AI & Marketing Radar

For a year that was supposed to bring the EU AI Act's toughest rules into force, 2026 delivered something closer to a reprieve. On 27 July 2026, the so-called "Digital Omnibus" package entered into force, and it pushed back exactly the deadlines that had been keeping compliance teams up at night — though not as many of them as some headlines suggested. The obligations for high-risk AI systems under Annex III — think recruitment tools, credit-scoring models, law-enforcement systems, education and border-control AI — were due to bite on 2 August 2026. The Omnibus pushes that to 2 December 2027, a delay of roughly sixteen months. AI embedded in regulated physical products (Annex I: machinery, medical devices, vehicles, lifts, toys) gets a similar reprieve, moving from 2 August 2027 to 2 August 2028. The deadline for EU member states to stand up regulatory sandboxes — supervised environments where companies can test AI systems under regulator oversight — also slipped a year, from August

The story in 2026 isn't that brands use AI now — that ship sailed years ago. It's that the bottleneck moved. For most of the last few years, the constraint on creative output was concepting: getting to a good idea. Now the tools that actually shipped this year target the step after that — turning an approved concept into finished, on-brand assets at speed, and increasingly without a specialist at every stage of the pipeline. Figma's Config 2025 additions landed properly in production workflows this year: Figma Make turns a prompt into a working prototype, Figma Sites publishes responsive pages straight from the design file, and Figma Buzz is the one that matters most for brand teams specifically — it lets a designer lock the brand elements (colors, type, logo lockups, spacing rules) into a template, then hands the variant-filling to marketers who aren't designers at

Germany, Austria and Switzerland are pouring real money into sovereign AI infrastructure this year, adoption is running ahead of the EU average across all three countries, and the region's own digital minister is publicly saying it isn't translating into results fast enough. All three of those things are true at once, and that tension is the actual state of AI in DACH in 2026. Germany's first "AI factory," the Industrial AI Cloud in Munich, went live on February 4, 2026. Built by Deutsche Telekom with NVIDIA and data-center operator Polarise, with SAP and Siemens as software and industrial partners, the Tucherpark facility occupies roughly 10,700 square meters and runs close to 10,000 NVIDIA Blackwell GPUs, delivering up to 0.5 ExaFLOPS of compute — capacity the announcement claims could theoretically serve all 450 million EU citizens simultaneously if needed. It's already running at over a third of capacity with real customers, including robotics and simulation companies, and hosts

Almost every company now says AI literacy matters. Far fewer have actually built something that deserves the name. The gap between those two facts is the real story of workplace AI training in 2026, and it shows up clearly once you look past the headline numbers. IDC estimates that AI skills shortages could cost the global economy up to $5.5 trillion by 2026, through delayed projects, quality problems, lost revenue and slower competitiveness. The workforce-readiness numbers behind that estimate are stark: more than 90% of global enterprises expect a critical AI skills shortage, only 35% of leaders believe they've adequately prepared their employees for AI-driven roles, and just one in three employees received any AI training in the past year at all. Half of employers say they're struggling to fill AI-related positions. Asked what's actually standing in the way, IT and business leaders point to a lack of qualified talent (46%), data privacy concerns (43%), poor data quality (40%), high

Every major data platform shipped some version of the same idea this year: a plain-language layer on top of your warehouse or lakehouse that lets someone who doesn't write SQL ask a real question and get a real answer. Snowflake has Cortex Analyst, Databricks has Genie (now "Genie One"), and Microsoft has Copilot built into Fabric and Power BI. They're solving the same underlying problem, but the bets they've placed on how are different enough that "which one is best" is the wrong question. The better one is which bet matches how your organization actually works. Cortex Analyst forces every natural-language question through a predefined semantic layer (Semantic Views) rather than letting an LLM guess at joins across raw tables, and it inherits Snowflake's row-level security automatically, so a question can't leak data a person isn't permitted to see. It's the pick for

If you want a single data point for how seriously the market now takes "vibecoding" — building software by describing what you want in plain language and letting an AI agent write and run the code — consider this: in August 2026, Cursor became a wholly owned subsidiary of SpaceX. The full sequence is worth walking through because it happened fast. Cursor's valuation had already climbed from $400 million in mid-2024 to $29.3 billion by a November 2025 funding round, on the back of annual recurring revenue that went from $100 million in January 2025 to over $1 billion by November that year. Then, in April 2026, SpaceX announced it held the option to acquire Cursor outright for $60 billion, or pay a $10 billion fee if the deal didn't happen. Four days after SpaceX completed what's been called the biggest IPO in history on June 12, 2026 — pricing at $135 a share and surging past $200 in pre-market trading within days, adding nearly $1 trillion to its own valuation — SpaceX exercised the

77% of small businesses now use AI regularly, up from 48% in mid-2024 — that's the headline from a QuickBooks/Intuit survey of over 34,000 small and midsize business owners across the US, Canada, the UK and Australia, built with economists at the University of Chicago and cross-checked against anonymized data from 5.3 million QuickBooks businesses in those same four markets. 41% say it's increased revenue, 74% say it's improved productivity. Those are genuinely strong numbers. They're also, per an independent look at the same survey, mostly a feeling rather than a measurement: more than half of respondents described a general sense that things were better, and fewer than half were tracking any specific metric at all. Productivity claims relied on self-reporting rather than time studies; revenue increases were attributed to AI by correlation, not by anything that ruled out a new client, a seasonal bump, or a lucky quarter. That's worth sitting with before adding anything new to your

In November 2025, IBM said something bold in public: it expected the industry to reach a verified quantum advantage — a task a quantum computer does better than any classical computer, checkable well enough that skeptics can't wave it away — sometime in 2026. Eight months later, that happened. Not everything in quantum computing moved that fast this year. But that one claim is worth taking seriously, and understanding why it's different from the "advantage" claims that came before it is the most useful thing to take from the state of the field in 2026. On July 30, 2026, IBM and the Finland-based quantum software company Algorithmiq announced they'd used an IBM Quantum Heron processor to simulate heterogeneous quantum matter — materials with irregular, localized variations in how information, energy and particles move through them, the kind of structure that shows up in real catalysts and battery electrolytes, not textbook toy systems. The result itself isn't the headline. The headline

For the week of February 9–15, 2026, AI models built in China processed more enterprise API traffic than models built in the US, for the first time — 4.12 trillion tokens routed through OpenRouter, a platform enterprises use to route production AI workloads across providers. By July, Chinese-origin models held as much as 46% of that US enterprise token volume, up from just 4.5% at the start of 2025. DeepSeek alone accounted for 17.6% of all tokens on the platform — more than Anthropic's 14.8%. That's not a research curiosity anymore. That's a chunk of the actual enterprise AI market, and DeepSeek is the single biggest reason it exists. DeepSeek's headline 2026 release was V4, launched April 24 as two models rather than one. V4-Pro is the flagship: a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters per query, built for coding and complex reasoning. V4-Flash is the economy version — 284 billion total parameters, 13 billion active — built for speed and

Four days. That's roughly how long it took Seedance 2.0 — ByteDance's AI video model, released February 12, 2026 — to go from a genuinely impressive technical launch to a cease-and-desist letter from Disney. Both things are true about Seedance at once: it's one of the best AI video models available today, and it's also become the clearest cautionary tale yet of what happens when a generative model this capable meets a training-data question nobody's fully answered. Seedance launched as version 1.0 in June 2025, ByteDance's answer to OpenAI's Sora and Google's Veo. Version 2.0 arrived February 12, 2026, and a further update, Seedance 2.5, was announced June 23 at ByteDance's Volcano Engine FORCE conference — currently in enterprise beta with general availability targeted for July 2026. Each version has pushed the same direction: longer, cleaner single-pass generation rather than stitched-together clips. Seedance 2.5's headline feature is native 30-second video in one continuous

Amazon Bedrock

Azure AI Foundry

Vertex AI

The useful question about AWS Bedrock, Azure AI Foundry and Google Vertex AI isn't which one has the strongest model this quarter. Model quality changes fast, the three catalogues overlap heavily, and any serious application usually ends up calling more than one model anyway. The real decision is which control plane best fits the systems that already own your identity, data, networking, observability and operational responsibility — because moving a prompt between providers is trivial, and recreating everything around that prompt is not. Treat these three products as application platforms, not interchangeable model endpoints. Moving a prompt from one to another takes minutes. Recreating the IAM or Entra permission boundaries, the retrieval pipeline, tool authorization, audit trails, private connectivity, evaluation jobs and deployment workflows built around that prompt does not. AWS Bedrock fits most naturally when an application already lives inside AWS and its operational model

On September 8, 2026, the French AI lab Mistral closed a €3 billion Series D round — by its own description, and by most independent accounts, the largest equity fundraising round any European technology company has ever completed — at a post-money valuation north of €21 billion. That's a striking number on its own. What makes it the story of Mistral's 2026, rather than just a big cheque, is what the round is actually buying: not a better chatbot, but infrastructure, compute, and a specific bet that "built and hosted in Europe" is about to become a real commercial advantage rather than a talking point. Mistral's current model lineup dates to December 2, 2025, when it released the Mistral 3 family: Large 3, a 675-billion-parameter mixture-of-experts flagship with 41 billion active parameters per query and a 256,000-token context window, alongside the Ministral 3 line — nine smaller models across three sizes (3B, 8B, 14B parameters), each offered in base, instruction-tuned and reasoning

Ask a general-purpose LLM to write SQL against a raw database schema and it will often guess at joins or produce plausible-looking queries that quietly return the wrong numbers — a failure mode that's dangerous precisely because it doesn't look like a failure. The real advantage of Snowflake Cortex Analyst is that it's built specifically to close that gap. It pairs agentic AI with a warehouse-native semantic layer, called Semantic Views, that forces natural language to route through a predefined business definition of dimensions, metrics and table relationships rather than letting the model improvise against raw tables. Snowflake reports 90%+ query accuracy from that approach, and — just as importantly — the data never has to leave Snowflake's own security and governance perimeter to get there. Google's Looker/Gemini stack and Microsoft's Fabric/Power BI Copilot both offer genuinely capable AI tooling, but their semantic understanding is largely locked inside their respective BI

Nordic organisations have the infrastructure for AI. What they're missing, according to Deloitte's State of AI in the Nordics 2026 report — a survey of 170 senior executives across Denmark, Finland, Norway and Sweden, benchmarked against a global comparison set of 3,235 respondents — is nearly everything built around that infrastructure: the strategy, the talent, the clear ownership of whether any of it actually paid off. That gap, more than any single adoption number, is the real state of AI in the region this year. 55% of Nordic organisations report feeling highly prepared on technical infrastructure for AI — a solid number, and not the problem. Strategic preparedness tells a different story: it sits at 43%, down sharply from 61% the year before, a genuine reversal rather than a plateau. Talent preparedness is weaker still, at just 14% — the lowest of the three readiness measures Deloitte tracks, and the one most likely to actually gate how much of that infrastructure gets used well.

On August 19, 2026, Unitree Robotics listed on Shanghai's STAR Market and closed its first day of trading up 460%, for an implied valuation around $50 billion. It wasn't a chatbot company or a chip company. It was a robot maker, best known for a four-legged robot you can order online for roughly the price of a used car. If "physical AI" needed a single moment to mark its arrival as a real commercial category rather than a trade-show spectacle, this was it. Deloitte defines physical AI as systems that autonomously perceive, understand, reason about and interact with the physical world in real time, and NVIDIA has built its post-ChatGPT strategy around exactly this shift, releasing its full Cosmos and Isaac GR00T stack at CES 2026. But the hardware actually shipping and being deployed at scale right now looks nothing like the humanoids on stage. A September 2026 market report shows quadrupeds splitting into clear tiers: Unitree's sub-$2,000 Go2 against Boston Dynamics' Spot at $74,500, with ANYbotics, Deep Robotics and Boston Dynamics collectively running thousands of units in active field deployment. Humanoids, meanwhile, remain mostly pre-order and pilot-stage — Atlas has no public price tag yet, and even Unitree's own 2025 humanoid shipments totalled a bit over 5,000 units worldwide. For anyone advising on physical AI today, the practical starting point is the quadruped already on patrol, not the humanoid still on the keynote stage.