376: Google Gemini Agents Gemini, Gemini Confused

Welcome to episode 376 of The Cloud Pod, where the forecast is always cloudy! Justin, Ryan, and Matt are in the studio this week bringing you the latest in cloud and AI news, including a slew of new models (yes, again), the always impressive stats from Amazon’s Prime Days, and some interesting updates to Lambda. There’s so much more where that came from, so let’s get started!

Titles we almost went with this week

• AWS Buys DuckDB, Ducks the Distributed Systems Fight
• Yo Dawg, We Put Gemini in Your Gemini
• AWS Lambda Webs Its Way Into Your Servers
• Serverless Finally Speaks Fluent Express
• DynamoDB Counts to 59 Trillion, Doesn’t Even Break a Sweat
• Gremlin Teaches AI to Break Things Nicely
• Prime Day Flexes AWS Muscles, Trillions of Requests Deep
• AWS Graviton Carries Prime Day Like It’s No Big Deal
• Maia 300 Says Nvidia Who? Azure’s New Chip Flex
• Nadella Says Trust No AI, Keep the Kill Switch Handy
• LiveTaskCount Crashes the Party, RunningTaskCount Goes Free
• AWS Boxes In Agents Before They YOLO Your Files
• Gemini, Meet Gemini
• AWS saw your edge Cloudflare and said “hold my beer” with Lambda Web Functions

A big thanks to this week’s sponsors:

We’re sponsorless! Want to get your brand, company, or service in front of a very enthusiastic group of cloud news seekers? You’ve come to the right place! Send us an email or hit us up on our Slack channel for more info.

Follow Up

01:24 Licensing costs driving 90 percent of VMware users to explore options: Survey

• New Rimini Street survey data confirms the ongoing VMware exodus trend, with 90 percent of 300 surveyed organizations now actively exploring alternatives due to licensing costs, and 54 percent citing Broadcom’s termination of perpetual license support.
• Cost increases reported since the Broadcom acquisition vary widely, from 100 to 300 percent for most customers to 1,000 percent in extreme cases, reinforcing earlier reports rather than introducing new pricing-shock figures.
• The survey adds a more complete picture of why migrations remain difficult in practice: operational complexity (40 percent), multi-vendor management (38 percent), security concerns from an expanded attack surface (37 percent), and skills gaps (37 percent) are now quantified as the leading barriers.
• This update provides useful context for listeners tracking the broader VMware alternatives narrative, since it shows cost pressure hasn’t eased a year after the acquisition closed, while also tempering expectations about how quickly organizations can execute migrations.
• Worth noting is the source bias: Rimini Street sells third-party VMware support and benefits from customers feeling locked in or overcharged, though the involvement of independent research firm Unisphere and alignment with prior customer reports lend the findings some credibility.

01:51 Justin – “Congratulations, Broadcom, now becoming the new Oracle.”

AI Is Going Great – or How ML Makes Money

04:53Introducing Claude Haiku 5.5

• Anthropic released Claude Haiku 5.5, positioned as its cheapest, fastest, and most capable small model to date, with pricing around 75 percent lower than Haiku 4.5 on average (90 percent lower for requests under 100,000 tokens, which account for about 90 percent of typical usage).
• Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, letting developers tune for cost versus intelligence, and it’s designed for high-volume tasks like summarization, classification, database queries, and as a subagent alongside Sonnet 5.5 and Opus 5.5 for coding work.
• Anthropic cut Sonnet 5.5 cache read pricing by 50 percent, from $0.20 to $0.10 per million tokens, reducing overall cost on agentic workloads by about 20 percent since cache reads make up a large share of token consumption.
• New monthly API credits are rolling out to Claude Max and Team subscribers: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team accounts, intended for experimentation with agents and apps on the Claude Platform.
• Haiku 5.5 is now available on AWS, Google Cloud, and Microsoft Azure alongside the Claude Platform, and the SDKs are getting beta support for computer and browser use, with Haiku 5.5 noted as well-suited for these speed-sensitive, lower-complexity tasks.

05:51 Justin – “It took a year for this to come out, though. They’ve realized how many Opus and Sonnet models the last one and the release of 4.5 Haiku? It’s a lot.”

08:36 GPT-6 and Intelligent UI for everyone

• OpenAI is rolling out GPT-6 in ChatGPT with a new Intelligent UI feature that generates interactive interfaces (buttons, forms, charts, diagrams) rather than plain text responses, choosing format based on the question type.
• The system uses a library of native streamable components plus a compiler that renders the interface progressively as the model generates it, avoiding a wait for the full response before anything appears.
• GPT-6 introduces interleaved thinking and answering, letting the model start responding before it finishes reasoning. Per OpenAI, GPT-6 Instant begins answering 44 percent sooner on average than GPT-5.6 Instant for web search queries, and GPT-6 Extra High matches GPT-5.6 Medium’s response start time while scoring higher overall.
• Safety updates include stronger resistance to multi-turn jailbreak attempts and updated training around cyberattack, biological, and violence-related misuse, building on safeguards referred to internally as Astra.
• Rollout starts today for Plus, Pro, Business, and Enterprise tiers (Chat tab only, powered by GPT-6 Sol), expanding tomorrow to Free and Go tiers (powered by GPT-6 Luna); Work and Codex experiences are unaffected by this release.

10:21 Matt – “I do like the artifacts and the work environment; it’s useful when I’m building something.”

11:49Now GA: Building permission-aware Databricks Apps with on-behalf-of-user authorization

• Databricks Apps on-behalf-of-user (OBO) authorization is now generally available, letting apps act under a signed-in user’s identity, so Unity Catalog permissions, row filters, and column masks are enforced automatically without reimplementing governance logic in app code.
• Apps can mix two authorization models: app authorization (service principal, for shared or app-owned operations) and user authorization (forwarded user token, for permission-sensitive queries), choosing the right identity per request path.
• Scopes like sql:restricted-query limit apps to read-only SQL rather than full SQL access, reducing the attack surface since scopes act as a capability ceiling, not a data access grant; the user still needs underlying Unity Catalog permissions.
• Workspace admins can allowlist the API scopes developers can request (Settings > Development > Apps), defaulting to all scopes but letting them narrow the list or disable it entirely, giving organizations centralized control over what apps can do on users’ behalf.
• Design guidance emphasizes failing closed when a forwarded token is missing, rather than silently falling back to the app’s service principal, to avoid returning results generated under the wrong permission set; this pattern also applies to custom agents built on Databricks Apps.

12:37 Justin – “It’s a cool idea, but I want a guarantee that it’s not going to accidentally use the wrong identity when it’s not supposed to.”

14:46 Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents

• Meta released Muse Wonder and Muse Wonder Full, their newest agentic AI models built on the earlier Muse Spark foundation, targeting expert domains like law, finance, health, and coding tasks.
• Two new features stand out: Active Memory, which retains relevant context across conversations over time, and Active Focus, which allows the model to persist through difficult tasks, try alternative approaches, recover from errors, and check in before taking significant actions.
• Safety measures include training the models to avoid sharing private information without permission, refuse harmful requests, and support security teams in finding code vulnerabilities, all evaluated under Meta’s Superintelligence Scaling Framework; a detailed safety report is expected soon.
• Availability is immediate through the Meta Model API and Muse Code, with broader rollout planned for Muse subscribers, positioning this as infrastructure for agents handling real-world tasks like reviewing leases or managing insurance claims.
• For cloud and developer audiences, the emphasis on frontier coding performance and API access suggests Meta is competing directly for developer workloads currently served by other agentic coding and reasoning models.

15:43 Ryan – “It’s interesting because usually the memories and that sort of management is at the harness level. And so it sounds like these are built directly into the models themselves, which is kind of fascinating. It’s something I haven’t seen before.”

18:37 Satya Nadella says we should assume all AI models are ‘compromised’

• Satya Nadella posted on X arguing AI models should be treated as potentially compromised by default, rather than trusted as opaque black boxes whose outputs are simply accepted or rejected.
• Key recommendation: an emergency brake capability, meaning authorized personnel should always be able to pause or shut down a model mid-task, with more advanced models requiring standardized containment technologies.
• Other proposals include timely incident disclosure, independent audits, verifiable data, and tamper-proof human-readable evidence logs for model actions and decisions.
• These ideas align with broader industry discussion on AI governance and safety, though Nadella’s emphasis on assuming compromise and requiring kill-switch-style containment is a more specific technical stance than some competitors have stated publicly.
• Worth noting for listeners: Nadella’s continued use of the term super intelligence in the post, which contrasts with the more operational, risk-mitigation focus of the rest of his message.
• We want to know what he knows that made him post this tweet…
• Is this just more sensationalism? Or is there something of substance here to be worried about?

18:37 Real-time feedback on our new tracking feature – courtesy of Matt

• It works. Sort of.
• It found the thing, researched it, but didn’t input it into the show notes.

Cloud Tools

25:14 Announcing Gremlin Foresight AI

• Gremlin launched Foresight AI, an agentic add-on to its Chaos Engineering platform that identifies potential failure modes, recommends and runs tests, applies fixes, and reverifies reliability without requiring deep distributed-systems expertise from every engineer.
• The system uses four agent personas: the Analyst gathers environment details and recommends tests; the Tester runs reliability tests; the Operator provides remediation guidance; and a Technical Program Manager tracks reliability health, issues weekly reports, and manages commitments with assigned dates.
• A key differentiator is the Failure Atlas, a proprietary dataset built from millions of fault injection experiments across tens of thousands of real systems over the past decade; Gremlin states customer data is not used to train the underlying LLM.
• The tool relies on actual test execution rather than LLM inference alone, using the LLM to summarize and explain results while basing decisions on data from reruns on the Gremlin platform, with human approval required before running tests or applying fixes.
• This addresses a stated gap: AI-generated code increases shipping velocity but reduces human review, creating architectural decisions and failure modes that often aren’t caught until after deployment; Foresight AI is positioned to close that gap with automated detection and verification at a pace comparable to code shipping.
• Foresight AI is available now as an add-on to the existing Gremlin platform, with demos and trials offered for teams evaluating adoption.

AWS

26:48 Amazon S3 Vectors now supports metadata pre-filtering for higher recall on filtered searches

• S3 Vectors now supports metadata pre-filtering, which resolves filter conditions before running the similarity search rather than evaluating them in tandem, improving recall on scoped queries by up to 5x on highly selective filters.
• The feature targets multi-tenant applications like RAG systems, agentic workflows, and document search, where queries typically target a narrow subset of an index, such as one customer’s records out of millions of total vectors.
• Implementation requires switching an index from CLASSIC to ENHANCED mode via the UpdateIndexMode API; existing vectors don’t need re-ingestion and query syntax stays the same, making adoption straightforward for existing deployments.
• New $startsWith operator enables prefix matching on paths, URLs, and hierarchical keys, useful for scoping searches to folder structures or document ID patterns, joining existing operators like equality, range, and set membership.
• No additional cost for pre-filtering; customers pay standard S3 Vectors pricing for storage, PUT requests, and queries, and the feature is available now in all commercial AWS Regions plus AWS China Regions, with vector buckets created after September 30, 2026 defaulting to ENHANCED mode.

27:45 Justin – “I was hoping this would be something we could maybe add to Bolt, because he uses Amazon S3 vectors for all of its work. And you know, it really doesn’t help us actually at our size ’cause we don’t have enough data in the vector database to make it worthwhile, which is a bummer. But still, if you have a large RAG workload and you’re struggling with some of the narrow searches, this is a good feature for you.”

28:38 AWS Batch now publishes job metrics to Amazon CloudWatch

• AWS Batch now natively publishes job metrics to CloudWatch under the AWS/Batch namespace, eliminating the need for custom instrumentation to monitor batch workload health.
• Metrics include state transition counts (submitted, running, succeeded, failed) and duration tracking (time to runnable, total execution time), segmented by JobQueueName dimension for per-queue visibility.
• This addresses a common pain point for batch processing teams who previously had to build custom monitoring solutions to track job failure rates and queue performance.
• Available now across all AWS Regions where Batch operates, with no additional setup required since metrics are published automatically.
• Standard CloudWatch metrics pricing applies, meaning costs scale with the number of metrics and API calls, though basic monitoring is typically covered under the CloudWatch free tier for lower-volume use cases.

29:10 Matt – “Batch ‘operates’ is doing a lot of heavy lifting, as far as I’m concerned… it’s a service looking for a job…”

34:27 Introducing Strands Box: AI agent sandboxes powered by Dogwood

• AWS released Strands Box in developer preview, an open-source sandbox (Apache 2.0) that pairs OS-level isolation with fine-grained policy enforcement for AI agents, addressing the risk of agents running in “YOLO mode” with unchecked access to files, shell commands, and APIs.
• Available at github.com/strands-agents/box.
• The system combines two components: OS-level containment (currently macOS Seatbelt) for hard boundaries, and the Dogwood policy engine for granular rules about what agents can do within those boundaries, including temporal rules based on action history, like rate-limiting Slack posts to three per 10 minutes.
• Policy enforcement happens outside the agent process at four interception points: network egress, the Python interpreter, the Shell interpreter, and an MCP broker, so rules apply consistently regardless of which tool the agent uses and cannot be bypassed by the agent itself.
• A notable technical detail is credential handling: the egress gateway injects real API keys and AWS credentials on outbound requests, meaning agents work with placeholder tokens and never see actual secrets, reducing credential exposure risk.
• Currently macOS-only, with Linux and other OS support planned, alongside a CLI to auto-generate configs from existing agent harnesses and plans to extend Box to deployed agents on Bedrock AgentCore, ECS, and Kubernetes.
• Relevant for teams building production agents with frameworks like LangChain, Claude Agent SDK, or Strands that need auditable, portable governance over agent actions.

35:57 Ryan – “I don’t really want to dictate all the different commands that are okay to run in sandbox, and getting credential information into the sandbox is incredibly hard with a lot of these things. And I’m hoping that this solves some of those problems, which I think is pretty cool.”

37:36 All the numbers: Amazon Prime Day 2026 powered by AWS

• DynamoDB processed over 59 trillion requests across the four-day event, peaking at 192 million requests per second while maintaining single-digit millisecond latency, illustrating the scale AWS managed databases handle under extreme load.
• Serverless compute saw heavy use: Lambda handled 2.3 trillion invocations per day and ECS/Fargate launched an average of 158.3 million tasks daily, a 47.7% increase over last year, showing continued customer adoption of container and function-based architectures for variable workloads.
• Security and reliability tooling scaled significantly as well, with GuardDuty monitoring 14.08 trillion log events per hour (up 59% year over year) and AWS running over 44,000 Fault Injection Service experiments, six times the 2025 volume, to validate resilience before the event.
• Networking and messaging services posted notable growth too, with CloudFront delivering 2.1 trillion HTTP requests (up 5% from 2025) and SQS receiving a peak of 213 million messages per second, reflecting the interconnected microservices architecture behind Amazon.com.
• AWS Graviton powered up to 49% of the EC2 compute Amazon.com used during the event, reinforcing AWS’s push toward custom silicon for cost and performance efficiency at scale; AWS also promoted its Countdown Premium service for customers preparing for similar high-stakes scaling events.

39:12 Ryan – “These numbers always continue to astound.”

40:44 AWS IAM Identity Center now supports network access controls for Identity Store

• IAM Identity Center now lets admins lock down the Identity Store API and SCIM API to specific networks, addressing a common security gap where identity management APIs were accessible from anywhere with valid credentials.
• Admins can require Identity Store API traffic to originate only from approved VPC endpoints or source VPCs, while separately restricting SCIM API access to specific IP ranges, such as an external identity provider’s published ranges.
• Practical use case: organizations using Okta, Azure AD, or similar providers for SCIM provisioning can now whitelist those providers’ IP ranges, while forcing internal custom apps and provisioning workflows through private VPC endpoints only.
• The feature is opt-in and off by default, configured via the UpdateIdentityStore API through SDKs or CLI, with no additional cost beyond standard IAM Identity Center usage; available in all regions where Identity Center is offered.
• Notable exemption: AWS service-initiated requests bypass these network controls, so this is aimed squarely at tightening external and custom application access rather than internal AWS service communication.

41:31 Justin – “So, this is my time now to plead with Amazon to, hey, work with all your partner companies like Okta and Azure AD to, you know, provide a managed firewall rule for this so I don’t have to figure out what their IP addresses are in their terrible documentation. Because again, you just have to publish a rule that I could then plug into a security group or into this policy. I’d be really happy if you could do that. So please.”

43:31 Announcing AWS Lambda Web Functions

• AWS Lambda now splits into three form factors: Event Functions (the original model), Web Functions (new), and MicroVMs, all built on Firecracker isolation but targeting different programming models.
• Web Functions let developers run standard HTTP servers (Express, Fastify, Next.js, etc.) directly on Lambda with no code rewrites or adapters, using a single CLI command for deployment: aws lambda-web deploy.
• Each execution environment supports up to 128 concurrent requests, and billing is based on active CPU time rather than wall-clock time, so I/O wait during database or API calls doesn’t accrue high compute cost.
• Multi-region deployment and latency-based routing are built in; a single command can replicate the app across regions like us-east-1, eu-west-1, and ap-northeast-1, with traffic automatically routed to the lowest-latency region via a global URL.
• Integrates with CloudFront using origin access control (OAC) for private-origin security, enabling a pattern similar to S3 origins, combining CDN-cached static assets with serverless dynamic compute.
• Available now in five regions (US East N. Virginia and Ohio, US West Oregon, Europe Ireland, Asia Pacific Tokyo), running Node.js 24 on Arm64 with VPC connectivity support; pricing details are on the AWS Lambda pricing page and vary based on usage.

44:53 Amazon CloudFront now supports AWS Lambda Web Functions as an origin

• CloudFront can now front Lambda Web Functions as an origin, adding edge TLS termination, caching, and custom domains with managed certificates without requiring application code changes.
• Origin Access Control (OAC) locks down Lambda Web Functions so they’re only reachable through CloudFront, blocking direct requests and adding network-layer DDoS protection; AWS WAF can be layered on for application-layer security.
• OAC passes original credentials through unchanged, meaning existing authentication and sign-in flows continue to work without modification.
• Available globally except in the CloudFront China region, with no additional fees for the feature itself; setup is supported via Console, SDK, CLI, or CloudFormation.
• This fills a gap for teams running web apps on Lambda who previously had to build their own caching, TLS, and access-control layers, simplifying the path to a production-ready edge-delivered serverless app.

47:40 Amazon ECS adds LiveTaskCount and removes the charge for RunningTaskCount metric

• New LiveTaskCount metric gives a complete picture of active ECS tasks across ACTIVATING, RUNNING, and DEACTIVATING states, including tasks still draining connections during scale-in events, making it more useful than RunningTaskCount for autoscaling and capacity planning decisions.
• LiveTaskCount is free and automatically available without opting into CloudWatch Container Insights (we’d like some credits, please), lowering the barrier for teams who want task visibility without additional setup or cost.
• RunningTaskCount, previously a paid Container Insights metric, becomes free for all users starting October 1, 2026, reducing monitoring costs for existing ECS customers who rely on it as a stability signal.
• No migration or action is required: both metrics coexist, with RunningTaskCount serving health/stability monitoring and LiveTaskCount serving scaling and capacity use cases.
• Available in all AWS regions where ECS operates, reflecting AWS’s continued investment in making container observability more accessible and cost-effective.

48:24 Justin – “Nice to see this move over to the right place. The fact that it needed container insights – I always thought was dumb. And I use this metric probably all the time.”

GCP

49:30 Announcing MCP Toolbox Java SDK v1.0: Agentic data access for the enterprise

• Google released MCP Toolbox Java SDK v1.0, bringing the Model Context Protocol to enterprise Java environments, letting AI agents connect to databases and data sources through a single standardized protocol instead of custom integrations for each model-database pairing.
• Key v1.0 additions include decoupled client authentication with async credential resolution, automatic pruning of server-side bound parameters so LLMs can’t manipulate sensitive fields like tenant_id, and runtime warnings when credentials travel over plaintext HTTP.
• The SDK integrates with Spring Boot and LangChain4j, using annotations like @MemoryId to handle conversation state across HTTP sessions, reducing the need for custom intent-parsing logic in conversational agents.
• Google demonstrated the SDK with a transit booking example using AlloyDB, showing how an agent can combine unstructured data (policy lookups) with structured queries and transactional booking in a single conversation flow.
• Deployment is designed for Cloud Run, with the MCP Toolbox server and the Spring Boot agent scaling independently; the SDK is open source and available via Maven or Gradle, with documentation at mcp-toolbox.dev and code on GitHub under googleapis/mcp-toolbox-sdk-java.

5:58 Ryan – “Wait – you want AI interaction directly to the data store? Instead of an API? Ok… good luck.”

53:45 EmbeddingGemma 2 is a best-in-class open model for natively multimodal embeddings

• EmbeddingGemma 2 extends last year’s text-only model to a natively multimodal embedder covering code, images, video, and audio in a shared embedding space, built on the Gemma 4 architecture with 740 million parameters under an Apache 2.0 license.
• The model is modular, requiring as little as 270M parameters for text-only workloads with optional 170M vision and 300M audio encoders, and uses Matryoshka Representation Learning to truncate output vectors from 768 down to 128 dimensions for up to 6x storage reduction.
• On-device performance is a key focus: quantized versions need roughly 191MB RAM for text-only and 567MB for full multimodal support on a Pixel 11 Pro, with an 8K token context window supporting up to 5.5 minutes of audio or 58 video frames processed locally.
• Benchmark results show a 9.92-point improvement on MTEB Code (68.76 to 78.68) along with leading scores for sub-1B multimodal embedders on MTEB and MAEB, positioning it for semantic code search, coding agent retrieval, and cross-modal search applications.
• Availability spans Hugging Face, Kaggle, and upcoming Gemini Enterprise Agent Platform Model Garden integration, with deployment support across MediaPipe, LiteRT, transformers, vLLM, Ollama, and Qdrant for vector storage, enabling fully offline RAG pipelines when paired with Gemma 4.

55:10 Matt – “It’s nice to see that these things are improving, because I know a lot of the embedding models get a lot of updates; I don’t think Amazon’s been updated in a while.”

56:42 Gemini at Work 2026: Introducing Gemini agent

• Google consolidates its AI offerings into a single Gemini agent that handles knowledge work, content creation, and coding through one interface, with persistent memory and multi-agent orchestration for tasks that run across sessions and devices.
• It also supports model choice across Gemini and Claude, with more third-party models planned.
• The agent integrates directly into Workspace apps (Gmail, Docs, Sheets, Calendar) and can be deployed as a coworker agent with its own identity, email address, and directory presence, raising questions about how enterprises manage access controls for non-human team members.
• Governance features include per-agent cryptographic identity, audit trails, an Agent Sandbox with network boundaries, and an Agent Gateway that enforces organization-wide policies in real time, addressing enterprise concerns about accountability and data exposure in agentic workflows.
• Cost controls include multi-model orchestration, Smart Routing to match workloads to the cheapest suitable model, and real-time spend caps configurable in the Cloud Billing Console that pause agent work when limits are hit, letting teams charge AI usage back to specific departments.
• Industry-specific versions are in preview for Financial Services and Legal, with Government, Healthcare, and Retail coming soon; adoption metrics cited include Bradesco cutting document review from 1 hour to 5 minutes and DBS Bank running chains of 70-80 specialized agents for credit memo generation, indicating movement from pilot projects to production use at scale.
• Infrastructure improvements include the TPU 8i system, claimed to deliver 80 percent better price-performance than the prior generation, underpinning the economics of running agent workloads at enterprise volume.

58:27 Matt – “They’re slowly getting all those pieces into one platform; I just wish that I would be able to search for it, because I feel that at this point searching for Gemini is going to pull up nothing useful in my life.”

59:19 SRE book second edition

• Google released a fully rewritten second edition of the Site Reliability Engineering book, ten years after the original, updating guidance for AI, microservices, multi-cloud, and platform engineering environments.
• Key practice areas have evolved significantly: SLOs moved from basic availability metrics to multi-tiered, user-centric models with dynamic burn-rate alerting, and monitoring has shifted toward OpenTelemetry and distributed tracing for automated root-cause analysis.
• The book addresses reliability challenges specific to AI infrastructure, including GPU and TPU cluster management, monitoring non-deterministic model outputs, and maintaining pipeline reliability for continuous training and fine-tuning.
• It also covers using generative AI as an operational tool for SRE teams, such as incident triage assistance, natural language telemetry queries, and predictive models for identifying bottlenecks before outages occur.
• The second edition is now available in physical format via Google Books and in digital format through O’Reilly, targeting practitioners, engineering managers, and technology executives looking to modernize their reliability practices.
• It’s odd this book is 10 years old considering we’ve all only aged 4 years in the same time.
• Also: hat tip to Ryan for getting everyone back on track. It doesn’t help the show’s overall length, but I appreciate the effort.

1:04:00 Gemini 4 Argon available on Google Cloud

• Gemini 4 Argon is Google’s new frontier model, now available on Google Cloud via Gemini for Business, API access under model ID gemini-4 Argon, Google AI Studio, and Google Antigravity.
• It’s designed for long, multi-step enterprise tasks requiring sustained context and domain expertise, such as financial reconciliations, legal research, and software development.
• Benchmark results show measurable gains: Argon leads the Vals Index for economic impact across finance, coding, legal, and tax work, ties for first on CWE-Bench v1 at 68% for vulnerability remediation, and cuts indirect prompt injection attack success to 0.7% on the Gray Swan benchmark.
• Internal Google usage provides concrete performance data: the ADK team reported a 40% reduction in session prep time and 21% faster cold imports after an autonomous overnight optimization run, while teams using Argon in Google Antigravity saw a 35% increase in agentic code submissions and nearly a third fewer rollbacks.
• Partners including GitLab, Figma, Harvey, ServiceNow, and HubX are integrating Argon into existing workflows, citing improvements in legal agent accuracy, long-document analysis, and automated security audits with zero fabricated vulnerability claims in HubX’s testing.
• Pricing starts at an introductory rate of $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached input tokens discounted 95%.
• Standard pricing afterward rises to $4 per million input tokens and $20 per million output tokens. A restricted version without cyber guardrails is also available to vetted defenders through the Fairwind Program for cybersecurity testing.

1:05:30 Justin – “So this is available to you now. Last time we talked about this, they had just announced it, but it was restricted to cybersecurity items; now it’s available to all. If you could just keep that 2 & 10 price, that would be great.”

Azure

1:07:08 Generally Available: Exceptions in WAF for Azure Application Gateway and Azure Front Door

• Ryan hates this one almost as much as a bad cover band version of Hotel California.
• WAF exceptions are now generally available for Azure Application Gateway and Azure Front Door, giving customers a new way to bypass WAF inspection for specific requests at the rule, rule group, or managed ruleset level.
• This differs from existing exclusions, which omit specific request attributes like headers or cookies from evaluation while still inspecting the rest of the request; exceptions instead let entire requests skip inspection based on criteria like request URI, remote IP, or header name and value.
• The feature addresses a common pain point: WAF policies blocking legitimate traffic as false positives, which can disrupt application functionality for customers relying on these gateways.
• Configuration is straightforward through the Azure portal under the WAF policy’s Managed rules section, so it’s accessible without complex scripting or API work.
• Relevant for security and networking teams managing web application firewalls who need more granular control to reduce false positives while keeping broader protections active.

1:07:54 Ryan – “I look forward to the outage that’s caused by the exception for the corporate IP space, and then you deploy your application to production and then nothing works, because you didn’t actually test your WAF. There are use cases, but they are small.”

1:15:05 Microsoft makes ‘biggest bet yet’ on its own silicon, unveils new AI chip to reduce reliance on Nvidia

• Microsoft unveiled Maia 300, its third-generation custom AI chip, with large-scale deployments planned for 2027; Azure chief Scott Guthrie says it will become the primary AI inference system for Microsoft products and services.
• The chip delivers 40 percent more AI output per watt compared to leading GPU systems, and a full rack produces up to seven times the output of the current generation, targeting multistep AI agent workloads and long-context document or conversation processing.
• Microsoft designs the chips in-house and manufactures through TSMC rather than purchasing finished units from Nvidia, though Nvidia remains a key partner for data center GPUs and joint Windows PC launches.
• This follows a broader industry pattern of custom silicon development, with Google selling TPUs to select customers, Amazon offering Trainium to companies like Anthropic, and OpenAI partnering with Broadcom on its own chip.
• Adoption will be gradual: Evercore ISI estimates Maia will represent only about 10 percent of Microsoft’s AI chip spending by early 2027, and the current Maia 200 generation is still limited to a small number of U.S. data centers as of this summer, with expansion into Europe and Australia underway.
• Maia is not available for external customers to rent, unlike some competitor offerings. More details are expected at Microsoft’s Ignite conference next month.
• Listener note: Look up the Microsoft Surface so that you too can participate in the audio extravaganza that is this story of the guys reacting to a visual.

Oracle

1:17:47 Oracle trucks natural gas to AI data centers amid pipeline delays

• Oracle is trucking in compressed natural gas to power AI data centers in Utah and Texas as pipeline connections lag construction schedules, and it is now considering the same workaround for its 2.45-gigawatt Project Jupiter site in New Mexico after a pipeline reroute pushed the in-service date to next year.
• The economics are notable: trucking compressed gas costs roughly four times pipeline delivery once labor, equipment, and fuel are included, and an analyst estimates a large trailer would be depleted in about 40 minutes at just 4 percent of Project Jupiter’s planned capacity, raising questions about whether this approach scales to a gigawatt-class facility.
• Oracle sent a force majeure notice to the New Mexico project developer, a legal move that could limit its payment obligations if pipeline delays persist, suggesting the company is hedging against further timeline slippage.
• This comes amid broader financial strain: Oracle’s capex hit $28.5 billion for the quarter ended August 31 (up from $8.5 billion a year prior), with negative free cash flow of $5.4 billion and plans to raise about $40 billion in debt and equity this fiscal year.
• Oracle’s stock dropped 5.5 percent following the Bloomberg report, and the company notably did not respond directly to Bloomberg’s inquiry, instead posting social media credit to its gas supplier VoltaGrid, which may be worth discussing as a messaging choice.
• This should make all the environmentalists happy! We foresee no problems with trucking in that much natural gas.

Emerging Clouds

1:19:06 Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash

• Cloudflare shipped Clef-omni just one week after launching the original Clef and Clef-flash decision models, adding native audio and video support alongside text and images in a single API call, eliminating the need for separate transcription or frame-splitting pipelines.
• Clef-flash pricing dropped from 0.09 dollars to 0.038 dollars per million input tokens, now undercutting competitor Jev, though the hosted context window was reduced from 64k to 24k tokens to support the lower price point; self-hosted weights still support 256k context.
• Clef-omni processes modalities quickly: text decisions return in about 130 milliseconds, images in about 150 milliseconds, and a full 21-second video with audio scores in roughly 1.5 seconds, all within one API call.
• The underlying architecture builds on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts model, using LoRA fine-tuning and a two-stage attention routing approach to score schema-constrained options across text, image, audio, and video without generating output tokens like a traditional LLM.
• Internal Cloudflare teams already use Clef for practical classification tasks, including spam detection in GitHub issues, phishing checks in a content management system, PII scanning for data loss prevention, and malicious domain detection, showing how decision models can replace custom-trained classifiers for narrower use cases.

1:19:54 Ryan – “Maybe it’s just the new hotness for the 15 seconds it’s popular.”

Cloud Journey

1:19:06 Terraform is the wrong tool for building developer platforms

• Terraform lacks a native runtime; it’s a CLI that reads code, talks to providers, and exits, so you have to bolt on continuous reconciliation, multi-tenancy, and API exposure through CI/CD pipelines and custom glue code.
• The article argues this architectural gap, not tool quality, is why Terraform struggles as an Internal Developer Platform foundation.
• State file locking creates serialization bottlenecks because each workspace has its own state file, forcing teams into workarounds like workspace slicing, custom state routers, and manual lock-breaking after failed runs.
• This operational overhead compounds as platforms scale across teams and environments.
• Multi-provider transactions have no rollback semantics in Terraform, so a failure partway through provisioning an application environment (AWS account, DNS, Kubernetes namespace, Vault path, etc.) leaves a half-provisioned state with only an exit code as feedback.
• Control-plane architectures like Crossplane treat compositions as atomic units, with status conditions and events.
• The rise of autonomous coding agents raises the stakes, since agents need a queryable API and truthful status responses rather than a pull-request-based workflow, and a confused agent hitting an opaque pipeline at machine speed can create a larger blast radius than a human would. This makes the API-versus-pipeline distinction more operationally significant going forward.
• The recommended alternative is a Kubernetes-plus-Crossplane style control plane (with tools like ArgoCD for GitOps delivery and ClusterAPI for cluster lifecycle) that provides continuous reconciliation and a real API contract, while Terraform retains a legitimate role for bootstrapping stable, low-churn infrastructure like landing zones and DNS.
• Teams evaluating platform engineering investments should pick the runtime first and treat IaC tooling as a component that fits inside it.

After Show

1:41:31 Chick-fil-A takes a stand against fast-food drive-thru AI trend

• Chick-fil-A is opting out of voice AI for drive-thru ordering while McDonald’s and Wendy’s roll out their own systems, Archy and FreshAi respectively, showing a divergence in strategy among fast-food leaders.
• This highlights a broader distinction between customer-facing AI and backend automation. Chick-fil-A still uses conveyor systems and multi-lane layouts to improve throughput without replacing human order-takers.
• Wendy’s FreshAi runs on Google Cloud, a reminder that cloud providers are competing to power retail and QSR AI deployments, with wins in this space serving as proof points for broader enterprise adoption.
• The decision raises a practical question for cloud and AI vendors: how much of the value proposition is operational efficiency versus customer experience, and whether those goals always align.
• With Chick-fil-A reporting nearly $ 24 billion in 2025 sales without voice AI, the case offers a data point for evaluating whether AI adoption is necessary for competitive performance in this sector, or just one of several viable operational strategies.

Closing

And that is the week in the cloud! Visit our website, the home of the Cloud Pod, where you can join our newsletter, Slack team, send feedback, or ask questions at theCloudPod.net or tweet at us with the hashtag #theCloudPod