AI Infrastructure
GPUs, silicon, data centres and the data platforms underneath.
Arcee AI's Trinity: 400B open model takes on Llama 4
Arcee AI released Trinity, a 400B-parameter Apache-licensed open foundation model. The 30-person startup claims it rivals Llama 4 Maverick and other leading open models, trained in six months for $20M using 2,048 Nvidia Blackwell B300 GPUs.
How GPU Price-Performance Doubles Every 2.5 Years
Analysis of GPU price-performance trends using a dataset of 470 graphics cards from 2006 to 2021, finding that FLOP/s per dollar doubles every ~2.5 years, with variations for ML GPUs and top-performing models.
Axelera AI Banks $68M Series B to Push Metis Chips Beyond the Edge
Dutch AI chip maker Axelera AI closed a $68 million Series B, bringing total funding to $120 million, as it ramps Metis production and explores data center silicon.
Borderless Founder: How AI Levels the Playing Field for International Founders
a16z's Borderless Founder network initiative supports immigrant and international founders, arguing that having one foot in your home country and one foot in Silicon Valley provides a competitive advantage in the AI era
AI Training Compute Scales 4-5x Annually from 2010 to 2024
Verified trends in training compute growth for AI models from 2010 through 2024, including frontier models, language models, and leading companies' top models, all showing consistent 4-5x/year growth.
Stargate AI Data Center Initiative Projected 9 GW Capacity by 2029 Across Seven US Sites
The $500 billion Stargate AI infrastructure project's planned capacity, site locations, construction timelines, and power configurations across seven US data center locations.
Liquid AI's LFM2.5-2.6B: Small Models, Big Edge Promises
Liquid AI released LFM2.5-2.6B, a 2.6 billion parameter open-weight model optimized for edge device deployment. Running on CPUs from Raspberry Pi to Apple Silicon without cloud dependency, the model targets agentic workloads with native tool calling, a 128K context window, and a revenue-gated open license.
Mirendil's $100M Google Cloud Deal Powers AI Developer Tools Startups' Self-Improving AI Ambitions
TechCrunch exclusive: Mirendil signs $100M+ multiyear Google Cloud partnership to fund recursive self-improving AI research, giving the startup access to TPUs, Nvidia GPUs, and managed training clusters.
Lumilens Emerges from Stealth with $700M Series C to Solve AI Data Center Connectivity
Lumilens emerged from stealth with $700M in Series C funding at a $5.51B valuation, already shipping optical interconnect products into hyperscaler AI data centers under a multi-billion dollar customer agreement. Founded in early 2024 by Ankur Singla to solve GPU connectivity bottlenecks in AI infrastructure.
Higher Benchmark Scores Don't Mean Lower Cost: Qwen 3.8-Max, Claude Opus 5 and Why Cost Allocation for Shared GPU Clusters Is the Real Metric
Analysis of why raw benchmark scores don't predict actual AI model costs - examining token budgets, time limits, and how reasoning models can appear better or worse depending on measurement framework. Key finding: cost per successful task (total spend including failed attempts divided by tasks that pass acceptance) is the metric that matters for routing decisions.
OpenAI's First-Ever Influencer Retreat Stirs Backlash Over AI's Growing Pains
TechCrunch report on the controversy surrounding OpenAI's first influencer brand trip "Summer Club" held at a luxury New York resort, where creators faced intense social media criticism for attending while AI data centers face growing environmental scrutiny.
AI Developer Tools Startups India Investments: The Real Shift Isn’t in Code, It’s in Control
VCs are pivoting from AI hype to infrastructure: India’s developer tools, HCL’s sovereign datacenters, and the quiet revolution in execution over models.