Latest briefing
July 7, 2026 · 5 actus (site) · 5 actus (base)
Imagine a Chinese car maker launching a new electric car: 600 km of range, half the price of a Tesla, and the manufacturing recipe released open source - meaning anyone can copy it, improve it, sell it. That is exactly what Tencent (the 'Chinese Google', owner of WeChat) did on J...
π₯ Top story
01
Tencent launches a 'home-made' AI model that is free, open source, as good as a model four times its size - and 45% cheaper than Anthropic
Imagine a Chinese car maker launching a new electric car: 600 km of range, half the price of a Tesla, and the manufacturing recipe released open source - meaning anyone can copy it, improve it, sell it. That is exactly what Tencent (the 'Chinese Google', owner of WeChat) did on July 6 with Hunyuan Hy3, its new AI model. The model is built to power agents inside a company - AI assistants that run actual tasks: answering customer support, helping to code, analysing a spreadsheet. Tencent has already deployed it across 50 of its own products, including WeChat and QQ Browser, and released it to the world under Apache 2.0 (the most permissive licence). The usage cost is a few cents per million words read, versus 25 cents at Anthropic. For a European company wanting to diversify its AI providers without depending only on Americans, this is a serious additional option.
02
xAI launches a 'no code' voice agent builder: a phone that answers your customers on its own, for 3 cents a minute
Imagine you want a virtual receptionist who calls your customers, answers their questions about your product, and books an appointment. Until now, you either had to hire someone or pay a technical provider a lot of money. On July 1, xAI (Elon Musk's house behind the Grok chatbot) released Grok Voice Agent Builder: an online tool that lets you describe what you want in plain language, and it gives you a phone that answers on its own in less than two minutes. No code needed. You upload your documents, write a few instructions in English, and your phone agent is live. Cost: about 3 cents per minute of conversation - cheaper than competitors like Vapi or Retell, which charge 3 to 5 times more. It works in 25 languages, and you can even clone your own voice from two minutes of audio. For a small business that wants to test a voice product without investing tens of thousands of euros in tech, this is a realistic entry point.
03
Anthropic, Amazon, Microsoft and Google want to create a common 'severity scale' for AI jailbreaks - just like for computer viruses
When a new computer virus is identified, security experts have long used a shared severity scale - from a simple cold (level 1) to a global pandemic (level 5). It lets everyone speak the same language and react proportionately. On July 2, 2026, Anthropic (the OpenAI rival behind Claude) proposed, with Amazon, Microsoft and Google, a similar framework - but for 'AI jailbreaks': the techniques by which someone convinces an AI to do something it should not. The framework is called CJS (Cyber Jailbreak Severity) and goes from level 0 - a harmless trick - to level 4 - a flaw that gives an attacker a power they did not have before with any tool. Their worked example: if an AI had revealed the Log4Shell flaw (a real cyber disaster) before its public discovery, that would have been level 4. Today, the same trick would be worth zero, because everyone knows it. For a company deploying agents in production, this is a common scoring grid to decide when an incident is serious and when it is not - and to anticipate upcoming legal obligations.
04
SoftBank (owner of T-Mobile) prepares a 10-gigawatt 'mega data center' in Ohio - the equivalent of 10 nuclear plants - to run tomorrow's AI agents
Imagine a field the size of a medium-sized city, covered in thousands of specialised computers, burning the electricity equivalent of ten nuclear plants. On July 2, SoftBank (the Japanese group that owns T-Mobile and is one of OpenAI's largest shareholders) announced a 10-gigawatt (10 GW) project in Ohio, on a site called 'Portsmouth'. This is a new class of dedicated 'AI clouds' - places where startups and big companies will be able to rent computing power to run their agents, instead of building their own data centers. The service begins in 2028 with 800 MW, and will ramp progressively to 10 GW. Direct competition: CoreWeave, Nebius, AWS, Google Cloud. Why this matters for ordinary people: when available computing power explodes, the price of each request sent to an agent drops - exactly like when electricity gets cheaper, the whole economy benefits. For a French startup that wants to deploy an agent at very large scale without going broke, the 2028-2030 horizon just became more reachable.
05
An open-source American start-up reserves $6.3 billion of 'compute credit' at SpaceX - to build its own 'open' AI model against OpenAI and Anthropic
To train a cutting-edge AI model, you need thousands of high-end graphics cards running for several months. It has become so expensive that only a handful of companies can afford it - OpenAI, Anthropic, Google. On June 22, a little-known start-up, Reflection AI, signed a blank cheque for $6.3 billion with SpaceX (Elon Musk's rocket company) to reserve computing power. In return, SpaceX supplies it with the new Nvidia GB300 processors at its huge Memphis data center. The goal: build an 'open source' model (so free and modifiable, unlike GPT or Claude) that rivals the Americans. For a European company that today depends on paid APIs to run its agents, this is one more possibility, by end of 2026 or early 2027, of having a high-performing, open, probably cheaper model. The flip side: Reflection, like its SpaceX neighbours (Anthropic, Google, Cursor), depends on the same data center - so an outage or natural disaster in Memphis could hit several major players at the same time.
π‘ To watch
Will Tencent's Chinese model take hold in Europe?
Hy3 is open source (free to reuse) and already battle-tested by 50 internal Tencent products. The question is whether European publishers - especially French and German ones, keen on diversification - will adopt it instead of staying on OpenAI or Anthropic. To watch over the next 4 to 8 weeks.
Will xAI Voice Agent Builder spark a voice AI price war?
xAI charges 3 cents a minute, voice included. Its direct competitors (Vapi, Retell, Bland) charge 3 to 5 times more. If customers migrate, these players will have to align or pivot to niches (medical, legal). The agentic voice market is young: a price cut could open it up to many more users.
Will other major labs sign the jailbreak severity scale?
Anthropic, Amazon, Microsoft and Google signed on July 2. OpenAI, Meta, Mistral and DeepSeek have not yet published their adherence. If the standard becomes unanimous, this is a new form of self-regulation that precedes government regulation. To watch in the next 30 days: official statements from OpenAI and Meta.
Is a 10 GW data center credible or just marketing?
SoftBank targets 10 GW in Ohio, but the first 800 MW block only comes online in 2028. The remaining 9.2 GW depend on energy partnerships (natural gas) and permits. The market has already burned itself twice on this kind of announcement - competitors CoreWeave and Nebius dropped on July 1 on another announcement. Caution on the roadmap.
π Trend
July 7, 2026 shows that the race for AI agents is now being fought on four grounds in parallel. First ground: price - Tencent sells the use of its model at 14 cents of a euro for one million words, when its American rivals charge 3 to 5 times more. Second ground: access - xAI offers a voice agent builder for 3 cents a minute, no code needed, which opens the door to millions of SMEs who could not afford a tech studio. Third ground: trust - four major players (Anthropic, Amazon, Microsoft, Google) co-sign a common scale to assess the severity of AI jailbreaks, like medicine has its pandemic scale. Fourth ground: power - a start-up (Reflection AI) reserves $6.3 billion of compute at SpaceX to build an open source model to rival the Americans. Three lessons emerge: (1) the marginal cost of a conversational agent is falling sharply, which makes consumer-grade products viable; (2) the quality of an agent depends as much on the tools you give it as on its model, and 'no-code' platforms are taking over; (3) source diversification - Chinese publishers, American open source, Japanese infrastructure - is becoming a strategic stake for European companies.