The Pentagon has entered into agreements with Nvidia, Microsoft, and AWS to deploy artificial intelligence capabilities on classified networks. This move reflects the Department of Defense's strategy to diversify its AI vendors, following a dispute with Anthropic over the usage terms of its models. The initiative underscores the importance of data sovereignty and infrastructural control for critical applications.
Companies are reclaiming control over their data to customize AI, balancing ownership with the secure flow of quality information. "AI factories" emerge as a solution for scalability, sustainability, and governance, making data control a strategic imperative for governments and enterprises. Experts from HPE and Oak Ridge National Laboratory discuss how these architectures support secure, scalable AI capabilities, from exascale systems to enterprise deployments.
Anthropic promoted Mythos Preview as a model with exceptional cybersecurity capabilities, restricting its access. However, new research from the UK's AI Security Institute (AISI) reveals that OpenAI's publicly released GPT-5.5 achieves a similar performance level in cyber evaluations. Both models demonstrated advanced abilities in Capture the Flag challenges and complex attack simulations, with GPT-5.5 slightly outperforming Mythos in some tests.
Luce-Org introduced PFlash, a C++/CUDA solution optimizing LLM prefill for long contexts. On an RTX 3090, PFlash achieves a 10x speedup over llama.cpp for quantized models like Qwen3.6-27B at 128K tokens. This innovation significantly improves user experience and efficiency for on-premise deployments, addressing latency and VRAM challenges on consumer hardware.
A recent experiment showcased the effectiveness of DFlash speculative decoding in llama.cpp for running a 35-billion-parameter LLM on a GPU with only 8GB of VRAM. By combining DFlash with MoE expert CPU offload, a token generation speedup of approximately 33-34% was achieved, increasing from 26.8 to around 35.7 tokens/s. This outcome highlights the potential for efficient on-premise deployments.
SAP emphasizes that robust AI governance is crucial for enterprises, transforming statistical estimates into deterministic control and safeguarding profit margins. Adopting agentic AI systems, managing proprietary data, and integrating with existing architectures demand a clear strategy to address operational risks, costs, and data sovereignty requirements, elevating governance to an executive priority.
For the first time in 25 years, over half of the world's countries now fall into the 'difficult' or 'very serious' categories for press freedom, according to Reporters Without Borders. This figure, up from 13.7% in 2002, highlights a profound shift in the global information landscape, influenced by digital infrastructures and the underlying tech platforms.
The announcement of new editions of iconic hardware, such as the Commodore 64C, offers a starting point to reflect on the "return" of established approaches in the technology landscape. In the context of Large Language Models, this translates into a growing focus on on-premise deployment. Companies are increasingly evaluating self-hosted solutions to ensure data sovereignty, optimize TCO, and maintain granular control over AI infrastructure, balancing cloud benefits with specific security and performance needs.
For IT professionals, true fears aren't ghosts, but the pitfalls of deploying complex AI systems. This article explores the challenges and anxieties associated with managing on-premise Large Language Model (LLM) infrastructure, from hardware selection to data sovereignty, highlighting the importance of meticulous planning to mitigate risks and operational costs.
The New York Blood Center, the world's oldest cord blood bank, is developing a stem cell management platform. This initiative raises crucial questions about handling sensitive biological data, the need for robust infrastructure, and the implications for data sovereignty, central aspects for those evaluating on-premise deployments in highly regulated sectors.
Meta has ended its contract with Sama, a Nairobi-based outsourcing company, following reports in February 2026. Sama's workers were tasked with labeling footage from Meta's Ray-Ban smart glasses, which included highly private and sensitive user content, raising serious concerns about privacy management and data sovereignty.
Anthropic's Mythos product, launched just three weeks ago, is sparking an intense debate among state actors. Governments cannot agree on its use or who should regulate it. An unnamed Trump administration official revealed to the Wall Street Journal the White House's opposition to Anthropic's plans to expand access to the system, highlighting growing geopolitical tensions related to new technologies.
The UK's Driver and Vehicle Licensing Agency (DVLA) is facing significant delays, exceeding fourteen weeks, in processing driving license applications that require medical checks. To tackle this backlog and enhance operational efficiency, the agency has implemented new technological solutions. This initiative aims to streamline processes and reduce the prolonged waiting times experienced by applicants, underscoring technology's critical role in resolving public service challenges.
Thomas Reardon, known for creating Internet Explorer and co-founding CTRL-labs, is embarking on a new challenge: developing artificial intelligence capable of "thinking" while consuming just 20 watts. This ambitious goal aims to redefine energy efficiency in the sector, with significant implications for on-premise deployments and edge AI, promising to reduce TCO and enhance data sovereignty.
A recent anecdote from The Register about an unexpected Excel malfunction, where even Oracle ERP was not the cause, offers a starting point for reflection on the complexity of enterprise systems. This incident highlights the importance of robust infrastructure and a deep understanding of interdependencies, crucial aspects for the deployment of Large Language Models on-premise, where control and data sovereignty are paramount.
Over 240 news publishers across nine countries, including The New York Times and CNN, have begun blocking the Internet Archive's Wayback Machine crawlers. The move aims to prevent AI companies from using their content for LLM training. The Archive's director called the decision "collateral damage" in a dispute not directly about them, highlighting growing tensions over data ownership and usage in the AI sector.
Large enterprises are preparing to manage thousands of AI agents by 2028, an exponential increase from today. Without adequate governance, this rapid growth could lead to uncontrolled management and significant operational risks. Gartner's analysis highlights the urgency of robust strategies to maintain control over these autonomous systems.
A development team has achieved significant results in running the Large Language Model Qwen3.6-27B on a single NVIDIA RTX 3090 GPU. The optimization allowed extending the context window up to approximately 218,000 tokens, while ensuring greater stability for tool-agent workloads. This advancement, achieved by resolving a memory management issue, is crucial for self-hosted deployments requiring high context capabilities and reliability.
Meta has ended its contract with Sama, a Kenyan firm providing data annotation services for Ray-Ban Meta's AI systems. The decision follows reports from workers who allegedly viewed private and sensitive footage recorded by the smart glasses. This incident raises crucial questions about privacy management and data sovereignty in AI development pipelines, especially when critical processes are outsourced. Sama stated that the termination affected over a thousand workers.
OpenAI has introduced "Advanced Account Security" for ChatGPT accounts, an opt-in feature that eliminates passwords in favor of hardware keys. This system, requiring two passkeys for authentication, does not allow email recovery or customer support assistance for lost access, elevating security to a level comparable to online banking.