AI News Roundup: Is Nvidia’s New Chip a 1000-watt Power Hog?

AI Business brings you the latest news and insights from across the AI world.

To keep up to date with coverage of all things AI, subscribe to the AI Business newsletter to get content straight to your inbox and follow the AI Business Podcast on Apple and Spotify.

A glimpse of Nvidia’s next flagship chip?

Nvidia is reportedly working on a GPU that consumes 1000 watts of power - and does not need liquid cooling to keep it in check.

Nvidia’s B100 chip is expected to launch later this year and will likely usurp the H100 in terms of performance.

The Register reports that a Dell Technologies executive may have revealed details of the next-gen hardware on an earnings call.

COO Jeff Clarke said that liquid cooling will not be needed to handle chips that consume vast amounts of energy, specifically referring to a unit called the “B200.”

No B200 is expected from Nvidia - the B100, or Blackwell series of GPUs were teased late last year. The units use TSMC's 3-nanometer process which is designed to reduce power reduction and yield performance increases.

The Register reports that if Clarke is correct, Nvidia’s new AI accelerator would consume 42% more power than the H100.

AI Business has contacted Nvidia for clarification.

Google AI secrets secretly sold to Chinese companies

A Chinese software engineer has been charged with stealing AI trade secrets from Google.

Ding Linwei joined Google in 2019 to work on its supercomputing data centers. The U.S. Justice Department alleges Linwei passed along files containing details of Google’s AI infrastructure to entities based in China.

Linwei allegedly supplied more than 500 documents that held “confidential information” on Google AI models, applications and hardware infrastructure between May 2022 and March 2023.

He is then alleged to have founded his own AI startup, acting as CEO, without Google knowing. The Chinese startup allegedly admitted it had “experience with Google's ten-thousand-card computational power platform; we just need to replicate and upgrade it.”

If found guilty, he faces up to 10 years in prison and fines of up to $250,000 for each count.

Attorney General Merrick Garland said in a statement the Justice Department “will not tolerate the theft of artificial intelligence and other advanced technologies that could put our national security at risk.”

Musk sued by former Twitter CEO

Elon Musk is being sued by the former Twitter CEO and other executives who claim they have not been paid millions of dollars in severance.

Former CEO Parag Agrawal, CFO Ned Segal, general counsel Sean Edgett and chief legal officer Vijaya Gadde claim Musk refused to pay them severance as “revenge” after he was forced to purchase the social media platform for $44 billion in 2022.

The former executives are trying to recover more than $128 million and claimed Musk “made up fake cause” to fire staff without reason.

“He claimed in his termination letters that each plaintiff committed ‘gross negligence’ and ‘willful misconduct’ without citing a single fact in support of this claim,” the lawsuit reads.

The action from former Twitter leaders is the latest in a string as former employees seek severance. Musk reduced the headcount at the company to just 20% of what it was when he took over.

Mistral models are coming to Snowflake

AI startup Mistral has extended its latest string of partnerships to bring its models to Snowflake.

Mistral models including the new Mistral Large will be accessible to Snowflake customers via its Data Cloud platform.

Snowflake users will also gain access to Mixtral 8x7B and Mistral 7B, which is optimized for low latency applications.

Snowflake’s venture arm has also invested in Mistral as part of its series A funding round.

Mistral’s deal with Snowflake follows recently struck partnerships with Microsoft, Amazon and Capgemini.

Cohere models to power Accenture solutions

Cohere is teaming up with Accenture to bring its AI models to enterprises.

Accenture clients will get access to Cohere’s large language model Command as well as its search technology, Embed.

Cohere will work with Accenture to offer clients customized AI solutions to meet their specific business needs.

An early project saw Cohere’s Command model power a knowledge agent for Accenture’s finance and treasury teams to help detect financial variances.

Lan Guan, chief AI officer at Accenture, said in a blog post that the partnership will “accelerate how we can help companies incorporate AI into their operations at scale, prioritizing privacy and security.”

DARPA is funding AI chip projects

The U.S. Defense Department’s research agency is investing $78 million in AI chip research projects.

DARPA is looking to fund hardware projects to power the next generation of AI computing.

Among those receiving funding is a project at Princeton University building chips that use less energy.

Naveen Verma, professor of electrical and computer engineering, is leading the project through his startup, EnCharge AI, which will attempt to find out how fast a compact and power-efficient chip can get.