Intel raises capital. Stroustrup to SIG. Meta pushes AI to the edge.
Intel raises $20bn as investors back its foundry ambitions, Bjarne Stroustrup joins Susquehanna as a Technical Fellow, and Meta releases a smaller open-weight model designed to run locally.
MARKETS
Intel raises $20bn to fund its semiconductor comeback
Intel upsized its planned equity offering from $15bn to $20bn, pricing 210.5 million shares at $95. The capital will support the company's manufacturing and foundry ambitions, including advanced process technology and packaging. Read more: Reuters
Bjarne Stroustrup joins Susquehanna as Technical Fellow
The creator of C++ has joined Susquehanna International Group as a Technical Fellow, helping shape and optimise the firm's codebase. Read more: Imperial College London
Compute becomes a financial market
Nvidia and a group of major financial institutions are developing financing platforms that could mobilise more than $500bn for AI infrastructure, with Nvidia potentially backstopping as much as $125bn. Separately, CME is preparing compute futures, pointing towards a market where GPU capacity can be priced, financed and hedged like a commodity. Read more: Reuters / Barron's
MACHINES
Intel and Together AI build inference cloud together
IBM and Together AI have agreed a $240m, multi-year deal to build a large AI inference cluster on IBM Cloud. Read more: Reuters
Memory remains the bottleneck in AI inference
As models become larger and inference workloads become more demanding, memory capacity and bandwidth are increasingly determining how efficiently accelerators can be used. The bottleneck has changed from simply acquiring GPUs to feeding them with enough data at acceptable cost and latency. Read more: HPCwire
Custom inference silicon moves further into the mainstream
The inference market is opening up beyond general-purpose GPUs as specialised architectures target particular combinations of latency, throughput and efficiency. The emergence of wafer-scale systems, custom accelerators and increasingly sophisticated memory architectures suggests the long-term inference stack will look considerably more heterogeneous than the training stack. Read more: TechCrunch
MODELS
Meta releases a smaller open-weight model designed to run locally
Meta has released Muse Glimmer, a smaller open-weight model designed to run directly on consumer hardware rather than requiring a cloud connection. The move reinforces Meta's bet that capable AI should increasingly be distributed across PCs and other devices, rather than concentrated entirely in hyperscale data centres. Read more: Reuters
DeepSeek releases a new model as pricing becomes part of the competition
DeepSeek has released its V4 Pro model, with pricing reportedly as much as 14 times higher than its previous generation. Token pricing, inference efficiency and willingness to pay are becoming part of the model race. Read more: DeepSeek
AI scientists move from answering questions to reproducing research
Inherent has introduced Faraday, a 27B-parameter AI Scientist agent designed to replicate scientific research. Rather than evaluating models primarily through questions and answers, these systems are being tested on whether they can execute multi-step research workflows and produce reproducible results. Read more: Dealroom
