Nemotron Labs Dory 1.0 4B on an 8 GB GPU: What Fits and What Does Not
Nemotron Labs Dory 1.0 4B fits an 8 GB GPU in Q5_K_M or FP4 formats, but BF16 does not. Weight sizes, loop trade-offs and what is still unverified.
GPU Cloud Guides
GPU Electricity Cost: What a 700-Watt H200 Adds at July 2026 U.S. Prices
GPU electricity cost at 700 watts is $0.0684 an hour at the July 2026 U.S. industrial price and $0.1017 at the commercial price, before the building.
Server Useful Life: What Five AI-Infrastructure 10-Ks Actually Depreciate
Server useful life in five 2025–2026 10-Ks runs from 2 to 6 years. Meta lengthened, Amazon shortened a subset, and the dollar gap is recomputed.
KV Cache Memory: How Many Llama 3 Tokens Fit on One H200
KV cache memory on one H200: Llama 3 70B at 16-bit weights leaves 1.0 GB, about 3,052 tokens. Byte formula, unit trap, and a worked example.
Edge AI vs Cloud GPU Inference 2026: The Data Sovereignty and Latency-Cost Framework
A construction-site monitoring system that ships raw video to the cloud for object detection is not primarily an engineering decision in Europe in 2026 — it is a GDPR liability, because video of a workplace is video of identifiable people, and every frame that leaves the site becomes a data-protection question a works council can … Read more
vLLM vs SGLang vs TensorRT-LLM 2026: The Enterprise Inference Engine Decision
A platform engineer migrating a production chatbot from vLLM to TensorRT-LLM in early 2026 discovered the throughput gain everyone promised — roughly 13% more tokens per second — came bundled with a 28-minute engine compilation step that had to re-run on every model update, turning a five-minute deploy into a half-hour ritual multiplied across dozens … Read more
The Enterprise GPU Procurement Playbook 2026: Contracts, Capacity Guarantees, and Provider Financial Risk
A mid-market AI company signed what its infrastructure lead described as a “capacity guarantee” for 200 GPUs with a growing neocloud in early 2026. Six months later, during a regional shortage, the provider invoked a clause defining that guarantee as “best commercially reasonable effort” — language buried in an appendix nobody on the buying side … Read more
Sovereign AI and US Export Controls 2026: How the Chip Security Act, BIS Rules, and Bifurcated GPU Stacks Reshape Enterprise Procurement
Chip Security Act 2026 is no longer a Beltway abstraction for ML infrastructure teams. Between the January 2025 AI diffusion framework (Federal Register 2025-00636), its May 2025 rescission, the January 2026 China licensing pivot, and bipartisan momentum behind H.R. 3447, the US regulatory landscape has shifted from binary embargoes toward tiered access, location verification, and … Read more
AI Networking at Hyperscale: InfiniBand vs Ultra Ethernet for 32,000 to 100,000 GPU Clusters in 2026
A 50,000-GPU training run does not fail because tensor cores ran out of FLOPS. It fails because one rail of the fabric fell behind during an all-reduce, a checkpoint storm saturated metadata paths, or a straggler NIC retransmitted through a congested spine. The InfiniBand vs Ultra Ethernet 2026 decision is therefore not a religious war … Read more
AI Data Center Power Infrastructure 2026: Nuclear SMRs, Grid Interconnection, and the Energy Crisis Facing US GPU Clusters
The binding constraint on frontier AI scale in 2026 is not H100 allocation or B200 yield—it is AI data center power infrastructure: megawatts, interconnection queues, and the physics of getting electrons to racks before GPU purchase orders ship. Hyperscalers can sign multi-gigawatt nuclear partnerships; a 40 MW enterprise training hall may still sit dark for … Read more