Three companies at the core of Europe's AI stack changed shape in a single week. NVIDIA's board accepted Jensen Huang's resignation on November 11, three days after opening an independent investigation into reports that he had been "the target of a foreign intelligence effort." On Tuesday its interim chief executive, CFO Colette Kress, reported quarterly revenue of $115.8 billion and said that "Vera Rubin is off to an explosive start." On Wednesday Anthropic priced what Renaissance Capital called the largest IPO ever at $83 a share, above its range, selling 980 million shares for about $81.3 billion, alongside a $26 billion private placement. The shares closed their first day at $97.17, up 17%. Its founders keep 61.1% of the votes. At Ignite, Microsoft made its Azure AI Gateway generally available and filled out a control plane for AI agents.
The European terms did not move with them. Amazon Bedrock lists no EU inference profile for Anthropic's newest model, Claude Fable 5.2; Microsoft Foundry offers it only as Global Standard or US Data Zone; and it requires 30-day data retention. Microsoft's new web search service for agents, Web IQ, "does not make EU Data Boundary, region-specific data residency, or in-region processing commitments today," and Microsoft acts as an independent controller of the queries. A US rule issued on Monday extends export licensing to data centres anywhere in the world that serve companies headquartered in China and other restricted countries.
Satya Nadella supplied the line for the week: "what the model tells us about itself is not independent evidence." The same is true of launch posts. The terms page, not the keynote, says where your data goes. This edition maps the week's new services against the boundary most European firms have promised their regulators, and argues that the newly available AI gateway is the place to enforce that map.
TL;DR
NVIDIA's numbers held; its governance didn't. Revenue rose 103% on the year, but the founder is gone, the board has added a national-security committee and the 10-Q is not yet filed. Treat key-person and export-control risk as live in your GPU supply chain.
Anthropic is now a public, founder-controlled company. It raised about $81 billion in the IPO, alongside a $26 billion private placement. It carries $900.9 billion of compute commitments and will report material events by 8-K. Its newest models still offer no documented EU-only inference route.
Microsoft shipped the control plane. Azure AI Gateway is generally available in five European regions, with billing from January 1. Most of the agent-governance features around it are still in preview.
New services, old boundary. Web IQ is outside Microsoft's Data Protection Addendum and the EU Data Boundary. Claude in Microsoft 365 Copilot is reportedly switched on by default on December 9 for tenants that have not chosen. Check each service's terms before it reaches personal data.
The Brief
1. NVIDIA's founder left; the numbers held
On November 8 NVIDIA's board opened "an independent investigation in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort." It added that it had "no reason to believe that either national security or NVIDIA secrets were compromised." The same day, Huang apologised to staff in a memo about his personal conduct. On November 11 the board "accepted the resignation of Jen-Hsun Huang, as NVIDIA's President, Chief Executive Officer and member of the Board, effective immediately." It appointed CFO Colette Kress as interim chief executive and Harvey C. Jones as its first-ever chair. On Tuesday's call Jones said: "We will not be addressing or taking questions on today's call regarding the circumstances of Jensen's resignation and the ongoing investigation." The board also created a National Security and Export Control Oversight Committee.
The numbers held. Third-quarter revenue was $115.8 billion, up 103% on the year, with data-centre revenue of $107.6 billion. "Vera Rubin is off to an explosive start, our roadmap is on track and our ambition has never been greater," Kress said. Three direct customers accounted for "14%, 11%, and 11% of total revenue." The fourth-quarter outlook of $130.0 billion "includes an additional week." NVIDIA said: "We plan to file our 10-Q in the coming weeks."
Why it matters: The supplier with the most concentrated position in your AI stack has an interim CEO, an open investigation and an unfiled 10-Q. Ask your GPU and cloud providers how they are assessing it. Source: NVIDIA — board statement, November 8 · SEC — NVIDIA Form 8-K, November 11 · NVIDIA — Q3 FY27 results · NVIDIA — national security committee · The Motley Fool — Q3 call transcript
2. NVIDIA's pipeline is set; the financing and the rules are tightening
NVIDIA says "Vera Rubin NVL72 and Vera CPUs began volume shipments in the third quarter," and its presentation says Vera Rubin "added nearly $20B in revenue q/q." On the call, NVIDIA said: "We remain on track to launch Vera Rubin Ultra next year and Feynman in 2028." Rubin HPC, a GPU for scientific computing, "is expected to be available in 2028." Co-packaged optics are "in volume production," and networking revenue exceeded $21 billion. For Europe, the DSX POD, a factory-built AI data-centre module, cuts deployment "from six months to as little as 14 days," and the company said: "We plan to ship our first pods to customers in the U.S. and Central Europe early next year." NVIDIA's investor deck puts its content per gigawatt at "Hopper $18B/GW to Rubin $40B/GW." Supply commitments rose to $390 billion, "primarily related to the multi-year procurement of memory supply." On financing, the CFO commentary says neoclouds "are growing faster than their balance sheets and long-term credit profiles can support," and NVIDIA's commitments "take effect only if the neoclouds secure financing." On allocation, Kress said: "All are getting the same pricing."
Then the rules. A US Commerce Department rule issued November 16 requires a licence to supply advanced chips to data centres anywhere that give companies headquartered in China and other restricted countries more than a set threshold of compute. NVIDIA's outlook assumes no data-centre compute revenue from China.
Why it matters: The roadmap is dated and on track, but supply is allocated to buyers who are ready and financed. European data centres serving Chinese-parented customers now carry US licensing exposure. Source: NVIDIA — supply chain partners · NVIDIA — DSX POD for Europe · NVIDIA — CFO commentary (PDF) · BIS — remote-access rule
3. Anthropic priced its IPO above the range
Anthropic priced its initial public offering on November 18 at $83 a share, against a filed range that stayed at "$70.00 and $78.00." Renaissance Capital called it "the largest IPO ever." It sold 980 million shares, upsized from 825 million, for about $81.3 billion. Of those, 956,351,354 were new shares. A concurrent private placement raised "$26.0 billion of Class N common stock." The shares began trading on Nasdaq under the ticker ANT on Thursday. According to Nasdaq data, they traded as high as $115.10 and closed at $97.17, up 17% on the offer price. The prospectus shows what the money is for. Anthropic has "purchase commitments and lease agreements … with an aggregate committed spend of approximately $900.9 billion," and "Approximately $70.7 billion of these commitments are due in the 12 months following September 30, 2026." Speaking with JPMorgan's Jamie Dimon at a livestreamed event days before the listing, Dario Amodei called compute "the deepest business dilemma that the company faces." A retail tranche of up to 80,135,135 shares was offered in nine European countries.
Why it matters: Your AI vendor now has more than $100 billion of new capital and a near-trillion-dollar compute bill to cover. Expect pricing discipline, not discounts, at renewal. Source: SEC — Anthropic final prospectus (424B4) · SEC — pricing term sheet · Anthropic — IPO pricing · Renaissance Capital — Anthropic IPO · Unusual Whales — Amodei on compute
4. A listed Anthropic is still a founder-controlled one
Going public changes what Anthropic must disclose, not who controls it. The prospectus says: "We are considered a 'controlled company' within the meaning of the Nasdaq Global Select Market corporate governance requirements." Its founders will hold "61.1% of the total voting power of our outstanding capital stock." Most of that comes through a Class F share that carries a fixed share of the vote: "Our Class F common stock will not have any economic rights." The Long-Term Benefit Trust, Anthropic's safety-oriented governance body, "intends to fill the vacant Class T director seat once a candidate is identified, which would result in four Class T directors and a full board of seven directors." Amazon and Google will hold non-voting Class N stock "representing 34.4% and 31.5% of our outstanding Class N common stock." These are the two cloud platforms through which many European firms buy Claude. The first lock-up release comes on the "27th day after the date of this prospectus (December 15, 2026)." What changes for customers is visibility. As a US-listed company, Anthropic now files quarterly reports and reports material events on Form 8-K.
Why it matters: Your third-party risk file gains a new evidence source: public filings. It also gains a governance fact. The vendor's control sits with its founders and trust, and its biggest shareholders are its distribution channels. Source: SEC — Anthropic final prospectus (424B4)
5. Fable 5.2 and Mythos 5.2: read the terms behind the benchmarks
Claude Fable 5.2 and Claude Mythos 5.2, launched November 9, come with data terms that European buyers need to read before the benchmarks. On the Claude API, you set inference_geo to "global", the default, where inference "may run in any available geography", or to "us". Anthropic's documentation for Amazon Bedrock lists Fable 5.2 with inference profiles for global and US, and "No" for the EU. In Microsoft Foundry, "Claude is available in Global Standard and US Data Zone Standard deployment types," and Microsoft's own documentation adds that these Anthropic-hosted versions "run on Anthropic's infrastructure (outside of Azure)." Anthropic's release notes say the 5.2 models "require 30-day data retention" and "aren't available under zero data retention unless expressly authorized by Anthropic." Enterprise Frontier Safeguards, which would let that monitoring run in the customer's own infrastructure, is still not generally available. Mythos 5.2 remains limited to verified programmes, and from December 15 individual "Defense Access" also "requires phishing-resistant MFA and no longer allows static or long-lived API keys." In Microsoft 365, Claude models in Copilot "are currently excluded from the EU Data Boundary." Microsoft has promised to move them inside by the end of 2026, and an admin guide citing Microsoft's message centre reports that "on December 9, 2026, it turns Claude on for every EU tenant that hasn't made a selection yet."
Why it matters: For workloads that touch personal or regulated data, Fable 5.2 today means processing outside the EU and 30 days of retention. That is a transfer and DPIA decision, not a model choice. Source: Claude docs — data residency · Claude docs — Bedrock · Claude docs — Foundry · Microsoft Learn — Claude models in Foundry · Microsoft Learn — Anthropic as a subprocessor · LazyAdmin — Claude and the EU Data Boundary
6. Anthropic's new features point to customer-controlled deployment
The most important launch from Anthropic in the IPO week got little attention. Self-hosted Claude Managed Agents, in private beta since November 12, lets customers "run agents in your own environment, whether that's in your own cloud or on-premise. You host the agents, credentials, and session history, and choose where requests to Claude are sent." Anthropic adds: "Model requests go where you choose, whether that's Anthropic directly, your cloud provider (Amazon Bedrock, Google's Agent Platform, or Microsoft Foundry), or your own gateway." It also launched Claude Meetings in beta. While in beta, Meetings "isn't available to some organizations, including those with zero data retention (ZDR), HIPAA-ready configuration, or customer-managed encryption keys (CMEK)." On security, Anthropic is "shortening the disclosure window in our coordinated vulnerability disclosure (CVD) policy from 90 days to 30 days for high- and critical-severity issues," for reports made from January 1, 2027. On November 16 its text watermark was extended: "Text generated by Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5 carries Anthropic's text watermark on supported platforms." That supports the AI Act's Article 50 marking duty, with the earlier caveat that changing sampling settings can weaken it.
Why it matters: Self-hosted agents put the agent's credentials and history inside your perimeter, and routing through "your own gateway" lets you choose a route that stays inside your boundary. That is last week's mandate in product form. Source: Claude — self-hosted Managed Agents · Claude — Meetings · Anthropic — disclosure timelines · Claude docs — API release notes
7. Microsoft made the AI gateway a product
Azure AI Gateway, "previously available as the AI Gateway tier of API Management, is now generally available as a fully managed Azure service for connecting users, applications, and agents to AI models and MCP tools across providers." It routes across Foundry, OpenAI, Anthropic, Bedrock, Vertex and other compatible endpoints, with token, request and cost limits and OpenTelemetry export. It is available in 17 regions, including West Europe, North Europe, UK South, France Central and Sweden Central. Pricing is free up to 10,000 requests a month, then $500 per gateway per month including 10 million requests. "Paid billing begins January 1, 2027." Customers on the old resource type have a deadline: existing AI Gateway SKU resources "continue to work until January 15, 2027." There is one inconsistency to check before migrating. The launch post promises to "Screen prompts and responses with Azure AI Content Safety," but the migration guide says: "Content safety policy isn't supported." Agent Grid, in preview, applies network policy to agents, and its "Enforcement operates outside the agent's trust boundary, without agent code changes."
Why it matters: A gateway is where you can enforce which models, regions and tools each workload may use. Now that it is a GA product with European regions, it can carry that boundary policy, provided you confirm which safety features survive the migration. Source: Microsoft Tech Community — Azure AI Gateway GA · Microsoft Learn — AI Gateway versus API Management · Microsoft Tech Community — Agent Grid
8. Agent governance became a control plane, mostly in preview
Nadella described Agent 365 as "a control plane that extends your existing governance, identity, security and management systems to the agents in your enterprise," adding: "We already have over 50 million agents in Agent 365." The Ignite announcements fill in the layers around it, though most are in preview. Entra Agent ID gives agents their own identities and is extending to agents built on other platforms. "MCP Firewall, in preview, will control access to MCP servers, tools, resources, prompts, protocols and transports." Agent 365's governance of the Azure AI Gateway and the open-source LiteLLM gateway "is in preview." Defender's runtime protection for local AI agents such as coding assistants "will be available to Defender for Endpoint P2 customers." Purview's "Enhanced data classification" is "generally available in January 2027." Nadella also announced an AI Transparency Initiative with more than 40 organisations: "Today, we can see almost everything models touch. But we cannot see what happens inside the model. And what the model tells us about itself is not independent evidence. These nested black boxes are simply unacceptable." The launch release adds: "Blaming the AI isn't an answer, and 'trust the model' isn't a control." The initiative has no products or dates yet.
Why it matters: The agent registry, identities and network controls that last week's mandate needs are arriving from Microsoft, but on preview timelines. Plan for 2027, and keep your own record in the meantime. Source: Microsoft — Ignite 2026: building the frontier ecosystem · Microsoft — Ignite 2026 Book of News · Microsoft — AI Transparency Initiative · Microsoft Security — Perception
9. Microsoft's new search for agents sits outside the boundary
"Today we're announcing that Web IQ is available to all Azure customers with new, lower pricing," Microsoft said on November 17. Web IQ is its web-search service for grounding agents in fresh information, priced from $1 to $9 per 1,000 calls depending on the endpoint. Its data terms differ from the rest of Azure. Microsoft's FAQ is explicit: "Section 4.1 of the Web IQ Terms of Use excludes the service from the Microsoft Products and Services Data Protection Addendum." Instead, "Microsoft and the customer act as independent data controllers for any Personal Data," and "Microsoft operates as a data controller under the Microsoft Privacy Statement." The FAQ adds that "Web IQ does not make EU Data Boundary, region-specific data residency, or in-region processing commitments today," and that it "doesn't detect or mask personal data in queries." A "Zero Query Logging" option exists, but "It isn't self-serve." Microsoft's enterprise search service is different: "Azure AI Search is renamed to Foundry IQ." It remains an Azure resource, but its new serverless tier "doesn't support private execution." The Sovereign Control Panel, in public preview and hosted in Sweden, assesses sovereignty controls but "uses non-enforcing policy assignments."
Why it matters: An agent that sends a customer's name in a Web IQ query discloses it to an independent controller, outside your processor contract and outside the EU Data Boundary. Block it for agents that handle personal data until legal has assessed it. Source: Microsoft Bing — Web IQ broadly available · Microsoft Learn — Web IQ FAQ · Microsoft Learn — Foundry IQ what's new · Microsoft Learn — Sovereign Control Panel
10. Brussels was breached, and Frankfurt named the asymmetry
On Thursday the European Commission confirmed an AI-assisted cyberattack. "I can confirm that the Commission has recently taken necessary steps in response to a detected cyber campaign," its spokesperson said. "AI assisted tools can be used and especially misused to detect vulnerabilities. Indeed we can confirm that this was done with AI tools." According to POLITICO, the EU institutions' cyber agency CERT-EU found that the attackers used ARTEX, "a publicly available AI-powered tool designed to identify vulnerabilities in IT systems." Sysper, the Commission's HR platform for payslips and leave, was among the systems disrupted for more than ten days. The same day, ECB Banking Supervision set its 2027–29 priorities. "Supervisors will assess banks' responses to AI-enabled cyber threats and their preparedness for post-quantum cryptography migration," it said, noting "92% of significant institutions revealing deficiencies in cybersecurity, third-party risk management and ICT risk controls." On the CEO action plans, Claudia Buch said: "We are now reviewing the plans, which had to be submitted by the end of October, and we will share good practices with the industry." She added: "Around 40% of incidents reported under DORA originate from third-party providers." The ECB's Financial Stability Review named the asymmetry behind this edition. Frontier models "are predominantly developed outside the EU and EU entities have often faced restrictions in accessing the most capable models." The AI Office also published a draft template for fundamental rights impact assessments. "Deployers will need to comply with this obligation as of 2 December 2027." Feedback is open until January 15, 2027.
Why it matters: The boundary cuts both ways. European firms get the newest models later and on non-EU terms, while attackers use whatever they can download. Supervisors will now ask how you are closing that gap. Source: Infosecurity Magazine — Commission confirms AI-assisted attack · POLITICO — Brussels Playbook · ECB Banking Supervision — priorities 2027–29 · ECB Banking Supervision — Buch speech · ECB — "Decoding the cyber-AI nexus" · European Commission — draft FRIA template
Deep Dive: Outside the Boundary
Most European firms have made a promise to their regulators, customers and works councils: our AI processes data here, under these terms, with these controls. This week the three suppliers that matter most to that promise changed shape, and the new services they launched did not all fit inside it. Here is how to read the week, what enterprises miss, and why a gateway, now generally available, is the right place to hold the line.
What Changed
At the chip layer, NVIDIA lost its founder in circumstances its board is still investigating. It added a board committee for national security and export controls, and is waiting on a US rule that reaches data centres worldwide. At the model layer, Anthropic became a listed, founder-controlled company with $900.9 billion of compute commitments. It released its strongest models on terms that, today, mean processing outside the EU and 30 days of retention. At the control layer, Microsoft shipped an AI gateway with European regions and announced the registry, identities and firewalls that agent governance needs, most of them in preview. It also made a new search service broadly available, under terms that make Microsoft an independent controller of the queries.
None of this is unusual on its own. Together, it shows how quickly the boundary a firm has promised can be crossed by a routine product update.
Why It Matters
Europe's rules make the boundary contractual. DORA Article 30(2)(b) requires financial entities' ICT contracts to state "the locations, namely the regions or countries, where the contracted or subcontracted functions and ICT services are to be provided and where data is to be processed, including the storage location." Article 28(4) requires them to assess an arrangement before entering into it. Under GDPR, a processor contract must ensure the processor acts "only on documented instructions from the controller, including with regard to transfers of personal data to a third country or an international organisation." Article 44 requires that "Any transfer of personal data" outside the EU meet the conditions of Chapter V.
Each of this week's launches lands somewhere on that map. A model whose default inference geography is "global" sits outside it unless you choose otherwise. A search service where the vendor is an independent controller is a disclosure, not processing on your behalf. A sovereignty panel with "non-enforcing policy assignments" is an assessment tool, not a control.
What Enterprises Usually Miss
The first thing is that defaults decide. Fable 5.2's API default is "global." Claude in Copilot is reported to switch on by default on December 9 for EU tenants that have not chosen. Most teams will never read the setting. The boundary is held only if someone owns those defaults.
The second is that the data role changes with the feature. The same vendor can be your processor for one service and an independent controller for another. Web IQ shows that inside a single cloud account. An agent's search query can carry personal data without anyone deciding to share it.
The third is that preview is not a control. MCP Firewall, Agent 365's gateway governance and the Sovereign Control Panel's policies are all in preview or non-enforcing. They are worth testing, but not worth citing to a supervisor as controls in place.
The fourth is that supplier governance is now material. NVIDIA's founder exit and its new national-security committee are a key-person and export-control question for every GPU contract. Anthropic's founders keep control after listing, and its largest outside holders are also its largest distribution channels. Both belong in the third-party risk file, not just the press summary.
The Governance / Infrastructure Implication
The answer is not to avoid the new services. It is to route them. A gateway that sits between every application, agent and model can hold the boundary in policy. Personal data goes only to models and regions inside it, while other work can use whatever the business prefers. Azure AI Gateway is now a GA product in five European regions. Anthropic's self-hosted Managed Agents explicitly support sending model requests to "your own gateway." Last week's mandate said what each agent may do. The gateway decides where its requests may go.
This is where Europe's rules become an advantage. A firm that already has to state data locations in its contracts, keep a register of its ICT providers and justify every transfer has, in effect, already written the routing table. The gateway turns that paperwork into enforcement. The new evidence streams help too. Anthropic now files 8-Ks, NVIDIA's board publishes its oversight structure, and Microsoft documents data roles service by service. Nadella's point about models applies to vendors as well: what a launch post says about itself is not independent evidence. The terms page is.
What Leaders Should Do Next
Before December 9, decide your position on Claude in Microsoft 365 Copilot. Before January 1, decide whether Azure AI Gateway, or another gateway, will be your routing and policy point, and confirm which safety features it enforces. For every new AI service this quarter, complete one line of the Boundary Map in the Artifact: where it runs, under whose terms, and what you will allow it to see.
Enterprise Playbook
For the CIO and CTO: Make the boundary map a gate. No new AI service, model version or feature reaches production data until someone has recorded where it runs, under which terms, and with which defaults. Start with the services in this edition's Artifact.
For the DPO: Assess Web IQ as a disclosure to an independent controller, not as processing. Decide whether agents that handle personal data may call it. Document the transfer and retention position for Fable 5.2 and Mythos 5.2 before any team uses them on personal data.
For the Microsoft 365 administrator: Set an explicit position on Claude in Copilot for each tenant before December 9, so that the reported default does not decide for you.
For the Head of Platform Engineering: If you run the old API Management AI Gateway SKU, plan the move before January 15. Confirm whether content-safety policies carry over, and use the gateway to route personal data only to EU-resident models.
For the Head of Third-Party Risk: Update the files for NVIDIA-dependent providers (key-person and export-control exposure) and for Anthropic (controlled company, public filings, compute commitments). Add Anthropic's 8-K feed to your vendor monitoring.
For the CFO: Budget for Azure AI Gateway billing from January 1, firm GPU pricing into 2027, and per-call charges on search and agent services that are easy to switch on and hard to see.
For the board: Ask, "Which AI services did we switch on this quarter that process data outside the boundary we promised our regulators, and who approved that?"
Artifact: The Boundary Map — This Month's New AI Services Against Your EU Commitments
Check each new AI service against the boundary your contracts and regulators expect: where it runs, under whose terms, and what to do before switching it on. The clauses behind it are DORA Articles 28(4) and 30(2)(b) for financial entities, and GDPR Articles 28 and 44. This is a starting point, not legal advice. Statuses are as published on November 19 and change often.
Service | Status | Where it runs | Data terms | Inside EU commitments? | Before you switch it on |
|---|---|---|---|---|---|
Claude Fable 5.2 — Claude API | Launched Nov 9 | "global" by default, or "us" | 30-day retention; no ZDR unless Anthropic authorises it | No EU-only option documented | Keep personal data off it, or route that work to an EU-hosted model |
Claude Fable 5.2 — Amazon Bedrock | Available via one endpoint | Global or US inference profiles; EU: "No" | Anthropic's 30-day retention applies | No | Pin the inference profile in every call and log it |
Claude Fable 5.2 — Microsoft Foundry | Preview, hosted by Anthropic | Global Standard or US Data Zone; "outside of Azure" | Anthropic's terms, not Azure's | No | Treat Anthropic as a separate provider in your register |
Claude in Microsoft 365 Copilot | Moving inside the EU Data Boundary by end-2026 | Today "excluded from the EU Data Boundary" | Anthropic as Microsoft subprocessor | Not yet | Decide per tenant before the reported Dec 9 default |
Web IQ | Available to all Azure customers | No in-region processing commitments | Microsoft is an independent controller; Zero Query Logging on request | No (outside the DPA and the EU Data Boundary) | Block for agents handling personal data until legal approves |
Foundry IQ (Azure AI Search), serverless | New consumption tier | Your Azure region | Azure terms | Yes, but no private execution | Use the Dedicated tier where private networking is required |
Azure AI Gateway | GA; billing from Jan 1, 2027 | 17 regions, five in Europe | Azure terms | Yes, in European regions | Make it the routing and policy point; confirm content safety |
Sovereign Control Panel | Public preview | Hosted in Sweden | Assessment only | Non-enforcing | Use it to assess, not as evidence of control |
Self-hosted Claude Managed Agents | Private beta | Your cloud or on-premise | Your terms, plus those of the model route you choose | Depends on the route | Pilot it with an EU-only model route through your gateway |
One line to keep: The boundary is a routing decision. Write it down once, then let the gateway enforce it.
What to Watch Next
November 20–24. The EU–UK summit meets in Brussels and the ICO's call for evidence on agents closes on November 20. The Commons considers the Lords' amendments to the Cyber Security and Resilience Bill on November 24.
Late November. The first sector work from the Commission's special envoy for industrial AI, Jim Snabe, is due "in two weeks" from November 13, according to von der Leyen.
November 30. Fast mode for Claude Opus 4.8 is removed.
December 1–2. Jensen Huang and NVIDIA's general counsel are due for transcribed interviews with US House committees, according to AP. The AI Act's Article 50(2) marking grace period ends on December 2.
December 9. Claude in Microsoft 365 Copilot is reported to switch on by default for EU tenants that have not made a selection.
December 15. Anthropic's first lock-up release; Mythos "Defense Access" requires phishing-resistant MFA.
January 2027. Azure AI Gateway billing begins on January 1, and legacy AI Gateway SKU resources stop on January 15. Anthropic's shorter vulnerability-disclosure window applies from January 1. Feedback on the AI Office's draft FRIA template closes on January 15.
What to Read Now
Market
SEC — Anthropic final prospectus — Search for "controlled company" and "aggregate committed spend". Those two passages say more about your vendor than the roadshow did.
Infrastructure
NVIDIA — CFO commentary, Q3 FY27 (PDF) — Read the neocloud paragraph. It says more about the financing of European GPU capacity than any analyst note.
Enterprise AI
Microsoft Learn — Web IQ FAQ — A rare example of a vendor stating its data role plainly. Give it to your DPO.
Microsoft Tech Community — Azure AI Gateway GA — Read alongside the migration guide; check which policies make the move.
Regulation
Claude docs — data residency — One table that answers where your prompts are processed.
Supervision
ECB — "Decoding the cyber-AI nexus" — The first ECB text to treat access to frontier models as a resilience issue. Read the section on third-party concentration.
The One Call to Make
This week, ask your DPO one question: "For every AI service we enabled or plan to enable this quarter, do we know where it processes data, under whose terms, and which defaults change on December 9?"
Why this one: Three suppliers changed shape in a week. The newest frontier model offers no documented EU-only route, Microsoft's new search service sits outside its data terms, and a Copilot default is reported to flip in three weeks. The question tests whether your boundary is a policy or a hope.
If the answer is "we'd need to check with IT": That is the finding. Start with the Boundary Map for the services in this edition.
Reply with one word: "inside" or "outside." It joins the mapped, ready, sent, contained, revocable, clocked, tagged, verified, failed-over, recorded, hosting and bounded tallies in this newsletter's picture of where European AI operations actually stand.
That’s it for this week.
Thirteen weeks into the autumn, the questions have moved from what the model can do to who controls it, who is told, who counts, who checks, what lies beneath, who keeps the record, who owns the weights and who gives an agent permission to act. This week the question was where the boundary sits. The suppliers will keep changing. The boundary should not change with them.
Until next Thursday, João
OnAbout.AI delivers strategic AI analysis to enterprise technology leaders. European governance lens. Vendor-agnostic. Actionable.
If this landed in your inbox from a forward — subscribe here to get the full picture every week.
