The New Stack Podcast

GPUs dominate today’s AI landscape, but Google argues they are not necessary for every workload. As AI adoption has grown, customers have increasingly demanded compute options that deliver high performance with lower cost and power consumption. Drawing on its long history of custom silicon, Google introduced Axion CPUs in 2024 to meet needs for massive scale, flexibility, and general-purpose computing alongside AI workloads. The Axion-based C4A instance is generally available, while the newer N4A virtual machines promise up to 2x price performance.

In this episode, Andrei Gueletii, a technical solutions consultant for Google Cloud joined Gari Singh, a product manager for Google Kubernetes Engine (GKE), and Pranay Bakre, a principal solutions engineer at Arm for this episode, recorded at KubeCon + CloudNativeCon North America, in Atlanta. Built on Arm Neoverse V2 cores, Axion processors emphasize energy efficiency and customization, including flexible machine shapes that let users tailor memory and CPU resources. These features are particularly valuable for platform engineering teams, which must optimize centralized infrastructure for cost, FinOps goals, and price performance as they scale.

Importantly, many AI tasks—such as inference for smaller models or batch-oriented jobs—do not require GPUs. CPUs can be more efficient when GPU memory is underutilized or latency demands are low. By decoupling workloads and choosing the right compute for each task, organizations can significantly reduce AI compute costs.

Learn more from The New Stack about the Axion-based C4A: 

Beyond Speed: Why Your Next App Must Be Multi-Architecture

Arm: See a Demo About Migrating a x86-Based App to ARM64

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

Enterprises are racing to deploy AI services, but the teams responsible for running them in production are seeing familiar problems reemerge—most notably, silos between data scientists and operations teams, reminiscent of the old DevOps divide. In a discussion recorded at AWS re:Invent 2025, IBM’s Thanos Matzanas and Martin Fuentes argue that the challenge isn’t new technology but repeating organizational patterns. As data teams move from internal projects to revenue-critical, customer-facing applications, they face new pressures around reliability, observability, and accountability.

The speakers stress that many existing observability and governance practices still apply. Standard metrics, KPIs, SLOs, access controls, and audit logs remain essential foundations, even as AI introduces non-determinism and a heavier reliance on human feedback to assess quality. Tools like OpenTelemetry provide common ground, but culture matters more than tooling.

Both emphasize starting with business value and breaking down silos early by involving data teams in production discussions. Rather than replacing observability professionals, AI should augment human expertise, especially in critical systems where trust, safety, and compliance are paramount.

Learn more from The New Stack about enabling AI with silos: 

Are Your AI Co-Pilots Trapping Data in Isolated Silos?

Break the AI Gridlock at the Intersection of Velocity and Trust

Taming AI Observability: Control Is the Key to Success

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

Rob Whiteley, CEO of Coder, argues that the biggest winners in today’s AI boom resemble the “picks and shovels” sellers of the California Gold Rush: companies that provide tools enabling others to build with AI. Speaking onThe New Stack Makersat AWS re:Invent, Whiteley described the current AI moment as the fastest-moving shift he’s seen in 25 years of tech. Developers are rapidly adopting AI tools, while platform teams face pressure to approve them, as saying “no” is no longer viable. 

Whiteley warns of a widening gap between organizations that extract real value from AI and those that don’t, driven by skills shortages and insufficient investment in training. He sees parallels with the cloud-native transition and predicts the rise of “AI-native” companies. As agentic AI grows, developers increasingly act as managers overseeing many parallel AI agents, creating new challenges around governance, security, and state management. To address this, Coder introduced Mux, an open source coding agent multiplexer designed to help developers manage and evaluate large volumes of AI-generated code efficiently.

Learn more from The New Stack about AI Parallelization 

The Production Generative AI Stack: Architecture and Components

Enable ParallelFrontend/Backend Development to Unlock Velocity

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

Many enterprises now run workloads across multiple IT infrastructures rather than a single environment. According to Nutanix, about 60% of organizations deploy this way, creating significant operational challenges. In an episode ofThe New Stack Makers, Deepak Goel, CTO for cloud native at Nutanix, outlined three major issues: operational complexity combined with a shortage of cloud native skills, the difficulty of migrating legacy VM-based workloads to microservices-oriented platforms, and the challenge of running virtual machines and containers side by side, often in silos.

To address these problems, organizations are adopting platform engineering, where specialized teams abstract infrastructure complexity and provide developers with standardized “golden paths” to deployment. Internal developer platforms and IDEs also help by handling observability, security, and infrastructure concerns. Nutanix contributes through its hyperconverged platform, which integrates compute and storage to support both VMs and containers. At KubeCon Atlanta, Nutanix announced NDK 2.0, adding advanced data protection, fault-tolerant replication, and enhanced security, including a partnership with Canonical to deliver a hardened operating system for Kubernetes environments.

Learn more from The New Stack about operational complexity in cloud native environments: 

Q&A: Nutanix CEO Rajiv Ramaswami on the Cloud Native Enterprise

Kubernetes Complexity Realigns Platform Engineering Strategy

Platform Engineering on the Brink: Breakthrough or Bust?

Join our community of newsletter subscribers to stay on top of the news and at the top of your game.

Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Nvidia Distinguished Engineer Kevin Klues noted that low-level systems work is invisible when done well and highly visible when it fails — a dynamic that frames current Kubernetes innovations for AI. At KubeCon + CloudNativeCon North America 2025, Klues and AWS product manager Jesse Butler discussed two emerging capabilities: dynamic resource allocation (DRA) and a new workload abstraction designed for sophisticated AI scheduling.

DRA, now generally available in Kubernetes 1.34, fixes long-standing limitations in GPU requests. Instead of simply asking for a number of GPUs, users can specify types and configurations. Modeled after persistent volumes, DRA allows any specialized hardware to be exposed through standardized interfaces, enabling vendors to deliver custom device drivers cleanly. Butler called it one of the most elegant designs in Kubernetes.

Yet complex AI workloads require more coordination. A forthcoming workload abstraction, debuting in Kubernetes 1.35, will let users define pod groups with strict scheduling and topology rules — ensuring multi-node jobs start fully or not at all. Klues emphasized that this abstraction will shape Kubernetes’ AI trajectory for the next decade and encouraged community involvement.

Learn more from The New Stack about dynamic resource allocation: 

Kubernetes Primer: Dynamic Resource Allocation (DRA) for GPU Workloads

Kubernetes v1.34 Introduces Benefits but Also New Blind Spots

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

At KubeCon North America 2025, GitLab’s Emilio Salvador outlined how developers are shifting from individual coders to leaders of hybrid human–AI teams. He envisions developers evolving into “cognitive architects,” responsible for breaking down large, complex problems and distributing work across both AI agents and humans. Complementing this is the emerging role of the “AI guardian,” reflecting growing skepticism around AI-generated code. Even as AI produces more code, humans remain accountable for reviewing quality, security, and compliance.

Salvador also described GitLab’s “AI paradox”: developers may code faster with AI, but overall productivity stalls because testing, security, and compliance processes haven’t kept pace. To fix this, he argues organizations must apply AI across the entire development lifecycle, not just in coding. GitLab’s Duo Agent Platform aims to support that end-to-end transformation.

Looking ahead, Salvador predicts the rise of a proactive “meta agent” that functions like a full team member. Still, he warns that enterprise adoption remains slow and advises organizations to start small, build skills, and scale gradually.

Learn more from The New Stack about the evolving role of "cognitive architects":

The Engineer in the AI Age: The Orchestrator and Architect

The New Role of Enterprise Architecture in the AI Era

The Architect’s Guide to Understanding Agentic AI

Join our community of newsletter subscribers to stay on top of the news and at the top of your game.

Jonathan Bryce, the new CNCF executive director, argues that inference—not model training—will define the next decade of computing. Speaking at KubeCon North America 2025, he emphasized that while the industry obsesses over massive LLM training runs, the real opportunity lies in efficiently serving these models at scale. Cloud-native infrastructure, he says, is uniquely suited to this shift because inference requires real-time deployment, security, scaling, and observability—strengths of the CNCF ecosystem. 

Bryce believes Kubernetes is already central to modern inference stacks, with projects like Ray, KServe, and emerging GPU-oriented tooling enabling teams to deploy and operationalize models. To bring consistency to this fast-moving space, the CNCF launched a Kubernetes AI Conformance Program, ensuring environments support GPU workloads and Dynamic Resource Allocation. With AI agents poised to multiply inference demand by executing parallel, multi-step tasks, efficiency becomes essential. Bryce predicts that smaller, task-specific models and cloud-native routing optimizations will drive major performance gains. Ultimately, he sees CNCF technologies forming the foundation for what he calls “the biggest workload mankind will ever have.” 

Learn more from The New Stack about inference: 

Confronting AI’s Next Big Challenge: Inference Compute 

Deep Infra Is Building an AI Inference Cloud for Developers 

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

The Cloud Native Computing Foundation has introduced the Certified Kubernetes AI Conformance Program to bring consistency to an increasingly fragmented AI ecosystem. Announced at KubeCon + CloudNativeCon North America 2025, the program establishes open, community-driven standards to ensure AI applications run reliably and portably across different Kubernetes platforms. VMware by Broadcom’s vSphere Kubernetes Service (VKS) is among the first platforms to achieve certification.

In an interview with The New Stack, Broadcom leaders Dilpreet Bindra and Himanshu Singh explained that the program applies lessons from Kubernetes’ early evolution, aiming to reduce the “muddiness” in AI tooling and improve cross-platform interoperability. They emphasized portability as a core value: organizations should be able to move AI workloads between public and private clouds with minimal friction.

VKS integrates tightly with vSphere, using Kubernetes APIs directly to manage infrastructure components declaratively. This approach, along with new add-on management capabilities, reflects Kubernetes’ growing maturity. According to Bindra and Singh, this stability now enables enterprises to trust Kubernetes as a foundation for production-grade AI.

Learn more from The New Stack about Broadcom’s latest updates with Kubernetes: 

Has VMware Finally Caught Up with Kubernetes?

VMware VCF 9.0 Finally Unifies Container and VM Management

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

The etcd project — a distributed key-value store older than Kubernetes — recently faced significant challenges due to maintainer turnover and the resulting loss of unwritten institutional knowledge. Lead maintainer Marek Siarkowicz explained that as longtime contributors left, crucial expertise about testing procedures and correctness guarantees disappeared. This gap led to a problematic release that introduced critical reliability issues, including potential data inconsistencies after crashes.

To rebuild confidence in etcd’s correctness, the new maintainer team introduced “robustness testing,” creating a framework inspired by Jepsen to validate both basic and distributed-system behavior. Their goal was to ensure linearizability, the “Holy Grail” of distributed systems, which required developing custom failure-injection tools and teaching the community how to debug complex scenarios.

The team later partnered with Antithesis to apply deterministic simulation testing, enabling fully reproducible execution paths and easier detection of subtle race conditions. This approach helped codify implicit knowledge into explicit properties and assertions. Siarkowicz emphasized that such rigorous testing is essential for safeguarding the sensitive “core” of large open source projects, ensuring correctness even as maintainers change.

Learn more from The New Stack about the etcd project

Tutorial: Install a Highly Available K3s Cluster at the Edge 

Join our community of newsletter subscribers to stay on top of the news and at the top of your game. 

Helm — originally a hackathon project called Kate’s Place — turned 10 in 2025, marking the milestone with the release of Helm 4, its first major update in six years. Created by Matt Butcher and colleagues as a playful take on “K8s,” the early project won a small prize but quickly grew into a serious effort when Deus leadership recognized the need for a Kubernetes package manager. Renamed Helm, it rapidly expanded with community contributors and became one of the first CNCF graduating projects.

Helm 4 reflects years of accumulated design debt and evolving use cases. After the rapid iterations of Helm 1, 2, and 3, the latest version modernizes logging, improves dependency management, and introduces WebAssembly-based plugins for cross-platform portability—addressing the growing diversity of operating systems and architectures. Beyond headline features, maintainers emphasize that mature projects increasingly deliver “boring” but essential improvements, such as better logging, which simplify workflows and integrate more cleanly with other tools. Helm’s re-architected internals also lay the foundation for new chart and package capabilities in upcoming 4.x releases.

Learn more from The New Stack about Helm: 

The Super Helm Chart: To Deploy or Not To Deploy?

Kubernetes Gets a New Resource Orchestrator in the Form of Kro

Join our community of newsletter subscribers to stay on top of the news and at the top of your game.