Cambridge's CuspAI closed a $450 million Series B on July 20 at a $2.6 billion valuation, its second headline in a month after June's $400 million raise. The round comes with an "AI Materials Foundry," a coalition of over 45 organisations including Nvidia and Meta that pools compute, lab access and atomistic modelling to speed up materials discovery.
Jeff Bezos and Britain's Sovereign AI Venture Fund both backed the round, a rare instance of a European deep-tech AI company drawing both American venture money and UK state capital into the same cap table. (Bloomberg)
Research & Labs
Google DeepMind maps out its approach to AI-era biosecurity
DeepMind and Isomorphic Labs published a joint bioresilience strategy on July 16, built around a four-step process (threat modelling, evaluations, mitigations, monitoring) developed with biologists and security specialists. It adapts the SynthID watermarking system to screen AI-generated biological sequences at DNA synthesis facilities, and repurposes tools like AlphaEvolve and AlphaGenome to speed up outbreak detection instead of only pathogen design.
The framing is deliberate: AI as pandemic-preparedness infrastructure, not only a source of biological risk. Whether DNA synthesis firms outside DeepMind's 15-plus partnerships adopt the same screening is the open question. (Google DeepMind)
CuspAI's Nvidia partnership points its materials engine at chipmaking
Inside this week's Foundry launch, Nvidia is providing compute and Meta's Fundamental AI Research team is contributing its Universal Model for Atoms, aimed squarely at discovering new semiconductor materials. It's a different application than CuspAI's water-treatment work with Finland's Kemira earlier this year, extending the same MIRA platform into chip supply chains.
For a continent trying to reduce dependence on Taiwanese and American chip materials, an AI-native discovery pipeline is a more direct lever than another fab subsidy. (CNBC)
Deployment & Applications
Munich's Microagi raises Germany's largest-ever seed round to fix robots' "last mile"
Microagi closed a $55 million seed round on July 16, led by Hummingbird, to scale its industrial robotics deployment platform, Atlas. The five-person founding team, drawn from Red Bull Racing, Mercedes-AMG Petronas and the Alan Turing Institute, targets a specific failure mode: robots that work in demos but break down on real production lines.
Rather than build hardware, Microagi buys robots from Chinese manufacturers Unitree and UBTech, then collects footage and data from live factory floors to fine-tune each deployment. It's a bet that the bottleneck in industrial AI isn't models, it's integration. (Sifted)
The European Parliament builds its own AI hub after MEPs go rogue with public chatbots
About 20% of Parliament staff already use AI chatbots daily to draft speeches and amendments, some of it landing as what officials call "AI slop." The Parliament's answer, EPGenAI Hub, gives MEPs sanctioned access to models from OpenAI, Meta, Anthropic and Mistral, hosted internally, with mandatory labelling of AI-assisted work, targeting a September launch.
It's an unusually candid admission from an EU institution: rather than ban staff use of AI, manage it. (Politico Europe)
Governance & Safety
The high-risk AI classification guidelines close for comment July 23
The Commission's draft guidelines on which systems count as "high-risk" under Article 6, covering biometrics, employment, education and credit scoring, finish their consultation window this week after stakeholders won a four-week extension. Final guidelines are due by year-end, but the draft already leans toward classifying more embedded, product-integrated AI systems than industry expected.
That reading is the single biggest determinant of which European AI products need full conformity assessment from December 2027. (European Commission)
The Digital Omnibus is racing the clock to beat its own deadline
The revised AI Act text, agreed by Parliament and Council on May 7 and formally signed on July 8, has to appear in the Official Journal by July 30 for its delayed high-risk deadlines to take effect before the original August 2 cutoff. Publication was expected in the second half of July, which means this week decides whether the postponement Ireland's presidency has been managing actually lands in time.
Miss that window, and companies planning around the softer deadlines face the original, stricter dates instead. (Consilium)
Compute & Infrastructure
Prague-coordinated OpenEuroLLM is ten days from its first real test
OpenEuroLLM, the 20-institution consortium building open multilingual European models, is due to ship its first models, dataset and evaluation code by July 31. The models won't be frontier-scale, but they're the first empirical proof point for a project that has so far delivered mostly infrastructure and partnerships rather than a usable model.
Compute remains the binding constraint. The consortium's Prague coordination and reliance on shared EuroHPC capacity, rather than a single national programme, makes it a genuine test of whether pooled European infrastructure can produce something researchers actually use. (OpenEuroLLM)
AI & Society
O2 Telefónica cuts 1,400 jobs months after promising AI would mean no cuts
O2 Telefónica Germany announced on July 14 it will eliminate around 1,400 roles, 18% of its workforce, under a restructuring it calls the "Future Operating Model." In March, the same CEO had ruled out job cuts, framing AI tools as a way to boost existing staff's efficiency instead.
The reversal followed the loss of 1&1, O2's largest wholesale customer, to Vodafone. AI-linked layoffs in Europe rarely arrive as a clean substitution story; usually they're tangled up with unrelated commercial pressure like this one. (TelcoTitans)
One to watch
GPAI enforcement powers activate August 2
Twelve days from this issue, the AI Office's supervision and enforcement powers over general-purpose AI models become operational, the moment scientists including Yoshua Bengio and Stuart Russell pushed Brussels not to blink on two weeks ago. Combined with transparency obligations taking effect the same day, it's the point where the AI Act stops being a text under negotiation and starts being a law that gets tested against real frontier models.
Whether the AI Office treats OpenAI, Anthropic and Google's submitted documentation as adequate in the following weeks will signal how seriously enforcement is actually meant. (European Commission)