The UN has launched its first Global Dialogue on AI Governance, created by the General Assembly as a follow-up to the Global Digital Compact. The goal is to give every country, including the least technologically advanced, a seat in the discussions on the rules, uses and risks of AI.
What you need to know
- Every country at the table: the Dialogue is a UN cooperation platform bringing together all 193 member states, the private sector, researchers, the technical community and civil society around AI governance.
- A first session in Geneva: it took place on 6 and 7 July 2026, alongside the ITU’s AI for Good summit, co-chaired by the ambassadors of El Salvador and Estonia. A second session is planned in New York in May 2027.
- Concrete proposals on the table: Secretary-General Antonio Guterres put forward four priorities, including a global AI fund for developing countries and an environmental transparency initiative inviting major players to publish their carbon footprint.
Stakes and outlook
Everything hinges on aligning rules and practices internationally. This Dialogue can help countries compare their approaches on very practical topics, such as the use of AI in public services, education or healthcare. It can also narrow the gap between countries with strong digital infrastructure and those still lacking access, data or skills. But its reach remains that of a discussion forum, with no binding power. The Geneva announcements are still proposals. The real test will be whether states and standards bodies take up these commitments before the next session.
On 7 July 2026 in Geneva, Salesforce, the American customer-relationship software company, announced a one-billion-dollar investment plan in Switzerland over five years to accelerate the adoption of agentic AI. The announcement, made by CEO Marc Benioff on the sidelines of the AI for Good summit, aims to strengthen customers, partners and local skills, at a time when Switzerland is establishing itself as an international hub for AI governance.
What you need to know
- A targeted investment: Salesforce wants to support the transformation of Swiss companies with Agentforce, its AI-agent solution able to automate tasks such as answering a customer or drafting a document.
- Deployments already live: Oviva handles more than 300,000 customer messages a month with Agentforce, while FREITAG reports over 95% satisfaction on recurring requests.
- A long-standing local presence: in Switzerland since 2004, Salesforce claims more than 1,000 customers, over 100 partners and offices in Zurich and Lausanne.
Stakes and outlook
The immediate lever is operational: delegating simple requests, such as a forgotten password or routing a customer, to free up human time for support, sales or advice. The real impact will depend above all on skills and on the quality of internal data, without which AI agents stay limited. The announcement should also be read for what it is: a corporate commitment whose precise breakdown remains to be detailed, while Salesforce stock has lost nearly a third of its value this year amid doubts about the return on its AI spending. Switzerland joins a series of European pledges, two billion in France and one billion in Italy, which is as much about positioning as about business.
In Switzerland, job postings tied to AI reached a record in 2025, with 25,000 positions counted by PwC’s AI Jobs Barometer 2026. The real change is not the appearance of new jobs, but the need for employees able to use AI day to day.
What you need to know
- A sharp rise in volume: in 2025, AI-related positions grew by about 9,000 to reach 25,000 listings, according to consulting firm PwC.
- Still a small share of the market: these listings accounted for just 1.8% of published jobs, or roughly one posting in fifty-five.
- Strong demand for applied skills: demand for profiles able to apply AI day to day jumped by about 8,400 positions, while demand for AI developers rose by only 220.
Stakes and outlook
The shift is playing out in the spread of AI within existing jobs. Employees able to use these tools on concrete tasks, such as analyzing data, automating documents or supporting a decision, are becoming more sought after. Technology, media and telecoms remain the most exposed sectors, but the use is spreading to other industries. PwC also notes that AI skills are associated with above-average salaries, notably in healthcare and energy. The takeaway for Swiss companies: training your teams in AI will matter as much as, if not more than, hiring experts.
Data centers are becoming a point of tension in Switzerland, driven by demand for cloud and artificial intelligence. Their consumption of electricity, water and land has a direct impact on local grids and planning projects.
What you need to know
- An already high density: Switzerland has more than 120 data centers, one of the highest concentrations per capita in the world.
- Direct pressure on electricity: in 2024, data centers accounted for 3.6% of Swiss electricity consumption, almost as much as the energy needed to run all the country’s trains for a year.
- Projects harder to accept locally: the Beringen project, in the canton of Schaffhausen, could use at full capacity the equivalent of 75% of the cantonal power supply and consume as much water as about 344 households.
Stakes and outlook
Everything depends on the grid’s ability to absorb demand concentrated in one place. Municipalities must weigh digital appeal, infrastructure costs and local acceptance. The large centers known as “hyperscalers”, run by groups such as Amazon, Microsoft, Google or Meta, are not mere server rooms: they are industrial facilities built to run cloud services and, increasingly, AI compute. Expected growth could push their share above 5% of Swiss electricity by 2030, a rise comparable to the annual consumption of about 180,000 households. One point stays little discussed: these sites take up vast areas for very few jobs and little local tax revenue, which fuels opposition as much as the water and energy questions.
On 8 July 2026, the European consortium ÆTHER, initiated notably by 2CRSi and SiPearl, announced its bid for the European Commission’s upcoming “AI Gigafactory” call for projects. The plan targets two industrial sites near Strasbourg to host sovereign AI compute capacity.
What you need to know
- A start planned for 2027: the first site, FR-SXB1, could go live in 2027 if its acquisition is finalized by the end of October 2026, followed by FR-SXB2 a few months later.
- A gradual ramp-up: the two sites would first offer 42 MW, with a target of 82 MW within twelve months and then more than 400 MW in the long run. The scale stays modest next to the leaders, when the largest American site, xAI’s Colossus 2, already runs around 950 MW and OpenAI’s Stargate project aims for several gigawatts.
- A full European chain: 2CRSi supplies high-performance servers, SiPearl designs processors without manufacturing them itself, Axelera AI provides AI accelerators, and Outscale, a Dassault Systemes subsidiary, handles the sovereign cloud.
Stakes and outlook
ÆTHER seeks to prove that Europe can build, power and run its own AI infrastructure, instead of depending solely on American or Asian platforms. The immediate challenge is operational: securing the sites, the electricity and the grid-connection permits, with RTE’s response on power capacity expected as early as late July 2026. Longer term, the project could support local skills in data centers, energy, high-performance computing and industrial maintenance, for uses ranging from research to healthcare, industry and public services. One caveat, though: ÆTHER is only a candidate, the Commission’s tender has not yet launched, and a rival consortium, AION, is pursuing the same goal with around thirty members. AI sovereignty will be decided as much in power cables and political timing as in processors.
On 13 July 2026, Helsing announced it had raised 1.8 billion dollars in a round that values this German defense group at 18 billion dollars. The deal, the largest ever by a European defense startup, strengthens the weight of military-AI players on the continent, but also underlines a growing reliance on private capital to fund defense innovation.
What you need to know
- A triple jump in valuation: founded in 2021 in Munich, Helsing went from about 12 billion euros in summer 2025 to 18 billion dollars with this Series E, a pace that makes it the best-funded defense startup in Europe.
- An AI player, not a classic manufacturer: Helsing designs software first, such as its Altra decision-support platform, but also HX-2 kamikaze drones deployed in Ukraine and an autonomous combat-aircraft project. It calls itself a “neo prime”, a software-driven defense company.
- A gap with the United States: its American rival Anduril raised 5 billion dollars in May at a 61-billion valuation, more than triple Helsing’s.
Stakes and outlook
This funding gives Helsing the means to hire, develop its platforms and answer public contracts, on skills tied to AI, embedded software and autonomous systems. The deal can speed up the structuring of a European technological-defense sector, carried by NATO’s rearmament. One contradiction is worth noting, though: Helsing presents itself as a sovereign European champion, but three of its biggest new investors are JPMorgan, Goldman Sachs and a Canadian pension fund. The displayed sovereignty thus rests in part on non-European capital. In defense, the edge will not come from the amount raised, but from the ability to deliver fast, reliable and truly sovereign.
Alibaba has banned its employees from using Claude Code, Anthropic’s coding assistant, after the discovery of a mechanism able to detect use from China.
What you need to know
- Claude Code flagged as high-risk: according to Reuters and the South China Morning Post, Alibaba added Claude Code to its list of high-risk software, asked employees to uninstall Anthropic products (Sonnet, Opus, Fable) and to switch to Qoder, its in-house coding assistant.
- Two readings of the same code: researchers found hidden logic checking time zones and Chinese proxies. An Anthropic engineer acknowledged the feature, presenting it as an anti-abuse experiment launched in March and removed on 1 July. Alibaba sees it as a surveillance risk.
- Against a backdrop of open conflict: the ban comes the day after an accusation by Anthropic, which blames operators tied to Alibaba’s Qwen lab for the largest known distillation attack against Claude. Alibaba did not respond on the substance.
Stakes and outlook
The affair crystallizes the question of control over the tools developers use every day. A coding assistant reads files, suggests fixes and sometimes touches sensitive tasks such as a payment module or a security flaw. If it is seen as able to spot certain countries, it becomes a matter of industrial sovereignty. Beyond the migration to Qoder, the episode speeds up the separation between American and Chinese AI ecosystems, with increasingly local tool chains on each side.
OpenAI is said to have cut the cost of running some of its models by more than 50%, according to sources cited by The Information. The immediate effect would be a massive drop in the number of Nvidia GPUs needed to handle part of ChatGPT’s requests.
What you need to know
- Cheaper answers to produce: software optimization is said to have taken compute needs from tens of thousands of high-end GPUs to just a few hundred for part of ChatGPT’s traffic.
- An in-house chip to go further: on 24 June 2026, OpenAI and Broadcom unveiled Jalapeno, their first chip specialized in inference, designed to reduce reliance on Nvidia’s general-purpose GPUs. Broadcom makes semiconductors and is not a cloud provider.
- Gains from several levers: the paths mentioned include model compression, reuse of computations already done, batching of requests and routing each query to the most suitable model.
Stakes and outlook
The principle is simple: each generated answer would cost far less to serve. For OpenAI, whose inference compute is said to represent about half of the revenue it spends, the stakes are almost existential. For companies, the topic is red-hot, because 2026 is marked by an explosion of AI costs with no clear ceiling, and any drop in cost per request finally makes large-scale deployments sustainable. The battle could also depend less on the raw volume of GPUs bought and more on software efficiency, custom chips and optimization skills. One sizable caveat remains: these figures are self-reported. OpenAI has disclosed nothing about the technical mechanism of its optimization, and Jalapeno’s performance has not been independently benchmarked. The industry has already seen real efficiency gains in the lab that proved far more limited in production.
The past month was the busiest for AI model releases in two years. Between 13 June and 13 July 2026, nearly every major lab shipped something. Here are the main releases, with their primary use, to help you spot what fits your needs.
| Provider | Model | Primary use |
|---|---|---|
| OpenAI | GPT-5.6 Sol | The most powerful: demanding reasoning, code and science |
| GPT-5.6 Terra | General enterprise use, a good cost-performance balance | |
| GPT-5.6 Luna | High volume, low cost, general-purpose assistants | |
| Anthropic | Claude Sonnet 5 | Code and writing, the new default model on claude.ai |
| Claude Fable 5 | Complex reasoning and strategy (back on 1 July) | |
| xAI | Grok 4.5 | Research and reasoning with real-time data from X and the web |
| Meta | Muse Spark 1.1 | Agentic work and computer use, very large context window |
| Mistral | Robostral Navigate | Robot navigation using a simple camera and natural-language instructions |
| ByteDance | Seedream 5.0 Pro | Multilingual image generation and editing, precise localized retouching |
| Meituan | LongCat-2.0 | Software development, open source, trained on Chinese chips |
Note: Google’s much-awaited Gemini 3.5 Pro was still not available to the public in mid-July, with its launch having slipped from June to July. Its faster version, Gemini 3.5 Flash, stays available in the meantime.
✳ App under the prism: LM Studio
1. What is it?
LM Studio is a desktop app (Mac, Windows, Linux) to install and run AI models directly on your machine, without the cloud. Everything stays local, no data is sent to a third party. Its strength for beginners: it analyzes your hardware and recommends models suited to its power.
2. Why it’s fascinating
Since June 2026, the LM Link feature lets you use remotely, from a phone or another computer, a model running on a machine left at home. A freelancer can thus host a large model on their PC and query it from their iPhone on the train, through the Locally mobile app. The connection goes through a private, end-to-end encrypted network, without exposing the computer to the internet.
3. Why it’s limited
Power depends on your hardware, but you need no exceptional gear to get started. Models up to 9 billion parameters run smoothly on a recent laptop, and their level is already impressive for correcting a text, rephrasing it or sorting notes. The scale gap remains: a giant like Opus is estimated to approach a thousand billion parameters, a power that even several tens of thousands of francs of equipment cannot reach at home. Local AI does not replace these frontier models, but it gives a concrete taste of the future, while keeping your data with you. Worth trying today.
