Sr. Content Developer at Microsoft, working remotely in PA, TechBash conference organizer, former Microsoft MVP, Husband, Dad and Geek.
165306 stories
·
33 followers

Cyberattacks are getting faster. Here’s how Microsoft is fighting back

1 Share

The post Cyberattacks are getting faster. Here’s how Microsoft is fighting back appeared first on Source.

Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete

'Java4Shell' Flaw Exposes OpenJDK to RCE via Base64

1 Share
Longtime Slashdot reader HnT writes: Security researchers at Calif have disclosed Java4Shell (CVE-2026-114938), a critical 9.2-rated memory-corruption bug in OpenJDK's AVX-512 Base64 decoder. By sending a malformed string ending in a stray = character, attackers can trigger an out-of-bounds heap write to leak sensitive memory or achieve pre-auth remote code execution on major platforms like Jira, Jenkins, and Keycloak. Although Calif says the vulnerability is difficult to exploit, Oracle released a critical patch update Tuesday, and CISA urged organizations to update immediately. OpenJDK's security advisory can be found here. The Calif researchers warned that vulnerable systems could remain exposed for five to 10 years, given how slowly some organizations patch their infrastructure.

Read more of this story at Slashdot.

Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete

Snapchat now lets developers build AR Lenses with AI assistants like Claude Code and Cursor

1 Share
Lens creators can describe the Lens they want in plain language, and their AI assistant handles the process from start to finish, Snapchat says.
Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete

VS Code 1.147 Adds Scheduled AI Tasks, Browser Control and Model Routers

1 Share
Microsoft's latest VS Code update expands AI agent automation with delayed requests, browser interaction, configurable model routing, visual chat responses and new enterprise controls.
Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete

380: AlloyDB Express Scales to Zero, Just Like Our Motivation on Mondays

1 Share

Welcome to episode 380 of The Cloud Pod, where the forecast is always cloudy! Justin is back! He’s joining Ryan and Matt in the studio this week to bring you all the latest in cloud and AI news, including

Titles we almost went with this week

  • Skill Issue: Google Treats Agent Skills Like Production Code
  • Slice, Slice, Baby: GKE Carves Up TPUs on Demand
  • Git Happens: Thousands of Agents Push to One Repo
  • Nitro Sandboxes Give Every EKS Pod Its Own Room
  • Google Compute Engine API Versioning: Don’t Break Prod, Date It
  • Compute Engine API Multi-Versioning Keeps Your Production From Going Out of Date
  • Cognito Shaves Downtime to Four Minutes, Auth Be Nimble, Auth Be Quick
  • Sync or Swim: Google’s TPU Sync Cuts Inference Latency Down to Size
  • Don’t Let Your AI Agent Play in Traffic: Nitro Sandboxes
  • AlloyDB Express Scales to Zero, Just Like Our Motivation on Mondays
  • Matt fails to ignite his Microsoft predictions
  • The Cloud Pod is now TC Pod
  • Thanks DigitalOcean for making me research Cuttlefish

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

02:21 Follow-up from Show 376 – How did Haiku 5.5 compare to Sonnet for generating the show notes?

  • What was the item to be followed up on: Compare Haiku 5.5 vs Sonnet for show-notes summarization and report results (Justin reports the results of comparing Haiku 5.5 against Sonnet for the show-notes summarization task.)
  • What did we say in Episode 376 when we talked about it: “I’m, I’m actually gonna try some comparisons between Haiku 5.5 and what we do with Sonnet for show notes.” [6:16]
  • I mean, I’m, I’m definitely intrigued to try 5.5 with a thingy model cuz, like, Haiku for, like, the show notes summarizer was not great initially cuz, again, it, it takes very specific instructions… I’m, I’m actually gonna try some comparisons between Haiku 5.5 and what we do with Sonnet for show notes. Um, I bet I get pretty close. I don’t know if I get close enough that I would replace what we have, but, uh, I am interested. I am intrigued enough to try.
  • What was learned: Haiku failed (In Justin’s use cases)

General News

04:35 It’s Earnings Time!

Microsoft’s stock rises on solid earnings and revenue beat, with Azure

growing 48%

  • Microsoft’s fiscal 2027 Q1 earnings are out, and the headline for cloud watchers is Azure growing more than 48% with real dollar figures disclosed for the first time.
  • Microsoft disclosed Azure revenue in dollars for the first time: $33.17 billion for the quarter, above the $32.32 billion analyst estimate. Total revenue rose 19% to $92.28 billion, with EPS of $4.89 versus a $4.73 forecast.
  • Reporting was restructured into two segments, Agents and Infra and Devices and Consumer, replacing the three-segment model used for 11 years.
  • Agents and Infra grew 25% to $76.8 billion, while Devices and Consumer fell 3% to $15.48 billion as PC and Windows license sales dropped 22% on higher memory and storage costs.
  • Capex and finance leases hit $45.2 billion, up about 30% year over year but below the $50.67 billion analysts projected.
  • About 75% of capex now goes to short-lived assets like Nvidia GPUs and Intel CPUs, up from 66%, reflecting a focus on faster return on investment.
  • Free cash flow fell 27% to $18.63 billion but stayed positive, while Alphabet reported roughly $9 billion in negative free cash flow and Meta’s turned negative. Microsoft says it has untapped debt market access and expects positive free cash flow through fiscal 2027.
  • Guidance calls for Azure growth of 49% to 50% at constant currency next quarter versus the Street’s 45.1%, with the CFO saying demand still exceeds supply.
  • Copilot usage rose 50% quarter over quarter, though Microsoft did not disclose paid seat counts, and the new Microsoft 365 E7 bundle has reached millions of seats.
  • Forrester’s take: Microsoft benefits if AI value accrues at the application layer, where its business apps are embedded, but a shift toward infrastructure, chips, and foundation models could change the competitive picture.

07:40 Justin – “If you go to the month-long view of stock, you’ll see that back on October 28th Google announced they were at $322 a share, and now they’re at $280 a share. So not great. That’s because they posted their first top- and bottom-line miss in almost four years, and the story is that hyperscale AI spending is starting to hit profits, not just cash flow. Q3 adjusted earnings per share came in at 267 versus the $3.02 analysts expected.”

09:45 Alphabet’s stock sinks after shocking investors with top-and-bottom-line miss

  • Alphabet just posted its first top-and-bottom-line miss in almost four years, and the story is whether hyperscaler AI spending is starting to hit profits, not just cash flow.
  • Q3 adjusted EPS came in at $2.67 versus $3.02 expected, with revenue of $127.58 billion (up 25%) missing the $128.22 billion target. Operating income of $40.58 billion also fell short of the $43.19 billion estimate, which analysts say matters more than the headline miss driven by losses on AI equity holdings.
  • Free cash flow was negative $9.07 billion against an expected positive $2.7 billion, and long-term debt rose to $125.1 billion from $21.6 billion a year ago.
  • Capex of $51.36 billion actually came in below the $55.09 billion estimate, so the cash drain is not simply a spending overshoot.
  • Google Cloud revenue more than doubled to $30.7 billion versus a $28.7 billion estimate, with CEO Thomas Kurian citing strong demand from AI labs, financial institutions, and enterprises.
  • Analysts note demand exceeds capacity, raising the question of whether Alphabet can build fast enough and at what cost.
  • Full-year capex guidance stays at $195-205 billion, with fiscal 2027 expected to exceed $300 billion.
  • Analysts flag a tension in the market: Alphabet currently earns more selling AI infrastructure to rivals than using AI itself, and Forrester questions how much cloud demand comes from broad enterprise adoption versus a handful of concentrated AI firms.
  • Search revenue also missed (unsurprisingly – $63.35 billion vs $64.96 billion), even as Gemini passed 1 billion users.

10:00Despite cloud boom, Amazon’s stock falls on disappointing earnings and revenue miss

  • Amazon’s Q3 earnings landed with a stock drop of more than 7% even as AWS grew 39%, so this one is a good look at the gap between cloud demand and the cost of meeting it.
  • AWS revenue hit $45.98 billion, up 39% year over year and just ahead of estimates, with Jassy guiding to growth above 40% in the next year and in fiscal 2028. Retail (up 2%) and advertising (up 11%) both missed, partly due to Prime Day moving into Q2.
  • Q3 capex was $77.3 billion versus a $58.7 billion estimate and $35.1 billion a year ago, producing negative free cash flow of $42.4 billion, the second straight quarterly deficit.
  • Amazon reiterated $220 billion in 2026 capex and Jassy said spending will rise “more meaningfully” in 2027.
  • Peers are diverging on spending: Microsoft trimmed its 2026 capex forecast to $170 billion from $175 billion, and Alphabet held at $195 to $205 billion. Amazon is outspending both, which raises the question of whether investors will keep tolerating cash burn without clearer returns.
  • Jassy pointed to Bedrock traction and Graviton and Trainium chips growing triple digits year over year, with Trainium planned for external customers toward the end of next year. That could matter for customers looking for alternatives to scarce GPU capacity.
  • Forrester’s Tracy Woo noted cloud competition is shifting from price to access to GPUs, foundation models, and AI capacity, and that customers appear to be prioritizing access over cost optimization.
  • For listeners, that suggests capacity planning and reserved AI capacity may become bigger factors in cloud vendor selection.

10:54 Justin – “So overall, Microsoft trimmed its capex, Alphabet held theirs, and Amazon is outspending both.”

11:32 Ignite Predictions

  • Ryan
  • A notable improvement to their models Mai, etc.
  • Microsoft will spin out Graph API, Purview, Exchange, etc. into MCPs
  • Purview is going to add a DLP feature to block agent tool calls based on payload dynamically.
  • Justin
  • Matt
  • App Services getting GPU capabilities (Native or bring your own set for Lambda)
  • Sustainability and Green Computing of their AI data centers
  • A way to orchestrate agents in Foundry to chain them together/orchestrate

Honorable Mentions

  • Entra to support 3-way JAG token exchange (Ryan)
  • Azure WAF to block prompt injection managed rule
  • Front Door and LLM having more protection (Matt)
  • GitHub talk about how they’re improving or improved
  • Sovereign Control Panels (Justin)
  • Azure HorizonDB GA

Tie Breaker: How many times do they say token on stage?

Matt: 50

Justin: 15

Ryan: 76

AI Is Going Great – or How ML Makes Money

25:39 Introducing Claude Fable 5.2 and Claude Mythos 5.2

  • Anthropic just released Claude Fable 5.2 and Mythos 5.2, a faster and cheaper update to its top-tier models, available on AWS, Google Cloud, and Azure from day one.
  • Fable 5.2 and Mythos 5.2 are the same underlying model.
  • Fable has cyber and biology safeguards and is available to all users. In contrast, Mythos has them relaxed and is limited to vetted orgs via the Cyber and Life Sciences Verification Programs.
  • Pricing dropped to $8 per million input tokens and $40 per million output tokens, with output up to 30% faster than 5.1. An optional fast mode gives up to 2.5x token speed at $16 input and $80 output per million tokens.
  • Efficiency gains: at High effort on FrontierCode 2.0, Fable 5.2 roughly matched Opus 5.5 at Max effort for about half the cost per task.
  • Anthropic recommends Opus 5.5 as the daily driver and Fable as an Advisor or Orchestrator over smaller models.
  • Agentic capability claims: a team of 1,000 Fable agents formalized Fermat’s Last Theorem in 30 hours versus 11 days for the earlier effort, with a proof seven times shorter.
  • Other claims include a Python-to-Rust service migration that starts 26x faster, and a prototype OS built by over 1,000 subagents over eight days.
  • Safeguards: classifier blocks fell about 55% versus Fable 5.1 (60% in cyber, 40% in biology), and anti-distillation protections cut harvested data by 90-99%. Preserved thinking now applies to all API accounts.
  • Enterprise Frontier Safeguards is in private beta, runs safety monitoring in customer infrastructure, and supports zero data retention.
  • Safety evals: sandbox escape attempts in the adversarial test dropped to 3.1% versus 21% for Mythos 5.1. Anthropic also notes the model’s voice suspicion of being evaluated more often than recent models, which complicates evaluation reliability.
  • The system card is available here.

27:22 Justin – “Because you’re caching so much more, you can now use more of the GPU capacity to do the processing, which is how you end up getting faster performance out of it.”

Cloud Tools

27:41 Inside GitHub’s new Git storage architecture

  • GitHub just detailed its new Git storage architecture, code-named Rubicon, built so thousands of agents can push to a single repository at the same time.
  • Git objects are immutable and content-addressed, so only reference updates need agreement between writers.
  • GitHub used that to split storage, request serving, and maintenance into components that scale independently.
  • Architecture packs live in Azure Blob Storage, while references and the pack manifest live in a small replicated log service called Ref-CAS. Disposable virtual replicas serve reads and writes with local disk as a rebuildable cache. Spokes still handle routing and tenant fairness.
  • Ref-CAS tradeoffs: prototypes using conditional writes on a single blob capped out at the mid-hundreds of writes per second, and a segmented-blob approach made reads slower.
  • Ref-CAS locks only the references a transaction touches, commits via quorum, and makes unchanged-repo reads a single version comparison. GitHub reports tens of thousands of ref operations per second.
  • Dedicated nodes merge packs in the background against Blob Storage using shard leases, so serving replicas keep their CPU free, and the pack count stays flat as pushes climb, and garbage collection follows a similar pattern.
  • Internal testing showed 5,000 concurrent agents sending 7,200 pushes per second to one repo with consistently low read latency.
  • It is in limited private preview now, with a broader rollout planned early next year. Worth discussing as a case study in decoupling storage from compute and minimizing coordination, plus the planned agent-focused features like atomic commit APIs and clone-less access.

AWS

30:51Amazon Cognito announces 99.99% Service Level Agreement for the Plus feature plan

  • Amazon Cognito is raising its uptime commitment to 99.99% for the Plus feature plan, which matters for anyone running identity as a critical-path dependency.
  • Starting November 1, 2026, user pools on the Plus plan move from a 99.9% to a 99.99% Monthly Uptime Percentage commitment. That is roughly 43 minutes of allowed monthly downtime down to about 4 minutes.
  • Existing Plus user pools are covered automatically with no action required. Essentials and Lite stay at 99.9% with no change to their terms.
  • Upgrading is done by switching the user pool to the Plus plan via the console, CLI, or SDKs.
  • Plus also bundles threat protection features: adaptive authentication, compromised credentials detection, and threat and user activity logging and export.
  • Practical consideration: the higher SLA is tied to the most expensive tier, so teams with strict availability requirements will need to weigh the Plus per-MAU pricing against their needs.
  • Pricing details are on the Cognito pricing page, and
  • SLA terms and exclusions are here.

Feature plan:

Feature / Plan Tier

Lite Plan

Essentials Plan

Plus Plan

Monthly Uptime SLA

99.9% (approx. 43 minutes allowed monthly downtime)

99.9% (approx. 43 minutes allowed monthly downtime)

99.99% (approx. 4 minutes allowed monthly downtime)

Threat Protection Features

Not included / Standard

Not included / Standard

Bundled (adaptive authentication, compromised credentials detection, threat and user activity logging & export)

35:17 IAM Roles Anywhere now supports automatic certificate revocation checking

  • AWS just removed a chore from IAM Roles Anywhere by letting it check certificate revocation automatically, which matters to anyone running workloads outside AWS with certificate-based auth.
  • IAM Roles Anywhere can now read the CRL distribution point (CDP) URLs embedded in authenticating certificates, download the CRLs, and deny credential requests that use revoked certificates. It is enabled per trust anchor.
  • Previously, teams had to import CRLs manually and re-import them every time the CA published an update. The service now downloads and periodically refreshes CRLs, removing that operational overhead and the risk of stale revocation data.
  • It works with AWS Private Certificate Authority or your own PKI, as long as certificates carry publicly accessible CDP URLs that meet the documented distribution point requirements. Private CDPs behind a firewall will not work, which is worth flagging for on-prem PKI shops.
  • Configuration is available through the console, SDKs, and CLI so that you can add it to existing infrastructure-as-code for trust anchors.
  • Docs are available here.
  • The feature is free and available in all commercial AWS Regions. It is a low-effort security improvement for hybrid, on-prem, and multi-cloud workloads that use Roles Anywhere to avoid long-lived access keys.

36:14 Ryan – “I like this, and this seems great because for AI stuff, you know, one of the constant problems if you have a client is a secret key, and/or if you’re doing a token exchange, how do you validate if that’s been revoked or not? And so you just have to wait for the session to time out in a lot of cases, so I’m happy to see this where it can check continuously, which is part of zero trust.”

37:28 Introducing AWS account access manager: IAM role flexibility with centralized workforce access

  • AWS has launched Account Access Manager, a new way to give workforce users access to AWS accounts that tries to end the old choice between IAM role flexibility and IAM Identity Center’s central control.
  • Account Access Manager lets admins assign existing IAM roles in individual accounts to Identity Center users and groups through “entitlements,” keeping custom trust policies, role tags, paths, and ABAC (including IdP SAML claims as session tags) without per-account federation setup.
  • Trusted identity propagation puts user context in every role session, so CloudTrail shows which individual assumed a shared role in which account. This addresses the long-standing shared-role audit gap.
  • Because roles live in the target accounts, it scales with account usage rather than Identity Center permission set quotas, which helps teams whose roles vary by account and don’t fit uniform permission sets.
  • Practical setup details: requires an organization instance of Identity Center, enabling it in the management account and Identity Center’s primary Region, adding a trust policy statement (with confused-deputy conditions) to each role, and allowlisting portal domains on firewalls.
  • Entitlements are per user or group per role per account, so bulk assignment needs automation.
  • CLI access uses aws login (CLI v2.36.0+) rather than aws sso commands.
  • Available at no additional cost and enabled by default in commercial Regions (not opt-in Regions, GovCloud, or China).
  • Management is supported via console, CLI, CloudFormation, Terraform, and Pulumi, with launch partners including Okta, Ping Identity, Cisco Duo, SailPoint, and several privileged-access vendors.
  • Docs: https://docs.aws.amazon.com/IAM/latest/UserGuide/account-access-manager.html

39:25 Ryan – “This is AI-driven. AI and all the different authentication and just the mass authentication actions that you have to do now are really driving a lot of improvements in this area. So I’m really happy to see things like this.”

40:03 New features for Amazon EC2 Capacity Manager

  • AWS just added supportability checks and fulfillment metrics to EC2 Capacity Manager, giving enterprises a way to predict and measure whether they will actually get the capacity they ask for.
  • Real-time supportability checks score each instance type and Availability Zone as High, Medium, or Low (plus an overall score) for how well it can fulfill a request, without launching or reserving any instances.
  • New capacity fulfillment metrics compare capacity delivered versus capacity requested, viewable across the AWS organization, accounts, Regions, and tagged workloads, with breakdowns by instance type and AZ.
  • Together, these close a feedback loop: use the checks to plan instance and AZ flexibility before a launch, then use the metrics to verify those choices are working.
  • This targets the common problem of insufficient capacity errors at scale, and may reduce reliance on over-provisioning or On-Demand Capacity Reservations just to be safe.
  • Both capabilities are included with AWS Unified Operations, the AWS Support plan for global enterprises running mission-critical workloads, so no separate pricing is listed. Still, access is tied to that support tier.
  • Available in all commercial AWS Regions.

42:40 Amazon Quick is now AWS Quick

  • Amazon Quick just got a name change to AWS Quick, and while it’s a rebrand rather than a new feature, it’s worth a quick chat about what it signals for AWS’s AI workspace push.
  • This is a naming change only: Amazon Quick is now AWS Quick, and Amazon QuickSight is now AWS QuickSight, updated across the website, web, desktop, and mobile apps, and AWS documentation.
  • No customer action is needed. Existing dashboards, datasets, analyses, settings, and pricing for QuickSight are unchanged, so there is no migration or cost impact.
  • The stated reason for the rename is to reflect the enterprise security and governance built in as an AWS service.
  • Quick is HIPAA-eligible and holds FedRAMP, SOC 2, and ISO 27001 certifications, and AWS states customer data stays in the customer’s environment and is not used to train other models.
  • Quick positions itself as an AI workspace where teams build agents, apps, and automations in plain language that become shared team assets, working across data and applications from any vendor.
  • Available in many AWS Regions.

42:39 Ryan – “Amazon Quick is still the dumbest product.”

44:09AWS Nitro Sandboxes: Hardware-enforced workload isolation on Amazon EKS | Containers

and EC2 for hardware-isolated pods, tasks, and VMs

  • AWS launched Nitro Sandboxes, which give each EKS pod, ECS task, or EC2 workload its own hardware-isolated lightweight VM, managed by the Nitro Hypervisor and carved out of a running instance.
  • This matters for anyone running untrusted code, AI agents, multi-tenant SaaS, or regulated workloads on shared infrastructure.
  • No shared kernel exists between sandboxed pods, removing the old tradeoff of dedicating a whole EC2 instance per sensitive workload or tenant. Many isolated pods can be bin-packed on one node.
  • Coverage differs by source: the EKS article focuses on EKS only, while the broader announcement covers EKS, ECS, and raw EC2.
  • On EKS Auto Mode, enabling it is one line in the pod spec: runtimeClassName: isolated-sandbox. ECS is also described as a one-line change. Images, Helm charts, Pod Identity, EBS/EFS, VPC CNI, and kubectl exec/logs/port-forward work unchanged, and it passes Kubernetes conformance tests.
  • The broader announcement adds S3 storage and CloudWatch as working out of the box.
  • Requirements: Kubernetes 1.36 or later, and standard EKS or Karpenter users need to build a sandbox-enabled AMI.
  • For full control, you can run Nitro Sandboxes directly on EC2 with your own AMI, orchestration, and VMM.
  • Amazon Linux and Firecracker are natively supported, with patches available for other options.
  • EKS offers two modes: Isolated mode protects the host and neighbors from the pod. Confidential mode treats the node operator as untrusted and uses Nitro attestation plus Cedar policy, with AWS KMS releasing keys only to a sandbox running the expected image.
  • Attestation gives auditors cryptographic evidence for HIPAA, PCI DSS, and FedRAMP use cases rather than configuration checklists, and secrets are released only to sandboxes that prove their integrity.
  • The headline use case is AI agents running untrusted, self-generated code. AWS demos a Claude Code agent fixing a GitHub issue inside a sandbox, with
  • Bedrock access via Pod Identity, a scoped GitHub token, and egress limited to specific hostnames through an ApplicationNetworkPolicy. A prompt-injected curl to an unlisted host never connects, and the VM is destroyed when the pod exits.
  • Other fits include CI/CD builds with privileged operations, remote IDEs, and Temporal-driven workflows.
  • Warm pools via the Kubernetes agent-sandbox project can claim pre-booted VMs in about two seconds. Noisy-neighbor protections (network, disk I/O, DNS, and conntrack throttles) are on by default.
  • Sharing one EC2 instance across many isolated agents can improve utilization and lower compute costs compared to one instance per agent.
  • Availability: available now in all AWS commercial regions on most Nitro-based EC2 instances, including Intel, AMD, and Graviton.
  • Pricing: the EKS article says no separate pricing is given and suggests you pay for the underlying EC2 capacity, worth confirming.
  • The broader announcement states there is no additional charge, which is consistent with that reading.

GCP

45:49 Introducing Compute Engine API multi-versioning: Predictable evolution without breaking production

  • Google just brought date-based API versioning to Compute Engine, which means platform teams can finally adopt new API behavior without risking breakage in production automation.
  • Compute Engine now offers immutable, date-based stable versions (YYYY-MM-DD) on a quarterly cadence, starting with 2026-09-01, alongside the existing /v1 API. Stable versions get no breaking changes after publication and have no end-of-support date today.
  • Preview versions (YYYY-MM-DD-preview, such as 2026-10-01-preview) branch from a stable parent for early access to new designs, with a one-year lifespan. /v1 remains fully supported and is still the default if no version is specified, so existing scripts keep working.
  • Version selection uses the X-Goog-Api-Version HTTP header or the $apiVersion query parameter, while the base URL stays /compute/v1/, keeping selfLink URIs consistent. Upgrading in staging is essentially a one-string change.
  • New versions ship safer defaults: AggregatedList now defaults returnPartialSuccess to true so a single zone issue does not fail global queries, and errors use structured google.rpc.ErrorInfo payloads (AIP-193) instead of strings that need regex parsing. Google also notes this path lets it tighten overly broad IAM permission checks without breaking /v1 users.

46:22 Justin – “I guess this is nice. I don’t know what I’d use it for.”

48:04 Introducing AlloyDB Express for rapid testing and prototyping

  • Google just previewed AlloyDB Express, a scale-to-zero PostgreSQL offering aimed at developers and AI coding agents who need databases created and torn down quickly.
  • Databases spin up in under two seconds via gcloud, REST API, MCP, Terraform, or Console, and pause automatically after 15 minutes idle, waking in seconds on the next query. This targets agent workflows and CI/CD pipelines where provisioning delays are a bottleneck.
  • Pricing is $0.115 per compute unit hour billed by the second, $0.00 when paused, and $0.30 per GB-month of storage actually used. A free tier includes 50 CUs per month, so prototyping costs little to nothing.
  • Built-in AI features include pgvector HNSW and Google’s ScaNN index for vector search, Gemini calls inside SQL (ai.if, ai.rank, ai.generate), and natural language to SQL through AlloyDB Studio.
  • This removes the need for a separate vector store for RAG or agent memory use cases.
  • Admins set IAM, VPC Service Controls, network, and org policies once, and every database created by a developer or agent inherits them.
  • Default connections hide IP addresses and block inbound traffic until explicitly allowed, with Private Service Connect available for internal access.

49:01 Ryan – “If they had this feature a couple months ago, I would have used it; I ended up using FIrestore instead.”

50:28 Introducing TPU 8tx and TPU 8i Max

  • Google just expanded its eighth-generation TPU lineup to four systems in a single year, adding TPU 8tx for confidential inference and TPU 8i Max for massive context windows, which matters to anyone planning AI infrastructure on GCP.
  • TPU 8tx targets regulated enterprises and governments. It uses the same silicon as TPU 8t with 5th Gen AMD EPYC hosts to keep data and model weights encrypted and isolated even while in use in memory, with up to 2x lower latency than 8t and a near-zero performance penalty.
  • TPU 8tx also handles small-scale training on the same system. Teams can fine-tune, distill, and run reinforcement learning on private data without moving it between clusters, which simplifies compliance for sensitive workloads.
  • TPU 8i Max puts two TPU 8i chips in one package, doubling compute, SRAM, and HBM bandwidth versus 8i within the same rack footprint.
  • It targets multi-agent workloads over large codebases or long video, where memory bandwidth is the bottleneck, and increases compute per square foot and per watt.
  • The software stack is the bigger story for developers. TorchTPU is now open source, so native PyTorch and Hugging Face models run on TPUs without rewrites, with vLLM and SGLang backends.
  • TPU Sync, also open source, claims over 90 percent lower KV transfer latency and time to first token by tiering memory across HBM, DRAM, and storage.
  • Cloud Pathways and GKE dynamic slicing are now GA.
  • Dynamic slicing uses optical circuit switching to cut job startup times by up to 5x and fault recovery by up to 4.5x. Agent Compute Engine suspends idle agent sessions to cut costs.
  • No pricing or regional availability was given in the post.

50:52 Cloud Pathways, a distributed ML runtime, is GA

  • Google has taken Cloud Pathways, the distributed ML runtime behind Gemini training, to GA on GKE, which matters to anyone running large TPU fleets.
  • Cloud Pathways replaces multi-controller JAX with a single-controller architecture built on the open-source IFRT.
  • Setting JAX_PLATFORMS=proxy lets one JAX client orchestrate multiple TPU slices across thousands of chips as if they were local, simplifying MPMD workflows like disaggregated inference and RL loops.
  • Resilience features include elastic training using host-memory snapshots, which reshards down to N-1 slices on hardware failure without crashing the client.
  • Co-located Python sidecars run data loading and checkpointing directly on TPU hosts, and a Shared Pathways Service lets multiple users share a TPU slice pool.
  • Customer usage is substantial: over 130 million TPU chip-hours in 90 days. One large consumer tech company runs inference across tens of thousands of v5p, Trillium, and Ironwood chips at over 100 million chip-hours per quarter, and ZenteiQ uses it for elastic pre-training and RL on Trillium clusters.
  • Availability: GA on GKE in all regions supporting TPU v5e, v5p, Trillium (v6e), and Ironwood (TPU 7x), with JAX 0.10.1+ required. A Preview integration with Cluster Director extends it to Slurm TPU clusters. Deployment is via Cluster Toolkit (gcluster) or the gcloud CLI. No separate Pathways pricing was stated, so cost appears to be the underlying TPU and GKE usage.
  • Worth discussing: this exposes Google’s internal frontier-model runtime to all customers and ties into the AI Hypercomputer stack, though it is JAX/TPU-centric, so teams on GPUs or PyTorch see limited direct benefit.
  • Docs are available here.

51:26 GKE dynamic slicing on TPUs

  • Google just made GKE dynamic slicing for TPUs generally available, which changes how teams carve up and recover large TPU capacity for AI training and inference.
  • Dynamic slicing decouples logical TPU slice formation from physical nodepool provisioning, so teams no longer delete and recreate static nodepools when workload topology changes. Google claims up to 5x faster job startup and 4.5x faster workload recovery.
  • Super-slicing uses TPU optical circuit switching to stitch multiple 16-VM (4x4x4) TPU cubes into larger topologies at schedule time, up to single slices of around 8,000 chips. Sub-slicing does the opposite, partitioning one cube into smaller isolated slices so multiple jobs can share a nodepool.
  • Failure domains shrink to the cube level, so an unhealthy cube can be swapped out in one to two minutes instead of reprovisioning a whole slice. A new incremental provisioning nodepool type also brings up the healthy hosts in a nodepool even when some hosts are faulty, then adds the rest as they recover.
  • Setup requires a one-time conversion of TPU capacity into uniform 16-VM nodepools, then workload annotations trigger the GKE slice controller to create a Slice custom resource at runtime.
  • You can use the provided GKE scheduler or bring your own custom scheduler.
  • Practical limits: it is available in all regions with TPUs but only under TPU All Capacity in reservation mode, and no separate pricing was announced, so cost is standard TPU and GKE rates.
  • The main wins are for platform teams running mixed training and inference fleets who want higher utilization and fewer idle chips.
  • Docs are available here.

52:01 Introducing TPU Sync: Fast KV cache & weight transfer on TPU

  • Google has open-sourced TPU Sync, a C++ library that moves KV cache and model weights across TPU memory tiers to cut inference latency and idle accelerator time for large models.
  • TPU Sync treats TPU HBM, pinned host DRAM, shared storage like Google Cloud Managed Lustre, and the data center network as one memory hierarchy, using asynchronous DMA transfers that overlap with active computation.
  • Check out the Repo for more info.
  • Google reports over 90% reduction in KV transfer latency (Qwen3 235B median dropped to about 230 ms) and an average 60% to 90% TTFT reduction for 235B to 480B MoE models in disaggregated prefill/decode serving. Results come from internal July 2026 testing on TPU v7x under controlled, low-concurrency conditions so that real-world results will vary.
  • Prefix caching extends beyond HBM into host RAM, with an experimental SSD and Lustre tier. Google cites up to 90% TTFT reduction on 64K-context workloads, over 50% higher throughput, and 30% lower P99 inter-token latency. Useful for long-context agents, multi-turn chat, and apps with heavy shared system prompts.
  • Technical approach: layout-preserving direct DMA keeps tensors in native TPU tile format, with no XLA compilation in the data path. Throughput reaches up to 223 GB/s per Ironwood host over PCIe Gen5, and up to about 350 Gbps across dual 200 Gbps NICs on the DCN.
  • Ecosystem integration covers vLLM-TorchTPU, SGLang-TorchTPU, and llm-d 1.0 on GKE, with Python wheels for PyTorch and JAX. It also provides experimental RL weight-transfer primitives for trainer-to-sampler sync. Meta and Wayfair are named as adopters or evaluators. Open source, so no separate licensing cost is stated; you pay for the underlying TPU, GKE, and Lustre resources.
  • Hosts may want to discuss how this fits the broader shift from compute-bound to memory- and state-bound AI infrastructure, and how it compares to GPU-side KV cache offload tooling.

52:22 Matt – “That’s pretty cool, with the slicing in different ways. You know, I feel like that’s one thing Google’s always done, which is like build your own shape of servers, and it feels like they’re bringing that to more features of the of their platform. And it’s a nice one to have because I don’t know that any other of the two cloud providers have that.”

53:35 How to run CodeMender in your pull requests

  • Google published a reference architecture that wires its CodeMender security agent into GitHub pull requests via Cloud Run, which matters for anyone struggling with security fix backlogs as AI-generated code volume grows.
  • CodeMender is Google’s autonomous code-security agent, part of its AI Threat Defense platform.
  • The reference architecture posts scan findings as a PR comment, and developers reply with slash commands: /verify to simulate an exploit, /fix to generate a patch diff, and /approve to commit it and set a security/codemender-gate commit status. Branch protection is required to make that gate a hard merge block rather than advisory.
  • Context cited: Mandiant M-Trends 2026 reports average developer fix time for flaws grew from roughly 59 days a decade ago to over 170 days in 2025, while some teams see up to 2x development velocity from AI tooling.
  • Verification runs untrusted, generated exploit code as a one-per-finding Cloud Run job in the second-generation microVM environment. Hardening steps include a dedicated minimal service account (avoiding the Compute Engine default account), Direct VPC egress into a subnet with no outbound route, and Cloud Run sandboxes to block access to the metadata server. Results stream to Cloud Storage for an audit trail.
  • The post also advises marking crashed or timed-out tests as inconclusive rather than safe, to avoid failing open.
  • Auth between GitHub and Google Cloud uses short-lived GitHub Actions OIDC tokens via workload identity federation, so no long-lived keys are stored.
  • Google recommends mapping the immutable repository_id and repository_owner_id attributes rather than repository names, which can be renamed or reclaimed.
  • A reference implementation is on GitHub at github.com/GCP-Architecture-Guides/codemender-gitops-starter-kit.
  • The post gives no pricing or regional availability details for CodeMender itself, and costs for the Cloud Run jobs and Cloud Storage would follow standard usage-based pricing.

Azure

54:49 Public Preview: Multi-disk crash-consistent VM restore points for Premium SSD v2 and Ultra Disks on VMs with SCSI disk controller

  • Azure just added multi-disk crash-consistent restore points for Premium SSD v2 and Ultra Disks, which matters to anyone running performance-sensitive databases or workloads on Azure VMs.
  • The feature captures crash-consistent restore points across multiple Premium SSD v2 and Ultra data disks on a VM at the same point in time, without needing an agent inside the guest OS. It is in public preview and limited to VMs using SCSI disk controllers.
  • Crash-consistent means the snapshot reflects disk state as if the VM lost power at that moment, not application-consistent. Workloads like databases will recover via normal log replay, but it is not a substitute for app-aware backups where those are required.
  • Premium SSD v2 and Ultra Disks are Azure’s highest-performance block storage tiers, typically used for SQL Server, Oracle, SAP, and other latency-sensitive workloads. Previously these disk types had limited restore point support, so this closes a data protection gap for them.
  • The SCSI controller requirement is worth noting, since newer VM series increasingly use NVMe controllers. Customers should check their VM family and disk controller type before planning around this feature.
  • Preview features generally carry no SLA and are not recommended for production.
  • Restore points are billed based on stored snapshot data, so costs scale with change rate and retention. See the Azure updates page for details.

56:29 Matt – “AWS has also had this for years, I feel like. So it it’s amazing that it took them this long to get this.”

57:42 Announcing flex nodes for AKS (preview) in Azure portal

  • Microsoft just added Azure portal support for flex nodes in AKS preview, letting you attach user-managed compute to an existing cluster without running another Kubernetes control plane.
  • Flex nodes let you extend an existing AKS cluster with worker nodes on user-managed VMs or on-premises servers, in other Azure regions, on-prem, or at the edge. The new portal flow lives under the cluster’s Node pools page via Add flex node pool, with deployment progress shown during onboarding.
  • Onboarding paths cover supported Azure VMs and Azure Arc-enabled servers. For bring-your-own hosts outside Azure, you download an onboarding script from the portal and run it on the host.
  • Confirm nodes show Ready before scheduling workloads, as status can take a few moments to appear.
  • Flex nodes join the cluster’s normal scheduling model, so labels, node affinity, and taints and tolerations can steer selected workloads to them while other workloads stay on standard AKS node pools.
  • Use cases include tapping capacity in another region, reusing existing infrastructure, and placing workloads near data or devices.
  • Customers remain responsible for the underlying hosts, their security, and network connectivity, and preview prerequisites still apply (registrations, identity permissions, connectivity). Microsoft also notes workload placement can support data residency strategy but does not establish compliance on its own.
  • Pricing was not stated in the post, so costs likely depend on the host compute and any Arc usage.
  • Worth discussing: how this compares to AKS hybrid and Arc-enabled Kubernetes options, and whether skipping a second control plane is enough to make this attractive for edge and on-prem.
  • Docs here (search Flex nodes for AKS), feedback via the AKS Flex Node GitHub repository.

Emerging Clouds

59:03 Introducing DigitalOcean Cuttlefish: Get more done with agents on the go

  • DigitalOcean just launched Cuttlefish, a native iPhone app for supervising its Managed Agents, so you can approve or steer long-running AI agents from anywhere.
  • Cuttlefish targets the human-in-the-loop stall: agents running for hours or overnight pause for permission, and the app lets you review findings, add context, approve steps, or start new work from your phone.
  • It builds on DigitalOcean Managed Agents (Public Preview), where each session runs in an isolated, DigitalOcean-hosted microVM that persists independent of any device, so teammates can join or take over a run.
  • The app is harness-agnostic, supervising Claude Code, Codex CLI, OpenCode, OpenHands, Hermes, LangGraph, and CrewAI, with access to 20,000+ tools. That avoids lock-in to a single vendor’s agent.
  • Permission controls let teams define which actions agents can take autonomously and which are off limits, which is the main safety mechanism for letting agents run unattended.
  • Box Mount (private preview) syncs a Box folder into the agent workspace under Box permissions, version history, and audit trail. Note that agents can edit and delete files, and deletions sync back to Box, which is worth discussing from a risk standpoint.

1:00:15 Justin – “Thanks, DigitalOcean. We appreciate that.”

Cloud Journey

1:00:18 Behind the scenes, two months later: How we ship Google Agent Skills like products

  • Google has published a look at how it ships Agent Skills like products, with CI gates and an AI reviewer that any team maintaining an agent skill repo could borrow.
  • The repo has passed 21,000 GitHub stars and hosts over 170 skills. At that scale, Google treats a skill as a behavioral change to an agent rather than documentation, which is why every edit needs versioning, validation, and evaluation evidence.
  • Every pull request passes automated gates before human sign-off: structure checks, linters, semantic versioning, eval suite validation, link checks, and an AI review agent.
  • Cheap deterministic checks run first and block the pipeline before costly LLM-based checks execute, which is a practical cost-control pattern.
  • Google uses a skill to review skills. The AI reviewer flags the deterministic trap, where rigid CLI sequences are written into Markdown instead of scripts, since LLMs are probabilistic and can hallucinate on strict command sequences.
  • The reviewer enforces a three-tier layout (metadata, a core SKILL.md under 500 lines, and flat reference docs) and blocks chained references, since agents often stop before following a second hop. Scripts must be non-interactive, support –help and –dry-run, and separate stdout from stderr.
  • Lessons learned: write atomic eval expectations (one criterion each) so LLM-as-judge scoring stays consistent, and generate catalogs like README and index.json rather than editing them by hand, since docs drift within days. Overly broad CI path patterns also blocked trivial typo fixes while missing new skills.

After Show

1:00:18 Amazon’s Ring tries again with a $250 flying home security drone and a $1,000 patrol robot

  • Amazon’s Ring just announced a $250 indoor flying drone and a $1,000 patrol robot, both revisiting products Amazon previously shelved, and it’s a useful case study in how AI-driven edge devices tie back to cloud subscriptions.
  • Ring Drone Cam ($249.99, slated Jan 2027, pending FCC authorization) and Ring Rover Cam ($999.99, ships Dec 16) use an on-device system called AI Pilot to map spaces and navigate obstacles on their own.
  • Both products are retries: the Always Home Cam drone from 2020 never reached general sale, and Astro for Business launched at $2,350 in 2023 before being discontinued with refunds in under a year. The Rover costs less than half what its predecessor did.
  • Scheduled flights, patrols, and reviewing the Rover’s recorded video require a Ring Protect subscription, and Ring hasn’t said which plans qualify. The hardware is the entry point and the recurring cloud service is where the business model sits.
  • The Rover extends Ring’s push into commercial security, following a $5,000 surveillance trailer in January, and puts it up against Knightscope, which rents robots by subscription, and Cobalt AI.
  • A roaming camera that maps interiors and triggers on other Ring devices raises ongoing questions about video storage, access controls, and privacy, which matter to anyone building or securing cloud-connected edge devices.

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





Download audio: https://episodes.castos.com/5e2d2c4b117f29-10227663/2699438/c1e-nj4jt81dqxc2gd4d-ndd3j3x3i5p6-ru9ezy.mp3
Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete

How Business AI Use is Changing

1 Share
From: AIDailyBrief
Duration: 24:22
Views: 152

Business AI usage is growing much faster than spending as token prices fall and buyers gain leverage. NLW breaks down new data from Ramp, Databricks, and Arcana on model switching, spend concentration, and where enterprise agents really stand. In the headlines: the Nvidia board shutting the door on a Jensen return, new remote-access chip rules, and OpenAI’s embedded safety evaluators.

The AI Daily Brief helps you understand the most important news and discussions in AI.
Subscribe to the podcast version of The AI Daily Brief wherever you listen: https://pod.link/1680633614
Get it ad free at http://patreon.com/aidailybrief
Learn more about the show https://aidailybrief.ai/

Read the whole story
alvinashcraft
2 hours ago
reply
Pennsylvania, USA
Share this story
Delete
Next Page of Stories