NVIDIA's NVLink Fusion Invites Custom Silicon In, Then Charges a Toll

1. NVIDIA's NVLink Fusion Invites Custom Silicon In, Then Charges a Toll

TL;DR: NVIDIA expanded NVLink Fusion with NVHBM, a custom high-bandwidth-memory design that lets hyperscalers plug their own accelerators and CPUs into NVIDIA's rack-scale architecture — but every qualifying rack still needs at least one NVIDIA part, and NVIDIA alone decides who gets a license.

Key Points:

• NVHBM moves the memory controller off the XPU die and into the HBM stack itself: up to 30% more bandwidth per stack, 15% lower power, and up to 67% less PHY/support-area overhead versus stock HBM4e — by NVIDIA's own comparison, not an independent benchmark.
• Sixth-generation NVLink links up to 72 XPUs in a single domain; NVIDIA claims 3x lower XPU-to-XPU latency and 10x higher packet rates than "standard Ethernet alternatives" without naming the Ethernet configuration it's measured against.
• CPU partners: Intel, Fujitsu, Qualcomm. Custom-silicon partners: Alchip, Astera Labs, Marvell, MediaTek. Samsung Foundry just joined for custom-silicon manufacturing, and NVIDIA put $2 billion into Marvell tied directly to the partnership.
• Every NVLink Fusion platform requires at least one NVIDIA product — CPU, GPU, or switch — and NVIDIA is the sole gatekeeper for NVLink IP licenses.
• UALink, the industry's nominal open alternative, is compromised as a genuine counterweight because several of its backers now hold NVIDIA equity.

Deep Dive: NVLink Fusion isn't new — NVIDIA introduced it in 2025 as a way for hyperscalers to plug custom compute into NVIDIA's rack-scale platform instead of building a fabric from scratch. What's new this week is NVHBM, and it's genuinely clever engineering: moving the memory controller off the XPU die and into the HBM stack frees up compute-die area on every accelerator that adopts it — up to 30% more room for logic, per NVIDIA's own numbers — while claiming a bandwidth and power win in the same breath. If those figures hold up under independent testing, that's a real advance, not just a slide.

But read past the partner list NVIDIA is showcasing — Intel and Fujitsu and Qualcomm for CPUs, Alchip and Astera Labs and Marvell and MediaTek for custom silicon, Samsung Foundry just added for manufacturing — and the actual terms of entry look less like interoperability and more like membership dues. Every NVLink Fusion platform has to include at least one NVIDIA part. NVIDIA alone decides who gets a license. And the $2 billion NVIDIA put into Marvell isn't charity — it's buying loyalty from exactly the company best positioned to build a genuinely independent interconnect instead. Contrast that with the two moves we've covered in the past week: the 19-company Open Silicon Photonics coalition, organized explicitly to keep co-packaged optics from becoming another NVIDIA-controlled standard, and Meta's MetaRoCE, an open RDMA transport that runs — on purpose — on someone else's NIC. Both are hedges against exactly the lock-in NVIDIA just built a friendlier-looking version of.

The tell to watch is whether Meta's own MTIA line — its fourth generation just claimed to beat Blackwell at training (see Datacenter, below) — ever shows up on an NVLink Fusion partner list. If Meta, which has both the scale and the motive to build a truly independent interconnect, still ends up paying the NVIDIA toll, that tells you how much real leverage even the biggest buyers have left.

So What? If you're evaluating a semi-custom AI rack in the next two years, treat "NVLink Fusion compatible" as a lock-in clause, not an interoperability feature — budget for the NVIDIA toll regardless of whose accelerator does the actual compute.

SourcesNVIDIA Developer Blog, NVIDIA Blog, Tom's Hardware, The Next Web, ServeTheHome

2. NetBox and Nautobot Both Ship MCP Servers — Source of Truth Becomes the AI Grounding Layer

TL;DR: NetBox Copilot has passed 200 organizations across 25+ countries since its February GA, and Network to Code now ships an official Nautobot MCP Server too — source-of-truth platforms have quietly become the layer AI agents query directly, instead of a database humans occasionally export from.

Key Points:

• NetBox Copilot is embedded across every edition — Community, Cloud, Enterprise — and reuses existing role-based access control rather than adding a parallel permission model.
• NetBox's Platform MCP Server picked up field-level filtering in a November update that cut token usage roughly 10x for a 50-device query — about 5,000 tokens down to around 500.
• Network to Code's Nautobot MCP Server lets any MCP-compatible agent query infrastructure state in natural language instead of through the UI or GraphQL, arriving alongside Nautobot's broader push toward a paid apps marketplace.
• Neither vendor has published specifics yet on write-back guardrails for agents that want to make changes, not just query state.

Deep Dive: This is the quiet version of the AI-infrastructure story — no keynote, no benchmark chart, just two competing source-of-truth platforms independently landing on the same architecture within months of each other. NetBox got there first: Copilot went GA in February, and NetBox Labs says adoption already spans 200-plus organizations in 25-plus countries — a real number, not a pilot-program number. The more interesting engineering detail is the November update to the underlying MCP server: field-level filtering that cuts a 50-device query from 5,000 tokens to about 500 is the difference between an agent you can afford to run continuously and one you only invoke for occasional lookups.

Nautobot's version arrives more quietly, bundled into Network to Code's broader push to commercialize Nautobot 3.1 with a paid apps marketplace. But the strategic signal is the same either way: if you're building any internal copilot or AIOps tool against your network's source of truth, you no longer need to hand-roll a query layer against a REST or GraphQL API. Point an MCP client at the vendor's own server and you inherit whatever access control the platform already enforces — which is also exactly the caveat. Both vendors are shipping natural-language write access as a real, near-term capability, not just read-only lookups, and neither has published detail yet on how an agent's blast radius gets scoped when it's the one making the change. That's the same gap Meta just learned about the hard way (see AI & Machine Learning, below), and it's precisely the problem the new IETF agent-identity draft in this week's Security section is trying to standardize an answer to.

So What? If you're piloting an internal copilot against NetBox or Nautobot, point it at the vendor's own MCP server instead of building a bespoke query layer — but audit the RBAC scoping before you let it write anything back, not after.

SourcesNetBox Labs, NetBox Labs Blog, Nautobot Docs, Access Newswire

3. PJM's Capacity Shortfall Forces a First-of-Its-Kind Emergency Power Auction

TL;DR: PJM asked FERC for permission to run an emergency capacity auction after data-center-driven demand growth left the grid operator 6,831 megawatts short of its own reliability requirement for the 2028/29 delivery year.

Key Points:

• PJM's latest capacity auction cleared 149,181.6 megawatts against a requirement it missed by 6,831.3 megawatts.
• The proposed Reliability Backstop Procurement would lock in new generation on contracts up to 15 years, at a maximum weighted-average willingness-to-pay of $555 per megawatt-day.
• PJM wants FERC approval by September 29, wants the procurement open by September 30, and wants results by December 2 — ahead of the regular 2029/30 base auction.
• PJM's own board projects roughly 70 gigawatts of new large-load demand by 2038 against only about 15 gigawatts of generation retired since 2022.

Deep Dive: This is the concrete, dollars-and-megawatts version of a story we've mostly covered in forecasts up to now. Dell'Oro told us datacenter capex could triple by 2030 — that's a projection. This is a grid operator filing an actual emergency procurement with its federal regulator because the normal multi-year planning cycle already broke. PJM's reliability requirement isn't a nice-to-have target; it's the number that determines whether the grid holds during a stress event, and PJM is telling FERC it's 6.8 gigawatts short of it for a delivery year that's only two years out.

The backstop mechanism is worth understanding on its own terms: 15-year contracts at a capped price is PJM offering long-term certainty to any generator willing to build fast — exactly the kind of underwriting new gas or nuclear capacity needs to pencil out on a timeline shorter than the interconnection queue normally allows. Whether that gets paid for by data center tenants specifically, socialized across all ratepayers, or some hybrid, is the fight actually happening inside this FERC docket right now, and it's the fight every hyperscaler with a build in PJM territory needs to be watching.

For anyone specifying power and interconnection timelines anywhere near PJM's 13-state footprint, the takeaway isn't the megawatt number — it's the calendar. September 29 is the FERC decision, September 30 is when the procurement opens, December 2 is when results land. If your project's power assumptions depend on PJM capacity clearing normally, that calendar is now your calendar too.

So What? If your infrastructure roadmap depends on PJM capacity in the next three to five years, track the September 29 FERC decision directly rather than assuming your utility contact will flag it — this backstop procurement changes what "committed capacity" means in that market.

SourcesData Center Knowledge, PJM Inside Lines, Utility Dive

netlab 26.08 Adds SONiC Containers, ArcOS, and a Fully Open Dataplane

TL;DR: Ivan Pepelnjak's netlab shipped its August release with first-class support for SONiC as containers, Arrcus's ArcOS, and VPP/FD.io — plus new generic ACL and DNS-service modules.

Key Points:

• New device support: SONiC (containerized), ArcOS, and VPP/FD.io with either FRR or BIRD as the control plane.
• Also added: GRE tunnels on Arista EOS, GRE/WireGuard on Mikrotik RouterOS7 and OpenBSD, and Podman support in the containerlab provider (see the Automation section for the debugging story behind that last one).
• Some template/config breakage flagged by maintainers in this release — check the changelog before upgrading an in-use lab.

So What? If you want a team to pilot SONiC before anyone signs a purchase order, point them at netlab — it now runs SONiC, several commercial NOSes, and a fully open dataplane side by side in the same topology, for free.

SourcesipSpace.net

BGPay Proposes Paying Bounties for BGP Hijack Filtering

TL;DR: A new arXiv paper proposes fixing the decade-old incentive problem behind BGP hijack filtering — the network best positioned to filter a hijack bears the cost, while the victim prefix owner gets all the benefit — by having prefix owners post bounties, with public route-collector data acting as the trust root that releases payment through escrow.

Key Points:

• Authors: Sadowy, Doumanidis, and Apostolaki.
• Filterers and monitors have to commit before revealing what they found, so payouts are decided by public route-collector evidence, not the prefix owner's say-so — closing the obvious "just don't pay" failure mode of a naive bounty scheme.
• Pure mechanism-design proposal at this stage — no deployment, no reference implementation.

So What? RPKI and Route Origin Validation adoption has stalled for exactly this incentive problem for a decade — worth a skim if "please just deploy ROV" advocacy has stopped moving the needle for you, though the escrow layer is the part most likely to get simplified into a plain clearinghouse before anyone ships it.

SourcesarXiv

AI-Generated Network Configs Need an Audit Process, Not a Peer Review

TL;DR: On Packet Pushers' Day Two DevOps, Overmind's Dylan Ratcliffe argued that peer review as practiced today assumes a human author whose judgment you're checking — and that assumption breaks once a meaningful share of commits, including network config changes, come from an AI.

Key Points:

• Core reframe: treat AI output as "a machine to be audited," not a colleague to rubber-stamp.
• The practical shift is procedural: move from "a human skims the diff" toward systematic verification against explicit policy and tests.
• Maps directly onto policy-as-code and pre-merge validation pipelines — Batfish- or pyATS-style CI gates — rather than eyeball review of YANG configs or playbooks.

So What? If your GitOps pipeline still gates merges on a human skimming the diff, that's the control to replace first as AI-generated changes increase — move the gate to automated policy and intent checks instead.

SourcesPacket Pushers

NetBox Analytics Turns Source-of-Truth Data Into Dashboards, No Scripting Required

TL;DR: NetBox Labs put NetBox Analytics into public preview — 10-plus pre-built, auto-refreshing dashboards for capacity, utilization, lifecycle, and data quality, sitting directly on NetBox data.

Key Points:

• Requires NetBox 4.5+ on the Cloud edition; general availability targeted later this year.
• The flagship "Executive Overview" dashboard combines capacity, utilization, data quality, and per-site change activity on one screen.
• Additive SaaS feature — no changes for self-managed or Community NetBox.

So What? If you've built a bespoke Grafana layer on top of NetBox's API just to answer capacity and lifecycle questions for management, this is NetBox Labs telling you that layer is about to become a first-party feature — hold off on further investment in the homegrown version.

SourcesNetBox Labs

A Six-Year-Old Doc Comment About netlab and Podman Was Finally Tested — It Needed Five Separate Fixes

TL;DR: Ivan Pepelnjak dug up a never-true claim buried in netlab's docs that Podman support just worked, tried it for real this week, and found five independent, each individually plausible, root causes standing between "the docs say it works" and "it actually works."

Key Points:

• Containerlab now probes for the Docker socket and refuses to start without an explicit runtime flag.
• The runtime environment variable silently failed to survive a sudo call on newer Ubuntu releases.
• Podman scopes container visibility per user, but containerlab needs root to create virtual interfaces — so non-root users couldn't see their own labs.
• Podman's Docker-compatible CLI returns different JSON shapes than Docker's, which broke containerlab's parsing.
• Podman defers DNS resolution to the host instead of running its own embedded server, which broke DNS inside lab containers.
• All five fixes ship in netlab 26.08, due early September, along with a sudoless operating mode and a one-line install helper.

So What? If you'd previously written off Podman as a containerlab runtime because of a locked-down corporate laptop or a rootless CI runner, netlab 26.08 is the first release where that path is actually real instead of aspirational.

SourcesipSpace.net

Alibaba's Qwen3.8-Flash-Next Previews Qwen4 With Open Weights and a Huge Context Window

TL;DR: Alibaba's Qwen team released Qwen3.8-Flash-Next, an open-weight multimodal mixture-of-experts model billed as an early preview of the Qwen4 architecture — 125 billion total parameters with only 6 billion active per token, and a native 262,144-token context window extensible to 1 million via YaRN.

Key Points:

• Hybrid Gated DeltaNet plus Qwen Sparse Attention architecture keeps attention compute and KV-cache memory manageable at long context lengths.
• NVIDIA shipped Day-0 validated inference support across SGLang, vLLM, and TensorRT-LLM on GB300 NVL72.
• Weights published under a "Qwen Community 1.0" license — worth reading the actual terms before assuming unrestricted commercial use, since Alibaba's prior community licenses have carried usage-scale thresholds.
• Alibaba reports 62.5% on SWE-bench Pro and 91.7% on GPQA Diamond; its claimed wins over Claude Opus 4.6 Max on CoWorkBench and JobBench should be read cautiously, since those two benchmarks are far less standardized or independently audited than SWE-bench or GPQA.

So What? A capable long-context coding and agentic model at only 6 billion active parameters is genuinely cheap to self-host relative to its footprint — worth an evaluation slot if you're benchmarking open-weight options, license terms permitting.

SourcesNVIDIA Technical Blog, MarkTechPost, Simon Willison, The Decoder

Meta's Internal AI Agents Took Actions "Humans Are Unlikely to Execute" Before a Layoff Plan Collapsed

TL;DR: A Reuters investigation found Meta explored cutting some teams by up to 60% to go "AI native," then shelved the plan after autonomous agents with broad internal access caused a rising rate of production-impacting incidents.

Key Points:

• Technical and security incidents tied to agent deployment reportedly rose 40% year-over-year; time employees spent responding to those incidents rose 70%.
• Meta reportedly installed keystroke- and mouse-tracking software on some employees' devices specifically to generate training data for the agents meant to replace them.
• Zuckerberg reportedly pulled back after internal data showed the approach wasn't working — not primarily after external or employee pressure.

So What? Before granting any agent broad internal system access, apply the same least-privilege discipline you'd apply to a new automation service account — Meta's incident-rate spike is what happens when that step gets skipped at scale, and it's exactly the gap the IETF's new agent-identity draft (see Security, below) is trying to standardize an answer to.

SourcesTimesLIVE (Reuters), Computerworld, The Globe and Mail

Meta's MTIA 400 Claims to Beat Blackwell at Training — On Meta's Own Benchmark, Against an Unnamed Configuration

TL;DR: Meta's fourth-generation MTIA chip, now aimed at LLM training as well as ad-serving inference, is reported to run about 20% faster than NVIDIA's top-specced Blackwell parts at training-relevant precisions, at similar power draw — a claim worth exactly as much skepticism as every other hyperscaler's internal-silicon benchmark.

Key Points:

• Spec: 8× 36GB HBM3e stacks for 288GB total memory, roughly 9.2 TB/s bandwidth.
• The "20% faster than Blackwell" figure is Meta-reported, on internal-only silicon, with no disclosed Blackwell SKU, precision format, or workload mix beyond "higher precisions more commonly used to train models."
• The same coverage indicates Meta's chip still lags NVIDIA and AMD's latest parts specifically on inference — the "split personality" framing reflects real architectural trade-offs, not a universal win.

So What? Treat "beats Blackwell" claims from any hyperscaler's internal silicon the same way regardless of vendor — ask which SKU, which precision, and which workload before the number means anything. The interconnect choice each of these chips makes matters more long-term than this quarter's throughput claim; see NVLink Fusion, above, for why.

SourcesThe Register, Tom's Hardware

Japan's "Shunkai" Neutral-Atom Quantum Computer Goes Live, Built Like a Server Stack

TL;DR: Japan's Institute for Molecular Science, with Hitachi and US-based Infleqtion, brought online "Shunkai" on August 24 — Japan's first full-stack neutral-atom quantum computer, running roughly 50 qubits today with a funded roadmap to 10,000 error-corrected physical qubits by March 2031.

Key Points:

• Neutral-atom computers trap individual atoms in laser "optical tweezers" and manipulate their quantum states with microwaves or laser pulses; readout is done by imaging each atom's fluorescence.
• Hitachi built the software and control stack; Infleqtion built the QPU hardware — the notable part is that the two were integrated end-to-end as one operational system, the way a server stack is layered, rather than shipped as a single-vendor black box.
• Infleqtion is the only non-Japanese vendor admitted into Japan's Quantum Moonshot program.

So What? No near-term infrastructure action item — this is bookmark-tier — but the framing is genuinely familiar to anyone who thinks in stacks: hardware layer, control layer, orchestration layer, integrated across two vendors and shipped as one operated system. Worth watching as neutral-atom architecture's answer to "can this scale past a lab bench."

SourcesThe Quantum Insider, PR Newswire

Physicists Predict a New State of Matter That Shouldn't Be Able to Exist

TL;DR: Monash University physicists published a peer-reviewed prediction in Physical Review Letters that mixtures of bosons and fermions — two fundamentally different particle types that were expected to fight each other into collapse — can form a stable, self-bound "quantum droplet" when fermionic pressure exactly balances the attractive force between particles.

Key Points:

• Lead author Sam Foster (Monash PhD candidate) and colleagues calculated the specific conditions under which the two particle types stabilize each other rather than dispersing.
• This is a theoretical and computational prediction — not yet observed in the lab.
• The authors explicitly frame it as groundwork for future ultra-precise sensors and quantum computing hardware.

So What? Bookmark-tier, theoretical-prediction-awaiting-confirmation science — but a genuinely surprising result worth knowing about before an experimental group (possibly) confirms it.

SourcesPhys.org, ScienceDaily

An IETF Draft Is Quietly Standardizing How AI Agents Prove Who They Are

TL;DR: draft-klrc-aiagent-auth, now at revision 03 and under active discussion on the OAuth working group mailing list, composes WIMSE, SPIFFE/SPIRE, and OAuth 2.0/2.1 into a reference model — called AIMS — for agent-to-agent and agent-to-service authentication and authorization.

Key Points:

• First published March 2026; three revisions and live OAuth-WG debate in five months is real standards-track movement, not vendor-blog noise.
• Treats agent identity as distributed across an identity provider, an attestation service, an authorization server, a policy engine, and a runtime enforcement point — not a single product.
• This is the same thread flagged in prior checks as "interesting but stale" after February's NIST NCCoE agent-identity concept paper — it's now actually moving.
• Meanwhile: fourth consecutive clean check against CISA's Zero Trust Maturity Model (still version 2.0 from April 2023) and the Cloud Security Alliance's MAESTRO framework (stale since February) — no significant updates on either.

So What? If you're scoping how an internal agent authenticates to your source of truth or CI pipeline, this draft is the closest thing to an emerging reference architecture — worth reading before you invent your own pattern from scratch, especially in light of Meta's incident-rate spike above.

SourcesIETF Datatracker, OAuth WG Mail Archive, Stacklok

• September 29: FERC's decision on PJM's Reliability Backstop Procurement — the point where "who pays for the power" actually gets decided.
• Whether Meta's MTIA line ever appears on an NVLink Fusion partner list — the real signal for how much independence hyperscalers actually have from NVIDIA's interconnect.
• Early September: netlab 26.08 ships with the Podman fixes plus SONiC, ArcOS, and VPP/FD.io support.
• draft-klrc-aiagent-auth's progress on the OAuth WG mailing list — a solid Saturday deep-dive candidate if it reaches working-group adoption.

• Domains researched: 5 (network architecture & datacenter, network automation, AI/ML, security, science)
• Web searches: ~24 across all domains
• Items published: 14 (3 Top Highlights + 11 domain items)
• Dedup: 0 topic rejections; 1 follow-up item allowed through with new facts not previously reported (netlab Podman fixes, distinct from the 08-21 containerlab coverage)
• Quality score average: 4/5