A roundup of AI news from the evening of September 16, 2026. While OpenAI signaled a new direction for its advertising business, AI infrastructure news was dominated by physical-layer stories: coordination with the power grid and Apple's potential return to the server market. In the local LLM community, techniques for handling long contexts within limited VRAM are drawing attention.

OpenAI 'Reimagines' Advertising With Agent Ads: 'Sponsored Agents'

On September 16, OpenAI made an announcement titled 'Reimagining advertising with AI,' unveiling new AI-powered advertising experiences. At the center is 'Sponsored Agents,' accompanied by a suite of tools for marketers and integrations with HubSpot and Shopify.

Concerns that ads in chat UIs degrade the user experience are a perennial theme, but OpenAI has positioned itself as redefining advertising in the context of 'agents.' As of the announcement, details such as pricing and the rollout schedule remain unclear.

Emerald AI, Google, and NVIDIA Launch 'AEMA,' an Industry Alliance for Power-Flexible Data Centers

On September 16, Emerald AI, Google, and NVIDIA announced the formation of the 'AI Energy Management Alliance (AEMA),' an industry alliance that aims to enable AI data centers to dynamically adjust their own power consumption in response to grid conditions.

The backdrop is the worsening power crunch facing AI infrastructure expansion in the United States. According to NVIDIA's announcement, power has become the defining factor in scaling US AI infrastructure, and legacy interconnection processes — designed for facilities with flat, static demand — cannot accommodate computing infrastructure capable of responding intelligently when the grid is stressed.

Flexible data centers can modulate the power they draw from the grid by shifting computing workloads, discharging stored energy, leveraging on-site generation, and responding to emergencies. This is said to allow existing infrastructure capacity to be used more efficiently, reduce demand during grid stress, and avoid or defer costly infrastructure upgrades.

AEMA is characterized by technology-neutral, performance-based operation. Rather than favoring specific hardware or software, it focuses on 'measurable services' such as response speed, duration, predictability, and emergency behavior. Its stated aims include defining obligations before connection — such as riding through short-duration grid disturbances, output curtailment, and emergency response — standardizing technical requirements, performance metrics, and operational data sharing, and establishing faster interconnection pathways for commitments to flexibility. The alliance seeks to bring together the entire value chain, from AI platforms to utilities and local grid operators, to build AI infrastructure that doesn't just connect to the grid but works with it.

Apple May Be Eyeing a Market Return With AI Servers Combining Nvidia Technology

A report from The Information revealed that Apple is exploring the idea of offering AI servers built around its 'M8'-series chips and is in talks to adopt Nvidia's networking hardware. MacRumors reported the news on September 16, noting that the system would be intended for sale to external customers — potentially marking Apple's return to the server business for the first time since it discontinued the Xserve in 2011.

The envisioned customers are companies that want to run AI models on their own premises, with a particular focus on inference — generating responses from trained models. One of the options under discussion is the use of Nvidia's 'NVLink Fusion' for chip-to-chip data transfer, and there are plans to use it to connect M8 processors to one another. That said, neither the Nvidia deal nor the servers themselves are set in stone; if the project goes ahead, the systems are expected to arrive around 2029, though it could still be canceled before then.

A 1-Million-Token Context on 3x RTX 3090 — Offloading Qwen3.8-Flash-Next's KV Cache to RAM

A finding gaining attention in the LocalLLaMA community shows that offloading most of Qwen3.8-Flash-Next's KV cache to system RAM barely reduces decoding speed. According to the poster, they got this configuration running with vLLM and are operating a 1-million-token context on 3x RTX 3090. Short contexts run at about 80 tokens per second, and even after the sparse selection mechanism (QSA) exhausts its 2,048-token budget, it sustains about 60 tokens per second — decoding speed stays nearly flat as the context grows. With 4 parallel requests, throughput reportedly reaches 150 tokens per second.

The reason this configuration holds up lies in the architecture. Only 12 of the model's 48 layers carry a KV cache; the remaining 36 use linear attention (gated delta-net) with a fixed-size recurrent state, so the state does not balloon as context length grows. Even the layers that do have a KV cache only reference what the sparse indexer selects — up to 2,048 positions from compressed keys — so per-step reads are determined by the number of selections rather than the context length. By the poster's estimates, this works out to roughly 4MiB per layer, or about 48MiB per token across the 12 layers combined, using only a small fraction of PCIe 4.0 x16 bandwidth (~32GiB/s).

The practical appeal is that even quantized models that just barely fit in VRAM can handle their maximum context length without quantizing the KV cache, and some suggest the same technique should work with 'qwen4exp'-family models, which are expected to form the basis of Qwen's next local models.

Perplexity Releases 'CobbleDB' — Built by 2 Engineers and AI Agents

Perplexity has released the storage engine 'CobbleDB.' Its distinguishing feature is that it was built by '2 engineers and AI agents,' and it has become a talking point on Hacker News. As a real-world demonstration of how far a very small team plus AI agents can go in building complex systems software like a database, it may become a telling example of how development gets done from here on.

SK Hynix Reported to Be in Talks With Intel on US Memory Chip Production

On September 16, reports emerged that SK Hynix is in discussions with Intel over memory chip manufacturing in the United States. Responding to an inquiry from TechCrunch, the company commented that no plans or arrangements have been finalized, so for now this remains at the report stage.