Lablup Inc.

Two days at The AI Summit London 2026, now done. This was the first time Lablup and Boston Limited shared a stand, pairing our Backend.AI with Boston's high-performance computing so a team can run its own AI infrastructure end to end. Thank you to everyone who came to say hello. It will not be the last time we share a stand.

I’m pleased to be speaking at the @AI for Good Global Summit 2026 in Geneva from 7-10 July. As the leading United Nations’ platform on #AI, the Summit brings together global leaders, innovators and experts to explore how AI is being applied across sectors and societies. The programme is now live! Discover what’s ahead across 150+ sessions, 50+ workshops and four days of AI in action. I will participate in “The future of AI native communication networks” workshop to share and disucss AI infrastructure challenges and experiences from Korea's sovereign AI foundation project and Lablup’s technical expertises. Learn more: https://lnkd.in/gkAyVpTQ #AIforGood

𝗟𝗮𝗯𝗹𝘂𝗽 𝗶𝘀 𝗷𝗼𝗶𝗻𝗶𝗻𝗴 𝗡𝗲𝘅𝘁𝗥𝗶𝘀𝗲 𝟮𝟬𝟮𝟲, 𝗦𝗲𝗼𝘂𝗹 (𝗝𝘂𝗻𝗲 𝟭𝟴-𝟭𝟵, 𝗖𝗢𝗘𝗫). We're exhibiting in the NVIDIA pavilion as an NVIDIA Inception member. What we build is why we're there. Backend.AI, our AI infrastructure operating platform, helps teams get more out of their NVIDIA stack. We operate NVIDIA GPU clusters across the full AI lifecycle. The same platform offers NGC™ catalog images and full support for NVIDIA NIM™ containers, letting teams seamlessly run, manage, and scale inference microservices while fully automating resource optimization. Patented container-level GPU virtualization pushes utilization higher, so a given cluster does more work. At NextRise (넥스트라이즈) 2026, we'll show what that looks like in practice. Come find us in the NVIDIA pavilion.

𝐃𝐚𝐲 𝟏 𝐨𝐟 𝐓𝐡𝐞 𝐀𝐈 𝐒𝐮𝐦𝐦𝐢𝐭 𝐋𝐨𝐧𝐝𝐨𝐧 𝟐𝟎𝟐𝟔 𝐢𝐬 𝐮𝐧𝐝𝐞𝐫𝐰𝐚𝐲. Find us at Booth 124, where we've joined Boston Limited for our first shared stand. Working through GPU orchestration, hardware that won't lock you in, or keeping your AI stack under your own control? Come talk to the team. We're here today and tomorrow.

Tomorrow, The AI Summit London opens at Tobacco Dock, and we are exhibiting with our partner Boston Limited for the first time. As the headline AI event of London Tech Week, the Summit brings together thousands of AI leaders from across industries, all focused on turning AI into real business results. This is precisely the audience this partnership is built for. Boston Limited contributes three decades of high-performance computing across 40+ countries, while Lablup contributes Backend.AI, allowing teams to share the same machines and keep training and inference running. Brought together, we form a single vertically integrated platform from the compute hardware to the orchestration software, giving your team AI infrastructure it can actually operate. This combination is why we are in London. See it running live at Booth 124 on 10-11 June.

KV cache offloading moves an inactive session's KV cache off the GPU. This frees memory for more concurrent work. When the session comes back, its cache can be reloaded instead of recomputed from scratch. Reloading the cache can be far cheaper than recomputing a long prefill. However, when the reload costs more than the recompute it would have saved, the offload turns into a regression. We mapped out where the line falls: when offloading pays off, when it backfires, and how to find the crossover for a given setup. If you run long-context LLMs in production, the full note is in the comments.

𝗢𝘂𝗿 𝗳𝗼𝘂𝗿 𝗱𝗮𝘆𝘀 𝗮𝘁 𝗜𝗻𝗻𝗼𝗩𝗘𝗫 𝟮𝟬𝟮𝟲 𝗶𝗻 𝗧𝗮𝗶𝗽𝗲𝗶 𝗰𝗹𝗼𝘀𝗲 𝘁𝗼𝗱𝗮𝘆, 𝗵𝗼𝘀𝘁𝗲𝗱 𝗮𝘁 𝗡𝗩𝗜𝗗𝗜𝗔 𝗜𝗻𝗰𝗲𝗽𝘁𝗶𝗼𝗻 𝗦𝘁𝗮𝗿𝘁𝘂𝗽𝘀 𝗣𝗮𝘃𝗶𝗹𝗶𝗼𝗻. Being invited to bring Backend.AI and AI:GO into this pavilion placed our work among the startups and partners shaping where AI goes next. This made the week meaningful for us. More than anything, the conversations across the floor are what we will carry home. They are why we came to Taipei in the first place. Today is our final day. If you are at InnoVEX, come find us at the NVIDIA Inception Pavilion before it closes.

🤝𝗟𝗮𝗯𝗹𝘂𝗽 𝗼𝗻 𝗡𝗩𝗜𝗗𝗜𝗔'𝘀 𝗸𝗲𝘆𝗻𝗼𝘁𝗲 𝘀𝗰𝗿𝗲𝗲𝗻 𝗮𝘁 𝗚𝗧𝗖 𝗧𝗮𝗶𝗽𝗲𝗶 𝟮𝟬𝟮𝟲 During Jensen Huang's GTC Taipei 2026 keynote, our logo appeared on NVIDIA's Ecosystem Partner wall, among the companies building the future of AI on NVIDIA. Here is what we bring to it. Backend.AI integrates deeply with the NVIDIA stack from container-level GPU virtualization to the world's first cluster-scale GPUDirect Storage. One site operates 2,000+ GPUs with it, more than 120 organizations run on it, and a sovereign foundation model was trained on it. This is the track record that earned our place in NVIDIA's Inception Startups Pavilion this week. And today is Day 1 of InnoVEX 2026 in Taipei. 📍 Find us in the NVIDIA Inception Startups Pavilion, open through June 5.

𝗜𝗻𝗻𝗼𝗩𝗘𝗫 𝟮𝟬𝟮𝟲 𝘀𝘁𝗮𝗿𝘁𝘀 𝘁𝗼𝗺𝗼𝗿𝗿𝗼𝘄. Find us at NVIDIA's Inception Startups Pavilion in Taipei through June 5, during COMPUTEX week. Come meet our team and talk about Backend.AI and AI:GO.