Engineering notes
How we size, build and run NVIDIA GPU infrastructure, written by the people who do it. Hardware comparisons, cluster design, data centers and the details that decide whether a deployment works.
All articles
8 x 96 GB for rendering and transcoding on RTX PRO 6000 Blackwell What fits in 96 GB per GPU for rendering, VFX and transcoding, why a renderer sees 96 GB and not 768 GB, and where 48 GB runs out. GPUs 4 min read Five questions to ask before you use a GPU benchmark to pick a node Five questions to ask before you trust an MLPerf result: division, scenario, power, software and system. Worked example: H200 on Llama 2 70B. GPUs 4 min read 70GB of weights on four GPUs: KV cache room at 80, 96, 141 and 270GB Load a 70B model in FP8 on 80, 96, 141 and 270GB GPUs and see how much memory each leaves for the KV cache, with the sizing arithmetic worked out. GPUs 4 min read 8 x 270GB: what fits on a B300 node What fits on an NVIDIA B300 node: 270GB memory math for 405B and 70B model inference, KV cache sizing, bandwidth ceilings, and where a single GPU stops. GPUs 6 min read 8 x 96GB: what fits on an RTX PRO 6000 Blackwell node What fits on an RTX PRO 6000 Blackwell node: the 96GB memory math for 70B model inference and QLoRA fine tuning, KV cache sizing, and where the limit sits. GPUs 4 min read Foundation model training bottlenecks: memory bandwidth, not GPU count Foundation model training throughput depends on GPU memory capacity, memory bandwidth and interconnect, not GPU count. Here is the data behind that claim. Infrastructure 6 min read What's driving RTX PRO 6000 Blackwell prices higher in 2026 Why the RTX PRO 6000 Blackwell costs more to buy in 2026, how the GDDR7 memory shortage drives it, and what it means for buy versus rent decisions. GPUs 4 min read Why Series A AI companies are moving off AWS GPU before they have to Why Series A AI companies move off AWS GPU before a forcing function decides for them: price spreads, quota walls, no renewal checkpoint and runway math. Infrastructure 3 min read B300 vs. H200: how to decide which node your workload actually needs B300 vs H200: a technical decision framework for dedicated bare metal GPU nodes, covering memory capacity, FP4 throughput, power, cooling and workload fit. GPUs 5 min read What a 36 to 60 month bare metal contract means for your planning What a 36 to 60 month bare metal GPU contract locks in, what stays flexible, and how it changes capacity planning and budget predictability for AI teams. Infrastructure 4 min read 10 questions that reveal the real cost of your cloud GPU setup Ten questions to run against your cloud GPU bills, covering utilization, egress, overruns, storage sprawl, support fees and what you really pay per job. Guides 5 min read B300 for AI model labs: sustained training on dedicated bare metal What sustained pretraining needs from infrastructure, what dedicated NVIDIA B300 bare metal delivers for AI model labs, and how to configure clusters for it. GPUs 5 min read Why long-duration training runs fail on shared cloud GPUs Preemption, noisy neighbors, storage contention and checkpoint egress are built into shared cloud. What long training runs face, and what bare metal fixes. Infrastructure 5 min read 10 questions to ask your GPU infrastructure provider before signing Ten questions to ask a GPU infrastructure provider before you sign: tenancy, performance guarantees, egress, storage, scaling, drivers, SLAs and exit terms. Guides 4 min read Why the B300 is the GPU serious AI teams are planning around right now The NVIDIA HGX B300 is shipping now. What Blackwell Ultra delivers technically, how it compares with H100, and why dedicated bare metal is the way to run it. GPUs 4 min read B300 for production inference: what changes beyond training What the NVIDIA B300's 270GB of memory, FP4 compute and latency profile mean for teams serving LLMs at scale on dedicated bare metal infrastructure. GPUs 4 min read Why ISV certification matters for media GPU infrastructure What ISV certification means for GPU infrastructure, and why it matters for broadcast, VFX and media SaaS pipelines running professional software at scale. Infrastructure 4 min read From H100 to B300: what changes for teams running AI at scale Running H100 in production and evaluating B300? Where each GPU makes sense, what the B300 resolves, and why the infrastructure layer matters for both. GPUs 4 min read B300 on bare metal: what the architecture means for LLM training The NVIDIA B300 brings 3.4x the memory of the H100 and native FP4 compute. What that means for LLM training at scale, and why bare metal makes the difference. GPUs 3 min read ComfyUI and AI video on dedicated bare metal Run ComfyUI, AI video generation and image pipelines on dedicated bare metal GPU servers, with full VRAM per job, no shared tenants and no egress fees. Infrastructure 1 min read LLM inference on dedicated bare metal Run LLM inference on dedicated bare metal GPU servers with full VRAM and bandwidth, FP4 and FP8 support, no shared tenants and no virtualization layer. Infrastructure 1 min read 1Legion at New York Tech Week 2026, meeting AI and media teams 1Legion attended New York Tech Week 2026, June 1 to 7, where CEO David Vargas met AI startups and media tech teams evaluating dedicated bare metal GPU infrastructure. News 1 min read 1Legion and Classover form Ousia Compute, a $50M GPU joint venture In May 2026, 1Legion and Classover formed Ousia Compute, a dedicated GPU joint venture. What was announced, and what a venture of this kind involves. News 2 min read Why LLM inference breaks on shared GPU infrastructure Shared GPU cloud adds latency variance you cannot control. Why dedicated bare metal suits production LLM inference, and when it makes sense to switch. Infrastructure 4 min read 1Legion and StreamVX: dedicated GPU infrastructure for media at scale 1Legion and StreamVX partnered to bring dedicated bare metal GPU infrastructure to broadcast and OTT workflows: predictable performance and no egress fees. News 2 min read Bare metal vs shared GPU infrastructure: what matters in production Bare metal vs shared GPU infrastructure: the real impact on performance, consistency and scalability for production AI, and when dedicated hardware wins. Infrastructure 7 min read The real cost of running RTX 5090 at scale: what AI teams need to know The real cost of running RTX 5090 GPUs at scale: how data movement, idle compute and infrastructure design shape AI training economics for generative AI. GPUs 6 min read 1Legion adds RTX 5090 and RTX PRO 6000 servers on Inflect 1Legion expanded GPU availability on Inflect with RTX 5090 and RTX PRO 6000 servers, offering dedicated bare metal GPUs for AI, VFX and media workloads. News 2 min read NVIDIA RTX PRO 6000 Blackwell servers now available in Canada 1Legion adds NVIDIA RTX PRO 6000 Blackwell servers in Montreal, Canada: dedicated bare metal for AI inference, VFX and media workloads on 36 to 60 month terms. News 2 min read When to choose an H100-class GPU for AI, VFX and media workloads When to move from consumer GPUs like the RTX 5090 to H100-class data center GPUs for frontier AI, multimodal models, distributed training and media pipelines. Guides 6 min read Moving video transcoding to the cloud with the NVIDIA RTX 4000 Ada Why the NVIDIA RTX 4000 Ada suits cloud transcoding, with hardware AV1, high stream density and stable performance for live and VOD media workflows. GPUs 4 min read The real cost of training AI models in the cloud Why AI training in the cloud costs more than the hourly rate suggests: storage, egress, idle GPU time, total cost of ownership and the bare metal alternative. Infrastructure 3 min read The smarter way to scale generative AI workloads How to build scalable, high-performance generative AI infrastructure with bare-metal RTX 5090 GPUs: transparent, predictable, and built for speed. Infrastructure 4 min read 1Legion showcases next-gen AI and HPC cloud infrastructure at IBC 2025 1Legion showcased its GPU and CPU infrastructure for AI, HPC and media workloads, with NVIDIA RTX 5090 and 4090 GPUs, at IBC 2025 in Amsterdam alongside DVEO. News 1 min read 1Legion and Cloudrift launch NVIDIA RTX 5090 for AI in the cloud 1Legion partnered with Cloudrift, a developer-first GPU platform, to make NVIDIA RTX 5090 GPUs on 1Legion infrastructure available to developers on cloudrift.ai. News 1 min read RTX 5090 vs RTX 4090: the next leap in AI GPU performance RTX 5090 vs RTX 4090 for AI: throughput, memory and efficiency gains for LLM inference and training, benchmark results, and how to choose between the two GPUs. GPUs 2 min read How to choose a cloud GPU for deep learning Choosing the right cloud GPU for deep learning is essential for ensuring that your AI initiatives are not only efficient but cost-effective. Guides 3 min read How to run Stable Diffusion models on cloud-based GPU How to run Stable Diffusion models on cloud GPUs: why the cloud suits image generation, how to choose a GPU, install dependencies, load the model and generate images. Guides 3 min read On-premise vs. cloud GPUs: what’s best for you? On-premise vs cloud GPUs for AI: control, cost, security and scalability compared, with questions to guide the choice and when a hybrid approach makes sense. Infrastructure 4 min read Using the NVIDIA RTX 5090 to maximize real-time LLM inference How the NVIDIA RTX 5090 speeds up real-time LLM inference: Blackwell Tensor Cores, low latency and multi-GPU scaling for chatbots, content generation and search. GPUs 5 min read
Tell us what you need to run.
Share the hardware, the number of servers, the term and the start date. You receive a written proposal built on our delivery specification.