← /editions
#77

Nvidia ships the Groq deal: decode becomes a product line

// Nvidia put Groq 3 LPX into full production at Hot Chips, formally splitting decode from prefill and turning the $20B Groq LPU license into rentable silicon. Plus the first critical CVE in a vendor-backed MCP server, A2A joining MCP under Linux Foundation governance, and Microsoft making agents an inventoried, blockable asset.

$ today_in_10_seconds

Nvidia put Groq 3 LPX into full production at Hot Chips — the $20B Groq license from last Christmas is now silicon you can rent, and it targets one thing: decode speed for agents. Meanwhile the MCP ecosystem took its first critical enterprise CVE, and A2A moved under the same governance roof as MCP. Infrastructure day.

$ lead_story

Nvidia Groq 3 LPX enters full production

Nvidia announced at Hot Chips that Groq 3 LPX — an inference accelerator extending the Vera Rubin platform — is now in full production, with Nebius the first AI cloud to adopt it. The headline number: 3,400 output tokens per second running Gemma 4 31B at 100,000-token context, which Nvidia calls the fastest recorded for that model, and 4x faster responsiveness than the nearest alternative platform. The architecture matters more than the benchmark: Rubin GPUs handle large-scale context processing while LPX takes the latency-sensitive decode workload — a formal split of prefill from decode into separate products. That dataflow design is Groq's LPU IP, licensed non-exclusively in a roughly $20B deal signed December 24, 2025 that also brought Jonathan Ross and his team inside Nvidia, and this is the first time it ships as a rack you can buy. For builders, the practical read is unglamorous: agents emit tokens across hundreds or thousands of steps, so decode rate is simultaneously your latency budget and your COGS. But one vendor benchmark, on one open model, at one context length, is a marketing artifact — not a capacity plan. Your agent loop is probably bottlenecked on tool calls and retries, not sampling.

// why_it_matters: Decode is now a separate product line — which makes agent latency an infrastructure purchasing decision, not just a model choice.

read more →

$ also_worth_attention

  • A2A joins the Agentic AI Foundation

    Google's Agent2Agent protocol formally moved under the Linux Foundation-directed Agentic AI Foundation on August 20, placing it alongside Anthropic's MCP in neutral governance — a body that has grown from under 40 members to more than 250 since launching in December 2025. The interop story is the obvious one; the interesting one is identity. A2A v1.0 shipped in March 2026 with multi-tenancy and signed agent cards for cryptographic identity verification, meaning cross-organisational agent authentication is now being settled in a standards body rather than by whoever ships first.

    link →
  • Splunk patches a critical RCE in its MCP Server app

    CVE-2026-76404, rated CVSS 9.1, let an admin-role user execute arbitrary OS commands via unsafe deserialization in the app's credential-management component; it's fixed in MCP Server 1.2.1 and shipped alongside 16 other flaws. Ten of the seventeen were in the Splunk AI Toolkit, including another deserialization bug in its model-loading REST API. Note where the bug lived: credential management. MCP servers are privileged secret brokers wearing a plugin costume — inventory them as tier-0 infrastructure, not as integrations.

    link →
  • Microsoft makes agents an inventoried, revocable asset

    Multi-tenant agent management hit public preview in the Microsoft 365 admin center, giving admins a consolidated agent inventory across connected tenants, the ability to install or block agents per-tenant, and tenant-specific permission review — with agent risk and activity gated behind an Agent 365 license. The verb that matters is block. Shadow agents are following the exact arc shadow SaaS did: someone useful turns one on, nobody has a list, then procurement shows up with a registry.

    link →

$ bigger_picture

The trend behind the headlines

Today's four stories are one story told at four layers. Nvidia is industrialising the token economics, the Linux Foundation is industrialising the wire protocols and agent identity, Microsoft is industrialising revocation — and the Splunk advisory is the invoice for building that surface area at speed. None of this is about capability; no model got smarter this week. It's about the boring machinery that decides whether an agent deployment survives its first audit, its first incident, and its first cloud bill. The teams that win the next twelve months will be the ones who treated decode cost, agent credentials, and a kill switch as day-one requirements rather than day-400 retrofits.

$ tool_of_the_day

Artificial Analysis

The independent benchmarking outfit Nvidia cited for its 3,400 tok/s record. Worth bookmarking precisely because you can use it against the press release: it tracks output speed, time-to-first-token, and price per million tokens across providers and models, so you can check whether a '4x faster' claim survives contact with your model, your context length, and your budget. Verify vendor throughput claims yourself, in the shape of your actual workload.

explore Artificial Analysis →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime