Latest / UpNext AI / Stability AI’s $76 Million Round, OpenAI’s Full Stack, and Better RAG Evaluation | UpNext AI – August 26, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, August 26th, 2026, and here's what matters in AI today. Stability AI, the company behind the image generator Stable Diffusion, has raised $76 million in Series B funding, according to TechCrunch. The new capital brings its total fundraising to $232 million. What stands out is who joined the round. Stability says participants include Universal Music Group, Sony Music Group, Warner Music Group, Electronic Arts, AMD Ventures, and Pacific Alliance Ventures. That brings content companies, a major gaming company, and a chip-focused investment arm into the financing of a business built around generative media. Stability says it will use the money to keep building its creative-production product suite and expand professional services. The company already offers models for image, video, and music generation, and it has been working with entertainment partners on tools for creative workflows. The funding does not establish product performance or commercial success on its own. But it is a tangible vote of support for a generative-media company whose future depends on more than model quality: it also depends on its relationships with the industries whose content and distribution channels shape the market. That broader systems question is also central to OpenAI’s latest infrastructure pitch. In a post by Chief Financial Officer Sarah Friar, OpenAI describes AI progress as a connected stack: data centers and chips, frontier models, developer tools, products, and devices. Its argument is that improvements at one layer make the others more productive. The immediate proof point is Jalapeño, OpenAI’s first custom inference chip. The company says that on the public InferenceX benchmark, using GPT-OSS 120B, the chip delivered higher peak throughput per kilowatt and lower token latency than the commercial systems in the comparison. OpenAI also says it performed strongly with DeepSeek R1 and Kimi K2. Inference is the work of running a trained model for users, and it is where cost, speed, and power use become daily operating constraints. OpenAI says building the model, serving software, chip, memory, and network together gives it more control over those tradeoffs. It also frames first-party silicon as an addition to, rather than a replacement for, its broader supplier portfolio, which includes Microsoft, NVIDIA, AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank. For customers, the promised outcome is faster responses, fewer retries, and lower cost for completed work. The important shift is strategic: frontier labs are increasingly treating compute architecture and energy delivery as part of the product, not merely back-office infrastructure. For the research note, consider a familiar enterprise problem: a RAG system gives a bad answer, but the team cannot tell whether it retrieved the wrong documents, generated poorly from good documents, or should have declined to answer. RAG, short for retrieval-augmented generation, retrieves source material before producing an answer. A new arXiv preprint, “The RAT: A Unified Bayesian Model for RAG Evaluation,” proposes evaluating that chain as a connected system rather than relying only on an overall accuracy score. The framework jointly models retrieval success, abstention behavior, and answer correctness, with the goal of tracing how mistakes move through the pipeline. The researchers applied it to 27 RAG configurations across three datasets, three retrievers, and three generators. They report that systems which looked equivalent under broad, marginal metrics showed substantial behavioral differences once the evaluation separated retrieval outcomes from generator behavior. They also found that annotations of retrieval success were more informative than final task-success labels for estimating whether a system followed the intended policy. That is a useful direction for teams operating document-grounded assistants: measure whether the system found the right evidence and handled uncertainty appropriately, not just whether an answer happened to be correct. Still, this is an arXiv preprint, and its production value will depend on how well its assumptions and diagnostics hold up across real enterprise workloads. ...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. ... A Tech Startups funding roundup counted $747.9 million across ten financings announced during its 12-hour research window. Its thesis is that strategic investors, including Aramco Ventures, Samsung Ventures, Siemens, Salesforce Ventures, and ARK Invest, are concentrating on AI deployment bottlenecks such as power, security, logistics, and specialized software. At travel company loveholidays, OpenAI says Codex is moving software creation beyond engineering teams. The company says more than ten new search experiences have been built through its internal Search Playground, most by non-engineers, with at least three now running on its website. Google has launched Gemini Enterprise for Legal, according to The Decoder. The product is intended to automate contract and legal-research workflows and connects with systems including iManage, DocuSign, and Everlaw. 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!