AT&T SHIFTS MORE AI WORKLOADS TO OPEN-SOURCE MODELS AT&T says open-source models now handle 40% of employee AI queries and expects that to rise to 60%-70% as it looks to keep spending on OpenAI and Anthropic models roughly flat. The company says model routing through LiteLLM cut AI coding costs by 56% with only a 2% drop in performance quality. AT&T’s internal AI platform now processes around 45 billion tokens per day, using models including Nvidia Nemotron, Meta Llama and Google Gemma alongside frontier models for more complex tasks.

FT reports $META and BlackRock’s $14B, 1GW Sopaipilla data center in Texas carries only partial insurance coverage, reflecting how difficult and expensive it has become to fully insure gigawatt-scale AI infrastructure. The project has about $450M of property coverage once operational, $645M of terrorism coverage and additional protection for construction delays and liability. Rather than insuring against a complete loss, lenders are relying on modeled “probable maximum loss” scenarios, an approach becoming more common as data center projects grow into the tens of billions. Importantly, the debt still received strong investment-grade ratings, including A+ from S&P and AA- from Fitch and KBRA. BlackRock has also argued that severe weather risk in El Paso is relatively low and that modern data centers are unlikely to suffer a total loss even in a major catastrophe.