AI News Week 24: Microsoft turns Foundry into an app store, Apple hires Gemini for Siri AI, OpenAI and Anthropic file for IPOs

1. Microsoft Build: Access to 11,000 Models – and Claude Directly in Excel

At Build 2026 (2–3 June), Microsoft set out its AI platform strategy for the year. Two points matter to you.

Foundry becomes the app store for models. Microsoft Foundry’s model catalogue now includes over 11,000 models. These range from frontier models (OpenAI GPT-5.5, Anthropic Claude Opus 4.8, Sonnet and Haiku, Google Gemini) to open source and specialised vision, multimodal and timeseries models. All run via a single Azure endpoint with single billing. For businesses, this means one contract, one login, one bill—instead of connecting five individual providers. The Agent Framework (stable orchestration, multi-agent patterns, integration of GitHub Copilot SDK and Claude Agent SDK) is now in stable release. (Microsoft Foundry Blog)

Claude is coming to Excel. Microsoft and Anthropic have expanded their partnership. You can now call Claude directly in Excel Agent Mode. Start writing and explaining formulas, cleaning data, drafting analyses and building workflows without leaving the spreadsheet. Excel has an estimated 750 million users worldwide. This puts Claude inside the tool that remains open in almost every office. (A Guide to Cloud – Build 2026 Recap)

The Takeaway: Microsoft is not selling the best model, but the most convenient access to all of them. This is the stronger position. If you use Microsoft 365, you get AI where you already work: Excel, Copilot, Teams. For most Swiss SMEs, this is more relevant than any benchmark arms race. A word of warning: Microsoft now bills Copilot coding based on token consumption. Forrester calls this "the new gateway drug model". Keep an eye on costs before entire teams start agentic coding.

2. Apple WWDC: "Siri AI" Runs on Google – but is Delayed for Europe

At the WWDC keynote on Monday, 8 June, Apple revealed the biggest Siri overhaul in years. This was Tim Cook’s final keynote as CEO, before he hands over to hardware head John Ternus on 1 September. Apple is renaming Siri to Siri AI. The new Siri is a standalone app that handles natural back-and-forth conversations, understands personal context and on-screen content, and takes actions across apps. In one demo, it navigated to a location from an Instagram post. The Passwords app now automatically changes insecure passwords agentically. (TechCrunch)

The Google partnership, decoded. Apple is not building its own frontier model. Instead, it is building a second generation of Apple Foundation Models in partnership with Google. The top model, AFM Cloud Pro, is on par with Gemini Frontier according to Apple, and runs on Nvidia GPUs in Google's cloud. The deal was finalised in January 2026, reportedly worth around $1 billion per year. Processing is three-tiered: on-device, Apple Private Cloud Compute, and the cloud models. "Siri AI is not Gemini with an Apple logo" – but it cannot function without Google. (CNBC)

You choose your model. Through the new Extensions framework in iOS 27, users can set a third-party model as their default assistant—whether Claude, ChatGPT, or Gemini. Apply is ending its exclusive tie-up with ChatGPT, a move prompted partly by the EU Digital Markets Act. For Anthropic, this is the biggest consumer opportunity in the company's history: Claude becomes directly selectable on the iPhone for the first time. Apple has around 2.2 billion active devices. (The Next Web)

Important for Switzerland and Europe: Siri AI will not roll out in Europe or China yet due to regulatory hurdles. Private developers can access a beta starting Monday (requiring the latest devices). For everyone else, the new Siri arrives in autumn via software update alongside new hardware. Concurrently, the new Siri is unlikely to launch in Switzerland for the time being. Apple shares dipped into negative territory during the keynote.

The Takeaway: A familiar pattern emerges: the most exciting AI features arrive later in Europe. Time will tell if this is a drawback or a buffer. What price are we willing to pay for security and data privacy?

3. Anthropic: Confidential IPO Filing, Claude Everywhere, Market Share Triples

Anthropic made three strategic moves this week that fit together perfectly.

Confidential IPO filing. On 1 June, Anthropic confidentially filed for an IPO, shortly after its Series H round valued the company at $965 billion (see week 23). Key reported data: around $44 billion in annualised revenue, with a first operating profit (approx. $559 million) anticipated for Q2 2026. (Anthropic Newsroom)

Claude becomes an ecosystem. Anthropic launched the Claude Partner Hub and a Services Track. This formalised $100 million partner programme supports firms deploying Claude productively at enterprise scale. The logic: a successful pilot is not the same as a system running operations. Anthropic is building the partner network that Salesforce established with AppExchange for CRM. In parallel, Project Glasswing (Claude autonomously locating software vulnerabilities) has expanded to around 150 organisations. (Anthropic – Expanding Project Glasswing)

Market share triples. According to the Momentic report (based on Similarweb data), ChatGPT accounts for 54.7% of worldwide web traffic among the seven largest AI chatbots—down from 76.5% in February 2025. Google Gemini follows at 27.4%, and Claude at 8.2%. However, Claude's volume is growing fastest: up 306% in a quarter, from 203 million web visits in January to 824 million in April 2026. In the US, Claude sits at 12.5%. An important caveat: this only measures browser visits. Many use Claude via the desktop app—especially for Cowork—which is completely missed by these metrics. It also excludes mobile apps, integrations in Windows and Microsoft 365, and API usage. Reports indicate Anthropic's enterprise revenue is growing faster than web traffic suggests. (Momentic)

The Takeaway: The trend from week 23 continues. While OpenAI makes louder demonstrations, Anthropic embeds Claude into the tools people actually work with: Excel, the iPhone, and enterprise partner integrations. Model benchmarks and actual distribution are two different things. Anthropic is winning the distribution game.

4. Regulation: Trump Favours Voluntarism, Colorado and the EU Set Deadlines

This week highlights the deep divide in AI regulation: voluntary in Washington, binding in Colorado and Brussels.

Trump's executive order (2 June): voluntary. The US is not imposing mandates on AI firms. The core policy: companies may voluntarily present new models to the government 30 days before launch. There is no licensing, and no mandatory auditing. Only one element is binding: agencies must strengthen their cyber defences, and a new Treasury unit will coordinate the discovery and patching of software vulnerabilities. In short, Washington relies on incentives rather than rules—the polar opposite of the European approach. (White House, NPR)

Colorado, 30 June – the first real US deadline. The Colorado AI Act takes effect in 22 days, governing high-risk AI in employment, healthcare, financial services, education, housing, and legal sectors. Obligations include risk-management programmes, annual impact assessments, disclosure around automated decisions, and a right to appeal. A federal bill ("Great American AI Act") aims to suspend state laws for three years, but it has not passed committee, and the Colorado deadline stands regardless. (buildfastwithai)

EU AI Act, 2 August – 55 days to go. Most provisions of the EU regulation apply in 55 days. This triggers full compliance duties for high-risk AI (employment, education, critical infrastructure, financial services, law enforcement, migration, justice). General-purpose models like GPT-5.5, Claude Opus 4.8, Gemini, and Grok 4.3 face additional transparency and auditing demands. Non-compliance fines reach up to €35 million or 7% of global annual turnover. (buildfastwithai)

The Takeaway for Switzerland: Businesses selling in the EU or handling EU customer data face binding duties in two months: risk assessments, documentation, and human oversight must be ready. If you prefer not to send your data to the US, watch the sovereignty stack: Mistral will launch its new data centre near Paris (13,800 Nvidia GB300 GPUs) at the end of June, positioning itself clearly as the European alternative to OpenAI and Anthropic. The EU deadline speeds up this decision. (Euronews – Mistral)

5. Quick Fire: June Model Window, SpaceX IPO, S&P 500 Rejection, Dark Patterns

June will be packed with new models. Google announced a more capable Gemini (version 3.5 Pro) for this month, though a precise date is missing. Anthropic is also expected to release a new Claude model mid-month, though this remains unconfirmed. My advice: plan based on the models available today, not on rumours.

OpenAI also files confidentially (8 June). A week after Anthropic, OpenAI announced it has filed a confidential draft S-1 with the SEC. The statement was surprisingly candid: "We expect it to leak, so we are just announcing it. We haven't decided on timing; it could take a while." This means an IPO is prepared but not scheduled. In the span of seven days, the two largest AI labs have formally initiated their listings. (OpenAI)

SpaceX IPO on 11/12 June. SpaceX is going public: pricing on Thursday, trading from Friday under the ticker SPCX. The target is $75 billion at a valuation of $1.75 trillion or more—the largest IPO in history. The AI link: Elon Musk's AI firm xAI (Grok) is structured within the same group. Notably, 30% of shares are allocated to retail investors, compared to the typical 10%. Rather than a generous gesture, this suggests institutional investors view the stock as overpriced. The AI division xAI is operating at a loss, though SpaceX overall remains profitable thanks to Starlink. The greatest scam or the biggest IPO in history? We will find out. Regardless, SpaceX, Anthropic, and OpenAI are hitting the market almost simultaneously.

No AI boom in index funds. Passive investors in S&P 500 funds will miss out on the major private AI firms for now. The index provider decided on 4 June not to fast-track SpaceX, Anthropic, and OpenAI. They must trade publicly for over a year and show several quarters of profitability. This delays entry until mid-2027 at the earliest. They could enter the Nasdaq-100 tech index sooner. This is highly relevant for anyone assuming an index tracker automatically captures the AI boom. (Fortune)

37 manipulative patterns in chatbots. A US civil rights group documented 37 manipulative design patterns in ChatGPT, Gemini, Claude, Replika, and Character.AI—ranging from artificially extended sessions to complex deletion processes and emotional dependence. While some are standard app designs, patterns fostering emotional dependency are concerning. These are a preview of upcoming regulatory debates. My view: the AI Act must integrate user protections that we neglected during the social media era. Otherwise, corporations capture the profits while the public bears the health costs. (buildfastwithai)

Three Things to Action This Week

1. Review whether Foundry simplifies your provider contracts. If you currently pay three or four AI vendors separately, calculate the impact of a single Azure contract. Analyse whether token-based developer billing could become a cost trap. Calculate before you roll out.

2. Prepare for the EU AI Act deadline if applicable. Do you process EU customer data or sell into the EU? If so, you have 55 days until 2 August. Compile an honest inventory: where do you use AI to make decisions about people (recruitment, credit, access)? That is exactly where you need documentation and human oversight ready.

3. Expect Siri AI to arrive late in Switzerland. Apple is stalling the launch of the new Siri in Europe. Once the model options (ChatGPT, Gemini, Claude) become available here, test the same task across all three on your device. You will find that "the best model" depends entirely on your specific use case. Until then, continue evaluating the models in their standalone apps.

PS: My practical guide "AI in E-Commerce: When it Actually Helps – and When to Avoid it" has been posted this week. From data analysis to product descriptions and data privacy, it covers what actually works and what to stay away from. (To the Article)