## Sources

1. [Anthropic Blackout Forces Europe to Confront AI Reliance](https://awesomeagents.ai/news/anthropic-blackout-eu-ai-sovereignty/)
2. [Microsoft Drops Claude Code as AI Budgets Run Out](https://awesomeagents.ai/news/microsoft-drops-claude-code-ai-costs/)
3. [AgentPerf - First Infrastructure Benchmark for Agents](https://awesomeagents.ai/news/agentperf-first-agentic-ai-benchmark/)

---

### **AgentPerf - First Infrastructure Benchmark for Agents** by **Sophie Zhang**

*   **Launch of a Specialized Benchmark:** On June 12, Artificial Analysis introduced **AA-AgentPerf**, the first benchmark specifically designed to evaluate AI infrastructure under **agentic workloads**, which differ significantly from standard chatbot requests [1, 2].
*   **The Primary Metric:** Unlike traditional benchmarks that measure tokens per second, AgentPerf focuses on **"Agents per Megawatt,"** a density metric measuring the maximum number of concurrent agents a system can support while meeting production service-level targets [3, 4].
*   **Performance Gap:** In the first published results, **NVIDIA’s Blackwell platform (GB300 NVL72)** demonstrated a **20x gain over the previous Hopper generation (H200)**, scoring 61,400 agents per megawatt compared to the H200's 2,600 [2, 3, 5].
*   **Real-World Workloads:** The benchmark utilizes **real coding-agent trajectories** from public repositories across 12+ languages, rather than synthetic data [2, 6]. These sessions can extend up to **200 turns** and reach sequences of **100,000 tokens** [7].
*   **Production-Grade Testing:** To reflect actual usage, the benchmark allows for optimization strategies like **KV cache reuse and speculative decoding** and includes a **three-tier Service Level Objective (SLO) framework** based on output speed and time-to-first-token [8].
*   **Current Limitations:** The initial test results only cover **NVIDIA hardware** and a single model, **DeepSeek V4 Pro** [2, 9, 10]. Data for other hardware vendors like AMD, Google, and AWS are not yet represented [9].

### **Anthropic Blackout Forces Europe to Confront AI Reliance** by **Daniel Okafor**

*   **Abrupt Export Ban:** On June 12, the US Commerce Department ordered Anthropic to block foreign nationals from its most advanced models, **Fable 5 and Mythos 5**, citing national security concerns [11, 12].
*   **Immediate Disruption:** Anthropic chose **global suspension** over geo-blocking to ensure compliance, leaving European businesses, research labs, and hospitals without access mid-session and with **no prior warning** [11, 13, 14].
*   **Sovereignty Crisis:** The EU Commission labeled the move **"discriminatory against partners,"** and European leaders are now framing AI as a critical infrastructure that must not be controlled by foreign governments [15-17].
*   **Shift to Local Alternatives:** Politicians in France and the Netherlands are calling for accelerated support for domestic providers like **Mistral and OVHcloud** to reduce reliance on US technology [16, 18].
*   **Geopolitical Tension at G7:** Anthropic CEO Dario Amodei is joining Sam Altman and Demis Hassabis at the **G7 summit in France**, where European leaders intend to raise the shutdown as a specific case study of the risks of AI dependency [15, 19, 20].
*   **Potential for Resolution:** Anthropic technical staff have met with White House officials to challenge the security characterization and explore a licensing process or model narrowing [13, 19].

### **Microsoft Drops Claude Code as AI Budgets Run Out** by **Daniel Okafor**

*   **Microsoft License Cut:** Microsoft’s Experiences and Devices division is removing **Claude Code** access for thousands of engineers by June 30 [21, 22].
*   **The Problem with Token Costs:** While standard subscription seats are predictable, **agentic coding loops**—which autonomously read files, run tests, and spawn sub-agents—consume massive amounts of tokens [23, 24].
*   **Economic Mismatch:** Claude Code's billing model of **$20/seat plus unlimited token usage** resulted in monthly costs per engineer ranging from $150 to $2,000 [23].
*   **Budget Depletion:** Uber reported that its 5,000-engineer team **exhausted its entire 2026 AI budget in just four months**, with power users generating **$1,200 in costs during single two-hour sessions** [22, 25].
*   **Strategic Transition:** Microsoft is migrating its engineers to **GitHub Copilot CLI**, which offers a **flat rate of $39 per seat**, providing more predictable budget control [22, 23, 25].
*   **Continued Partnership:** This move is a **billing restructure** rather than a total abandonment of Anthropic; Claude models still power features within the Copilot CLI under existing agreements [22, 26].
*   **Revenue Growth vs. Cost Constraints:** Despite these corporate cuts, Anthropic’s annual recurring revenue (ARR) hit **$30 billion in April 2026**, suggesting that many other enterprises are still absorbing the high token costs [22, 27].