AI digest: Infrastructure gets serious
Big tech splashes $725 billion on AI infrastructure while NVIDIA pushes speculative decoding speeds and the Pentagon builds classified AI networks.
The money’s flowing and the infrastructure’s getting proper attention. Here’s what caught our eye this week.
Big tech commits $725 billion to AI infrastructure
Google, Amazon, Microsoft, and Meta are spending a combined $725 billion on AI data centres, chips, and infrastructure this year. That’s not marketing budget or R&D fluff, that’s proper hardware money. The scale suggests they’re betting everything on AI becoming genuinely transformative, not just a fancy chatbot layer.
NVIDIA cracks the inference speed problem
NVIDIA’s new research shows speculative decoding in NeMo RL delivers 1.8x speedup at 8B parameters, with projections of 2.5x at 235B scale. This matters because inference costs are still the biggest barrier to deploying large models in production. Lossless acceleration means you get the speed without sacrificing quality.
Pentagon builds classified AI networks
Eight tech giants including NVIDIA, Microsoft, and AWS just signed deals with the Pentagon to deploy AI on classified military networks. Anthropic notably isn’t on the list after rejecting usage terms and getting flagged as a security risk. The “AI-first fighting force” rhetoric suggests this isn’t just about better spreadsheets.
Meta turns AI models into data scientists
Meta’s new Autodata framework lets AI models become autonomous data scientists for creating high-quality training data. This could solve the data quality bottleneck that’s plaguing most AI projects. Self-improving data pipelines feel like the logical next step after self-improving code.