AWS Buries Fat-Tree — Random Graph Networks Are Now the Cloud Default

1. AWS Resilient Network Graphs — Expander Math Kills Fat-Tree as Cloud Default

TL;DR: AWS has quietly replaced the industry's default datacenter fabric with a flat topology based on expander graph theory. Resilient Network Graphs became the default for all new non-GPU AWS datacenters globally in April 2026, delivering 69% fewer routers, 33% higher throughput, and a projected 40% reduction in network equipment power consumption.

Key Points:

• ShuffleBox is a passive optical device with internally shuffled fiber wiring that creates quasi-random inter-rack connectivity. Because it's passive, it adds no latency, consumes no power, and has no failure mode of its own — randomness comes from the physical cabling topology, not electronics.
• Spraypoint is the custom distributed routing protocol that operates in two phases: source routers spray packets randomly across neighbors, while designated waypoints near each destination act as attractors guiding traffic to endpoints. This creates nearly twice as many independent paths as conventional shortest-path routing.
• Failure behavior is fundamentally different from fat-tree: losing 1% of routers results in roughly 1% capacity loss — proportional degradation rather than catastrophic spine collapse. In fat-tree, losing a spine switch cascades into severe bottlenecks.
• First production deployment went live in Ireland (end of 2024); Germany and Spain followed. Made global default in April 2026. GPU/AI clusters still use the dedicated UltraServer architecture since AI training generates centralized all-to-all traffic that is incompatible with RNG's random distribution assumptions.
• 530 processor-years of simulation validated the design before production. Amazon Science published the full technical writeup on May 28, 2026.
• Unit cost savings of 9–45% compared to traditional fat-tree designs.

So What?

This matters for every network engineer thinking about datacenter fabric design for the next five years. The "three-tier fat-tree is the safe choice" assumption is now empirically wrong at the largest scale anyone has ever built. Watch for the expander-graph design pattern to migrate downstream to enterprise and co-location vendors the same way disaggregated networking, merchant silicon, and SONiC did — with a 2–3 year lag. If you're specifying a fabric refresh today, ask your vendor what their roadmap looks like on flatter, randomized topologies. If they haven't heard of Spraypoint, that's a gap.

SourcesInfoQ, DataCenter Dynamics, Tom's Hardware

2. Anthropic Files for IPO the Same Week It Argues Someone Should Slow Down AI

TL;DR: Anthropic confidentially filed a draft S-1 with the SEC on June 1, 2026 — one of the largest AI listings ever attempted at a nearly one-trillion-dollar valuation — while simultaneously publishing a blog post arguing it would be "good for the world" to slow or temporarily pause frontier AI development. The timing is not accidental; it's an accurate picture of where the most capable AI lab in the world actually stands.

Key Points:

• Anthropic co-founders Jack Clark and Marina Favaro published a paper on recursive self-improvement arguing that Claude's code contributions to Anthropic's own codebase jumped from single-digit percentages to over 80% by May 2026. Task completion speeds doubled every four months; their latest model handles 12-hour tasks versus four-minute tasks for earlier versions.
• The paper argues enforcement of any pause would resemble nuclear treaties — requiring global agreement among all major AI labs plus policy support. Training runs are harder to verify than missile silos, creating enormous defection incentive.
• Anthropic's S-1 filing came one week after the company raised $65 billion in a Series H funding round, pushing its valuation to $965 billion.
• The White House issued an executive order on June 2, 2026 directing federal agencies to establish a voluntary review process for frontier AI models that could pose risks to critical infrastructure, financial networks, government operations, or national security before public release.
• The two events together — a lab arguing for a pause while racing for public markets, and a government creating a voluntary review framework that labs can largely self-certify — illustrate the structural gap between what AI safety requires and what competitive dynamics allow.

So What? For infrastructure teams, the Anthropic IPO plus the White House order signals something concrete: the governance conversation around frontier AI is now entering financial and legal frameworks, not just ethics papers. The White House order specifically calls out critical infrastructure — which includes the networks and datacenters this audience builds. The "agentic AI governance" arc that ran through AutoCon 5, NANOG 97, and the Copilot Cowork exfiltration stories this week isn't a theoretical concern. It's acquiring regulatory weight. If your organization is deploying AI agents with access to network infrastructure, the question "what does your approval gate at the action boundary look like" has taken on a new dimension.

SourcesThe Register, CNBC, Data Center Knowledge

3. NANOG 97 Closes — SSH Is Structural Debt, SRE Is the New Frame

TL;DR: NANOG 97 in Bellevue wrapped June 1–3 with site reliability engineering as the explicit conference theme. The most operationally useful session was Amazon's Brett Lykins laying out the structural problems of SSH at scale — no transactional semantics, no structured data output, no streaming telemetry, compounding failure modes at fleet size. This is the clearest operator-level articulation of why SSH is now automation debt, not a neutral tool.

Key Points:

• Brett Lykins (Amazon, Sr. Infrastructure Automation Engineer) presented "The Hidden Costs of SSH at Scale" — framing SSH not as a protocol limitation but as an architectural constraint: SSH produces unstructured text, has no native transaction model, provides no streaming telemetry, and accumulates operational risk nonlinearly as fleet size grows. When you're running automation at the scale Amazon operates, SSH is a bottleneck by design.
• Cogent Communications CEO Dave Schaeffer delivered the opening keynote "The SRE of AI — Engineering Network Reliability for the Tokenized Era," framing the next phase of network operations as applying SRE rigor (error budgets, reliability targets, postmortem discipline) to AI-driven infrastructure — not just to application uptime.
• A SONiC leaf-spine lab workshop and Meta's BPF-based RDMA driver bug detection session were among the technical highlights.
• AWS also ran a separate conference workshop on network observability at scale.

So What? If SSH-driven automation is still the primary method your team uses to push configs, the Lykins presentation is the business case for changing it. The argument isn't "SSH is bad" — it's "SSH doesn't scale, and you'll learn that the hard way at the worst possible time." The replacement path (gNMI streaming + structured config + NETCONF transactions + event-driven orchestration) is now well-documented, production-validated, and commercially supported across all major platforms. Watch the NANOG 97 session recordings when they post at nanog.org — the Lykins session should be in your queue this week.

SourcesNANOG 97, NANOG Stories

VRF-Lite Gets a VXLAN Upgrade — ipSpace Drops the Lab

Ivan Pepelnjak published a new netlab exercise implementing VRF-Lite over VXLAN — and it's a quiet architectural win worth understanding. Traditional VRF-Lite requires a dedicated VLAN on every link and every device for each VRF running a routing protocol. The VXLAN-based variant needs only IP routing on core switches: VXLAN tunnels carry the per-VRF routing sessions, dramatically reducing the per-VRF VLAN sprawl. Pepelnjak compares the result to DMVPN — without the IPsec and NHRP complications. The lab runs on free GitHub Codespace, on netlab-enabled infrastructure, or on an Apple Silicon Mac with Arista cEOS.

So What? If you're running VRF-Lite today and dreading the per-device VLAN management overhead, this pattern is worth an hour of lab time. The VXLAN-backed design is simpler to extend, easier to automate (fewer interface-level VRF bindings to track), and maps cleanly to declarative YAML configs in netlab. Run the lab. If the design fits your topology, you're looking at a meaningful reduction in per-change blast radius.

SourcesipSpace.net

AutoCon 5 Munich — First European Edition, June 8–12

The first European AutoCon lands next week at the Westin Grand Munich, June 8–12. Key sessions to watch when recordings post:

• Swisscom's SRv6 keynote: Martin Gysi presents ten thousand devices running SRv6, orchestrated by NSO with zero manual intervention, serving over one hundred thousand service instances — and a candid account of what the four-year journey actually looked like compared to what the architectural PowerPoints promised. This is the most credible large-scale SRv6 production case study available anywhere.
• Michael Bushong (Nokia) opening keynote — "The Cognitive Biases Impacting Network Automation": The argument is that the real obstacles to automation adoption aren't technical — they're psychological. Loss aversion, survivorship bias, and misframed business cases have quietly killed more automation initiatives than any protocol gap ever did.
• DE-CIX production GitOps case study: Lucas Immanuel Nickel from DE-CIX presents two live delivery patterns: a gNMIc config generation job driven from Zabbix and NetBox, and a geospatial network view web service — both shipping via GitOps pipeline in under ten minutes from commit.

So What? If you can get to Munich, go. If you can't, the Swisscom SRv6 keynote and the DE-CIX GitOps session are the two recordings to queue the moment they post. The Swisscom story specifically — ten thousand devices, NSO orchestration, zero manual intervention — is the production evidence that the "SRv6 is too complex to operate at scale" objection is now refuted by a tier-one European carrier.

SourcesAutoCon 5, Program Guide

White House Orders Pre-Release Federal Review of Frontier AI Models

The White House's June 2 executive order establishes a voluntary federal review process for frontier AI models that could pose risks to critical infrastructure, financial networks, government operations, healthcare, emergency services, or national security before public release. Federal agencies are directed to create the evaluation framework; participation is voluntary for labs, but the order signals that voluntary is the transitional state before mandatory.

So What? For infrastructure teams deploying AI in regulated environments or near critical systems, this executive order is the early regulatory signal to act on now. Build your audit and approval infrastructure before it's required, not after. The organizations already running governance frameworks at the tool-call level — not session level — will have a much easier time demonstrating compliance when reporting requirements arrive.

SourcesData Center Knowledge

AMD Takes a Third of Server CPU Units — AI Demand Bucking the x86 Decline

Mercury Research Q1 2026 data: AMD captured 33.2% of server CPU unit shipments in Q1 2026, up six percentage points year-over-year. On revenue, the lead is sharper — AMD holds 46.2% of x86 server CPU revenue, driven by EPYC pricing and AI workload density. Overall x86 processor shipments declined more than 6% in Q1 2026, but server CPUs bucked the trend with more than 10% unit growth year-over-year. Intel holds 18A process leadership on paper from the Clearwater Forest Xeon 6+ launch this week, but supply is constrained with AMD Venice EPYC expected in July.

So What? The competitive dynamics in server CPU have real infrastructure implications: AMD's AI inference density per watt (particularly for agentic workloads running on CPU, not GPU) is now meaningfully differentiated. The Intel Vera CPU and Clearwater Forest both make claims in the agentic workload space — but AMD's market trajectory means both belong in your next RFP evaluation. The Intel 18A process advantage has a narrow window before AMD Venice closes it.

SourcesThe Register

• FEIBN — Federated IBN Evaluation for IIoT (arXiv 2512.20627): A framework combining LLMs for intent translation with federated learning for distributed strategy evaluation, avoiding centralized deployment and rollback in tightly coupled industrial IoT workflows. Score 5.9 in today's RSS. Early-stage research but the federated evaluation pattern — validate intent satisfaction across distributed nodes without a central policy engine — is the right architectural direction for intent-based networking in constrained environments. Watch: when IBN vendors talk about "evaluation," ask whether their evaluation model is centralized or federated. At IIoT scale, the distinction matters operationally.

• ARM eats server share alongside AMD: Mercury Research Q1 data shows ARM-based server CPU shipments are also growing, with Graviton, Ampere, and NVIDIA Grace all contributing. The x86 duopoly is eroding from both sides — AMD from the high-end AI workload segment, ARM from the efficiency-sensitive inference segment. If your 2027 hardware refresh still assumes x86 by default, challenge that assumption explicitly.

SourcesarXiv, The Register

• AutoCon 5 Munich (June 8–12): First European edition. Queue the Swisscom SRv6 keynote and DE-CIX GitOps session recordings when they post. Conference recordings typically post to networkautomation.forum within a week of the event.
• NANOG 97 session recordings: Full recording archive expected at nanog.org within two weeks. Priority queue: Brett Lykins "Hidden Costs of SSH at Scale," Cogent CEO keynote "SRE of AI."
• Amazon Science RNG technical paper: The full technical writeup (ShuffleBox, Spraypoint, 530 processor-year simulation) was published May 28 on Amazon Science. Worth a careful read for anyone doing datacenter fabric architecture.
• Anthropic IPO process: Public filing will be the next signal. The S-1 will contain infrastructure cost disclosures and compute capacity commitments that are worth reading as an infrastructure engineer, not just as an investor.

• Articles processed: 62 (RSS digest, 22 feeds) + 8 supplemental web searches
• Topics researched: networking, automation, ai-ml, datacenter, science
• Primary items published: 7
• Quick Takes: 2
• Quality score: 4.5/5
• Dedup rejections: 4 (all within 72-hour cooldown: June 4 items — intent drift arXiv 2606.05076 re-covered, Containerlab v0.76.0, NVIDIA CPO Q3450, PJM grid warning)
• Cold open variant: A (story lead — AWS RNG)
• Friday Week in Review: included in podcast