Nvidia's New AI Model Arrives as Chip Rental Costs Rocket 114%

Nvidia launches unified AI model as B200 rental prices double; industry faces $740B spend and 'tokenmaxxing' phenomenon.

The cost of running artificial intelligence is spiraling, but Nvidia is betting a new software architecture can help ease the pain. The chipmaker has launched Nemotron 3 Nano Omni, a unified model that processes vision, audio, and text in a single system, just as the price of renting its B200 hardware has nearly doubled in six weeks.

The new model packs 30 billion parameters but activates only a fraction per computational step, allowing it to run on 25 gigabytes of memory. Nvidia claims this delivers up to nine times higher throughput on video tasks compared with fragmented, multi-stack approaches. Early adopters include Palantir and Foxconn, while Oracle and Dell are evaluating the system. The entire Nemotron-3 family has racked up over 50 million downloads in the past year.

That software push comes against a backdrop of soaring hardware costs. According to the Ornn Price Index, renting a single Nvidia B200 chip now costs nearly $5 per hour — a 114% jump in just six weeks. On the secondary market, a B200 chip is estimated to fetch around $500,000. Google is reportedly snapping up most available capacity, leaving other players scrambling. Analysts don't expect relief until the second half of 2026.

The price surge is reshaping the economics of the entire AI industry. Nvidia Vice President Bryan Catanzaro has pointed to a new phenomenon where some companies' AI compute costs already exceed their payroll expenses. Industry insiders have coined the term "tokenmaxxing" for this dynamic. Tech giants are on track to spend $740 billion on AI infrastructure this year, straining balance sheets. Uber has reportedly exhausted its entire 2026 AI budget already.

Even the biggest names are feeling the heat. OpenAI is expected to burn through $25 billion this year and missed its internal revenue targets in the first quarter. The ChatGPT developer subsequently ended its exclusive partnership with Microsoft Azure and is now shopping for affordable compute capacity from other cloud providers.

Nvidia's own financials reflect the broader frenzy. On May 20, Chief Financial Officer Colette Kress will present first-quarter results for fiscal 2027, with management targeting around $78 billion in revenue. The company has locked in firm delivery commitments worth $95.2 billion, including massive capacity allocations for OpenAI and Meta.

The stock is trading at €180.34, just shy of its 52-week high, with a 30-day gain of nearly 20%. On a monthly basis, the advance is roughly 25%. CEO Jensen Huang continues to frame AI as the creator of entirely new industries, and the next hardware cycle — the Vera-Rubin chips — is slated for late 2026. Until then, the current Blackwell generation remains the expensive standard.