Radiant

Radiant is extending its @nvidia Vera Rubin deployment with the NVIDIA #OpenShell Reference Architecture. Our cloud customers get model compute, agent execution and security controls in a single stack. The work breaks down like this: 1. NVIDIA Vera Rubin NVL72 runs model inference and reasoning. 2. NVIDIA Vera CPUs run agent harnesses and the NVIDIA OpenShell secure runtime. 3. NVIDIA BlueField-4 DPUs run NVIDIA Sentry, enforcing policy from a hardware-isolated security domain that operates independently of the host. Enforcement sits outside the system the agent runs on. It does not depend on the host staying trustworthy. Our aim is to let customers deploy agentic services with clear boundaries on what agents can access and do, backed by coordinated runtime governance and independent security enforcement. Learn more here: lnkd.in/gnsz3nJP

#AgenticAI operates in a fundamentally different way than what we have seen to date. One prompt can set off hundreds of model and tool calls. This changes what your infrastructure needs to handle and more importantly how you build it. Radiant CTO, Chris Galloway, provides some key things to consider in this transition: Stop building one kind of rack. Prefill and decode stress compute and memory differently, so you build pools of node types, split where the workload justifies it, and let a scheduler place each step on the right tier. Treat KV-cache as a managed memory tier. Agent steps often share a model and prompt prefix, so KV blocks can sit in HBM, CPU memory or NVMe. That makes memory capacity and bandwidth a spec next to GPU count. Design for east-west traffic. KV-cache now moves between prefill and decode nodes, on top of retrieval and storage calls. NVLink covers the NVL72 rack, but the InfiniBand or Spectrum-X fabric needs sizing for cache movement, not just training collectives. Plan for confidential compute. Agents run generated code on private and sovereign data. Sandboxes contain the code, but some tenants need GPU and CPU TEEs, encrypted memory and attestation to protect them from the operator itself - budget that overhead from day one. Make telemetry the billing system. Selling completed tasks instead of tokens means metering compute, memory and attributed power per step, per tenant, or you can't bill or cap anything. Build it into the operating layer from the start. Size power and cooling for duty cycle. NVL72-class racks already pull well over 100 kW on liquid cooling, and agents keep them busy with few quiet periods, so plan for peak draw and the real duty cycle. More in the post: lnkd.in/gxxcGwnf

In his series on building repeatable #AIFactories, Manesh Patel turns to the power boundary for the first of three posts. What decisions can be made upstream, what downstream and what are the implications. The post also touches on what to measure as you build the third, fifth or tenth factory. This is practitioner, builder level stuff, but it is also a must read for anyone in the ecosystem, from power to financing. radiant.co/blog/the-boundary…

Securitization follows standardization. That standardization is what's creating the market for AI compute, with a futures market expected before the end of the year. It has come together remarkably quickly and that's a function of standardization, transparency and data. Standardization is based on @nvidia 's remarkable efforts around the #DSX AI Factory blueprint and the adjacent Exemplar status among other things. @Mahdiy Yahya breaks down the component parts, including new data streams from @silicondata and explains how compute is transitioning from custom project finance into an investable commodity class and which players that ultimately benefits. radiant.co/blog/ai-compute-a… #AI #NVIDIA #Infrastructure #CapitalMarkets

As GPU fleets and workloads grow, mean time to failure (MTTF) falls considerably. A 1024 GPU job can have an MTTF of about two orders of magnitude lower than an 8 GPU job. That means disruptions become more frequent, and a part of the operating reality. The answer is not to wait for something to break. It is to build detection, diagnosis, remediation and return to service into the way the cluster is run. This new article looks at how Radiant delivers reliable GPU operations with proactive workflows that blend maintenance intelligence, fine-grained hardware visibility and automated failover. Check out this peek into operating resilient clusters - radiant.co/blog/building-res…

"The control plane is the country. It is one thing in the stack you must have maximum control over." #raisesummit #sovereignai

Packed house at @RaiseSummit main stage for the #sovereignai panel. Mahdi Yahya defines the term - "Sovereignty is not an end state, rather is a continuum. Ultimately it is about control."

Radiant's Mahdi Yahya talks about European AI demand in his conversation with NYSE Wired's Gemma Allen. TLDR, it is accelerating and Radiant is building for that acceleration. #AIFactories #AIInfrastructure

Why are these guys smiling? Well they are about to pack this room full of @nvidia #GB300 #GPUs Do what you love they say.....

Great conversation between @JordanNanos of @SemiAnalysis_ and Patrick Wohlschlegel of Radiant on 400kW racks, @Brookfield's powered land portfolio and @nvidia's compute roadmap + it's implications for #AIfactory design: youtube.com/watch?v=SQtavfvi…