Latest / UpNext AI / Gemini Goes Shopping, OpenAI Agents Hit a UN Site, and China’s Nvidia Question | UpNext AI – September 28, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Monday, September 28th, 2026, and here's what matters in AI today. Google is testing a more transactional version of its AI products in India. TechCrunch reports that some users can see a Buy button on select Walmart-owned Flipkart listings in Gemini and Google AI Mode. Tapping it opens a Flipkart-branded checkout flow, without first sending the shopper back through conventional search results. The early test is limited: it covers some users and a small set of products, including smartphones, electronics, and mobile accessories. Other people still see ordinary Flipkart listings without a direct purchase option. Google plans a broader rollout later in October, ahead of India’s festive shopping season. That matters because the experiment shifts conversational AI from product discovery toward the point where commerce actually happens. Google has been developing its Universal Commerce Protocol for agent-to-retailer interactions, but this test appears to use Flipkart’s own checkout rather than a Google-hosted one. The company also has a financial relationship with Flipkart, having invested about $350 million in the retailer in 2024. For merchants and platforms, the question is no longer only whether an AI assistant recommends a product. It is whether the assistant becomes the front door to checkout—and who retains the customer relationship once it does. That closer connection between agents and real-world systems also raises a less glamorous but important operational question: what happens when automated systems behave like relentless web users? The Verge reports that security researcher Rowan Howard-Jones found OpenAI agents scanned the UN Conference on Trade and Development’s statistics site more than 16,000 times between April and June. The site is operated by UNCTAD, and the activity was reportedly connected to retrieving publicly available data related to its Productive Capacities Index. The report says the agents apparently did not have direct access to the site’s application programming interface, or API, and repeatedly probed the public-facing service instead. This was not described as a security breach, and the reporting does not establish the agents’ precise purpose or whether they disrupted the service. Still, it is a useful example of the new failure mode around autonomous software. A system does not need to steal data or break in to create a problem. Repeated automated requests can burden public infrastructure, bypass intended access patterns, and make ordinary rate limits and authorization checks part of the AI-agent safety stack. On the hardware side of the AI race, China may be considering a narrow opening for some domestic companies to buy a newly released Nvidia chip. The Information reports, citing two people familiar with the matter, that China’s Ministry of Industry and Information Technology has asked companies including Alibaba and ByteDance about plans to purchase Nvidia’s RTX Pro 5500 chip. The companies were reportedly asked how many chips they would want and what they would use them for. The report says officials have told some companies that purchases are intended to be approved. This is not a general reopening of access to Nvidia hardware. The chip in question is aimed at high-end professional computers, and key details remain unresolved: when approvals might arrive, how many chips would be allowed, and what criteria officials would use. But even a selective approval would matter. Chinese AI companies are seeking computing capacity for chatbots and agents, while access to advanced chips remains a central pressure point in the technology relationship between China and the United States. For Nvidia, its Chinese customers, and policymakers, the details of any approval will matter more than the signal alone. For the research note, a newly posted PREreview examines LLMs4All, a preprint survey of how large language models are being used across academic disciplines. The practical problem is that benchmark scores do not tell an institution much about deploying a model in legal analysis, financial auditing, economics, or scientific work. The reviewed survey organizes those domain-specific applications and highlights recurring constraints, including context-window limits, weaker reasoning in complex tasks, and hallucination risks that vary by field. Its useful takeaway is organizational rather than model-specific: deployment decisions need to be tied to the workflow and its consequences, not to a generic claim that an LLM performs well. A mistake in an academic draft, a compliance process, and a life-sciences workflow can have very different costs and require different safeguards. The caveat is important. This is a reviewer’s assessment of a broad preprint survey, not a controlled study establishing performance in those settings. It is best read as a map of the questions institutions should test locally before relying on AI in high-consequence work. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... The Information reports that inference providers Fal and Fireworks AI are drawing investor interest as developers spend more to run models quickly. Fal has reportedly discussed a funding round at a valuation of $15 billion, with talks still early and possible targets ranging from $17 billion to $20 billion. Goldman Sachs projects that Amazon, Alphabet, Microsoft, Oracle, and Meta could collectively spend $1.2 trillion on AI infrastructure in 2027, according to The Decoder. It is a projection, not committed spending, but it illustrates how aggressively the largest platforms are planning for compute demand. Nvidia is introducing an open-source software tool intended to keep AI agents from escaping containment, Wired reports. The move puts agent controls closer to the developer toolkit, rather than treating containment as only a policy concern. The Financial Times reports that security researchers are seeing a surge in so-called LLM-jacking: attacks that hijack companies’ AI accounts and servers to steal costly computing resources. As AI workloads become more valuable, account security and usage monitoring increasingly protect both data and the compute bill. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!