Walmart opens its doors to AI shopping agents while a major competitor locks them out. Walmart announced technology that lets AI agents browse and purchase items on a customer's behalf, with configurable spending limits and approval controls. The move is a direct contrast to Amazon, which has been actively blocking third-party AI agents from completing purchases on its platform. Walmart's system is designed so customers stay in control of what the agent can spend and approve. The split puts the two largest US retailers on opposite sides of a question every e-commerce platform will have to answer. (TechCrunch)
Brief
Thursday, October 22, 2026 · 5 min read
OpenAI models land inside Google Cloud Marketplace
Good morning. Two of the biggest names in AI infrastructure spent years positioning themselves as rivals. Today one of them is selling through the other's storefront. The AI industry has a way of making yesterday's competitive diagram look like a rough draft.
Today's reading time is 5 minutes.
Listen · Snippt Daily0:00 / 3:32
INDUSTRY
Until now, developers who wanted OpenAI's frontier models and Google Cloud's infrastructure had to manage two separate relationships.
Driving the news: OpenAI and Google Cloud announced that OpenAI's models, including Codex workflows, are now available through Gemini's Model Garden and the Google Cloud Marketplace. Developers can access them directly inside the same console they use for Google's own models. The deal makes OpenAI a distribution partner of the platform it competes with at almost every other level.
- Codex, OpenAI's coding-focused model family, is included alongside frontier models in the Model Garden listing.
- Google Cloud Marketplace handles billing and access, meaning enterprise customers can draw on existing cloud commitments to pay.
Zoom in: Google Cloud has spent two years building Model Garden as a one-stop catalog for third-party and proprietary models, a direct answer to the fragmentation developers complained about when juggling multiple AI vendors. OpenAI, meanwhile, has been expanding distribution aggressively, adding partners across cloud and enterprise channels. The two companies are also competitors in consumer AI, enterprise contracts, and the underlying model race, which makes this arrangement genuinely unusual.
- OpenAI separately published a policy piece this week on how it enforces consistent usage standards across distribution partners, suggesting the Google Cloud deal fits a broader push to grow through third-party channels.
- Google's own Gemini models remain in the same catalog, so enterprise buyers can now compare and mix models from both companies in one place.
Why it matters: For a developer already running workloads on Google Cloud, this removes the friction of a separate OpenAI account, separate billing, and separate API keys. Teams building coding tools with Codex can now wire it into existing Google Cloud pipelines without leaving the console. The practical effect is that OpenAI's models become easier to reach for the large slice of enterprise engineering that standardized on Google Cloud infrastructure.
Bottom line: The two companies most publicly competing to own the AI stack just agreed to sell through each other, which says more about where enterprise distribution is going than any product announcement either has made this year.
OpenAI ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Data centers are dragging air quality in Texas back toward 2000s levels. A new report finds that the surge of AI data center construction in Texas could reverse decades of air quality improvements, pushing pollution back to levels not recorded since the early 2000s. The problem is electricity demand: data centers require so much power that utilities are leaning on fossil fuel generation to keep up. Texas has become the densest concentration of new AI infrastructure in the US, which makes it the clearest case study for what the physical cost of generative AI actually looks like. The report names both new and proposed facilities as contributors. (The Verge)
Hugging Face released an open-source biology model that interprets DNA sequences as a kind of language. The model, published by Hugging Face, treats DNA the way a language model treats text, letting researchers query it about how specific genetic changes affect gene activity and health outcomes. It is open-source and aimed at biology researchers who want to explore the downstream effects of mutations without running wet-lab experiments for every variant. The framing as a conversational interface, where you describe a change and ask what it does, is new for this class of tool. Wired reports the model is available now. (Wired)
ChatGPT's ad platform expanded to 26 more countries after crossing 100,000 advertisers. OpenAI confirmed that ChatGPT Ads is now live in 26 additional countries, with the platform having already served more than 100,000 advertisers. The ad stack uses the underlying language model to infer conversational intent and serve ads separately from the model's answers, a distinction OpenAI has been careful to maintain publicly. OpenAI also published a technical breakdown of how the system was built, covering how it separates sponsored content from responses at the infrastructure level. The expansion makes ChatGPT one of the larger ad-supported AI surfaces globally. (OpenAI)
The thread
Distribution is the new moat
Two of today's stories turn on the same logic: the model is not enough on its own, the channel is where the leverage is. OpenAI listing its models inside Google Cloud's marketplace and Walmart opening its platform to AI shopping agents are both bets that whoever controls the purchase path controls the outcome. Developers and retailers are both being asked to pick a distribution posture, and the choices being made now are not easy to reverse.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- Anthropic launched Claude Duet, a standalone iOS and Android agent app aimed at professionals and small business owners, currently in research preview - it enters a category where competing agent apps from Meta and OpenAI already exist.
- Perplexity and NVIDIA published JOLT, a training method that uses one model as both teacher and student; on the MATH benchmark it matched a standard reinforcement learning baseline using 6.5 times fewer processed training tokens.
- AT&T built an internal legal tool called LegalEdge on top of OpenAI models to handle intake and scale access to legal judgment across the company, per an OpenAI case study published yesterday.
- A distributed system using thousands of AI agents, built by OpenAI, PerturbAI, and Scripps Research, was used to surface research leads from data covering millions of brain cells, per OpenAI.
- A new arXiv paper finds that LLMs struggle with planning tasks in ways that depend more on whether the symbolic vocabulary is interpretable than on how structurally complex the plan actually is - a useful distinction for anyone benchmarking reasoning models.
- Researchers proposed using compiler error messages from the Lean 4 proof assistant to guide LLM repair of failed proof steps, rather than simply resampling, which reduces wasted attempts on already-rejected paths.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy