AI Update Week 22: Google redesigns search, two trillion-dollar IPOs on the way

1. Google I/O 2026: More Than Just a Search Engine Update

Google I/O took place on 19 and 20 May in Mountain View. Over 100 announcements – here are the most relevant ones, complete with demo links.

New Models

Gemini 3.5 Flash (Keynote Playlist) outperforms Gemini 3.1 Pro in coding, agent, and multimodal benchmarks. It is four times faster than comparable frontier models, and twelve times faster on the Antigravity developer platform. Available worldwide immediately.

Gemini Omni processes and generates text, audio, images, and video in real time – in dynamic formats. This is Google's first serious step into generative media production.

Project Genie (YouTube Demo) is a DeepMind research model that generates interactive virtual worlds from simple text descriptions. Not a product yet, but a preview of procedural game world generation.

Search

The new smart search box accepts images and files alongside text. It guides users towards complex, longer queries rather than simple keywords.

Information Agents run 24/7 in the background, monitoring topics without requiring user input. Custom dashboards will follow in the coming months for AI Pro and Ultra subscribers.

Generative UI builds the layout of the search results page dynamically based on the specific query – offering a context-specific view instead of a fixed link list.

Personal Intelligence integrates Gmail and Google Drive directly into Search – across nearly 200 countries. Users of Gmail and Drive get search results that understand their personal context.

Workspace

Google Pics is a new AI image tool built directly into Docs, Drive, and Slides. It treats every image element as an independent object – allowing users to edit, replace, and regenerate foregrounds and backgrounds separately. Every output is watermarked with SynthID. Rolling out this summer for AI Pro and Ultra subscribers.

WebMCP is a new open web standard: browser agents can use it to not only read pages, but also act on them. Origin Trial in Chrome 149 has begun.

Analysis: This week, Google showed that Search is no longer just a search box. Personal Intelligence (integrating Gmail + Drive into Search) is the feature with the biggest immediate daily impact. Open question: The classic business model – search, click, ad – and the new model – agent completes task directly – do not fit together structurally. Google does not yet have a convincing answer.

2. Two Trillion-Dollar IPOs: OpenAI and SpaceX/xAI

This week, it became clear: the world's largest AI companies are going public – and the rules of the most important stock index have been changed specifically for this purpose.

OpenAI confidentially filed its S-1 prospectus with the SEC on 22 May. Goldman Sachs and Morgan Stanley are lead underwriters. Target valuation: $852 billion to $1 trillion. Planned timing: Autumn 2026.

The numbers: $25 billion annualised revenue (as of March 2026). 50 million consumers, 9 million business users. Microsoft holds 27%, and the OpenAI Foundation holds 26% of the new Public Benefit Corporation. At the same time, OpenAI lost $1.22 for every dollar earned in Q1 2026 – and is investing around $50 billion in compute power in 2026.

SpaceX/xAI: Elon Musk had SpaceX acquire xAI in February and dissolved it as a separate company. Grok and X are now under the new "SpaceXAI" unit of SpaceX. SpaceX plans its IPO for 11 June on the Nasdaq (ticker: SPCX), targeting a $1.75 trillion valuation. This would break Saudi Aramco's previous IPO record. SpaceX grew by 30% in 2025 to $18.7 billion in revenue – but lost $4.9 billion because xAI's losses amounted to $6.4 billion.

S&P 500 Rule Change: S&P Dow Jones Indices has eliminated the profitability requirement for IPOs large enough to rank in the market-cap top 100. This ensures OpenAI and SpaceX can join the index immediately after flotation – without waiting for profitability.

Analysis: Two companies that have not yet proven themselves in the market are going public at a combined valuation of nearly $3 trillion – and the world's premier stock index has changed its rules just to let them in. Consequently, pension fund and ETF capital will flow automatically into these companies, whether investors like it or not. This is highly questionable. Does this open the door to hype and speculation? Pension assets should not be sacrificed to hype. My recommendation: if you hold S&P 500 ETFs or have a pension fund, you should check where your money actually ends up – and whether this aligns with your investment strategy. Of course, it could work out: Amazon made no profit for years and still transformed global commerce. But few ETF holders knew that back then – and they had no choice.

3. Anthropic: Pricing Model Changes, KPMG Signs Up

Code with Claude took place on 6 May in San Francisco and on 19 May in London (Tokyo follows on 10 June). Core features – Dreaming (agents build their own long-term memory from past sessions), Multi-agent Orchestration, and Outcomes – were first showcased in SF, and detailed for EU developers in London.

What's new this week: From 15 June, Anthropic will separate programmatic agent usage from standard chat limits. Pro users will receive $20 in monthly credits for agent runs, Max 5x $100, Max 20x $200. The API grew 17-fold year-on-year.

KPMG is integrating Claude into Digital Gateway, the central workspace for all 276,000 KPMG employees worldwide.

Analysis: The pricing model update is structurally fair: those who run agents pay for agents. Those who use Claude for quick answers won't notice a difference. More important than pricing is direction: with Dreaming and the KPMG rollout, Claude is evolving from a tool into a platform embedded in corporate processes.

4. Switzerland: A Two-Speed AI Boom

The DigitalBarometer 2026 (published 24 May, Risiko-Dialog Foundation / digitalswitzerland, n=1,278): Three-quarters of the Swiss population regularly use AI tools like ChatGPT, Claude, or Gemini. 48% believe opportunities outweigh risks, but only 53% feel confident using them. 76% prioritise data protection over convenience. 59% want Switzerland to actively promote new technologies.

The UBS Study (19 May, n=2,500 companies): 60% of Swiss companies use AI – but selectively, not systematically. 97% have ongoing AI initiatives. Only 5% have a data foundation ready for scaled AI.

Infomaniak: Founder Boris Siegenthaler transferred majority voting rights to a new non-profit Swiss foundation on 13 May. The foundation holds special shares that can never be sold – structurally protecting Infomaniak from takeovers. The company remains operationally independent under CEO Marc Oehler. All 36 employees voted unanimously in favour.

Infomaniak is my personal recommendation for sovereign cloud and AI infrastructure in Switzerland. Swiss data centres, no data transfer to the US, GDPR compliant – and now with an ownership structure that guarantees long-term stability. For companies that cannot allow data to leave Switzerland, it is the logical alternative to AWS, Azure, and Google Cloud.

Analysis: 97% AI initiatives, 5% ready data foundation. That is the real challenge of the next three years – not adopting AI tools, but building the foundation for them.

5. Mistral Medium 3.5: Consolidation over Expansion

Mistral has released Medium 3.5, consolidation four previous products: Medium 3.1, Magistral, Devstral 2, and the coding assistant in Le Chat and Vibe CLI. The new model handles chat, reasoning, coding, and agentic tasks in a single system. Coding sessions now run asynchronously in the cloud – in parallel, without supervision.

The strategic move behind this: Mistral is following OpenAI and Anthropic towards consolidated, generalist models.

3 Things I am Watching This Week

1. Personal Intelligence is the Google feature with the greatest daily impact. Gmail + Drive directly in Search across 200 countries will change how people interact with their own information daily.

2. Two loss-making trillion-dollar companies are entering the S&P 500. This is an index-level experiment. If it fails, a large portion of global retirement savings will bear the loss.

3. The 5% data-readiness figure from the UBS study is the most relevant benchmark for AI consulting. Not "are you using AI", but "do you have the foundation for it" – that is the correct question.