AI Bills Triple, Tokens Vanish: The Enterprise Cost Crackdown
Atlassian, Adobe, Amazon, and Citi are cutting access to frontier models after monthly AI spend hit $15M. Inside the enterprise token crunch.
Tag
#inference
A 330 GB On-Die DRAM AI Chip Lands as a Whitepaper
PhantaField's Sophon PFG-1 whitepaper claims ~95x Nvidia HBM4 bandwidth via monolithic 3D stacking. No silicon yet. Here's why it matters anyway.
Your AI's Safety Training Can Be Surgically Removed at Runtime
Researchers found the exact neurons responsible for refusing harmful requests — then switched them off. No retraining. No fine-tuning. Just geometry.
Google's TurboQuant Could Let You Run Bigger AI Models on Your Hardware
New compression algorithm achieves 6x memory reduction with zero accuracy loss. No retraining required. This matters for anyone running local AI.
Cloudflare Enters the Big Model Game: Workers AI Now Runs Kimi K2.5 With 256K Context
Cloudflare adds its first frontier-scale model to Workers AI, claiming 77% cost savings over proprietary alternatives with new caching features.
Open-Weight LLM Showdown: GTC Pivots to Inference, DeepSeek V4 Still MIA
Jensen Huang bets on inference chips, Ollama adds multimodal support, and DeepSeek V4 remains the most anticipated release that hasn't happened yet.
Nvidia GTC 2026: Vera Rubin, NemoClaw, and the $20 Billion Groq Bet
Jensen Huang's keynote today marks Nvidia's biggest pivot in years - from training chips to inference, from cloud to edge, and from prompts to autonomous agents
Your AI Prompts Are a Security Liability. Most Companies Haven't Noticed.
While enterprises focus on training data and model safety, inference - where AI actually processes requests - has become an overlooked security frontier with critical vulnerabilities.
The Week Local AI Grew Up: Ollama 0.17 and llama.cpp's New Home
Ollama delivers 40% faster inference while llama.cpp finds a permanent home at Hugging Face. Two developments that secure the future of running AI on your own hardware.
Longer Isn't Smarter: Google Research Shows Token Count Predicts Failure, Not Success
New research from Google and UVA reveals that longer AI reasoning traces actually correlate with wrong answers. The fix: measure how deeply the model thinks, not how much it writes.
Taalas HC1: The AI Chip That Bakes the Model Into Silicon
A Toronto startup is etching LLM weights directly into transistors, achieving 17,000 tokens per second. The catch: you can't change the model.
OpenAI Just Deployed Its First AI Model on Non-NVIDIA Chips
GPT-5.3-Codex-Spark runs on Cerebras' wafer-scale chips at 1,000+ tokens per second. It's OpenAI's first production break from NVIDIA - and it won't be the last.
Tiiny AI's $1,399 Pocket Lab Claims to Run 120B Models Locally - Here's What We Actually Know
A 300-gram device promises GPT-4o-level performance without cloud or internet. The specs are real, but the benchmarks are missing.