Good morning. Did you know that computer-related jobs have been the most common occupation for workers with H1-B visas for more than a decade?
As the Trump administration figures out what it wants to do with the U.S. “specialty occupation” visa program, three fast facts about it, courtesy of Pew Research:
—Amazon is the largest employer of H1-B workers. It represents 3% of total applications.
—New York City is the largest metro area for H1-B workers. Silicon Valley ranks third, behind the nation’s capital.
—India is, by far, the largest supplier of H1-B workers. Three in four folks approved for the program were born there.
The more you know. Today’s news below. —Andrew Nusca
Want to send thoughts or suggestions to Data Sheet? Drop a line here.
Meta is testing its first in-house AI training chip
At Meta, it’s apparently “move fast and fabricate things.”
The Facebook, Instagram, and WhatsApp parent is testing its first homegrown chip for training AI systems, according to a Reuters report.
An in-house chip would allow Meta to reduce its reliance on suppliers like Nvidia and reduce its enormous infrastructure costs. Meta expects to spend $65 billion in capital expenditures in 2025, largely allocated to the buildout of AI infrastructure.
The new chip is reportedly a dedicated “accelerator” meant only for AI-specific tasks. That narrow application would allow it to use less electricity than the advanced GPUs currently used to power AI systems.
The in-house chip’s deployment is small but Meta plans to rapidly expand production if the test proves successful. That’s made possible through an agreement with Taiwan’s TSMC, the world’s largest contract chipmaker and supplier to Apple, Nvidia, and other leading tech firms.
Meta’s top brass have previously said they want to use their own chips by 2026 to train AI systems. After successfully training recommendation systems, Meta would use the chip for generative AI products like its Meta AI chatbot.
For now, though, it’s baby steps. Meta once tried to make its own inference chip only to see it fail a small-scale deployment; it scrapped the idea in 2022 and placed a big order with Nvidia. —AN
Google owns a 14% stake in Anthropic but has little control
Court filings have revealed the size of Google’s stake in one of the most promising AI startups of its generation.
For competitive reasons, Google has been quiet about many of its external AI investments. But we now know that the search giant owns 14% of Anthropic, the AI startup known for its Claude family of large language models, according to a New York Times report.
That’s not all. Google’s investment gives the company little control over Anthropic, according to the filings. Google is limited to owning up to 15 percent of Anthropic and it holds no voting rights, no board seats, and no board observer rights at the smaller company.
The legal documents show that Google plans to invest another $750 million in convertible debt in September, bringing its Anthropic investment total to more than $3 billion.
The filings are part of a major Google antitrust case in the U.S. A federal court ruled in August that Google had unlawfully monopolized online search.
As a remedy, the Justice Department first proposed that the court force Google to sell AI investments that could compete with search. The DOJ has since softened that position to requiring only notification of additional AI investments.
Google isn’t the only prominent investor in Anthropic. Amazon has invested $8 billion in the company. Cisco, Qualcomm, Intuit, Salesforce, and Zoom have also pitched in. And a who’s who of leading VC firms—Lightspeed, Bessemer, Menlo, General Catalyst—have greased the skids. —AN
Who cyber-attacked Elon Musk’s X?
More on this week’s X outage: It may well have been a massive attack that succeeded because the platform didn’t keep itself safe from harm.
Figuring out who was behind what Elon Musk claims was a cyberattack will be as tricky as it always is with such things. By their nature, distributed denial-of-service (DDoS) attacks, which involve flooding a target’s web connection with data from a so-called botnet, are routed from all over the place.
Musk insists that Ukrainian IP addresses were used in this instance, but that may not mean much. A pro-Palestinian hacker outfit called Dark Storm has claimed responsibility, but there’s no confirmation of that.
There is, however, an interesting wrinkle to this episode. As Wired reports, security experts such as Kevin Beaumont reckon that X accidentally left itself exposed by failing to secure servers behind Cloudflare’s DDoS protection. It has since corrected that mistake.
Beaumont said the botnet in question is made up of compromised webcams and video recorders, mostly in the U.S. (These botnets are typically for hire, meaning any “hacker” can commandeer them for the attack of their choice.)
Eleven11bot, as this one is known, has recently been throwing record amounts of data around, degrading services in a variety of sectors. And, per Beaumont and others, there’s absolutely no reason to suspect Ukrainian involvement. —David Meyer
More data
—Apple’s new Macbook Air, reviewed. Welcome updates, but not worth an upgrade unless your existing machine is long in the tooth.
—Who will own Intel’s foundry? TSMC reportedly pitches Nvidia, AMD, Broadcom, Qualcomm about a joint venture.
—Oracle reports $130 billion in future contracts. The sum doesn’t include anything related to the ambitious Stargate Project.
—Starlink will come to India. India telecom giant Airtel signs a deal with Elon Musk’s SpaceX to get it done, pending government approval.
—Spotify says it paid $10 billion in royalties in 2024. That’s more than 60% of its total revenues.
—Uber quits acquisition of Foodpanda Taiwan after regulators block the $950 million deal.
—Waymo expands in Silicon Valley. The company’s robotaxis are ubiquitous in San Francisco, but were limited further down the peninsula.
Endstop triggered
