Slurm Learns Topology — Rack-Scale Scheduling Becomes a Hard Constraint

1. NVIDIA GB200 NVL72 Forces Schedulers to Think in Racks, Not Nodes

When NVIDIA built the GB200 NVL72, they extended NVLink coherence across an entire rack — seventy-two Blackwell GPUs unified by fifth-generation NVLink at one hundred thirty terabytes per second aggregate bandwidth. That was the hardware story. This week the software story caught up: NVIDIA's technical blog published a detailed guide on Slurm's topology/block plugin and the new --segment argument, and it makes clear that rack-scale locality is no longer a preference — it is a hard constraint.

The core problem is straightforward. A scheduler that treats an NVL72 rack as a bag of independent nodes will fragment workloads across domain boundaries. When that happens, performance drops over ten percent in tokens-per-second throughput. At scale, that regression is not a rounding error — it is a meaningful operational cost. Slurm's block plugin solves this by treating each NVLink domain as an atomic scheduling unit. Jobs either land inside the block or they wait. The new IMEX (inter-process memory exchange) integration adds GPU memory isolation at the driver level, letting administrators run multi-tenant workloads on rack-scale hardware without the isolation risks that normally come with shared coherence domains.

What this means for infrastructure teams is that GPU cluster scheduling is now a topology-aware discipline, not a compute-allocation exercise. Every tool in the stack — from Slurm to Kubernetes operators to bare-metal orchestrators — needs to understand that rack boundaries are performance boundaries. This is the same principle that made AI fabric design a networking problem in the first place, and it is migrating further up the stack into the scheduler layer.

So What? Audit your GPU cluster scheduler configurations against GB200 NVL72 topology today. If your Slurm deployment does not have the topology/block plugin configured and NVLink domains declared as atomic blocks, you are leaving meaningful throughput on the table and potentially causing unpredictable job performance. Add IMEX configuration to your NVL72 runbook before your first multi-tenant workload lands.

SourcesNVIDIA Technical Blog, NVIDIA Technical Blog — Rack-Scale Workloads

2. Cloudflare Cuts Twenty Percent of Workforce, Cites Agentic AI Restructuring

Cloudflare announced layoffs of eleven hundred employees — roughly twenty percent of its total workforce — citing AI-driven changes to how the company operates internally. CEO Matthew Prince and COO Michelle Zatlyn told staff in an internal memo that "the way we work at Cloudflare has fundamentally changed," and that Cloudflare's own use of AI had increased by over six hundred percent in the past three months. The company estimates restructuring charges of between one hundred forty million and one hundred fifty million dollars. Stock fell eighteen percent on the announcement.

This is not a demand or revenue story. Cloudflare is growing. This is an internal productivity restructuring: the same headcount can now produce more, and the company is resizing accordingly. The mechanism is agents replacing roles that were previously human-speed — code generation, operations, customer support tooling. That six hundred percent internal AI usage spike in ninety days is the real signal here. It suggests the inflection from "AI as productivity aid" to "AI as headcount substitute" is happening faster inside mature software companies than most outside observers expected.

The infrastructure implications are real. Cloudflare is one of the most important network infrastructure companies on the planet — zero-trust fabric, DDoS mitigation, Workers AI, R2 object storage. If the engineering teams shrink while the product surface stays constant or grows, the bet is that agentic tooling compensates for headcount. That bet either validates at scale or it surfaces quality and reliability regressions over the next twelve to twenty-four months. Worth watching closely.

So What? If Cloudflare is in your infrastructure stack for zero-trust, DDoS, or CDN services, monitor their service quality metrics over the next two quarters. The restructuring is a signal to pressure-test your contingency plans. More broadly: the six hundred percent internal AI usage figure is the most concrete data point yet for how fast organizations are actually shifting operational work to agents.

SourcesCNBC, Bloomberg

3. Anthropic Secures All of Colossus 1 — Two Hundred Twenty Thousand GPUs, Environmental Record Included

Anthropic announced at its Code with Claude event that it has secured access to all compute capacity at SpaceX and xAI's Colossus 1 data center — over two hundred twenty thousand NVIDIA GPUs including H100, H200, and GB200 accelerators, representing more than three hundred megawatts. The agreement also includes a stated interest in partnering on orbital AI compute development. For Claude Pro and Max subscribers, this means near-term capacity relief. Anthropic separately noted API volume is up seventeen times year-over-year.

The story has a second dimension that Simon Willison flagged prominently: Colossus has a documented environmental record problem. The gas turbines installed to power the facility initially operated without Clean Air Act permits or pollution control equipment, classified as "temporary" to avoid permitting requirements. Credible reporting links the facility to increases in hospital admissions related to air quality in the surrounding area. Anthropic acknowledged being "severely compute-constrained" as context for the decision.

This deal completes a picture of Anthropic's compute strategy: up to five gigawatts from Amazon with nearly one gigawatt by end of twenty twenty-six, a five gigawatt agreement with Google and Broadcom coming online from twenty twenty-seven, thirty billion dollars of Azure capacity via a Microsoft and NVIDIA partnership, and now Colossus's three hundred megawatts as an immediate bridge. The speed of the capacity agreements signals how far behind current API volume is running against available GPU supply.

So What? The Anthropic compute story matters for anyone building on Claude APIs — capacity constraints that drove the deal are the same ones that have caused rate limiting and availability issues. The near-term relief is real. The environmental dimension is a legitimate factor for organizations with ESG procurement criteria and should be evaluated explicitly rather than ignored.

SourcesAnthropic, Simon Willison, DataCenter Dynamics

NANOG 97 Lands in Bellevue June First — SRE Is the Spotlight Topic

NANOG 97 convenes at the Hyatt Regency Bellevue, Washington, June first through third, twenty twenty-six. The program committee chose site reliability engineering as the conference's spotlight topic, which makes it a natural follow-on to the automation-centric thread running through AutoCon 5 (Munich, June eighth through twelfth). The combination of NANOG 97 and AutoCon 5 in the same month creates the densest concentration of SP-scale automation and operations content the calendar has seen. Over twenty-six hours of presentations, panels, tutorials, and workshops are scheduled.

The SRE framing at NANOG is notable because it bridges the operator culture gap that the AutoCon 5 AutomationMap tool is diagnosing. AutoCon identified organizational barriers as the primary adoption gap for network automation — not tooling. NANOG SRE content is likely to surface the same tension: production reliability practices that assume human-speed response hitting friction as automation tooling changes the response time contract.

So What? Register for NANOG 97 now if you can attend — the SRE focus is directly relevant for anyone operationalizing network automation. If you cannot attend in person, the presentations publish publicly after the conference. Plan the June calendar: AutoCon 5 workshops start June eighth, four days after NANOG closes.

SourcesNANOG

Segment Routing 101 Videos Now Public — Free Entry Point for the SR Journey

ipSpace.net released Jeff Tantsura's twenty seventeen "Introduction to Segment Routing" webinar videos as free public content — no account required. Ivan Pepelnjak's framing note is blunt: at ITNOG 10 in mid-April twenty twenty-six, approximately one hundred engineers attended an SR workshop, suggesting the technology remains unevenly adopted nearly a decade after the foundational work was done. The free release is a deliberate on-ramp for teams early in their SR journey.

This is a small story with a pointed subtext. Segment Routing is not new. SR-MPLS has been in production at hyperscalers for years. SRv6 uSID is live in Microsoft's AI fabric (covered April thirtieth) and is shipping in SONiC 202505 (May thirty-first). The gap between where the bleeding edge is and where most enterprise operators are standing is wide. These videos are the first rung of a ladder that ends at SRv6-enabled AI fabric design.

So What? If anyone on your team is still treating Segment Routing as a future consideration, point them at the ipSpace.net free videos as a starting point. The practical destination — SRv6 uSID in SONiC for AI fabric — is already in production. The gap is not technology readiness; it is familiarity.

SourcesipSpace.net

NVIDIA and IREN Announce Up to Five Gigawatts of DSX AI Infrastructure

NVIDIA and IREN announced a strategic partnership to deploy up to five gigawatts of AI infrastructure aligned to NVIDIA's DSX AI factory architecture. The flagship deployment starts at IREN's two-gigawatt Sweetwater campus in Texas. NVIDIA received a five-year warrant to purchase up to thirty million IREN shares at seventy dollars per share — a potential two-point-one billion dollar investment — subject to conditions including regulatory approval. IREN had separately acquired Mirantis OpenStack and Kubernetes distribution earlier this week, signaling a move toward full-stack platform rather than raw GPU rental.

The DSX architecture angle is the technically interesting part. DSX (short for data center-scale accelerated computing) is NVIDIA's reference design for AI factories: NVL72 rack units as compute atoms, high-speed Ethernet or InfiniBand fabric between racks, and integrated DPU-level security and observability. A five-gigawatt commitment to DSX architecture from a major neocloud is a validation signal for the reference design itself — and it standardizes the stack on which future NVIDIA hardware generations will be deployed.

So What? NVIDIA is not just selling hardware — it is locking in architecture commitments. The DSX reference design embedded in this five-gigawatt deal shapes what Sweetwater's fabric, scheduling, and networking stack will look like for years. If you are evaluating AI cluster architectures, treat DSX as a real reference point alongside the hyperscaler-specific designs.

SourcesNVIDIA Newsroom, Data Center Knowledge

Neocloud Full-Stack Buildout Accelerates — IREN Plus Mirantis as Case Study

This week's IREN activity — acquiring Mirantis on Wednesday, announcing the NVIDIA five-gigawatt DSX partnership on Thursday — tells a coherent story. Neoclouds are under pressure from two directions: hyperscalers offering GPU capacity as part of integrated cloud suites, and enterprise buyers who want managed platforms rather than raw accelerated compute. The Mirantis acquisition gives IREN an OpenStack and Kubernetes management layer. The NVIDIA partnership locks in hardware and reference architecture. Together they represent a neocloud attempting to build a differentiated full-stack platform instead of competing on GPU spot pricing.

The broader neocloud supplement published this week by DataCenter Dynamics reinforces the trend: the GPU rental model is compressing in margin as hyperscaler AI infrastructure products mature. The neoclouds with durable positions will be those that add operational software value on top of the hardware. IREN is making explicit bets on both the software layer (Mirantis) and the architecture layer (DSX).

So What? When evaluating neocloud providers for AI workloads, ask explicitly about their managed platform layer. Raw GPU availability is table stakes. The question is what operational tooling, observability, and architecture support they provide above the bare metal. IREN's moves this week define one answer to that question.

SourcesNVIDIA Newsroom, The Register

Q-CTRL and IBM Achieve Three Thousand Times Speedup on Quantum Materials Simulation

Q-CTRL and IBM achieved a three-thousand-times speedup in simulating the Fermi-Hubbard model on one hundred twenty qubits using runtime error suppression — demonstrating practical quantum advantage over classical methods for materials science simulation. The Fermi-Hubbard model describes strongly correlated electron behavior in materials and is classically intractable at the scales where quantum effects matter. This result runs on existing hardware without fault-tolerant quantum error correction, which makes it categorically different from the threshold demonstrations that have dominated quantum milestone coverage.

This is the "useful without fault tolerance" path materializing. IBM Heron's forty-qubit spin transport simulation (published mid-April in Physical Review Letters) showed the same pattern: mid-circuit measurement algorithms that work within current hardware's error budget by designing around the noise rather than correcting it. The Q-CTRL result extends that approach to one hundred twenty qubits and to a model with direct industrial applications in materials discovery and drug design. The speedup is not theoretical — it is measured against the best classical methods for the same problem.

So What? Quantum-classical hybrid applications in materials science are arriving before fault tolerance. The infrastructure implication is that quantum compute resources will be consumed alongside classical GPU and CPU resources in research workloads within a three-to-five year horizon. Start thinking about quantum-classical hybrid rack layout and network latency requirements now — the same planning horizon as any other compute infrastructure investment at that scale.

SourcesQ-CTRL, Quantum Computing Report

TrustFall — MCP Client Trust Model Is the New Attack Surface

Adversa AI disclosed TrustFall, a one-click remote code execution vulnerability class affecting Claude Code, Gemini CLI, Cursor CLI, and GitHub Copilot CLI. The mechanism is architectural: all four tools execute project-defined MCP servers immediately after the user accepts a folder trust prompt. A malicious repository ships an attacker-controlled MCP server and auto-approves it via the project's settings file. One Enter keypress to accept the trust dialog is sufficient to spawn the server as an unsandboxed OS process with the developer's full system privileges.

The architectural lesson here is distinct from the MCP server identity and cryptographic signing story from earlier this week (the MDPI paper on Sigstore-based MCP trust). That paper addressed the question of how you verify a server is who it claims to be. TrustFall addresses a prior question: how does the client decide whether to execute a server at all? The answer across all four affected tools is "after one modal dialog," which is not a secure trust model when the artifact being evaluated is code from an unreviewed repository. Anthropic's response to The Register — that users should not have clicked OK on untrusted repos — is technically accurate but sidesteps the design question of why auto-execution is the default behavior.

So What? Apply the same trust model to MCP server execution that you apply to arbitrary code execution in a CI/CD pipeline: repositories from external contributors should not auto-execute anything without explicit review. Configure your MCP-capable tools to require explicit per-server approval, not per-project blanket trust. This applies to every developer on your team using Claude Code, Cursor, or Gemini CLI.

SourcesAdversa AI, The Register

• Quantizing With Randomized Hadamard Transforms proven optimal — arXiv paper demonstrates that composing two RHTs on any input vector gives provably near-Gaussian marginal distributions, formalizing what was previously a heuristic used in KV-cache compression, gradient compression, and model weight quantization. Important for teams building their own quantization pipelines.
• DataVita OpenClaw Challenge — data center operator DataVita is running a hiring contest: build an AI tool for data center operations, win a permanent role at thirty-five thousand pounds starting salary. Interesting signal that operator-side AI tooling talent demand is turning into direct hiring pipelines.
• NANOG 97 RSVP reminder — registration is open; June first through third in Bellevue. SRE spotlight topic, twenty-six-plus hours of content. This is the highest-density SP-scale operations event before AutoCon 5 Munich two weeks later.

SourcesarXiv, Data Center Knowledge, NANOG

• SONiC 202505 drops May thirty-first — run the 202504-to-202505 diff against your deployment templates before the stable release. DPU dark-mode, SRv6 uSID, and per-VLAN STP are the headline additions.
• AutoCon 5 Munich — June eighth through twelfth, workshops separate registration. Psychological barriers keynote and NAF Framework track are the two highest-signal sessions for teams mid-automation journey.
• Cloudflare service quality metrics — the eighteen-percent stock drop after the layoff announcement is not the story. The story is whether eleven hundred fewer engineers affects incident response time, feature velocity, or reliability SLAs over the next two quarters. Start your baseline measurement now.
• Q-CTRL quantum results — watch for independent replication of the Fermi-Hubbard three-thousand-times speedup. A second team confirming practical quantum advantage at one hundred-plus qubits on current hardware would change the timeline framing significantly.

Pipeline stats: five domains researched, eight web searches, nine primary stories, three quick takes, quality score four point five out of five.