured, real-time data as the foundation for agents that can actually take action.

Infrastructure & Hardware

8th-Generation TPUs

📺 Introducing 8th Generation TPUs: Purpose-Built for the Agentic Era · 1:22 — Announcement trailer.

📺 How Google’s 8th Generation TPUs Power the Agentic Era · 11:08 — Amin Vahdat’s technical deep dive. TPU 8t (training): 121 exaflops native FP4 compute, 2 petabytes shared HBM in a 9,600-chip superpod (3x peak performance per superpod over Ironwood), 2x ICI scale-up bandwidth, 4x scale-out DCN bandwidth. Scales to over 1 million chips in a single training cluster via JAX + Pathways, with 134,000 chips having nonblocking communication delivering 1.6 million exaflops on the Virgo networking fabric. Managed Lustre with 10 TB/s throughput eliminates storage bottlenecks. TPU 8i (inference): pioneered the Boardfly topology for ICI networking, achieving up to 50% latency improvement for communication-intensive workloads (MoE, chain-of-thought); 1,152-chip pod with 10x the floating-point exaflops and 7x the HBM capacity of the previous generation. Both chips available to Google Cloud customers by end of 2026.

📺 Google TPU 8t and TPU 8i: Purpose-built for the Agentic Era · 2:45 — Companion overview clip. Confirms that TPU 8i’s Boardfly topology is a direct response to the “latency wall” — where traditional architectures struggle with real-time autoregressive decoding and chain-of-thought reasoning at scale. Google emphasized this is not a one-size-fits-all approach: both chips handle the full AI lifecycle (pre-training, RL, fine-tuning, serving) but are purposefully optimized for their primary workloads.

📺 Create infrastructure from an intent with AI assistance · 5:19 — “Vibe clouding” demo: a single natural-language prompt (“convert this Cloud Run service to the equivalent on GKE with a Gemma 4 model hosted in the same cluster”) triggered Cloud Assist via MCP to generate a complete Kubernetes manifest — including optimal inference server configuration, scaling policies, and best practices — and deploy it to GKE. The Gemma 4 inference server and converted runner service both appeared in the GKE console within minutes. Cloud Assist connected to Google Cloud resources via MCP directly from the IDE.

📺 How Salesforce Research Cut AI Training Costs by 42% with Google Cloud Managed Lustre · 2:10 — Salesforce case study training a Llama 3.1 8B model. Problem: GPUs sitting at 40% utilization due to storage I/O bottlenecks. Solution: Google Cloud Managed Lustre alongside a Vertex AI training cluster. Result: IO latency dropped by 75%, GPU utilization increased by 70%, models now train 1.5x faster, and overall training cost dropped 42% by saturating hardware they already paid for.

Developer Sessions

AI Studio & Vibe Coding

📺 Vibe coding to production: Logan Kilpatrick on the evolution of AI Studio · 43:22 — Logan Kilpatrick (Google) described AI Studio’s evolution through three eras: prompt-to-prototype, then production (18 months ago), and now the “Build” tab as a full vibe coding platform. New features shown: Design Previews (multiple UI iterations generated before building so you can choose a direction); an “I’m Feeling Lucky” button that generates a first app idea connected to the Google ecosystem; and Tab Tab Tab (Flash-powered prompt autocomplete — start typing your app idea and keep pressing Tab to let the model complete and extend it). Apps built in AI Studio can be deployed to Cloud Run with Firestore and auth, largely for free. Millions of people are building in AI Studio, and the Build tab can now target specific models like Imagen for image generation.

📺 From idea to production: Building AI apps with GDE Tomek Porozynski · 10:34 — Google Developer Expert Tomek Porozynski (Deepsense AI, Poland) demonstrated building a multi-voice audiobook generator using Gemini text-to-speech APIs: the system detects first vs. third-person narration, assigns distinct voices to characters, splits text into character-attributed chunks, and concatenates the final audio. He used Gemini CLI with Skills (specifically the Live API skill and Vertex AI skill) for rapid prototyping, describing the Skills pattern as the key to building and testing components incrementally before composing them. The project is open-sourced and was shown at DevFest talks.

📺 Startup founder spotlight: Michele Catasta, Replit · 19:53 — Replit President and Head of AI Michele Catasta described how Replit has moved away from IDE-style interfaces entirely — most internal tools at Replit are now built with Replit Agent (50%+ of time in front of screen). Catasta’s view: developers are becoming “managers of agents,” expressing intent in natural language while a swarm of agents executes. Replit’s “Vibe coding” terminology was coined because users interact with the AI product — they don’t look at the code itself. He highlighted Gemini’s long context window and advanced reasoning as key enablers, and called out multimodality as a differentiator that has been strong in Gemini from early on.

Software Development & DevOps

📺 Automating the SDLC with LangChain, LangSmith, and Gemini · 20:44 — Harrison Chase (CEO, LangChain) introduced the concept of the agent harness layer — the scaffold connecting an LLM to its tools, environment, and memory. Key insight: tuning the harness (no model changes) can dramatically outperform fine-tuning; LangChain’s DeepAgents went from 30th to 5th on Terminal Bench just by optimizing the harness. Chase explained how LLMs are well-suited to interacting with file systems, so harnesses often use virtual file systems to expose knowledge from databases in a format models handle best. The session covered deploying LangGraph on Google Cloud’s Reasoning Engine — a secure managed environment for running LangChain/LangGraph apps as part of Gemini Enterprise Agent Platform, solving the statefulness and long-running challenges that prevent agent prototypes from reaching production.

📺 How Wiz & Google Cloud Build AI You Can Trust · 17:51 — Post-acquisition overview with Salmon Ladder (Wiz, product marketing for Wiz Code). Wiz has three product lines: Wiz Cloud (security posture and visibility), Wiz Defend (security operations, threat response), and Wiz Code (shifting security into the developer workflow). The new theme is “shifting down” — embedding security directly into agents rather than “shifting left” onto developers. Wiz Code plugins scan AI-generated code the moment it’s produced in any coding environment (Gemini CLI, Claude Code, Codex, VS Code) for misconfigurations, hardcoded secrets, and vulnerabilities before the first commit. The developer-to-security-team ratio is now 1,000:1 with AI-generated code, making automated scanning no longer optional.

📺 Scaling to Dozens of PRs a Day: Governance and Compliance in the Agentic Era · 10:26 — Jeff Whelpley (CTO, GetHuman; Google Developer Expert) described the governance challenge when agents are generating high volumes of PRs: traditional code review, DevOps pipelines, and compliance processes become bottlenecks. The session covered how enterprises are rethinking their review gates and CI/CD pipelines to handle agent-generated code at scale — separating human review of intent from automated validation of output, and using the Gemini Enterprise Agent Platform’s policy enforcement layer to ensure agent-generated PRs meet compliance requirements before they reach human reviewers.

📺 Write Dart everywhere: Support for Firebase Functions is here! · 21:06 — Firebase Functions now supports Dart (GA), meaning Flutter developers can write their entire stack — frontend, mobile, and backend — in a single language without touching Node.js or Go. The session noted that a plurality of new apps on the Google Play Store are Flutter apps. Dart in Firebase Functions also means AI/LLM code running in the Flutter frontend can be moved server-side with minimal refactoring, and Gemini CLI Skills (including the Live API skill and Vertex AI skill) work with Dart-based Firebase Functions for backend AI features.

📺 Delivery Navigator: Now in Public Preview · 3:24 — Delivery Navigator is a knowledge base and methodology framework for Google Cloud delivery projects — a comprehensive library of best practices, templates, and project-specific guidance. It integrates with existing project management tools (Jira, etc.) rather than replacing them. Now entering public preview for direct Google Cloud customers (previously partner-only). Roadmap includes AI-driven project insights, customized delivery plan generation, and specialized agents that execute delivery work directly in the GCP environment.

Models & Gen AI

📺 From smartphones to Raspberry Pi: Running Gemma 4 anywhere · 10:23 — Omar Sanseviero (Google DeepMind) covered the Gemma 4 family: models from 2B to 31B parameters, with over 40 million downloads in the three weeks since launch (largest open model release ever). The 2B models are designed for phones and support audio input and speech-to-text; larger models handle images and video but not audio. All models are multimodal and trained on over 140 languages. The Gemmaverse — community fine-tunes and derivatives — was highlighted as a key value of open weights: examples included Gemma fine-tuned for Quechua (an indigenous Peruvian language) and other low-resource languages. The largest (31B) model runs on a consumer GPU, making it accessible for local development.

📺 A closer look at Gemma 4 with Baseten and NVIDIA · 18:23 — Jay Rodge (NVIDIA) announced two Google Cloud / NVIDIA hardware partnerships: Google Cloud will be among the first providers to offer Vera Rubin (NVIDIA’s next-gen inference and training chip, arriving H2 2026), and Google Cloud is adding Blackwell GPUs (RTX PRO 6000) with 96GB VRAM, enabling multiple models on a single GPU. Philip Kiely (Baseten), one of the first NVIDIA Dynamo users, discussed running billions of inferences at scale and the challenge of building low-latency, high-reliability AI application experiences — emphasizing that inference is now arguably more important than training for the practical delivery of AI products.

📺 Nano Banana, Veo, and Lyria: Mastering the Google gen media stack · 19:25 — Khulan Davaajav (Product Marketing Manager, Gen Media) introduced the gen media model family: Nano Banana (image generation and editing), Veo (video generation), Gemini Audio (transcription, text-to-speech), and Lyria (music generation). New model updates are shipping nearly every week. The demo showed a complete multi-modal short film built from scratch using all four model types — the session walked through prompting techniques for each: character consistency across Nano Banana image frames, motion and physics in Veo, voice casting and narration in Gemini Audio, and emotional tone in Lyria music generation.

📺 From Raw Video to Real Physics: The Google Cloud AI Breakdown · 5:14 — Shaun White (three-time Olympic snowboarding gold medalist) joined Google Cloud to demonstrate a system built in collaboration with Google DeepMind that extracts physics data from flat 2D video. Given a clip of White’s 2017 switch cab double flip 1440 (four full rotations + two flips in under 3 seconds), the system generates a 3D spatial pose model from 2D footage, then tracks flight dynamics, rotational velocity, and tuck compression frame by frame using Gemini. The system was framed as a blueprint for any industry needing to extract structured physics or biomechanics understanding from raw video.

Customer Stories

Production deployments were front and center. These aren’t proofs of concept — they’re agents running at scale in real enterprises.

📺 Unilever designs and deploys agents at scale with Google Cloud AI · 1:43 — Unilever (serving 3.7 billion people daily) is building a multi-agentic procurement solution through their Horizon3 Labs, co-created with Google. The Competitive Buying agent orchestrates multiple specialized agents through a single UI using Gemini Enterprise, ADK, and underlying Gemini models — enabling procurement colleagues to conduct sourcing analysis in minutes instead of days. Their stated goal is “co-creating agentic foundations” so every agent is architected for performance, scale, security, and observability from day one.

📺 The Home Depot turns fragmented customer service into seamless journeys · 1:39 — Magic Apron is Home Depot’s digital agent covering the full customer journey: inspiration, research, shopping, buying, and post-purchase support. Built on Gemini Enterprise as the “bedrock” for speed and power. The agent handles wayfinding, product knowledge, and compatibility questions both in-store and on the website. Home Depot reports higher conversion rates when customers use Magic Apron vs. when they don’t, and sees Gemini Enterprise’s agentic backend as the path to blending physical and digital service channels.

📺 How Walmart is empowering store leaders to better serve customers · 1:29 — Walmart is deploying Gemini Enterprise connected to their enterprise data on a Pixel Fold device for store and supply chain leaders — giving them answers in seconds rather than hours while staying on the sales floor instead of at a desk. The framing: “people-led and tech-powered” with AI helping leaders make decisions with clearer insights without pulling them away from their teams.

📺 Citadel Securities runs workloads up to 4x faster with Ironwood · 1:29 — Citadel Securities (trading over $500 billion per day) built a cloud-based research environment on Google Cloud TPUs, specifically Ironwood. With TPU Ironwood, they can run thousands of parallel chips for a single workload. Workloads that previously took weeks or days now complete in hours or minutes, at 2–4x the speed and 30% lower cost. The stated goal: empower quantitative researchers to test all their ideas without being limited by platform scale or economics — only by their own creativity.

📺 Virgin Voyages empowers its crew with a world-class AI assistant · 1:45 — Project Ruby is Virgin Voyages’ AI assistant built on Gemini Enterprise, serving both crew (answering any question about operations, service, and the ship) and sailors (as “Rovey,” a personal concierge from inspiration through voyage end). A key infrastructure challenge: ships aren’t connected to fiber, so they use Google Distributed Cloud Edge for data resiliency — the intelligence stays online even when the ship is offline. Results: AI stack reduced production timeline by up to 60% and contributed to a 28% month-over-month increase for a record sales quarter.

📺 New Way Now: KeyBank saves 400+ hours on deployments · 2:39 — Azeem Sheikh (Lead Enterprise Architect, KeyBank) described using Gemini CLI across the development lifecycle: cutting proof-of-concept time from months to weeks or days, reducing PRs from 5–6 per change to fewer than 2, generating detailed documentation for legacy code (critical in a financial institution where documentation drift creates risk), and building a shared internal framework that saves 300–400 hours per deployment cycle.

📺 How Owl AI is improving sports officiating with Gemini · 2:35 — Owl AI (CEO Josh White) uses Gemini’s multimodal video understanding to assist sports officiating. Human review of complex calls can take 30 seconds to 5 minutes; Owl AI’s system completes the same review in seconds. Multiple Gemini instances analyze the same play simultaneously, combining outputs into an aggregate score for penalty/no-penalty decisions. At the 2025 X Games, Owl used Gemini voice models to commentate, provide data science predictions, and translate content into four languages in real time. In their first week, the team wrote 30,000 lines of code using Gemini.

📺 How UKG uses AlloyDB to scale its People Fabric platform · 2:59 — UKG (80,000+ customer organizations, 20 million daily transactions, 730 million shifts scheduled monthly, $1 trillion in annual payroll calculated) needed a unified data platform for their People Fabric HR/workforce product. They chose AlloyDB for transactional data and BigQuery for analytics. Results: 2x CPU speed improvement, sufficient scale for their largest enterprise customers, and meaningful cost efficiencies vs. their previous system.

📺 New Way Now: Honeywell paves the way from automation to autonomy · 3:01 — Suresh Venkatarayalu (SVP, CTO, and President, Connected Enterprise, Honeywell) presented Honeywell’s path from industrial automation to full autonomy. The transcript was minimal (primarily a music/title card clip), but the framing aligns with Honeywell’s broader industrial AI strategy: using Google Cloud to move from rule-based automation to AI-driven autonomous operations in industrial and building technology environments.

Special Content

Acquired Podcast Live

📺 Acquired’s Ben Gilbert and David Rosenthal live from Google Cloud Next · 20:25 — Ben Gilbert and David Rosenthal (Acquired podcast, ~11–12 hours of prior Google coverage across three episodes) discussed the conference from a business and market perspective. Key observations: the TPU 8t/8i split was called out as an “underrated” announcement — when the chips were designed, training was the dominant workload; by the time they ship, inference compute has eclipsed training compute (evidenced by every Google Search query now triggering an AI inference). The compute loads for inference have likely crossed the line for training across all major workloads. Ben noted this probably warrants a “Part 4” Google episode on Acquired. David framed it as: a few years ago the industry said “inference is fine, worry about that later” — that position is no longer tenable.

Google Antigravity (Gaming/Interactive)

Google launched an interactive game called “Antigravity” at the conference, with a series of tutorial and gameplay videos:

📺 Tutorial: Liftoff with Google Antigravity! · 1:29 📺 Level 1: Get started with Antigravity · 4:10 📺 Level 2: Enable services for AGY apps with Firebase · 4:15 📺 Level 3: Manage a team of agents · 4:34 📺 Beat the final boss · 2:05

Other

📺 AI Agent Challenge 2026 | Google for Startups · 3:28 — Google for Startups launches an agent-building challenge.

📺 Transform Studio: Behind the scenes at Google Cloud Next 2026! · 1:09 — Behind-the-scenes look at the conference production.

📺 What Will You Build? | Google Cloud Next ‘26 · 1:01 — Conference closing hype reel.

Key Takeaways for Developers

Gemini Enterprise Agent Platform is the centerpiece — Google is betting that enterprises need a unified platform for the full agent lifecycle (build, test, deploy, govern, scale), not a collection of point tools. The platform integrates with Cloud Run, Firebase, and the full GCP stack. It includes Agent Registry, Agent Gateway, Agent Identity, Agent Runtime with Sessions and Memory, Agent Observability with production evals, and the low-code Agent Studio. ADK supports Python, Go, TypeScript, and Java.
Every Google Cloud service is now MCP-enabled by default — This was announced in the Developer Keynote and is significant: any MCP-compatible agent (built on any framework) can now connect to Google Maps, Finance APIs, Cloud Storage, AlloyDB, BigQuery, and the full GCP service catalog as tools, with security included by default.
Agents CLI brings agent development to the terminal — Following the trend of CLI-first developer tools (Claude Code, Codex CLI), Google now has Agents CLI for building and testing agents locally before deploying to the cloud. It’s a global install that works as a context/skills provider for any connected coding agent — not a replacement for Gemini CLI or Claude Code, but a complement.
Skills replace giant system prompts — The Skills pattern (YAML metadata + markdown body, loaded at runtime based on context) is now a cross-industry standard and a first-class concept in the Gemini Enterprise Agent Platform. This is the architectural answer to “how do you build an agent that can do 50 things without context pollution.”
8th-gen TPUs split training and inference — TPU 8t for training (121 exaflops FP4, 9,600 chips, 2 PB HBM), TPU 8i for inference (Boardfly topology, 50% latency improvement for MoE/CoT workloads, 1,152 chips, 10x exaflops per pod). The split reflects the industry reality that inference compute has now eclipsed training compute by volume, especially with AI embedded in Search, YouTube, and every enterprise workflow.
Dart goes full-stack on Firebase — With Dart support for Firebase Functions (GA), Flutter developers can now write their entire stack in a single language. A plurality of new Play Store apps are already Flutter — this removes the last major reason for Flutter devs to drop into another language for backend logic.
Vibe coding gets a production path — AI Studio Build now supports deploying apps (with Firestore + auth) directly to Cloud Run, largely for free, with design previews and prompt autocomplete (Tab Tab Tab). The gap between “prototype in AI Studio” and “run in production” is closing fast.
The Wiz acquisition shapes the security story — Wiz Code embeds in every major AI coding environment and scans AI-generated code the moment it’s produced. With AI generating code at 1,000:1 developer-to-security-team ratios, runtime scanning in the agent itself (“shifting down” vs. “shifting left”) is the only viable security model.
ADK + MCP = multi-agent interoperability — The Agent Development Kit now supports MCP, which means agents built on Google’s platform can interoperate with the broader MCP ecosystem (Anthropic, OpenAI, open-source tools). Combined with every GCP service being MCP-enabled, this creates a bidirectional bridge.
The Agentic Data Cloud is the data strategy counterpart — Knowledge Catalog (auto-tagging PDFs/images on ingest, zero manual engineering), Cross-Cloud Lakehouse (Apache Iceberg, data stays in place across any cloud), Data Agent Kit (write data pipelines from natural language prompts), and Lightning Engine for Apache Spark (2x price-performance) collectively address the “my data isn’t agent-ready” problem that blocks most enterprise agent deployments.

Sources

[[Topics/AI]]
[[Daily Notes/2026-04-28]]

Google Cloud Next 2026