Welcome to P3 Media’s AI Commerce Brief, your daily update on the AI and commerce stories shaping how companies build, sell, and grow. It’s Thursday, June 25, 2026. Let’s get into it.
Our top story is OpenAI moving deeper into the hardware layer of the AI stack.
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom AI inference chip. OpenAI says the chip is designed for large language model inference and is the first accelerator in a multi-generation compute platform the companies are building together.
OpenAI is not the first AI company to pursue a custom chip. Google has TPUs, AWS has Trainium, Microsoft has Maia, and Meta has MTIA. But OpenAI is the first independent frontier model lab to publicly unveil its own named custom inference chip.
Frontier model labs are under resource pressure from three directions at once: exploding inference demand, limited access to advanced chips, and a hardware market still dominated by Nvidia. Custom silicon gives OpenAI a path to more vertical integration, more control over supply, and potentially better economics for serving products.
Reuters has reported that Anthropic is exploring its own chip design, but the company has not announced a produced Anthropic chip.
Next, Google moved the agent market forward.
The company said computer use is now built into Gemini 3.5 Flash. That means developers and enterprises can use the model to build agents that interact across browser, mobile, and desktop environments.
Google also highlighted safeguards, including user confirmation for sensitive actions and automatic stops for some indirect prompt injection risks.
Google is not the first company to release an OS-using agent, but its focus on guardrails speaks to a key challenge with agentic computer use: how to empower agents to use operating systems while safeguarding against catastrophic mistakes. Google is betting that the key differentiator will be the systems that can operate reliably, with permissioning, logs, and human oversight.
Meanwhile, Anthropic is pushing agents further into shared workplace context with Claude Tag.
The beta lets teams tag Claude inside Slack, delegate a request, and have Claude break the task into stages before replying in the thread. Anthropic says teams are using this pattern to chase down product metrics, work through support tickets, and help find the root cause of bugs.
Claude can also remember relevant channel context, work asynchronously, and, if enabled, proactively flag items that may need attention.
For operators, this is the pattern to watch: agents moving into the tools where teams already coordinate work, with persistent context and administrative controls.
In retail media, Walmart said it has agreed to acquire Vibe.co, a self-serve connected TV advertising platform.
Walmart says the deal will bring Vibe into Walmart Connect and help small and mid-market advertisers launch connected TV campaigns with better commerce measurement. Terms were not disclosed.
The announcement comes in the wake of Walmart’s shoppable TV demo at Cannes Lions and signals that Walmart Connect is running a full press to build smarter, more frictionless inventory for advertisers.
For Walmart, connected TV is not just a brand channel. It is becoming part of the retail media performance stack, where video exposure, audience data, and commerce outcomes can be tied more closely together.
Now the commerce pulse.
Prime Day has become a live AI shopping test.
Amazon says Alexa for Shopping can create personalized deal guides, track deals, auto-buy items at a target price, show price history, and support visual shopping through Amazon Lens.
The company has positioned Prime Day to investigate several key AI commerce questions: how much product discovery has shifted from search boxes and sponsored placements into conversational recommendations? And how much friction can AI remove from the buying process?
Finally, in global model watch, Reuters reports that Anthropic has accused Alibaba-linked operators of using almost 25,000 fraudulent accounts to extract more than 28.8 million Claude exchanges.
Anthropic described the alleged fraud as a distillation campaign: a black-hat method of using one model’s outputs to train or improve another model.
Frontier models routinely use AI’s ability to train itself to improve their own models, but the accusation points to a key challenge in keeping proprietary technology moated when it can act on its own.
What to watch: whether OpenAI publishes hard Jalapeño benchmarks, whether Google’s computer-use safeguards become a standard expectation for enterprise agents, whether Claude Tag pulls agents deeper into everyday team workflows, and whether Walmart can turn connected TV into a self-serve retail media channel for marketplace sellers.
That’s your AI Commerce Brief for today. Thanks for listening.
