The community interest came from a practical question: can a huge MoE model be useful on ordinary hardware? Colibri uses GLM-5.2’s sparse activation pattern to avoid loading the whole model into RAM or a GPU at once.
A LocalLLaMA build with five RTX PRO 6000 cards and a 5090 made the practical cost of serious local inference hard to ignore.
Etched came out of stealth with a working chip, $800 million raised and more than $1 billion in signed customer contracts. The bigger signal is that AI inference is becoming a full-stack systems race, not just a hunt for more general-purpose GPUs.
Together AI raised $800 million at an $8.3 billion valuation, a large bet that open-model infrastructure can undercut closed-model economics. The company says annual bookings topped $1.15 billion last quarter and plans to expand capacity about 50-fold over five years.
Enterprise AI bottlenecks are shifting from model access to operational control. NVIDIA says its internal Enterprise Inference Hub serves more than 100 model endpoints and processes trillions of tokens every week.
LocalLLaMA focused on the practical question: can a diffusion LLM keep quality while making generation meaningfully faster?
The AI bottleneck is shifting from model release cadence to inference infrastructure. OpenAI says Jalapeno was taped out in nine months and is planned for gigawatt-scale deployment beginning in late 2026.
The HN discussion focused less on model quality and more on cost control. As generative AI moves from experimentation into operating budgets, token pricing is becoming a buying constraint.
Alex Ellis’s post resonated because it framed local LLMs through business use, control, cost, and agent reliability instead of a simple benchmark ladder.
The LocalLLaMA angle is not just the 1000+ tps headline, but whether FP4, DFlash, and commodity GPU kernels can be reproduced outside Xiaomi’s hosted trial.
NVIDIA says its GB300 NVL72 delivered up to 20x more concurrent agentic coding capacity per megawatt than H200 on Artificial Analysis’ new AA-AgentPerf benchmark. The test measures concurrent AI agents under service-level objectives, not just raw token throughput.
Google DeepMind released DiffusionGemma, a 26B MoE open model that uses text diffusion instead of token-by-token decoding. The pitch is up to 4x faster generation on dedicated GPUs for local, interactive workflows.