Atop a newly-completed, 3.5-million-square-foot building that stands on 1,100 acres in the Arizona desert north of Phoenix is a giant logo of a microchip wafer and the letters TSMC.
CNBC first visited the fab in 2021, not long after TSMC broke ground. TSMC initially announced the plant would cost $12 billion and pump out 5-nanometer chips by the end of 2024. Three years later, that price tag has soared to $20 billion and full production is delayed until 2025.
Instead, the fab is in pilot production, making sample wafers and sending them to customers for verification. TSMC has committed to building two more fabs on the site by the end of the decade, for a total investment of $65 billion.
The project is “dang near back on the original schedule,” TSMC Chariman Rick Cassidy told CNBC during an exclusive first look at the completed fab in November.
“When we came to the U.S., we knew we were going to go through a learning process,” Cassidy said. “Whether it was permitting, learning how to work with the trades, learning how to work with the unions, local labor laws. Lots of learnings that went on. Now we’ve overcome those.”
TSMC chairman Rick Cassidy shows CNBC’s Katie Tarasov around its newly completed fab on November 7, 2024, where it will make advanced chips on U.S. soil for the first time.
Andrew Evers
With the help of some 2,000 employees, the fab is set to make more advanced chips than originally planned. It will produce 4-nanometer chips, at a rate of 20,000 wafers per month, TSMC said.
“We’ve seen TSMC be able to kind of name its price, and everyone’s going to pay it because right now it’s the dependability and the quality that is needed,” said Daniel Newman, CEO of The Futurum Group.
‘On par with our Taiwan compatriots’
The fab’s yields are anticipated to be “right on par with our Taiwanese compatriots,” Cassidy said. Still, some 92% of the world’s most advanced chips are currently made by TSMC’s Taiwan fabs, so the U.S. is far from self-reliant.
“It’s difficult or impossible for the U.S. or any country to be fully self-sufficient in everything that they need to build semiconductors,” said Stacy Rasgon of Bernstein Research. “That’s a pipe dream.”
Despite being the birthplace of microchips in the 1950s and remaining a top chip design hub, the U.S. now manufactures only 10% of the world’s chips and none of the most advanced ones. When supply chain chaos collided with booming demand for consumer electronics during the pandemic, the resulting chip shortage exposed the big risks of relying on outsiders for such a critical technology.
In the event of aggression between China and Taiwan, an earthquake or some other event that impacts Taiwan for a period of time, “the entire market, the entire world could suffer from lack of availability of leading edge nodes,” Newman said.
TSMC’s first fab in Arizona, shown in November 2024, where it will make advanced chips on U.S. soil for the first time.
TSMC
Other fears surfaced when President-elect Donald Trump expressed opposition to the $52 billion CHIPS Act in October during his campaign. Weeks later, the U.S. Commerce Department finalized TSMC’s allotted $6.6 billion from the bipartisan bill.
“Repealing the CHIPS Act would make Americans less safe,” Commerce Secretary Gina Raimondo told CNBC in an interview, adding that she doesn’t think the incoming administration would repeal it.
“I just don’t think they’ll do that,” Raimondo said.
Talks with TSMC about bringing advanced chip production to the U.S. began in 2018, during Trump’s first term.
“I set up a phone call between the chairman of TSMC and the head of Apple,” said Wilbur Ross, who was commerce secretary at the time. “Apple became very strongly supportive of the idea of TSMC coming.”
Rose Castanares, a 26-year company veteran and now president of TSMC Arizona, was also involved with the early conversations. Customers “wanted supply resilience,” Castanares said.
Relying on chips from Asia has also complicated the U.S. drive for technological dominance. That’s why President Joe Biden hit the chip industry with a complex web of export controls meant to keep China from pulling ahead with advanced tech.
In October, some TSMC chips were spotted in Huawei devices, despite bans on selling to the Chinese company.
“This problem is as old as time,” Newman said. “There’s a lot of complex rerouting of goods to get gray market to different countries that have limited access to leading edge or the most advanced technology.”
TSMC Arizona president Rose Castanares with CNBC’s Katie Tarasov in the newly completed fab on November 7, 2024, where it will make advanced chips on U.S. soil for the first time.
Andrew Evers
Workers, water and power
Nearby in Chandler, Arizona, Intel is also building two huge fabs.
The U.S. company has a far different business model, designing and manufacturing its own chips, while TSMC only makes chips for others. The relationship between the two companies is solid, Cassidy said.
“We meet with [Intel] weekly and the feedback is we’re helping them increase their ranks,” Cassidy said. “We’re helping them train on the most advanced stuff, so I think they’re pretty happy with what we’re doing.”
Both companies have delayed the timelines for full production at their new Arizona fabs. But where TSMC has remained the uncontested leader in advanced chips, Intel has stumbled time and again.
The two will also be competing for a scarce resource in the U.S. chip industry: workers.
“When we finished the construction of this fab, it was really the first advanced manufacturing fab that had been built in the United States for at least 10 years. Semiconductors is a very, very tough technology,” TSMC’s Castanares said. “The experience is just not here in the United States.”
At the beginning of the project, TSMC sent some 600 engineers to train in Taiwan. Process integration engineer Jeff Patz spent 18 months there starting in 2021.
“The purpose was to go and actually make things, right? And learn how they’re made,” Patz said. “You have to have a kitchen to cook.”
TSMC has also brought experts over from Taiwan on 3-year temporary assignments. TSMC plans to hire at least 6,000 workers by the time all three fabs are completed.
“For engineers, we are actively recruiting at universities in Arizona and all across the U.S.,” Castanares said. Arizona State University “even has what they call a TSMC day.”
Water is another scarce resource needed in abundance.
With Taiwan recently facing its worst drought in nearly a century, TSMC is no stranger to recycling the massive amount of water it needs to make chips. TSMC will take 4.7 million gallons of water daily to run the first Arizona fab, but it will bring that demand down to 1 million gallons a day, in part by recycling some 65% of that, the company said.
It also takes a massive amount of power to make chips.
TSMC built solar on site, but it’s not nearly enough to cover the 2.85 gigawatt-hours per day needed to run the first fab. That’s equivalent to the power used by roughly 100,000 U.S. homes. TSMC said it’s purchasing renewable energy credits to offset that. But amid the AI-fueled data center boom, Arizona’s largest utility warned that it could run out of transmission capacity before the end of the decade.
That’s also when TSMC plans to start production at its third Arizona fab, which Cassidy said is “probably going to be 2 nanometer and more advanced.”
TSMC is also broadening its global footprint. It opened its first fab in Japan in February and broke ground on an $11 billion fab in Germany in August.
Within the U.S., Cassidy said TSMC is also likely to keep expanding.
“There’s room for lots of fabs,” Cassidy said.
Watch the full video for never-before-seen footage inside TSMC’s Arizona fab: https://cnbc.com/video/2024/12/12/inside-tsmcs-new-chip-fab-where-apple-will-make-chips-in-the-us
Chief executive officer at Palo Alto Networks Inc., Nikesh Arora attends the 9th edition of the VivaTech trade show at the Parc des Expositions de la Porte de Versailles on June 11, 2025, in Paris.
Earnings per share: 93 cents adjusted vs. 89 cents expected
Revenue: $2.47 billion vs. $2.46 billion expected
Revenues grew 16% from $2.1 billion a year ago. Net income fell to $334 million, or 47 cents per share, from $351 million, or 49 cents per share in the year-ago period.
Palo Alto’s Chronosphere deal is slated to close in the second half of its fiscal 2026. The cybersecurity provider is also in the process of buying Israeli identity security firm CyberArk for $25 billion under CEO Nikesh Arora‘s acquisition spree.
He told investors in an earnings call that Palo Alto is making this simultaneous acquisition to address the fast-moving AI cycle.
“This large surge towards building AI compute is causing a lot of the AI players to think about newer models for software stacks and infrastructure stacks in the future,” he said.
Palo Alto guided for revenues between $2.57 billion and $2.59 billion in the second quarter, the midpoint of which was in line with a $2.58 billion estimate. For the full year, the company expects $10.50 billion to $10.54 billion, versus a $10.51 billion estimate.
Capital expenditures during the period were much higher than expectations at $84 million. StreetAccount expected $58.1 million. Remaining purchase obligations, which tracks backlog, grew to $15.5 billion and topped a $15.43 billion estimate.
The rise of artificial intelligence has also stirred up increasingly sophisticated cyberattacks and contributed to tools for customers. The Santa Clara, California-based company has infused AI into its tools and launched automated AI agents to help fend off attacks in October.
Tesla CEO Elon Musk (L) talks with Nvidia CEO Jensen Huang during the U.S.-Saudi Investment Forum at the Kennedy Center on Nov. 19, 2025 in Washington, DC.
Win McNamee | Getty Images
Nvidia and xAI said on Wednesday that a large data center facility being built in Saudi Arabia and equipped with hundreds of thousands of Nvidia chips will count Elon Musk’s artificial intelligence startup as its first customer.
Musk and Nvidia CEO Jensen Huang were both in attendance at the U.S.-Saudi Investment Forum in Washington, D.C.
The announcement builds on a partnership from May, when Nvidia said it would provide Saudi Arabia’s Humain with chips that use 500 megawatts of power. On Wednesday, Humain said the project would include about 600,000 Nvidia graphics processing units.
Humain was launched earlier this year and is owned by the Saudi Public Investment Fund. The plan to build the data center was initially announced when Huang visited Saudi Arabia alongside President Donald Trump.
“Could you imagine, a startup company approximately 0 billion dollars in revenues, now going to build a data center for Elon,” Huang said.
The facility is one of the most prominent examples of what Nvidia calls “sovereign AI.” The chipmaker has said that nations will increasingly need to build data centers for AI in order to protect national security and their culture. It’s also a potentially massive market for Nvidia’s pricey AI chips beyond a handful of hyperscalers.
Huang’s appearance at an event supported by President Trump is another sign of the administration’s focus on AI. Huang has become friendly with the president as Nvidia lobbies to gain licenses to ship future AI chips to China.
When announcing the agreement, Musk, who was a major figure in the early days of the second Trump administration, briefly mixed up the size of the data center, which is measured in megawatts, a unit of power. He joked that plans for a data center that would be 1,000 times larger would have to wait.
“That will be eight bazillion, trillion dollars,” Musk joked.
AMD will provide chips that may require as much as 1 gigawatt of power by 2030. The company said the chips that it would provide are its Instinct MI450 GPUs for AI. Cisco will provide additional infrastructure for the data center, AMD said.
Qualcomm will sell Humain its new data center chips that were first revealed in October, called the AI200 and AI250. Humain will deploy 200 megawatts of Qualcomm chips, the company said.
Yann LeCun, known as one of the godfathers of modern artificial intelligence and one of the first AI visionaries to join the company then known as Facebook, is leaving Meta.
LuCun said in a LinkedIn post on Wednesday that he plans to create a startup that specializes in a kind of AI technology that researchers have described as world models, analyzing information beyond web data in order to better represent the physical world and its properties.
“I am creating a startup company to continue the Advanced Machine Intelligence research program (AMI) I have been pursuing over the last several years with colleagues at FAIR, at NYU, and beyond,” LeCun wrote. “The goal of the startup is to bring about the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences.”
Meta will partner with LeCun’s startup.
The departure comes at a time of disarray within Meta’s AI unit, which was dramatically overhauled this year after the company released the fourth version of its Llama open-source large language model to a disappointing response from developers. That spurred CEO Mark Zuckerberg to spend billions of dollars recruiting top AI talent, including a June $14.5 billion investment in Scale AI to lure the startup’s 28-year-old CEO Alexandr Wang, now Meta’s new chief AI officer.
LeCun, 65, joined Facebook in 2013 to be director of the FAIR AI research division while maintaining a part-time professorial position at New York University. He said in the LinkedIn post that the “creation of FAIR is my proudest non-technical accomplishment.”
“I am extremely grateful to Mark Zuckerberg, Andrew Bosworth, Chris Cox, and Mike Schroepfer for their support of FAIR, and for their support of the AMI program over the last few years,” LeCun said. “Because of their continued interest and support, Meta will be a partner of the new company.”
At the time, Facebook and Google were heavily recruiting high-level academics like LeCun to spearhead their efforts to produce cutting-edge computer science research that could potentially benefit their core businesses and products.
LeCun, along with other AI luminaries like Yoshua Bengio and Geoffrey Hinton, centered their academic research on a kind of AI technique known as deep learning, which involves the training of enormous software systems called neural networks so they can discover patterns within reams of data. The researchers helped popularize the deep learning approach, and in 2019 won the prestigious Turing Award, presented by the Association for Computing Machinery.
Since then, LeCun’s approach to AI development has drifted from the direction taken by Meta and the rest of Silicon Valley.
Meta and other tech companies like OpenAI have spent billions of dollars in developing so-called foundation models, particularly LLMs, as part of their efforts to advance state-of-the-art computing. However, LeCun and other deep-learning experts, have said that these current AI models, while powerful, have a limited understanding of the world, and new computing architectures are needed for researchers to create software that’s on par with or surpasses humans on certain tasks, a notion known as artificial general intelligence.
“As I envision it, AMI will have far-ranging applications in many sectors of the economy, some of which overlap with Meta’s commercial interests, but many of which do not,” LeCun said in the post. “Pursuing the goal of AMI in an independent entity is a way to maximize its broad impact.”
Besides Wang, other recent notables that Zuckerberg brought in to revamp Meta’s AI unit include former GitHub CEO Nat Friedman, who heads the unit’s product team, and ChatGPT co-creator Shengjia Zhao, the group’s chief scientist.
In October, Meta laid off 600 employees from its Superintelligence Labs division, including some who were part of the FAIR unit that LeCun helped get off the ground. Those layoffs and other cuts to FAIR over the years, coupled with a new AI leadership team, played a major role in LeCun’s decision to leave, according to people familiar with the matter who asked not to be named because they weren’t authorized to speak publicly.
Additionally, LeCun rarely interacted with Wang nor TBD Labs unit, which is compromised of many of the headline-grabbing hires Zuckerberg made over the summer. TBD Labs oversees the development of Meta’s Llama AI models, which were originally developed within FAIR, the people said.
While LeCun was always a champion of sharing AI research and related technologies to the open-source community, Wang and his team favor a more closed approach amid intense competition from rivals like OpenAI and Google, the people said.