The Stack - SiliconNoon (Page 9)

The Stack

Sugon validates GPU-direct RDMA across a 100,000-card AI cluster

The most important AI infrastructure news out of China this week is not a model — it is the plumbing underneath one. Sugon says IBGDA, a networking technique that lets GPUs drive the network card directly instead of routing every transfer through the CPU, has completed full-scale validation on a cluster of 100,000 accelerator cards. The company's token-acceleration stack, built on its scaleFabric RDMA fabric, marks the first at-scale deployment of InfiniBand GPUDirect Async inside China — filli

The Stack

Open Source Radar — September 11: skills go open, models go home

The open-source story today is about ownership: who holds the model, who holds the agent's instructions, and who keeps the knowledge after the chat ends. Five picks from the daily board. Colibrì (C, 27,632 stars) — The most ambitious project on the board: an inference engine, in pure C with zero dependencies, that runs frontier mixture-of-experts models up to 2.8 trillion parameters on hardware you already own. The trick is treating your SSD, RAM, and VRAM as one memory hierarchy — the always-n

The Stack

Deep Dive — OpenAI's Agents API: renting out the machinery behind Codex

OpenAI spent this week handing out its plumbing. On September 10, the company put the Agents API into public beta: a managed service that exposes the same harness and infrastructure behind Codex — context management, tool use, multi-agent orchestration — through a single API call. As we covered in our brief on the launch, developers specify a task, a model, tools and an environment; OpenAI spins up the Codex harness, fans work out to parallel subagents that each keep their own context, compacts

The Stack

AI 101 — What is LoRA (low-rank adaptation)?

LoRA is a way of customizing an AI model by training a tiny add-on instead of rewriting the whole thing — the base model's weights stay frozen, and a small pair of "adapter" matrices carries the change. If you have read that someone "dropped a LoRA" to teach an image model a new style, or that a community fine-tune of an open-weight model was trained on a single GPU, LoRA is the technique underneath. It is the most common word in open-source model customization, and this is the ten-minute versio

The Stack

YuE2-3B:TaT3-billioZ

Two trends are colliding in open-source audio: frontier-quality music generation is moving downmarket, and the tooling to run it locally is getting good enough that hobbyists are benchmarking it like gamers. The latest proof is YuE2-3B, a new open-weights music model from the MAP team behind the original YuE — and the community is already running it on hardware most people actually own. MAP's YuE2-3B unifies songwriting and performance in one open model — and a community quantization lets it ru

The Stack

Ayar Labs adds $150M more as copper interconnects hit the wall

Two ends of the AI buildout moved today: the physical layer and the policy layer. An optical chipmaker pulled in another $150 million to rip copper out of AI clusters, and China's industry ministry put hard 2028 deadlines on rebuilding its software industry around AI. Ayar Labs, the startup building optical links between AI chips, has raised an additional $150 million — bringing its 2026 funding to $650 million. According to Reuters, the Series E extension adds Taiwanese server manufacturer Wiw

The Stack

Qualcomm's next NPU runs 30B-parameter models on a phone

Two very different bets on where AI actually runs landed today: Qualcomm says the next Snapdragon will carry a 30-billion-parameter model in your pocket, and China's industry ministry picked the car as the first industry where AI models and agents get to run at scale. Qualcomm has detailed its next-generation Hexagon NPU, and the headline number is 30 billion parameters on a phone. The company says mixture-of-experts models up to 30B total parameters will run on the NPU, with roughly 3 billion

The Stack

Adobe's AI revenue is real — the market just billed it as a cost

Adobe beat on revenue, more than doubled its AI-first recurring revenue, and still watched the stock slip. The disconnect is worth reading closely, because it tests a question every incumbent software company is now facing: does AI show up as growth, or as the reason customers expect to pay less? Adobe beat Wall Street on Thursday with quarterly revenue of $6.76 billion against a $6.70 billion consensus — but the stock fell about 2% in extended trading after Q4 guidance came in just short. The

The Stack

Pentagon in talks to lend $5 billion to AI cloud startup Fluidstack

Two money stories lead a heavy AI news day: Washington may become a direct lender to the AI cloud buildout, and a defense-AI startup just doubled its valuation in three months. The Pentagon is in talks to lend roughly $5 billion to Fluidstack, the neocloud that has become one of the fastest-growing AI infrastructure startups, according to the Wall Street Journal. The company has been advised on its loan application by Erebor Bank, the tech-focused bank co-founded by Palmer Luckey. Neither the P

The Stack

Microsoft plans to more than triple data center capacity to 38 GW

Microsoft's data center footprint is about to get unrecognizable. According to a Bloomberg report citing people familiar with the plans, the company intends to expand its global data center capacity from roughly 12 gigawatts today to more than 38 GW by 2032 — an addition of about 26 GW, or more than triple the current base. About a third of that 38 GW would be dedicated to racks of AI-specific chips, a signal of just how much of Microsoft's future build-out is aimed squarely at training and serv

The Stack

Slack's Surfaces turn chat into a dashboard builder

Slack is blurring the line between chat and the BI tool. Slackbot can now build interactive dashboards and reports inside your Slack chats. The feature, called Slackforce Surfaces, lets anyone in a workspace ask Slackbot for a tool in plain language — say, an arcade-themed visualization of AI token usage across sales, design, and engineering — and get back an interactive dashboard rather than a static chart. Surfaces can be shared with colleagues, pinned to channels, and commented on, and Slack

The Stack

OpenAI opens its Codex agent harness to every developer with the Agents API

Two days after the full-duplex voice model hit the API and a day after ChatGPT went free for 23 million US public servants, OpenAI is handing developers its agent machinery. The company introduced the Agents API in public beta: a managed service that exposes the same harness and infrastructure behind Codex — context management, tool use, multi-agent orchestration — through a single API call. OpenAI operates the harness and the sandbox; developers bring the tools and workflows. The mechanics are

See all