The Business AI Briefing

Thirty-five years after Hammer warned firms not to automate the old path, most AI programs still buy another place to type.

A gigawatt Ulanqab site, Huawei Ascend 950DT for running models, Nvidia for training. Beijing's quota is the competitor Nvidia cannot underprice.

Phase 3 melanoma data: tumor sequencing, up to 34 neoantigen targets, a custom mRNA shot plus Keytruda. The SME lesson is the pipeline, not the ticker.

Stripe agreed to acquire OpenRouter, the gateway across 400+ models. Token routing is now a vendor-risk event, not a convenience feature.

NVIDIA's Nemotron 3.5 Lightning scored 24 on Artificial Analysis's Intelligence Index, matching gpt-oss-120b at a quarter of the parameters. Remap the workhorse.

Cohere Labs released a 2.4B Apache-2.0 vision model that reads A4 pages at 200 dpi. Compact OCR is a model-selection decision, not a souvenir.

xAI's Grok 4.6 targets long-running agents with frontier benchmark scores and $2/$6 token pricing. Continuity planning beats IDE defaults.

Muse Glimmer is a 30B Apache 2.0 agentic model built for single-GPU local workflows. Privacy upside - and shadow-AI risk - arrive together.

The lab that dragged the industry into a price war resumed a mega-raise - then warned developers rates will rise significantly. Continuity planning beats bargain hunting.

StepFun put a 4B Apache 2.0 preview on Ascend. If your continuity plan only knows one GPU vendor, it is already incomplete.

KPMG says most APAC firms plan big AI budgets while only 5 percent can prove ROI. Productivity claims are not a finance-ready return.

Mistral open-sourced a 3B policy-adaptive guard that runs on a 16GB GPU. The model is cheap; writing and running the policies is the real work.

Cheaper tokens met agentic workflows and legal tech spend still rose. Without model routing and usage caps, the bill grows in the dark.

DeepSeek V4-Flash topped its own Pro preview on agent tests after post-training. Buying the biggest model by habit is how agent bills quietly inflate.

Dell's 2028 token estimate jumped 57× — and today's run rate is already higher. Agentic and reasoning workloads are quietly rewriting the bill.

LG AI Research released K-EXAONE 2.0, 750 billion parameters, Apache 2.0, on Hugging Face. If your model policy still lists only US vendors, your continuity plan is stale.

CEOs keep funding AI experiments. ROI stalls because nobody owns the outcome, the baseline, or the kill criteria.

July 31 public beta of V4-Flash-0731: vendor agent benches beat Pro preview at roughly one-third the output price—while Nikkei flags a heating price war and peak-hour pricing plans.

July 30 updates add production Gemini 3.5 Flash and preview feedback services—plus API breaking changes around sampling parameters and turn-structure validation for agent deployments.

A solo operator documents a production OpenClaw pipeline with isolated researcher, writer, and publisher agents—scoped tools and parallel execution after shared-context failures.