Tryfonov - stock.adobe.com Opinion Microsoft's ICC email block reignites European data sovereignty concerns Why Microsoft's rhetoric on protecting European users from US government actions does not quite ring true By Owen Sayers, Secon Solutions Published: 23 May 2025 During his recent visit to Brussels, Microsoft chief Brad Smith committed his company to defending European interests from ‘geopolitical volatility’, including the impact of potential US administration interventions. Suggesting that Microsoft is ‘critically dependent on sustaining the trust of customers, countries, and government across Europe’, anyone leaving his session with EU leaders should have reasonably felt buoyed up by his words; but might also have sensibly awaited evidence of the commitments being applied in practice before relying upon them. If so, the news that the International Criminal Court (ICC) chief prosecutor and his staff have had their Microsoft email and services cancelled in direct response to US government sanctions might come as an unwelcome reality check. According to media reports, ICC chief prosecutor Karim Khan had his Microsoft email and other services suspended after the US applied sanctions in February to all ICC staff in response to their investigations into key Israeli politicians. The circumstances of the situation that gave rise to those sanctions are outside the scope of this article, and largely irrelevant to the problems these service suspensions indicate, however. Regardless of the ‘why’, what the service suspensions demonstrate is that Microsoft has the means (and when it comes down to it also possess the will) to do the US government’s bidding and disrupt services to any party deemed to be unacceptable. This is almost exactly contrary to the assurances Brad Smith so very recently gave. The disconnection of prosecutor Khan is a mouse-click heard around the world, and will undoubtedly give anyone using or currently considering the adoption of Microsoft cloud technologies pause for thought. By disconnecting the ICC staff in this way, Microsoft has done themselves some serious damage, and how much may take some time yet to become clear. Immediately after the disconnection became public, the Dutch government and public bodies are reported to have accelerated their examination of non-Microsoft and EU-located alternative services. Meanwhile, several suppliers have indicated an uptick in requests for backup of key data to protect against possible Microsoft disconnections. Press coverage in Germany suggests these concerns are rippling out to them also, whilst the Nordics and France have long made clear that they see a future that is distinctly less Azure in colour. The likelihood or otherwise of further disconnections is unclear, and for most users it should be considered very unlikely that Microsoft will start switching off services for no good reason. With 25% of Microsoft’s global revenues coming from European customers, it is unlikely to act rashly to damage that market, and can generally be counted on to be sensible and not commit commercial suicide – so most customers should not be worried. Nonetheless much of the damage to the confidence of public sector bodies might well have already been done. Governments like to be in control of their own destiny and that extends to digital services and data. When a key supplier they have relied upon for many years shows themselves to be subject to the whims and foibles of a foreign government – friendly or otherwise – most public sector buyers intuitively know it’s time to find an alternative provider “just-in-case”. Having a plan B option is just common sense. The big problem for Microsoft is that in the IT sector “just-in-case” or plan B options, often become strategic plan A directions of travel. And a trickle of departures can quite soon become a flood. Governments are herd animals – when one turns they all tend to follow. I’m not by any measure suggesting we are going to see an overnight exodus. Even if that was technically feasible (which it isn’t in most cases), these organisations are a bit concerned, not panicked. However, these previously affirmed Microsoft user groups are now openly talking about the need for alternatives to the Redmond cloud provider, and that should have Microsoft worried. Concerns that US hyperscalers might be subjected to pressure from US authorities to disclose information have existed for some time but have been broadly assuaged by repeated promises and commitments from Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft that they would resist such requests and protect their customers. When it has come to the acid test, however, many clearly feel that Microsoft has failed, and that instead of protecting the ICC as a key pillar of the global legal community, instead acted as an instrument of US policy. To restore his own email access, prosecutor Khan reportedly turned to Proton Mail, the Swiss end-to-end encrypted mail service beloved of whistleblowers and other digital refugees. Proton Mail operate under its own constraints and obligations to disclose information to the Swiss government on demand, but this is limited to IP address info, rather than email payloads, which it is generally accepted they cannot access. In doing so it’s likely that Mr Khan has had to forgo some user functionality and ease of use – but he may feel that’s a small price to pay to protect his office and role from US government influence. That might be a choice others have to make in the months and years to come, since regardless of their choice of cloud provider, the lesson here is that we cannot always trust them to rigorously and strongly protect our data or our services, despite what they may say, or how often they do so. In this case, Microsoft’s actions sadly speak a lot louder than Mr Smith’s words. Read more about Microsoft Microsoft’s hold on government IT is under scrutiny, following a disclosure to a Scottish policing body that saw the software giant advise that it cannot guarantee data sovereignty in its cloud-based Microsoft 365 suite Documents show Microsoft’s lawyers admitted to Scottish policing bodies that the company cannot guarantee sensitive law enforcement data will remain in the UK, despite long-standing public claims to the contrary In The Current Issue: UK critical systems at risk from ‘digital divide’ created by AI threats UK at risk of Russian cyber and physical attacks as Ukraine seeks peace deal Standard Chartered grounds AI ambitions in data governance Download Current Issue SAP Sapphire 2025: Developers take centre stage as AI integration deepens – CW Developer Network Microsoft entices developers to build more Windows AI apps – Cliff Saran's Enterprise blog View All Blogs

Meta Platforms has decided to delay the public release of its most ambitious artificial intelligence model yet — Llama 4 Behemoth. Initially expected to debut at Meta’s first-ever AI developer conference in April, the model’s launch was pushed to June and is now delayed until fall or possibly even later. Engineers at Meta are grappling with whether Behemoth delivers enough of a leap in performance to justify a public rollout, The Wall Street Journal reported. Internally, the sentiment is split — some feel the improvements over earlier versions are incremental at best. The delay doesn’t just affect Meta’s timeline. It’s a reminder to the entire AI industry that building the most powerful model isn’t just about parameter count—it’s about usefulness, efficiency, and real-world performance. Sanchit Vir Gogia, chief analyst and CEO at Greyhound Research, interprets this not as a standalone setback but as “a reflection of a broader shift: from brute-force scaling to controlled, adaptable AI models.” He said that while Meta has not officially disclosed a reason for the delay, the reported mention of “capacity constraints” points to larger pressures around infrastructure, usability, and practical deployment. What’s inside Llama 4 Behemoth? Behemoth was never intended to be just another model in Meta’s Llama family. It’s intended to be the crown jewel of the Llama 4 series, designed as a “teacher model” for training smaller, more nimble versions like Llama Scout and Maverick. Meta had previously touted it as “one of the smartest LLMs in the world.” Technically, Behemoth is built on a Mixture-of-Experts (MoE) architecture, designed to optimize both power and efficiency. It is said to have a total of 2 trillion parameters, with 288 billion active at any given inference — a staggering scale, even by today’s AI standards. What made Behemoth especially interesting was its use of iRoPE (interleaved Rotary Position Embedding), an architectural choice that allows the model to handle extremely long context windows—up to 10 million tokens. That means it could, in theory, retain far more contextual information during a conversation or data task than most current models can manage. But theory doesn’t always play out smoothly in practice. “Meta’s Behemoth delay aligns with a market that is actively shifting from scale-first strategies to deployment-first priorities,” Gogia added. “Controlled Open LLMs and SLMs are central to this reorientation — and to what we believe is the future of trustworthy enterprise AI.” How Behemoth stacks up against the competition When Behemoth was first previewed in April, it was positioned as Meta’s answer to the dominance of models like OpenAI’s GPT-4.5, Anthropic’s Claude 3.5/3.7, and Google’s Gemini 1.5/2.5 series. Each of those models has made strides in different areas. OpenAI’s GPT-4 Turbo remains strong in reasoning and code generation. Claude 3.5 Sonnet is gaining attention for its efficiency and balance between performance and cost. Gemini Pro 1.5, from Google, excels in multimodal tasks and integration with enterprise tools. Behemoth, in contrast, showed strong results in STEM benchmarks and long-context tasks but has yet to demonstrate a clear superiority across commercial and enterprise-grade benchmarks. That ambiguity is believed to have contributed to Meta’s hesitation in launching the model publicly. Gogia noted that the situation “reignites a vital industry dialogue: is bigger still better?” Increasingly, enterprise buyers are leaning toward SLMs (Small Language Models) and Controlled Open LLMs, which offer better governance, easier integration, and clearer ROI compared to gargantuan foundation models that demand complex infrastructure and longer implementation cycles. A telling sign for the AI industry This delay speaks volumes about where the AI industry is heading. For much of 2023 and 2024, the narrative was about who could build the largest model. But as model sizes ballooned, the return on added parameters began to flatten out. AI experts and practitioners now acknowledge that smarter architectural design, domain specificity, and deployment efficiency are fast becoming the new metrics of success. Meta’s experience with smaller models like Scout and Maverick reinforces this trend—many users have found them to be more practical and easier to fine-tune for specific use cases. There’s also a financial and sustainability angle. Training and running ultra-large models like Behemoth requires immense computing resources, energy, and fine-grained optimization. Even for Meta, this scale introduces operational trade-offs, including cost, latency, and reliability concerns. Why enterprises should pay attention For enterprise IT and innovation leaders, the delay isn’t just about Meta—it reflects a more fundamental decision point around AI adoption. Enterprises are moving away from chasing the biggest models in favor of those that offer tighter control, compliance readiness, and explainability. Gogia pointed out that “usability, governance, and real-world readiness” are becoming central filters in AI procurement, especially in regulated sectors like finance, healthcare, and government. The delay of Behemoth may accelerate the adoption of open-weight, deployment-friendly models such as Llama 4 Scout, or even third-party solutions that are optimized for enterprise workflows. The choice now isn’t about raw performance alone—it’s about aligning AI capabilities with specific business goals. What lies ahead Meta’s delay doesn’t suggest failure — it’s a strategic pause. If anything, it shows the company’s willingness to prioritize stability and impact over hype. Behemoth still has the potential to become a powerful tool, but only if it proves itself in the areas that matter most: performance consistency, scalability, and enterprise integration. “This doesn’t negate the value of scale, but it elevates a new set of criteria that enterprises now care about deeply,” Gogia stated. In the coming months, as Meta refines Behemoth and the industry moves deeper into deployment-era AI, one thing is clear: we are moving beyond the age of AI spectacle into an age of applied, responsible intelligence.