Editor's note — Today's brief follows one thread: the agentic layer is consolidating and, for the first time, being governed. Regulation and this week's model releases sit in the quick hits, where a thin, aggregator-heavy model news cycle kept them. No pure regulation story led today.
Google donates A2A, and the agent economy's plumbing lands under one roof
Google has handed its Agent2Agent (A2A) protocol to the Agentic AI Foundation, the Linux Foundation body that already houses Anthropic's Model Context Protocol. The move, announced on 17 August and formalised on 20 August, puts the two standards that matter most to agent builders under the same neutral governance: MCP standardises how an agent reaches tools and data, A2A standardises how agents discover and talk to one another. For a market that spent a year arguing about whose framework would win, the answer is turning out to be nobody's — the protocols are becoming shared infrastructure rather than competitive moats.
The foundation has grown from fewer than 40 members at its December 2025 launch to more than 250, with AWS, Anthropic, Google, Microsoft, Bloomberg, Block and OpenAI among the backers. That breadth is the point. When the connective tissue of the agent stack is vendor-neutral, buyers can wire agents across clouds and labs without betting on a single supplier's roadmap — the same interoperability logic that makes a sovereign, provider-agnostic inference layer worth building rather than a walled garden.
IBM commits $240M to a dedicated inference cluster with Together AI
IBM has signed a multi-year, $240 million agreement with Together AI to stand up a large NVIDIA HGX B300 cluster on IBM Cloud, wired together with Spectrum-X Ethernet and aimed squarely at inference rather than training. Availability is slated for the first quarter of 2027, and Together — which says it serves 400 trillion tokens a month — will use the capacity to run open-weight models for enterprises chasing frontier performance at lower per-token cost.
The deal is a clean signal of where the infrastructure spend is rotating. Training clusters made the headlines for two years; the money now follows the tokens, and the tokens are inference. A named, single-tenant cluster built to serve open models is also a reminder that “run frontier-class models on dedicated, known hardware” is a product enterprises will pay nine figures for — precisely because most of them cannot see, let alone control, where their inference actually runs today.
The money keeps chasing whoever can govern the agents
Obsidian Security closed an $85 million Series D at a $1.1 billion valuation, crossing into unicorn territory on the strength of one pitch: enterprises are deploying agents faster than they can control them. Led by Crescent Cove with existing investors including Greylock and Menlo Ventures, the round pushes Obsidian past $200 million raised. The company says it counts 60 of the Fortune 500 as customers and is extending runtime governance to Anthropic's Claude Code and Cowork — letting security teams cap an agent's permissions around production data and block unsanctioned MCP or tool use.
That last detail closes the loop with the day's other stories. The same protocols now being handed to a neutral foundation are the ones security teams need to police at runtime, and the inference capacity being built out is what those agents will run on. Governance, not raw capability, is becoming the thing buyers are short of — which is the market a trust-first platform is built to serve.
Quick hits
- EU AI Act enforcement is now live. Since 2 August, the Commission's enforcement powers over general-purpose AI model providers have applied in full: the AI Office can request documentation, evaluate models directly, order corrective measures, and levy fines of up to €15M or 3% of global turnover under Article 101. Models placed on the market before 2 August 2025 have until 2 August 2027 to comply.
- Skan AI raises $63M. The Series C, co-led by Cathay Innovation and Dell Technologies Capital, funds a platform that grounds enterprise AI in “how work actually gets done,” shipping alongside the launch of Skan AI Blueprint and Skan AI Agents.
- Together AI's inference footprint keeps compounding. Beyond the IBM deal, the company says it is serving open-weight inference at 400 trillion tokens a month — a data point worth watching as a proxy for how fast open models are moving into production workloads.
- Model news stayed noisy but thin. Late August saw a heavy cadence of releases and benchmark claims across the frontier and open-weight fields, but much of it circulated through aggregator trackers rather than primary announcements. We're holding those items until they're confirmed at the source.