The H100 is a data-center GPU from NVIDIA used widely for training and running large models. Recent coverage has centered on how that compute is allocated and priced: Ai2 replaced priority scheduling with GPU time budgets across thousands of H100, B200 and B300 GPUs; a paper found the A100 cheaper than the H100 for gpt-oss-120b, with self-hosting at $0.12-$0.35 per million output tokens; and SupraLabs trained its 100M-parameter Supra2-IMG text-to-image model on a single H100 in under 10 hours.
Ai2 replaces priority scheduler with GPU time budgetsAi2 replaced priority-based scheduling with GPU time budgets, hierarchical fair-share allocation, and time slicing, shifting allocation from case-by-case debates to a transparent budgeting process. Its clusters host thousands of NVIDIA H100, B200, and B300 GPUs for about 150 researchers, with demand often 2-3x supply.
21M model with 6.4B lookup table matches 114M dense model, table on SSDDeveloper fehcyy released sparse-memory-lm, giving a 21M-parameter model a 16.8M-row lookup table (6.4B table params, 33M used per token), achieving performance comparable to a 114M dense model on 500M Wikipedia tokens. The table can be 4-bit quantized and memory-mapped from NVMe SSD, achieving 140 tok/s on RX 9070 with only 0.4 GB VRAM, but long prompts are slow due to 4KB page misses. Triton kernels run unchanged on Radeon, MI350X, and H100/H200. Bolting a table onto a pretrained model didn't work.
Jeeves: Reasoning Improves Jev-Like Decision ModelsA developer released Jeeves, a Jev-like decision model built on Qwen3.5-9B with LoRA, a pointer head, CISPO training, and a block-4 diffusion drafter. It scores 0.889 on unseen test data versus 0.822 for Kev-9B and 0.857 for Jev, and 0.935 on JevBench's public hard tier; latency on one H100 is about 0.3s without thinking and 3.3s median with it, with code, weights, and data open-sourced.
CoWindow and MassAlloc Attention Speed Up Long-Context ModelsAuthors present CoWindow Attention, which spreads distant context across KV heads via complementary windows without a learned router, and MassAlloc Attention, which uses softmax statistics to skip low-contribution post-score compute. At 128K context on 8 H100s, attention-operator speeds reach 7.4x/8.6x/3.0x for forward/backward/decode, and 14B models at 32K cut training FLOPs by 28.5% and 23.1% with comparable capability to FullAttn.
Economics of Open-Weight Inference: Older GPUs Retain ValueA paper finds open-weight inference costs about one-fifth of comparable closed models, with self-hosting at $0.12-$0.35 per million output tokens. On gpt-oss-120b, the A100 is cheaper than the H100, and A100 five-year rental prices retain 80% of one-month prices, showing older GPU families keep multi-year earning life.
SupraLabs Open-Sources 100M Text-to-Image Model Supra2-IMGSupraLabs released Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch in under 10 hours on a single H100. It generates high-quality 256x256 images and is openly available, taking about 20 seconds per image on CPU and 2 seconds on GPU.
The A.I. Daily