Today in AI — 5 March 2026

Today's AI news splits into two currents: a geopolitical tug-of-war over who builds AI and who controls it, and a relentless wave of model releases that makes last month's frontier look like yesterday's baseline. If you're building products, both currents affect you: the first shapes where your models can run, the second determines what they can do.

The Pentagon standoff reshapes market share

Anthropic's refusal to cross its safety red lines on the Pentagon deal is now a commercial event, not just a policy one. Claude hit #1 on the App Store while ChatGPT faces uninstalls, and OpenAI staff are publicly siding with Anthropic's position. When your competitor's employees say they respect you more than their own leadership, that's a brand problem money can't fix.

• Anthropic investors push to de-escalate Pentagon clash as supply chain risk designation looms — Reuters
• OpenAI staff fume over Pentagon deal as Altman admits it 'looked opportunistic and sloppy' — CNN
• Claude hits #1 on Apple's App Store as 1.5 million users sign up to leave ChatGPT — Vice

Four frontier models in one week

GPT-5.3 Instant triples context to 400K and cuts hallucinations. DeepSeek V4 goes trillion-parameter with a 1M context window on Chinese hardware. Qwen 3.5's 9B model beats gpt-oss-120B on key benchmarks. GLM-5 achieves the lowest hallucination rate ever tested by teaching a model to say "I don't know." The practical takeaway: if you're still hardcoding model choices, you're building on sand.

• OpenAI releases GPT-5.3 Instant with 400K context window and 27% fewer hallucinations — OpenAI
• DeepSeek poised to unveil trillion-parameter V4 model timed to China's NPC session — TechNode
• Qwen 3.5 small model series brings frontier-class reasoning to edge devices under Apache 2.0 — VentureBeat
• Zhipu AI's GLM-5 achieves 77.8% on SWE-bench with lowest hallucination rate of any model tested — LLM Stats

China's parallel stack accelerates

The NPC's 15th Five-Year Plan is expected to formalise what's already happening: China building an AI supply chain that doesn't touch American silicon. DeepSeek V4 running on Huawei and Cambricon chips is proof of concept. Proposed per-customer caps on H200 exports would accelerate the split further.

• China's NPC to unveil 15th Five-Year Plan with AI, robotics, and chip self-reliance at its centre — Reuters
• US weighs per-customer limits on NVIDIA H200 chip exports to China — Tech Startups

The infrastructure squeeze

NVIDIA is spending $4 billion to lock down photonic interconnects while RAM prices have surged 90% this quarter. If you're planning hardware for inference workloads, the cost assumptions from six months ago are already wrong.

• NVIDIA invests $4 billion in Lumentum and Coherent to secure photonic interconnect supply for AI — CNBC
• Global RAM shortage deepens as AI data centres consume 70% of memory production — Bloomberg

Platforms and the governance gap

Apple's white-labelled Gemini Siri ships this month, MCP is becoming the standard connector for agentic AI, and OpenClaw's security crisis shows what happens when agent ecosystems scale faster than security review. The pattern: platforms are moving fast, governance is not.

• Apple's Gemini-powered Siri overhaul enters final stretch ahead of March iOS 26.4 release — CNBC
• MCP adoption accelerates as enterprise agentic AI outpaces governance frameworks — TechCrunch
• OpenClaw supply chain crisis escalates with fake installer malware campaign — SecurityWeek

The signal for builders: model abstraction layers and security-first agent architectures are load-bearing infrastructure now, not optional extras.

---