Both Sides of the Counter

OpenAI shipped a model built to work a keyboard and mouse. A Visa-backed survey found that only 19% of small businesses can tell when an AI agent bought something from them. And Nvidia bought the warehouse where nearly every open AI model lives. Machines showed up on both sides of your counter this week.

The Bottom Line

Three stories, one shift. The software you run and the customers you serve are both getting machine operators, and the shelf they pull their models from now has a single landlord. None of that calls for a rebuild. It calls for a meter: know which visits and which orders came from an agent, make one boring workflow readable by a machine, and write down which model each of your tools depends on. Tell me in the comments: would you let an AI agent buy from you without a human ever seeing your storefront? Or does that trade your brand for a line item inside someone else's answer?

See what we build →

Top 3 Stories This Week

01 · OpenAI · GPT-6 Astra · Computer Use

OpenAI Shipped a Model That Works Your Keyboard

For three years the arrangement was simple: AI wrote the words, you did the clicking. On September 3, OpenAI released GPT-6 Astra and went after the clicking. Astra is built for computer use, meaning it reads a screen and drives the mouse and keyboard the way a temp worker would: filling web forms, updating customer records in a CRM, editing spreadsheets, checking that a website still works. On OSWorld 2.0, the standard test for real desktop work, it scored 72.6%, ahead of Claude Opus 5 at 70.2% and its own predecessor GPT-5.6 Sol at 65.7%. Speed moved further than the score did. Astra finished those tasks in roughly 40 minutes each against Sol's 75, about 47% less time per task. It ships as gpt-6-astra in OpenAI's API and through Microsoft Azure and Amazon Bedrock, priced at $10 per million input tokens and $50 per million output, with cached input at $1. Now read the number the way an owner should. 72.6% means it still fails one desktop task in four.

Microservice Opportunity: Pick the job nobody wants: the one where an employee retypes the same fields from an email into three systems. Permit applications, insurance intake, order entry, the vendor portal that never built an API. A microservice (one small AI tool doing one specific job) drives that screen on a schedule and hands anything unusual to a person. At a one-in-four failure rate, build the review step first and the automation second.

02 · PYMNTS Intelligence · Visa Acceptance · Agentic Commerce

Only 19% of Small Businesses Can See the AI Shopper

Your next customer may never look at your website. On September 2, PYMNTS Intelligence published the Global Digital Shopping Index: Merchant Edition, commissioned by Visa Acceptance Solutions, and the finding is a blind spot. Only 23% of merchants can clearly identify both AI-driven traffic and AI-driven purchases. Among small and mid-size businesses that falls to 19%, against 27% for large merchants. Another 21% can see agent traffic but cannot tie it to a sale. The same merchants expect the channel to matter: 61% think AI-generated results will influence purchases more than traditional search, 58% expect agents to choose the payment method by fees and rewards, 56% expect agents to complete transactions outright, and 38% expect agent purchases to pass 15% of sales within two years. Adobe's tracking says it started already. AI-referred traffic to U.S. retail sites has grown 1,324% since October 2024, and those visits convert about 60% better than the rest. You cannot price a channel you cannot count.

Microservice Opportunity: Install the meter before the makeover. A microservice tags every visit and every order with its referral source, separates agent traffic from human traffic, and writes one weekly line: sessions, conversion, average order, payment method. Then make the catalog machine-readable: structured product data, live prices, real stock counts, policies in plain language. The merchants who measure this channel first will be the ones who price it first.

03 · Nvidia · Hugging Face · Open Models

Nvidia Bought the Warehouse Where Open AI Models Live

Every "free" model your developer downloads comes from somewhere, and for most of the industry that somewhere is Hugging Face. On September 2 Nvidia signed a definitive agreement to buy it, confirming the deal publicly on September 3: $12.93 billion, structured as roughly $11.9 billion in cash plus up to $1 billion in equity to keep the staff. The platform hosts three million models, half a million datasets and a million applications, used by more than 18 million developers and over 200,000 companies. Nvidia now owns the chips, the software layer and the shelf where models get found. The company says the hub stays open and that Nvidia compute will not be required to build or deploy through it. The deal is expected to close in the first half of 2027, pending regulatory approval. Nothing breaks for a small business tomorrow. What changed is concentration: the cheap open model you self-host to avoid vendor lock-in now sits in a warehouse owned by the company that sells the hardware.

Microservice Opportunity: Write down your model list. For each microservice, note which model it calls, where that model runs, and what a swap would cost in hours. Keep prompts and business rules in your own repository instead of a vendor console, and pin model versions so an upstream update does not quietly rewrite your quotes. One page per service, one afternoon of work, and you keep the right to walk away.

This week in numbers: 72.6%: GPT-6 Astra's score on OSWorld 2.0, the desktop computer-use test. It still fails one task in four. · 19%: small and mid-size merchants who can identify an AI-agent purchase. Large merchants: 27%. · $12.9B: what Nvidia agreed to pay for Hugging Face, home to three million open AI models.

Sources

TL;DR — Key Takeaways

Top 3 Stories This Week

OpenAI Shipped a Model That Works Your Keyboard

Only 19% of Small Businesses Can See the AI Shopper

Nvidia Bought the Warehouse Where Open AI Models Live

Related answers