|
May 16, 2025
|
|
|
|
Supported by
|
|
|
|
|
TGIF! Meta Platforms delays the largest version of Llama 4. The Trump administration and the United Arab Emirate unveil a plan for a multi-gigawatt AI data center in Abu Dhabi. Nvidia is planning to build a new research and development center in China.
|
|
|
|
Meta Platforms has delayed the largest version of its flagship large language model Llama 4 to as late as this fall—or even later—over performance problems, according to a person familiar with the matter. Llama 4 has encountered other delays in the past. Meta previously pushed back Llama 4’s release date because, during its development, the model didn’t perform as well on technical benchmarks as the company had hoped. The company has also delayed a reasoning version of Llama 4, though the reason for that delay couldn’t be learned. Meta released smaller versions of Llama 4, known as Scout and Maverick, in April. Those models had a rocky debut, in part because Meta uploaded an experimental version of Llama 4 Maverick to the popular LLM leaderboard LMArena, which performed better than the version actually released to the public. A spokesperson for Meta did not immediately respond to a request for comment. The Wall Street Journal first reported on a delay for the largest version of Llama 4, known as Behemoth.
|
|
|
|
The Trump administration and the United Arab Emirates on Thursday announced a large data center campus for artificial intelligence in Abu Dhabi that aims to power American technology. President Donald Trump and Emirati president Sheikh Mohammed bin Zayed discussed the plan in front of a miniaturized model of the data center park, according to the White House and images from the event. The announcement came after Trump rescinded Biden-era rules that would have limited the UAE’s ability to buy large amounts of advanced AI chips over concern that U.S. adversaries such as China could access them there. Abu-Dhabi-based technology firm G42, which has financial or commercial ties to U.S. firms such as Microsoft and OpenAI, will build the data centers and operate them with “several U.S. companies” that weren’t named.
G42 initially plans to build a 1 gigawatt AI data center that would later consume 5 GW of power and span 10 square miles, though a timeline for the project wasn’t disclosed. (That’s enough power for multiple large cities and would likely cost hundreds of billions of dollars.) Companies such as OpenAI and Meta Platforms have plotted developing similar facilities in the U.S. According to the announcement, U.S. companies will use the data center capacity to offer fast AI cloud services to people who live within 2,000 miles of the UAE. Earlier this week, The Information reported that OpenAI was preparing to announce a new data center in the UAE as part of its Stargate data center effort with SoftBank. It isn’t clear whether Thursday’s announcement is related.
|
|
|
|
Nvidia is planning to build a new research and development center in Shanghai as part of efforts to maintain its competitive edge in China amid declining sales due to stricter U.S. export controls, the Financial Times reported, citing two people with knowledge of the matter. During a visit to China last month, Nvidia CEO Jensen Huang discussed the plan with Shanghai’s mayor, Gong Zheng, according to the FT. The Shanghai government has expressed preliminary support for the plan, but Nvidia will also need approval from U.S. authorities to go ahead with it. The move is Nvidia’s latest effort to maintain its foothold in China, one of its largest markets for artificial intelligence chips where the U.S. company is facing growing competition from local AI chip developers such as Huawei. Huawei is accelerating its effort
to build a competing ecosystem for its AI chips that could pose a significant threat to Nvidia. The new R&D center in China would work on projects that address the specific needs of Chinese customers while navigating complex technical requirements under U.S. regulations. Nvidia plans to keep its core design and production activities outside China to eliminate concerns about intellectual property transfer. The company told the FT that it is not sending any GPU designs to China.
|
|
|
|
Upstart AI cloud provider CoreWeave on Wednesday signed a new deal to rent Nvidia chips to OpenAI worth $4 billion through 2029, CoreWeave said in a securities filing. The deal is in addition to one worth $11.9 billion struck with OpenAI in March. That deal runs through 2030. As part of the deal, CoreWeave isued 8.75 million shares of its stock to OpenAI, valued at $350 million at the time of CoreWeave’s IPO in March. The shares are now worth $626 million. CoreWeave has become a go-to AI cloud provider because it has access to Nvidia chips, allowing it to offer cloud capacity when other cloud services have capacity constraints.
|
|
|
|
Artificial intelligence model developer Cohere fell well short of revenue projections the company prepared in early 2023, missing its target for last year by 85%. In a pitch to investors two years ago, the company forecast making around $450 million in annualized revenue by the end of last year. But as of February, it was on pace to generate just $70 million in annualized revenue, The Information reported Thursday. The company said it was overly optimistic about how quickly enterprises would adopt AI, and that the market for business AI tools grew more slowly than the consumer market. Cohere’s figures highlight how AI developers that didn’t focus on selling revenue-generating applications fast enough have been left in the dust. Cohere could be a bellwether for other startups, including Mistral and Reka, that raised
large amounts of money in recent years to capitalize on the hype around new AI models.
|
|
|
|
YouTube CEO Neal Mohan says revenue on YouTube Shorts is growing and, when measured by one metric, is comparable to revenue generated by longer videos on the main YouTube service. Speaking at a MoffettNathanson conference on Thursday, Mohan said that YouTube Shorts has reached parity with traditional YouTube videos when measured by the revenue each format generates per hour watched. This is true in the U.S. and some other countries, hes said. Of course, as YouTube Shorts are by definition much shorter than regular videos, YouTube has more opportunities to insert ads within the Shorts feed—something which Mohan acknowledged. But Shorts has momentum, he said, with views up 20% year over year during Q1. Mohan’s comments came a day after YouTube hosted its annual Brandcast event for advertisers in New York, where it
showcased top creators and podcasters on the service, as well as a new deal with the NFL to live stream a full game from Brazil in September. YouTube executives boasted about the unique nature of its video service, which can offer everything from live sports to podcasts to shows from creators to TikTok-like shorts—all in one place.
|
|
|
|
| |
