The week Europe chose sides — five stories that defined European AI

On Friday morning, Cohere, the Toronto-based enterprise AI company, announced it will acquire Aleph Alpha, the Heidelberg-based German AI startup that has been the most politically loaded symbol of European AI sovereignty since its 2023 fundraising. The announcement was made in Berlin. Germany’s Digital Minister Karsten Wildberger and Canada’s AI and Digital Innovation Minister Evan Solomon were both in the room.

The structure is a Cohere acquisition dressed in merger language. Cohere’s shareholders receive approximately 90% of the combined company; Aleph Alpha’s shareholders approximately 10%. The combined company retains the Cohere name, with Cohere co-founder and CEO Aidan Gomez leading as CEO. Global headquarters will be in Toronto; European headquarters in Germany, complementing Cohere’s existing Paris office. Cohere was last valued at $7 billion in September 2025 with $240 million in annual recurring revenue. Aleph Alpha was valued at approximately $3 billion in late 2023. The $20 billion combined figure represents a meaningful premium over both.

Schwarz Group — which already led Aleph Alpha’s $500 million 2023 Series B and owns the Stackit sovereign cloud platform — is committing €500M (~$600M) to Cohere’s upcoming Series E, expected to close later in 2026. The investment gives the combined company both capital and sovereign cloud infrastructure to serve EU government customers who demand on-premise data hosting.

The strategic logic is explicit: both Canada and Germany are alarmed by dependence on a handful of American AI providers, anxiety sharpened by trade tensions under President Trump and the broader reassessment of transatlantic tech dependencies. Aleph Alpha brings deep German public sector relationships, regulatory credibility, and a brand that carries symbolic weight in the European sovereignty debate — backed by SAP, Bosch, and the German federal government. Cohere brings global scale, model development capability, and enterprise reach.

On the same day as the Cohere-Aleph Alpha announcement, DeepSeek published preview versions of V4-Pro and V4-Flash on Hugging Face — exactly one year after its R1 model upended global AI markets. V4-Pro is a 1.6 trillion parameter mixture-of-experts model with a 1-million-token context window, making it the largest open-weight model ever released. V4-Flash is the smaller, faster, cheaper variant.

DeepSeek’s own benchmarks show V4-Pro as the strongest open-source model in coding and mathematics, trailing only Google’s closed-source Gemini 3.1-Pro in world knowledge. The company acknowledges a gap of “approximately 3 to 6 months” behind GPT-5.4 and Gemini 3.1-Pro overall. Independent benchmarking has not yet been completed.

On price, V4 undercuts everything. V4-Flash costs $0.14 per million input tokens — below GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5. V4-Pro costs $0.145 per million input tokens — below Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4.

The geopolitically charged element is the chip stack. DeepSeek optimised V4 for Huawei Ascend processors and Cambricon chips, not giving Nvidia or AMD early access — a reversal of standard industry practice. Semiconductor Manufacturing International Corp. shares jumped 10% in Hong Kong trading on Friday as a result.

What is agreed: Both Parliament and Council have converged on fixed postponement dates: December 2, 2027 for stand-alone high-risk AI systems (Annex III), and August 2, 2028 for AI embedded in regulated products (Annex I). A targeted ban on AI nudifiers — systems that generate non-consensual intimate imagery — is backed by both institutions. Both have also reinstated the registration obligation for self-assessed non-high-risk systems.

What is contested: Parliament proposes moving all Annex I-A product categories into Annex I-B and horizontally integrating AI Act requirements into sectoral laws covering machinery, medical devices, toys, and others. The Council does not cover this. The Standing Committee of European Doctors has already called for medical devices to remain under the AI Act’s scope. This is the biggest structural divergence going into Monday’s session.

The synthetic content deadline: Parliament wants watermarking requirements for AI-generated audio, image, video, and text by November 2, 2026. Council proposes February 2, 2027. A six versus three month gap. For generative AI providers already active in Europe, this timeline matters regardless of the Omnibus outcome.

What happens if there is no deal: If the trilogue fails or formal adoption is delayed past August 2, 2026, the original AI Act high-risk obligations apply as written — the very deadline that triggered this entire legislative exercise. Organisations should treat August 2 as the operational deadline until Official Journal publication confirms otherwise.

Nokia CEO Justin Hotard told Reuters on April 23: “The issue today is Europe doesn’t have the infrastructure... If you don’t build that infrastructure, then ultimately the business and the developers will move to where that is. The reality is right now, that’s in China and in the US.” Data centres already consume 3% of EU electricity, with demand expected to rise sharply. Amazon’s European expansion has been actively hindered by power grid connection delays.

EuroHPC JU contracted €320M in sovereign compute in 48 hours: €290M for the IT4LIA AI Factory in Bologna (E4 Computer Engineering + Dell Technologies, NVIDIA GB200 NVL4 liquid-cooled, 160+ exaflops peak AI inference) and €30M for the Mimer AI Factory in Linköping, Sweden (Bull, BullSequana XH3500, already serving 200+ companies). Both are part of a 19-factory, €10B EuroHPC programme running through 2027.

Anthropic, on the same day as Nokia’s warning, posted its first-ever European data centre role — a London-based Transaction Principal paying £225,000–£270,000 to source capacity across Frankfurt, London, Amsterdam, Paris, and Dublin plus the Nordics. This is the first time Anthropic has built a dedicated infrastructure team outside the US, as rival OpenAI has paused its Stargate projects in the UK and Norway.

Eurostat’s 2026 Digitalisation in Europe interactive publication, published Thursday, confirms that 20% of EU enterprises with 10 or more employees used AI technologies in 2025 — up from 13.5% in 2024. This is a 6.5 percentage point increase in twelve months, the fastest single-year jump since tracking began.

Country divergence is the real story. Denmark leads at 42%, followed by Finland (38%) and Sweden (35%). Romania sits at 5.2%, Poland at 8.4%, Bulgaria at 8.6%. The gap between the top and bottom of the EU is 37 percentage points. This is not one AI-adopting Europe; it is two different competitive realities operating under the same regulatory framework.

By use case: 31% of businesses used AI for administrative organisation, 23% for accounting and finance, 21% for production processes, and 20% for ICT security. Language analysis and content generation lead on technology type. The 2026 edition adds new subsections on ICT and the environment — 30% of EU businesses used ICT to reduce energy or material consumption in 2025.