## Sources

1. [SoftBank Commits €75B to French AI Data Centers](https://awesomeagents.ai/news/softbank-75b-france-ai-data-centers/)
2. [GitHub Copilot Goes Token-Based: Devs Report 25x Bills](https://awesomeagents.ai/news/github-copilot-token-billing-cost-shock/)
3. [IBM and Red Hat Bet $5B on AI to Secure Open Source](https://awesomeagents.ai/news/ibm-red-hat-project-lightwell-5b-open-source/)

---

### **GitHub Copilot Goes Token-Based: Devs Report 25x Bills | Elena Marchetti**

**Main Arguments**
*   **Transition to Usage-Based Pricing:** Starting June 1, 2026, Microsoft is moving GitHub Copilot from a flat-rate subscription model to a **token-based system** using "GitHub AI Credits" [1-3].
*   **Economic Sustainability:** GitHub argues that the shift is necessary because modern **agentic workflows**—which involve complex, multi-step autonomous tasks—consume significantly more compute than simple code completions [4, 5].
*   **Alignment with Actual Costs:** The new model aligns user pricing directly with the cost of running frontier AI models, such as GPT-5.5 and Claude Opus [4, 6, 7].

**Key Takeaways**
*   **Credit Conversion:** While monthly plan prices remain the same (e.g., $10 for Pro, $39 for Pro+), these fees now act as **spending limits** against API token rates [2, 6]. One credit equals $0.01 [6].
*   **Metered vs. Unmetered Features:** **Code completions and Next Edit Suggestions remain unlimited.** However, Copilot Chat, agentic coding sessions, and the use of premium models are now metered [2, 6].
*   **Drastic Cost Increases for Power Users:** Developers have reported projected monthly costs jumping from **$29 to $750**, and in extreme cases, from $50 to $3,000, particularly for those heavily utilizing agentic features [1, 8].
*   **Market Competition:** The change has sparked significant backlash (893 negative reactions in the official thread), leading many developers to explore alternatives like **Cursor ($20/month flat)** or **Windsurf ($15/month flat)** [2, 8-10].

**Important Details**
*   **Model Tiering:** Under the new system, accessing high-end models like GPT-5.5 ($30 per 1M output tokens) costs significantly more than smaller models like GPT-5 mini ($2 per 1M output tokens) [7].
*   **Agentic Session Costs:** A single agentic coding session with a premium model can consume **$30 to $40**, which instantly exhausts the monthly budget of a Pro or Pro+ subscriber [2, 11].
*   **Plan Flexibility:** Business and Enterprise accounts benefit from **pooled credits** and promotional limits for the first three months, whereas individual Pro users have no way to increase their $10 budget through admin tools [12].
*   **End of Downgrades:** Previously, Copilot would downgrade users to a cheaper model when limits were reached; now, **access stops completely** until the next cycle or until more credits are purchased [5].

***

### **IBM and Red Hat Bet $5B on AI to Secure Open Source | Sophie Zhang**

**Main Arguments**
*   **Securing the Open Source Supply Chain:** IBM and Red Hat have launched **Project Lightwell**, a $5 billion initiative to scan, verify, and patch vulnerabilities in the open-source software that powers enterprise infrastructure [13, 14].
*   **Version-Specific Remediation:** Unlike existing tools that force users to upgrade to the latest version, Project Lightwell focuses on **backporting fixes to exact deployed versions**, minimizing the risk of breaking certified production environments [15, 16].

**Key Takeaways**
*   **Human-AI Hybrid Triage:** The project utilizes frontier AI models alongside a massive team of **20,000 engineers** to triage vulnerabilities across more than 62,000 open-source packages [13-15].
*   **Strong Financial Sector Adoption:** Eleven major financial institutions, including **Goldman Sachs, JPMorganChase, and Mastercard**, are early adopters due to their complex, highly regulated software stacks [14, 17].
*   **Dependency Manifest Integration:** The system operates by processing dependency manifests (like `pom.xml` for Java/Maven); the user's actual **application code never leaves their environment** [18, 19].

**Important Details**
*   **The Scaling Problem:** The initiative is a response to the massive volume of vulnerabilities; Anthropic’s Mythos Preview recently identified **3,900 high-severity vulnerabilities** in a single scan, a volume impossible for human-only teams to manage quickly [14, 20].
*   **Commercial Model:** This is a paid subscription service where IBM acts as a security intermediary between upstream open-source maintainers and enterprise users [14, 19, 20].
*   **Current Limitations:** At launch, the project only supports **Maven/Java**, with other ecosystems like Npm and PyPI planned for later [19, 21].
*   **Vendor Concentration Risk:** Critics point out that centralizing patch delivery through IBM creates a **single point of failure** in the supply chain [22].

***

### **SoftBank Commits €75B to French AI Data Centers | Elena Marchetti**

**Main Arguments**
*   **Massive Infrastructure Expansion:** SoftBank Group is committing up to **€75 billion (~$87B)** to build 5 gigawatts (GW) of AI data center capacity in France [23, 24].
*   **Nuclear Power as a Strategic Advantage:** The decisive factor for choosing France over the UK or Germany was its stable, **nuclear-heavy energy grid**, which provides the reliable baseload power required for gigawatt-scale data centers [24, 25].
*   **Geopolitical Positioning:** By building in France, SoftBank secures first-party compute infrastructure within the EU, helping navigate **data residency rules** and the EU AI Act [26].

**Key Takeaways**
*   **Phased Commitment:** Phase 1 involves a **€45 billion investment** for 3.1 GW of capacity across three sites in the Hauts-de-France region (Dunkirk, Bosquel, and Bouchain) by 2031 [23, 24, 27].
*   **Scale of the Project:** At full buildout, the French footprint would exceed the capacity of the **Northern Virginia data center corridor**, which is currently the densest compute cluster in the world [24, 27].
*   **Personal Diplomacy:** The deal was brokered directly between SoftBank Chairman **Masayoshi Son** and French President **Emmanuel Macron** during a visit to Tokyo in May 2026 [24, 28].

**Important Details**
*   **Strategic Partnerships:** Partners include **EDF** for power connectivity and **Schneider Electric**, which will operate a robotized manufacturing plant at the Dunkirk hub [29, 30].
*   **OpenAI Connection:** SoftBank holds an **11% stake in OpenAI**; these data centers could potentially provide the compute infrastructure OpenAI currently lacks in Europe [26, 31].
*   **Comparison to US Projects:** This follows a similar SoftBank announcement for a **$500 billion, 10 GW data center campus in Ohio**, which will be powered by a new $33 billion natural gas plant [32].
*   **Implementation Challenges:** The project requires provisioning power equivalent to **five large nuclear reactors**, a massive undertaking for the French grid operator (EDF) that may face permitting and construction delays [30, 33].