NVIDIA new chips to cut costs by 35x as coding tools grab half of AI related searches

NVIDIA just put out on its newest GB300 NVL72 systems. They can handle 50 times more work per megawatt of electricity compared to the older Hopper platform. That means costs drop by 35 times for each piece of information processed.

Signal65 did separate testing on the GB200 NVL72 and found it processes more than 10 times the information per watt, cutting costs to a tenth of what they were.

The improvements keep coming. NVIDIA’s TensorRT-LLM library upgrades alone gave the GB200 a five-times performance boost in just four months for tasks needing instant responses. Teams working on Dynamo, Mooncake and SGLang tools are pushing efficiency even higher.

These AI tools break down if there’s lag time or they can’t remember enough context. Companies need them to actually work in real business situations, not just demos.

Artificial intelligence tools that write code and work as digital helpers now account for nearly half of all AI-related searches, up from just 11% a year ago.

The numbers come from OpenRouter’s State of Inference report and show how fast things are changing in this space.

This sudden jump has companies scrambling to build hardware that can keep up. These AI assistants need to respond right away and remember context from entire software projects, which puts serious demands on computing power.

The money involved is huge. The AI agent market was worth 4.92 billion dollars in 2024. Estimates put it at 6.016 billion dollars for 2025, then ballooning to 44.97 billion dollars by 2035. That’s 22.28% growth every year for the next decade. Banks, hospitals, stores and factories are the early adopters.

Businesses are putting these agents into customer management systems, planning tools and security setups to save money and get more done. What started as optional tech is becoming basic infrastructure.

As reported by Cryptopolitan, Alibaba just launched Qwen3.5 targeting China’s market, claiming 60% lower processing costs than before. The model can look at screens and do tasks across phones and computers. It’s going head-to-head with ByteDance’s Doubao app, and there’s a DeepSeek update coming too.

OpenAI hired Peter Steinberger on the 15th. He built OpenClaw, an open-source AI agent. CEO Sam Altman said Steinberger will lead work on next-generation personal agents and called him a genius with great ideas about smart assistants that can get useful stuff done.

The talent crisis nobody can solve

94% of business leaders say they’re short on AI skills. By 2028, 44% expect to still have 20 to 40% shortages. Workera says these gaps could cost the global economy 5.5 trillion dollars in 2026 from delayed products, quality issues and lost sales.

Right now demand for AI talent beats supply 3.2 to one worldwide. AI jobs pay 67% more than regular software positions. Yet 85% of office workers are learning about this stuff on their own time, and 83% say they’re mostly teaching themselves instead of getting formal training.

When companies buy AI tools from specialized vendors, they succeed 67% of the time. Internal builds only work about a third as often.

Salesforce saw 119% agent growth in early 2025 and crossed 500 million dollars in recurring revenue for these products. They added 6,000 enterprise customers in three months.

Businesses will probably buy solutions instead of building them. That points to the market concentrating around a few big players who can deliver what actually works.