Amaze Networks Morning Briefing

1. Meta Drops Its First Closed Model — And That's the Real Story

Meta released Muse Spark, its first new AI model in nearly a year and its first model that is not open weights. This is a strategic inflection point. Muse Spark scores 52 on the Artificial Analysis Intelligence Index — placing it behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 — while using over an order of magnitude less compute than Llama 4 Maverick. But the architecture story is secondary to the strategy story: Meta's hybrid open/closed posture, signaled when they launched Llama 4, is now reality. Muse Spark is hosted-only, API preview only, and requires a Facebook or Instagram login to try on meta.ai. The Register's headline said it best: "Meta's new model is as open as Zuckerberg's private school."

So what? Meta built its AI credibility on open weights. Muse Spark is their first signal that the highest-capability models will be proprietary. If you're building infrastructure that depends on local inference of Meta models, this is a data point worth tracking. The Llama family will likely remain open for the mid-tier; the frontier tier is closing.

Actionable takeaway: If you're in enterprise AI planning discussions, add "Meta's frontier models are going closed" to your vendor dependency map. Llama 4 Maverick and Scout remain open — plan around those. Muse Spark will mature before it opens (if it ever does).

2. The Agentic AI Inference Problem — Intel and SambaNova Bet on Heterogeneous Compute

As agentic AI workloads scale, the GPU-only inference stack is showing its limits. Intel and SambaNova announced a split inference architecture targeting H2 2026 that divides workloads across three processor types:

• GPUs handle the prefill phase (parallel matrix ops on the input prompt)
• SambaNova RDUs (Reconfigurable Dataflow Units) handle decode (token generation, latency-sensitive)
• Intel Xeon 6 CPUs handle agent orchestration — tool calls, API lookups, multi-step decision making

The key insight: GPUs are optimized for massive parallelism at prefill, but they're mismatched for the decode phase of agentic workloads, especially when agents are waiting on tool calls between tokens. The CPU layer for orchestration acknowledges that coordination-heavy agentic tasks don't need GPU parallelism at all.

So what? The inference economics shift is real. Inference now represents over 55% of AI compute spend, and agentic workloads push that further because they're continuous, interactive, and stateful. This architecture directly addresses idle GPU time — the silent cost multiplier in every agentic deployment. Available H2 2026, targeting enterprises, cloud providers, and sovereign AI deployments.

Actionable takeaway: When evaluating inference infrastructure for agentic AI, start asking vendors how they handle the prefill/decode split and what their CPU utilization story is. Monolithic GPU deployments for agentic workloads are increasingly hard to justify on cost.

3. Maine Becomes First State to Pass a Datacenter Moratorium

The Maine House passed LD 307 (82-62), and the Senate cleared it 19-13. Governor Mills has signaled support. The bill pauses new data center construction for facilities drawing 20 megawatts or more through November 1, 2027. Maine would be the first US state to enact a statewide pause on large-scale datacenter construction.

The driver: grid stability. A 20 MW datacenter can power more than 15,000 homes. Maine's grid infrastructure cannot absorb rapid datacenter expansion without significant upgrades, and the legislature moved before the infrastructure could catch up. This follows a broader trend: at least 12 states are reviewing similar measures, and localities in New Orleans, Chandler, and Bangor have already moved. On the federal side, Sanders and Ocasio-Cortez introduced the AI Data Center Moratorium Act in March.

So what? Site selection just got harder. Any infrastructure planning team looking at the US Northeast or other constrained grid regions needs to treat "legislative moratorium risk" as a first-class variable alongside power, land, and fiber availability. The industry's growth rate is now politically visible in a way it wasn't two years ago.

Actionable takeaway: For organizations evaluating new datacenter builds or co-lo expansions, add a regulatory risk layer to site scoring. States with acute grid constraints and active legislatures are real near-term blockers.

Kelsey Hightower's "Zero-Token Architecture" — The Best Automation Take of the Week

Speaking at Nutanix .NEXT, former Google distinguished engineer Kelsey Hightower delivered what might be the most useful framing for network automation practitioners this year. His argument: your existing Bash scripts, Ansible playbooks, and cron jobs already do what agentic AI is being sold to do — and they do it for zero tokens. His satirical suggestion: rename /etc/cron.d to /etc/agent.d and call it "zero-token architecture."

The bit lands because it's pointing at something real. Enterprises are being sold AI agents for password resets and config changes that burn millions of tokens per month, when proven automation tooling already handles those workflows reliably and cheaply. Hightower's deeper point (beyond the joke): IT professionals who understand the underlying architecture — who know when to use AI and when to use a shell script — will stay relevant. The engineers who paper over fundamentals with AI wrappers won't.

So what? For network automation practitioners, this is validation. Your Nornir playbooks, your GitOps pipelines, your source-of-truth workflows — these are not legacy. They're the zero-token architecture. The question to ask on any new AI-assisted automation pitch: "What does this do that Ansible with an API call doesn't?"

Actionable takeaway: Before adding an AI agent to any automation workflow, document what the non-AI equivalent would be, how reliable it is, and what the token cost of the AI alternative is per 1,000 runs. That math will clarify a lot of decisions.

Cisco Agentic Workflows — Meraki Gets a Low-Code Automation Layer

Cisco launched Agentic Workflows for Meraki in January, and it's worth a closer look now that deployments are being reported. The platform integrates directly into the Meraki dashboard and offers:

• Pre-built workflows from Cisco engineers via Workflows Exchange
• Visual drag-and-drop editor combining API calls, logic operations, and LLM prompts
• Multi-domain integration across Meraki, Catalyst Center, Catalyst SD-WAN, and ISE
• Natural language triggering via the Cisco AI Assistant — ask for an outcome, get a workflow recommendation

This is meaningful for organizations in the Meraki ecosystem. The AI Assistant integration means network teams can describe desired outcomes ("find all switches with stale configs and flag them") and get workflow recommendations rather than writing automation from scratch.

So what? This is Cisco's answer to the "automation requires Python skills" barrier. The LLM layer translates intent to action through a visual pipeline. It's not replacing Nornir for complex multi-vendor automation, but for Meraki-centric shops, it lowers the barrier significantly.

Actionable takeaway: If you manage a Meraki environment, the Workflows Exchange is worth an hour of your time. The pre-built Cisco-authored workflows alone cover common Day 2 operational tasks without requiring any code.

UK Goes Nuclear for AI — £370M Into SMRs and Fusion Startups

The UK's Advanced Nuclear Framework is pulling private capital into small modular reactors and fusion startups at an accelerating rate. £370 million has gone into the sector, with 2024 seeing a £170M surge. The activity is concentrated in "Nuclear Valley" around Oxford and Abingdon. Notable projects:

• X-Energy and Centrica: 12 advanced modular reactors planned for Hartlepool
• Holtec + EDF + Tritax: SMRs at the former Cottam coal-fired power station in Nottinghamshire
• £60M Sunrise supercomputer at the UK Atomic Energy Authority's Culham campus, using AI to accelerate fusion research

Context: US projects secured 12 GW of datacenter-targeted power capacity for 2026, but actual construction amounts to roughly 4 GW. The gap between planned and buildable capacity is the defining infrastructure constraint of the next three years.

So what? The nuclear-for-datacenters story isn't just a UK story — it's a global response to the same constraint. Where power lead times run 18-36 months and moratoriums are spreading, the only path to new large-scale compute capacity is either behind-the-meter generation or pre-negotiated power agreements at constrained sites.

Two-Phase Liquid Cooling — Beyond the 30 kW/Rack Ceiling

DataCenter Dynamics published analysis on two-phase liquid cooling as chip TDPs push beyond known limits. The argument: conventional single-phase direct-to-chip cooling, which has been the baseline response to Blackwell's 1,000W/chip demand, struggles above 100 kW/rack densities because the coolant temperature differential narrows.

Two-phase systems (where coolant boils and vaporizes, moving heat as latent energy rather than sensible heat) can handle higher heat flux at the same pipe diameter. The tradeoff is system complexity and fluid selection. For rack densities targeting 130+ kW — where next-generation GPU configurations are heading — two-phase is emerging as the only viable baseline.

So what? If you're speccing AI infrastructure for 2027 deployment, two-phase liquid cooling is no longer an exotic option — it's the trajectory. Single-phase direct-to-chip buys you through the current Blackwell generation; the generation after likely requires the upgrade.

Deloitte: Inference Economics Are the New Capex Story

Deloitte published analysis framing inference optimization as the primary AI infrastructure design challenge of 2026. Key numbers:

• Inference now consumes over 55% of AI-optimized infrastructure spending
• Projections put this at 70-80% of total AI compute costs by year-end
• Continuous agentic inference — agents that stay running, polling, reasoning — is the primary cost driver
• Memory extension technology (preserving context instead of re-prefilling) is emerging as the highest-ROI optimization

The report frames the enterprise challenge as "eliminating idle compute that inflates cost per token" — GPUs sitting idle between agent steps are the silent waste in agentic deployments.

So what? This directly connects to the Intel/SambaNova story above. The industry is arriving at the same conclusion from multiple directions: GPU-only inference for agentic workloads is economically unsustainable at scale. The architecture response is specialization — prefill/decode/orchestration as separate compute tiers.

• Cloudflare symbolic execution for BPF malware — Cloudflare published research using the Z3 theorem prover to reverse-engineer malicious BPF socket programs, automatically generating "magic packets" to trigger dormant malware. Turns hours of manual assembly analysis into seconds. Strong architectural-security read.

• Cryptographers' $5K quantum bet — Filippo Valsorda and Matthew Green have a public wager: Valsorda bets that ML-KEM-768 (a post-quantum algorithm) breaks before X25519 does. Green is essentially betting that quantum computers won't matter before classical cryptanalysis finds a flaw in the new standard. Matthew Green says he'd "bet huge amounts against a relevant quantum computer by 2029 or even 2035." The disagreement among experts is itself the story — smart people are miles apart on the timeline.

• ALTK-Evolve from IBM Research — Hugging Face published IBM's on-the-job learning framework for AI agents, enabling agents to learn from task feedback without retraining the base model. Early-stage research, but relevant to long-running network automation agents.

• Meta Muse Spark context — Meta's blog describes Muse Spark's architecture as a "natively multimodal reasoning model with tool-use, visual chain of thought, and multi-agent orchestration." It achieves this using over 10x less compute than Llama 4 Maverick. Efficiency at capability is the design goal — which matters for inference cost at scale.

• Maine Governor Mills signing decision on LD 307 — could come this week, making Maine the first US state with a statewide datacenter pause.
• Intel/SambaNova split inference architecture formal announcement — architecture details expected with H2 2026 product preview.
• Meta Muse Spark API access expansion — currently private preview; broad availability signals how serious Meta is about the proprietary model tier.

Amaze Networks Morning Briefing | Published under Beeston Labs | beestonlabs.dev