Marvell Says Connectivity Is the Next AI Bottleneck — and Jensen Huang Agreed

1. Marvell at Computex — Connectivity Is the Next AI Bottleneck

TL;DR: Marvell CEO Matt Murphy opened Computex day two with a keynote titled "The Future of AI Scaling Depends on Connectivity," with NVIDIA's Jensen Huang walking on stage to co-sign the thesis. Marvell's core argument: compute gains have outpaced networking gains for two consecutive generations, and the gap is now wide enough to reshape where investment goes.

Key Points:

• Marvell has invested roughly $36 billion in its data infrastructure platform over the past decade, including the acquisitions of Celestial AI (photonics) and XConn (scale-up switching)
• The connectivity bottleneck claim is not marketing — Marvell's internal customer data shows AI training runs increasingly stalling on cross-rack and cross-campus bandwidth, not GPU compute
• Marvell argues that the next phase of AI scaling — distributed training across multiple datacenters and eventually geographically separated facilities — is fundamentally a networking problem
• Jensen Huang's on-stage appearance with Murphy represents NVIDIA publicly validating the connectivity-first framing, which matters for how hyperscalers allocate their next buildout cycle

Why This Matters

Marvell has been making the connectivity-first argument quietly for two years. Having Jensen Huang walk on stage at Computex is the industry's clearest signal yet that this framing has won. When the person who sells the GPUs tells you the GPUs aren't the bottleneck anymore, that's the inflection point.

For network architects, this is the clearest invitation yet to position network design — not just GPU procurement — at the center of AI infrastructure conversations. Optical interconnects, scale-up switching fabric, and cross-datacenter bandwidth planning are no longer supporting roles. They're the critical path.

So What?

If you're involved in AI infrastructure planning, request a breakdown of your current training run bottlenecks by stage — compute, memory bandwidth, cross-node fabric, and storage I/O. The probability is high that cross-node or cross-rack bandwidth shows up as a constraint you haven't been measuring. Marvell's photonics play (Celestial AI acquisition) and XConn scale-up switching are bets worth tracking on a two-year horizon.

SourcesMarvell Newsroom, ServeTheHome, HPCwire

---

2. Itential FlowAI Goes GA at Cisco Live — Governed Agents for Production Infrastructure

TL;DR: Itential announced the general availability of FlowAI at Cisco Live US 2026, making generally available a platform for building, deploying, and running governed AI agents against production network and infrastructure systems. GA begins July 1, 2026, with early access now open for enterprise customers. This is the first infrastructure-automation platform to reach GA with governance baked into the agent architecture itself, not bolted on afterward.

Key Points:

• FlowAgents are task-oriented reasoning agents that preserve full reasoning traces for audit — every decision the agent makes is logged, inspectable, and replayable
• FlowAgent Builder enables role-based agent construction with defined purpose, toolset scope, and policy boundaries — an agent built for firewall policy automation cannot reach into BGP configuration
• Six enterprise customers validated FlowAI against production infrastructure across five use cases: incident triage, pre-flight change validation, fault remediation, firewall policy automation, and compliance evidence collection
• The platform builds on Itential's existing workflow automation engine, so organizations already running Itential have a migration path rather than a rip-and-replace
• Governance model: agents operate within scoped toolsets, reasoning traces are preserved for audit, and human-in-the-loop approval gates are configurable per action type

Why This Matters

The governance-first architecture here is the distinguishing technical choice. Most agentic automation frameworks give you capable agents and then ask you to figure out policy enforcement. FlowAI inverts that — the scope boundaries and audit requirements are the foundation, and the agent capability builds on top. For regulated industries and operators who need change-control compliance, that inversion matters enormously.

This also arrives the same week as AutoCon 5 (Munich, June 8-12), the world's largest dedicated network automation conference. The convergence of governed agentic infrastructure tools with the practitioner community's biggest annual gathering suggests this week is when "agentic NetOps" moves from conference topic to actual procurement conversation.

So What?

If you're evaluating AI-assisted network operations tooling, ask vendors specifically about governance architecture: how does the agent's scope get bounded, what's the audit trail for every action taken, and can approval requirements be configured per action type rather than per session? FlowAI's GA sets a reference architecture for those requirements. Request early access now — the July 1 GA timeline means you can run a pilot before Q3 planning.

SourcesItential, VMblog, Itential Blog

---

3. arXiv — "Move the Query, Not the Cache" Changes the AI Fabric Routing Question

TL;DR: A paper published today on arXiv argues that cross-instance KV cache systems have been solving the wrong problem: they move KV cache blocks to wherever the query is, when they should route the query to wherever the cache lives. Multi-head Latent Attention compresses key-value state to roughly 1 KB per query row — smaller than the cache chunks the query attends to — making query routing orders of magnitude cheaper than cache migration.

Key Points:

• Frontier LLMs increasingly use sparse-attention indexers that select a few KV-cache chunks per query; in agentic workloads, many sub-agents query the same large codebase simultaneously, creating massive KV reuse
• When the corpus exceeds one GPU's memory, the conventional reflex is to pull cache blocks to the requester — the paper shows this is backward for Multi-head Latent Attention (MLA) architectures
• In MLA (used in DeepSeek and derivative models), the key-value state compresses to a narrow vector: routing the query row (~1 KB) is cheaper than routing the cache chunk it attends to
• Implication: GPU fabric design for agentic workloads should optimize for fast query-routing across instances rather than high-bandwidth cache migration — a different traffic pattern with different switch requirements

Why This Matters

This paper shifts the conversation about what "right-sized" GPU fabric means for agentic AI workloads. Current fabric benchmarks (bandwidth-first, optimize for all-to-all collective traffic) were built for training workloads. Agentic inference with MLA architectures has a structurally different traffic pattern — many small query-routing operations rather than large tensor broadcasts. A fabric designed for one optimizes poorly for the other.

For anyone specifying or operating AI inference fabric today, this is the research that changes which traffic metrics you should be instrumenting.

So What?

When talking to network vendors about AI inference fabric, ask what traffic profile their switching optimization is designed for. The right answer is increasingly "we support both training all-to-all collectives and agentic small-packet query-routing" — vendors who only talk bandwidth are describing a training-era fabric. Instrument your current inference fabric to capture query-routing latency across GPU instances, not just bulk throughput.

SourcesarXiv 2606.01502, Packet Pushers NB577

---

Forward Networks Launches "Predict" — Deterministic Pre-Deployment Change Validation

TL;DR: Forward Networks announced Forward Predict, shipping fall 2026, which runs proposed network changes against a mathematically accurate digital twin before deployment. AI agents can propose a change, receive specific failure feedback from the twin, iterate until the change passes all validation gates, and only then submit for human approval.

Key Points:

• The validation checks connectivity, security policy, and compliance impacts simultaneously — not sequentially
• When a proposed change creates an unintended consequence, the agent receives structured failure feedback rather than a binary pass/fail
• Forward is explicitly positioning Predict as the pre-execution verification layer for agentic networking, solving the "AI agent makes change without knowing the downstream blast radius" problem
• Fall 2026 availability puts this in planning cycles for 2027 budget

So What?

Forward Predict is the cleanest answer currently available to the "how do I let an AI agent touch production without a human reviewing every change" question. Book a demo before AutoCon 5 (June 8-12) — the Munich conference will almost certainly surface case studies from early access customers.

SourcesConverge Digest, Help Net Security

---

NetBox Labs — Infrastructure Intelligence Starts with Your Data

TL;DR: NetBox Labs published a pointed editorial this week: the reason most AI infrastructure tooling fails to deliver is not the model or the agent framework — it's that enterprises don't have clean, structured data about their own networks.

Key Points:

• The piece argues that enterprises cannot automate what they don't understand, and most still can't answer basic questions (what's running, where, and what's connected to what) from a single authoritative source
• NetBox's framing positions the source-of-truth investment as the prerequisite for any agentic automation stack — without it, every agent is working from incomplete or stale data
• This connects directly to the FlowAI story above: Itential's governance model depends on agents knowing the actual state of the network they're operating on

So What?

If your NetBox or Nautobot instance has data quality problems (stale device entries, missing IP allocations, inconsistent site data), those problems become agent problems the moment you connect any AI tooling. Run a data quality audit before evaluating agentic infrastructure platforms.

SourcesNetBox Labs Blog

---

Tony Mattke's Networking CLI Tools — Worth a Look

TL;DR: ipSpace.net flagged a set of networking-focused CLI tools built and open-sourced by Tony Mattke, targeting gaps in existing tooling for network engineers doing day-to-day operations and automation work.

Key Points:

• The tools are available on GitHub and fill specific workflow gaps that existing tools (NAPALM, Scrapli, Nornir) don't address out of the box
• Ivan Pepelnjak's endorsement via ipSpace.net carries practitioner signal — this is not a vendor announcement

So What?

Check Mattke's GitHub for tools relevant to your automation workflows. The signal here is less about any specific tool and more about the practitioner community still solving real gaps in the open-source tooling stack.

SourcesipSpace.net

---

AutoCon 5 Opens in Munich This Week — Practitioner-First, No Pay-to-Play

TL;DR: AutoCon 5, the world's largest dedicated network automation conference, opens in Munich on June 8 with workshops June 8-9 and main conference June 10-12. The first European edition, built entirely around practitioner experience with no vendor-sponsored content tracks.

Key Points:

• Sessions include case studies from Deutsche Bahn and PlayStation, plus a dedicated NAF Framework track on Thursday morning
• Itential, NetBox Labs, and Nokia are all presenting — and notably, all three have active product announcements this week, suggesting intentional timing
• The NAF (Network Automation Framework) track represents the first time the framework has a dedicated conference slot — signals maturity in the automation methodology space

So What?

If you're in range of Munich, the June 8-9 workshops are the highest-signal days — that's where practitioners build things, not watch slides. If remote, the NAF Framework track sessions from June 11 morning are worth flagging for async viewing.

SourcesNetwork Automation Forum, Itential AutoCon 5 Post

---

Vector Core Compute — Disaggregated Inference Neocloud Goes Live

TL;DR: Vista Equity Partners and Cambium Capital launched Vector Core Compute, a purpose-built inference cloud that demonstrated disaggregated inference at Computex using Intel Xeon for orchestration, SambaNova SN40 for decode, and NVIDIA Blackwell for prefill — all operating from a Los Angeles datacenter with Together.ai as the first commercial customer.

Key Points:

• This is the first production demonstration of fully disaggregated inference (prefill and decode on different hardware from different vendors) at commercial scale
• SambaNova unveiled the SN50 chip, claiming five times the throughput of competing chips for decode-phase inference
• Vista Equity's 90+ portfolio companies get early access, giving the platform an immediate enterprise customer base across 750 million end users
• Intel and SambaNova also announced a planned collaboration with Foxconn for rack-scale AI infrastructure targeting data center and hyperscale deployments

So What?

Disaggregated inference at production scale changes the purchasing model: you no longer need to buy the same vendor's silicon for every stage of an inference pipeline. Get SambaNova on your vendor evaluation list if you're doing inference at scale — the decode-specialization argument is worth benchmarking against your actual workload token ratio.

SourcesDataCenter Dynamics, Intel Capital, BusinessWire

---

NVIDIA JetPack 7.2 — Agentic AI at the Physical Edge

TL;DR: NVIDIA released JetPack 7.2 at Computex, adding NemoClaw support (OpenClaw with privacy and security controls), Jetson-native agent skills, and Multi-Instance GPU support on Jetson Thor for deterministic multiworkload execution — bringing governed agentic AI to edge hardware.

Key Points:

• NemoClaw adds privacy and security controls on top of OpenClaw's agent skill architecture, deployable with one command on Jetson Orin
• MIG (Multi-Instance GPU) support on Jetson Thor means multiple independent workloads can share one edge accelerator without interfering — critical for real-time physical environments
• Yocto Project support enables custom Linux distributions optimized for specific edge deployments

So What?

If you have infrastructure monitoring or network telemetry running on edge compute, JetPack 7.2 is the first version where deploying a governed agent alongside operational workloads is practical rather than experimental. Test NemoClaw on a development Jetson Orin before the next hardware refresh cycle.

SourcesNVIDIA Technical Blog

---

Cloudflare Boot Time Fix — Engineering Transparency Worth Noting

TL;DR: Cloudflare published a deep technical post on how a firmware update caused their Gen12 fleet of nearly 2,000 core servers to take four hours to boot rather than minutes — and how they tracked it down to a UEFI handoff quirk that cascaded across maintenance windows.

Key Points:

• A routine firmware update introduced a UEFI initialization timeout that compounded with other hardware initialization steps
• The debugging path required isolating the boot sequence stage by stage across a nearly 2,000-unit fleet
• Resolution required coordinating with firmware vendors on a targeted fix rather than rolling back the entire update

So What?

Read this post with your infrastructure team before your next fleet-wide firmware rollout. The failure mode — a component interaction that only manifests at scale, in a maintenance window, under time pressure — is the failure mode every fleet operator should have planned for.

SourcesCloudflare Blog

---

SoftBank Bets €75 Billion on France as an AI Power Strategy

TL;DR: SoftBank announced up to €75 billion ($85 billion) in French AI infrastructure investment, targeting 5 GW of datacenter capacity by 2031. The investment thesis is explicit: France's nuclear-heavy, decarbonized electricity grid is the strategic differentiator, not real estate or labor costs.

Key Points:

• Three initial sites in Dunkirk, Bosquel, and Bouchain, targeting 3.1 GW by 2031 — the first phase alone is €45 billion
• Partnership with EDF at the former Bouchain power plant site — an idle grid-connected facility with existing transmission infrastructure becomes a datacenter campus
• Schneider Electric manufacturing cluster in Dunkirk will produce datacenter enclosures and integrate Schneider's power modules on-site
• CEO Masayoshi Son explicitly named France's electricity profile as the "decisive" factor — not French tech talent or regulatory environment

Why This Matters

This is the clearest signal yet that power availability has become the primary site selection variable for hyperscale AI infrastructure — ahead of connectivity, ahead of land costs, ahead of permitting timelines. Repurposing former power plant sites (Bouchain) is the emerging pattern: the transmission infrastructure that existed for generation also works for massive datacenter draw.

For datacenter planners and network architects, this signals that the next generation of large facilities will increasingly cluster around grid assets rather than traditional Internet exchange or fiber hub locations. Network design assumptions built around proximity to existing carrier hotels need revisiting.

So What?

If you're doing datacenter network design for facilities that will be built in the next five years, your connectivity planning should start from the power asset location outward — not from existing fiber routes inward. The EDF partnership model (former plant to datacenter campus) is being replicated elsewhere in Europe and will drive connectivity construction, not follow it.

SourcesDataCenter Knowledge, DataCenter Dynamics, CNBC

---

Ohio Pauses Datacenter Tax Breaks — $1.6 Billion Problem

TL;DR: Ohio Governor Mike DeWine paused the state's datacenter sales tax exemption program after the cost ballooned from the projected $142 million in fiscal 2025 to nearly $1.6 billion — an eleven-fold overshoot. The Tax Credit Authority will accept no new applications after its June 1 meeting.

Key Points:

• The original fiscal 2024 projection estimated $136 million cost in fiscal 2025; actual cost was $554 million in 2024 and $1.6 billion in 2025
• This is not a ban — existing exemptions remain; new applications are suspended pending a legislative review
• Ohio was a top-five US state for datacenter investment; the pause creates immediate uncertainty for projects in pipeline
• Virginia and Georgia are watching — both states have active incentive programs facing similar cost-overshoot pressure

So What?

If you have datacenter projects in Ohio pipeline that were counting on the tax exemption, verify their status before the June 1 cutoff date. More broadly, the regulatory backstroke on datacenter incentives is now happening in multiple states — factor this into any three-to-five year facility planning that assumed stable tax treatment.

SourcesThe Register, Ohio Governor's Office

---

Harvard 448-Qubit Milestone — Fault Tolerance Below the Error Threshold

TL;DR: Harvard, MIT, and QuEra Computing demonstrated scalable fault-tolerant quantum computing using 448 neutral-atom rubidium qubits, achieving logical error rates below 0.5% — the first time adding more qubits has reduced rather than increased error rates in a large-scale system. Mikhail Lukin now says fault-tolerant quantum computers are likely this decade, five to ten years ahead of previous timelines.

Key Points:

• The system combines all essential elements simultaneously: quantum teleportation, layered error correction, entropy removal, and mid-circuit measurements
• Previous systems showed that error rates scaled with qubit count; this work reverses that relationship — the error-correction architecture works as intended at 448 qubits
• Published in Nature; Lukin group at Harvard
• Timeline acceleration: Lukin previously estimated fault-tolerant systems in the 2030s; this result moves that estimate to "by end of this decade"

Why This Matters

The 0.5% logical error rate figure is the engineering threshold — below it, error correction actually helps rather than introducing more errors than it catches. This is the proof that fault-tolerant scaling is tractable, not just theoretically possible. Combined with IBM's Anderon quantum foundry (covered May 27) and the $2 billion CHIPS Act quantum investment (covered May 26), the hardware manufacturing and error-correction pillars are both advancing simultaneously.

So What?

Post-quantum cryptography (PQC) migration using ML-KEM and ML-DSA should be an active engineering project — not a 2030 problem. The Harvard result compresses the timeline for when cryptographically-relevant quantum computing becomes plausible. If your organization hasn't started PQC migration planning, start now.

SourcesQuantum Computing Report, Harvard Gazette, The Quantum Insider

---

Meta AI Support Bot — Agentic Permission Boundaries Fail in Production

TL;DR: Krebs on Security and Simon Willison documented a case where hackers instructed Meta's AI support bot to link target Instagram accounts to attacker-controlled accounts — and it worked. The accounts of the Obama White House and a senior Space Force official were briefly compromised. The mechanism: the bot had the ability to modify account ownership and no enforcement of requester legitimacy.

Key Points:

• The attack did not require exploiting any code vulnerability — it was purely social engineering against an AI agent with broad account-management permissions
• Simon Willison's framing: this is the same class of problem as the Copilot Cowork file exfiltration from last week — agents with the ability to take external or cross-account actions without per-action authorization
• The fix is architectural, not prompt-level: agents with write permissions to account systems need requester verification at the action boundary, not session-level trust

So What?

Map every AI agent in your organization that has the ability to make changes to systems — network devices, account directories, access control systems, ticketing systems. For each one, verify that authorization is enforced at the action boundary, not assumed from session identity. The Meta case makes this the practical standard, not a theoretical concern.

SourcesKrebs on Security, Simon Willison

---

• Space backbone network: The US Space Force awarded SpaceX a $2.29 billion contract for a satellite-based space backbone network — the same week the FAA grounded SpaceX following a launch mishap, creating an interesting procurement-vs-operations tension to watch.
• GitHub Copilot billing shift: As of June 1, all GitHub Copilot plans now bill on AI Credits (usage-based model), with new user-level budget controls and an upgrade path to Copilot Max. Enterprise teams should audit Copilot usage patterns before the next billing cycle.
• SONiC access layer: A new post on network-notes.com covers SONiC running on 1G commodity access switches, arguing the economics now work for the access layer — a meaningful expansion from the datacenter-spine-only framing that defined SONiC adoption for its first decade.
• Amazon serverless OpenSearch: Amazon re-engineered serverless OpenSearch with separated storage and compute, explicitly targeting agentic AI's bursty query patterns. Relevant if you're using OpenSearch for network telemetry or log analysis with AI tooling.

SourcesPacket Pushers NB577, network-notes.com, The Register

---

• AutoCon 5 opens in Munich, June 8-12. The NAF Framework track on Thursday June 11 morning is the one to prioritize. Watch for session recordings — this is the practitioner community's most candid space.
• Itential FlowAI early access is open now for enterprise customers ahead of the July 1 GA date. If agentic infrastructure automation is on your 2026 roadmap, this is the time to get in the queue.
• Forward Predict is targeting fall 2026 GA, with demos at Cisco Live Las Vegas. If you're evaluating digital twin platforms for pre-deployment validation, add this to your RFI.
• Ohio datacenter tax exemption review: The legislative committee reviewing the program will shape decisions for facilities across the state. If you have Ohio-based infrastructure, track this closely.

---

• Domains researched: 5 (networking/architecture, automation/programmability, AI/ML, datacenter, science)
• RSS digest: 86 articles from 22 feeds; top score 13.7 (arXiv cross-instance KV cache paper)
• Web searches: 14 targeted searches across all domains
• Primary items published: 12
• Quick takes: 4
• Dedup rejections: 6 (Cisco SONiC N9K — 72h cooldown May 27; MRLS networks — May 27; IBM Anderon — May 27; HPE Mist — May 27; SR Linux silent failure — May 27; DCP protocol — May 27)
• Quality score: 4.5/5