Wednesday · Sep 16, 2026
·
The Signal · Issue No. 49
Big day
Daily AI intelligence, translated for business operators.
Today's Signal
New AI Model Cuts Processing Cost 238x and Runs 40x Faster, Designed to Live Inside Your Software
TypeSafe, founded by a ChatGPT researcher, launched Jev, an AI system that doesn't generate text but makes fast decisions from preset options with near-zero hallucination. It costs $42 per billion input tokens with free output, runs in 70 to 500 milliseconds, and is built to handle judgment calls inside applications like request sorting, record scoring, and quality screening.
What this means for you
If this pricing and speed hold up in production, you'll see AI move from a chatbot layer bolted onto your stack into the operating system itself. Every routine decision in your workflow becomes a candidate for automation, not because it's sexy but because it's now cheaper and faster than human review or traditional rules engines.
See how we build it in →
02Platforms
Salesforce Builds Its Own Reasoning Model to Keep Your Data Inside Your Own House
Salesforce and NVIDIA launched Koa, a reasoning model designed specifically for sales, marketing, and support workflows. It runs through Salesforce's own AI gateway using synthetic data rather than your customer information, giving you a Claude or ChatGPT alternative without leaving your security perimeter.
For you: You get a choice: use Salesforce's model for sensitive workflows and keep data inside the platform, or route to Claude or ChatGPT through Claudeforce when you need it, all with the same admin controls.
Go deeper →
03Platforms
Meta Turns AI Tools Into Paid Tiers Across Facebook, Instagram, and WhatsApp
Meta launched Meta One, a subscription bundling expanded access to AI image, video, and editing tools. Instagram Plus revenue is now averaging $1.2 million daily, up 475% week over week, showing early traction in monetizing generative AI features across its social platforms.
For you: If you advertise or sell on Meta's platforms, expect your audience to have access to AI editing and creation tools. Consider how that changes the cost and quality of user-generated content that lands in your feed.
Go deeper →
04Models
Hardware Bottleneck Emerges for AI at Scale, Not Training but Inference Memory
TeRAM emerged from stealth with $37 million to build 3D SRAM for AI data centers, targeting a shift in where AI chokes. As models grow past two trillion parameters, memory capacity and bandwidth are now the constraint, not compute. The company is integrating 3D SRAM directly onto chips.
For you: If you're running AI in production or considering it, know that the economics are shifting. Inference cost and latency will improve as hardware catches up, which means your payback timeline on AI deployment projects should start improving this year.
Go deeper →
05The Market
Oil Company and Mistral Bet 100 Million Euros on Agentic AI for Geology
TotalEnergies and Mistral committed more than €100 million over three years to build frontier AI models for reservoir exploration and subsurface engineering. The models will work on a decade's worth of geoscience data to generate exploration scenarios and analysis.
For you: Domain-specific AI models trained on your industry's data and expertise will outperform generic tools. If you operate in a field with deep historical data and repeatable complex problems, this is the playbook: partner with a capable AI shop and invest in training, not buying off-the-shelf.
Go deeper →
06Signal Check
The Model Is Still the Engine. Your System Is the Car. Today Changed Only the Engine.
Three out of four stories today are about new models or model improvements. Jev is cheaper and faster. Koa is enterprise-focused. TeRAM makes inference cheaper. Mistral gets better oil-field data. But here is what did not change: how you decide which problems AI should solve, how it connects to your other systems, who owns the output, and whether your team can maintain it when the model updates. The model is the commodity now. The system you build around it is what runs your business.
InGen editorial · Sep 16