NVIDIA B300 servers retail around $350k min, crazy power/cooling requirements/etc. (bad maths) per GPU (8) is $43,750, if you can even find them (narrator: you can't) then pay for collocation etc... 9 PFLOPS dense FP4. DGX Spark: $4700, plug in anywhere, 1 PFLOP (fp4 sparse). B300 9ish PFLOP dense FP4 super expensive impossible to collocate anywhere etc. etc., $4861.11 per PFLOP. DGX Spark $4700 plug in any 15 amp outline in any house, $4700/PFLOP (but sparse). Memory is slower, so I guess use less of it/less manipulation of it... Yep, math checks out - pretraining on decentralized DGX Spark with native sparse FP4 (checks notes: yes we're using ternary weights which slot perfectly into sparse FP4 compute) should be a thing. Time to test it out...

10,000 @nvidia B300 GPUs. India's largest AI Factory. Together AI and @larsentoubro are building the country's biggest GPU cluster, backing open-source inference, fine-tuning, and training at scale for India's AI-native ecosystem.

The Compute Desk Nvidia B300 GPU-hour index is at its all-time high. As margins on training grow thinner, neoclouds are collectively shifting focus to inference. High-priced B300s are generating ROI by lowering the marginal cost per token generated through batch inference.

$NBIS Nebius is raising on-demand H100/H200/B200/B300 rates 17%–21% from Oct 1, which has definitely drawn significant attention to its ability to capitalize on all those GPUs. But the more interesting move is actually outside GPUs. Specifically, CPU-only instances rise 25% and memory about 41%. That's very interesting piece of evidence showing that AI-cloud scarcity is spreading into the host stack, not staying isolated in accelerators. And Nebius already produced a 50% adjusted EBITDA margin in AI Cloud in Q2, and its pricing page offers up to 35% commitment discounts for large clusters. So the model is now clearly turning into a two-tiered layer. That is, monetize scarcity on-demand while using discounts to lock in longer-duration capacity. If that pricing holds without slowing demand, incremental capacity should arrive with better unit economics of scale, which in theory should significantly boost the revenues. Into the longer term, this is easily a $500+ stock given its ability to scale and its capacity to expand without incurring significantly more cost.

IREN has achieved @nvidia Exemplar Cloud status on NVIDIA HGX B300 for training workloads. This status confirms that IREN's infrastructure performs within NVIDIA's reference performance targets across its full suite of benchmarking recipes, validated against NVIDIA reference architecture. "IREN's achievement of NVIDIA Exemplar Cloud status reflects deep engineering collaboration between our teams and the quality of infrastructure behind IREN's AI Cloud, giving enterprises confidence to run their most demanding training workloads at scale." — Warren Barkley, VP Product Management, NVIDIA Read full blog:

MI355X IS UP TO 1.7X BETTER 💰️PERF PER DOLLAR 💰️THAN DGX B300. The AMD Mainland China UMBP team co-designed, in collaboration with Alibaba & the @sgl_project community, a new feature in SGLang that removes the duplicated KVCache contained between local L2 DRAM & distributed L3 DRAM, allowing for up to 2x more KVCache to be stored in DRAM. This feature is called UnifiedRadixCache external cache. But importantly, this marks the trend of AMD increasingly being first-class co-designed for new features in widely used top production engines like SGLang.

Kimi K3 serving in vLLM now delivers 2.2–2.8x throughput on our B300 benchmark vs v0.27.1. We break down the work across scheduling, KDA state handling, and MoE kernels, with benchmarks and commands to reproduce the results. Thanks to the vLLM community for pushing Kimi K3 performance forward! Read the deep dive: