Today in AI — 7 March 2026

Three separate teams shipped models this week that do the same work with roughly half the resources. The arms race is shifting from scale to efficiency, and the implications for cost, deployment, and what you can run locally are immediate.

Efficiency is the new frontier

GPT-5.4's tool search cuts agent token usage by 47% by loading only the definitions a model needs. OLMo Hybrid matches its predecessor's accuracy with 49% fewer training tokens. Phi-4-reasoning-vision skips chain-of-thought when it doesn't help. Same conclusion from three different directions: the next performance gains come from spending compute more carefully.

• OpenAI launches GPT-5.4 with native computer use and tool search that cuts agent tokens by 47% — TechCrunch
• AI2 releases OLMo Hybrid, achieving same accuracy as its predecessor with 49% fewer tokens — Allen Institute for AI
• Microsoft open-sources Phi-4-reasoning-vision-15B, a model that decides when thinking helps — Microsoft Research

The Pentagon standoff escalates

The DoD formally designated Anthropic a supply chain risk — the first time an American company has received that label. Anthropic will challenge in court. Meanwhile, Claude is adding over a million users a day. The designation is meant to freeze Anthropic out of government contracts, but it's doing more for consumer loyalty than any marketing spend could.

• Pentagon formally designates Anthropic a supply chain risk; company will challenge in court — NPR
• Claude adds over a million new users daily as Anthropic's Pentagon stand drives consumer surge — TechCrunch

Infrastructure economics

Oracle and OpenAI scrapped the Stargate data centre expansion after financing stalled, with Meta circling the site. Oracle is also preparing mass layoffs as AI infrastructure spending hits $50 billion for fiscal 2026. Foxconn, meanwhile, posted record revenue from building the servers everyone else is buying. The returns from AI infrastructure are concentrating in hardware suppliers, not the companies writing the cheques.

• Oracle and OpenAI scrap Stargate Texas data center expansion as Meta circles the site — Bloomberg
• Oracle planning thousands of job cuts as AI data center spending creates cash crunch — Bloomberg
• Foxconn reports 21.6% revenue surge to record $41.9B as AI server demand accelerates — Data Center Dynamics

Developer ecosystem shifts

GitHub's Octoverse data shows AI coding tools creating convenience loops: languages that work well with AI attract more developers, generating more training data, which improves AI support further. TypeScript surged 66% to the top spot. OpenClaw passed React in GitHub stars as demand for self-hosted AI agents accelerates. And OpenAI shipped ChatGPT for Excel with live financial data feeds, a sign that AI copilots are moving beyond code into domain-specific tooling.

• GitHub data reveals AI tools creating 'convenience loops' that reshape developer language choices — InfoQ
• OpenClaw surpasses React with 250,000 GitHub stars as self-hosted AI assistants surge — Yahoo Finance
• OpenAI launches ChatGPT for Excel with real-time financial data from FactSet, S&P Global, and more — OpenAI

Google's mixed week

Google gave developers three days to migrate off the Gemini -latest alias, shorter than its own 14-day deprecation policy. Pin your model versions. Separately, DeepMind launched an equity-free robotics accelerator for European startups, with applications closing 25 March.

• Google switches Gemini -latest alias to 3.1 Pro, giving developers 3 days to migrate — Google
• Google DeepMind launches equity-free robotics accelerator for European startups — EdTech Innovation Hub

The cost of running agents is dropping fast; the stability of the platforms you run them on is not. Build for portability.

---