Broadcom said Monday that it’s agreed to produce future versions of artificial intelligence chips for Google, and signed an expanded deal with Anthropic that will give the AI startup access to about 3.5 gigawatts worth of computing capacity drawing on Google’s AI processors.

Shares of Broadcom rose 3% in extended trading.

The disclosure in a securities filing underscores the surging demand for infrastructure that can run generative AI models. Anthropic’s popularity has soared this year, with its Claude app becoming the top free U.S. app listed in Apple’s App Store in February after a dispute between the company and the Pentagon became public.

On an earnings call last month, Broadcom CEO Hock Tan said that “for Anthropic, we are off to a very good start in 2026” in providing 1 gigawatt of compute from Google’s homegrown tensor processing units (TPUs). Broadcom helps Google make its TPUs.

“For 2027, this demand is expected to surge in excess of 3 gigawatts of compute,” he said.

In a note following the earnings call, analysts at Mizuho led by Vijay Rakesh estimated that Broadcom would pick up $21 billion in AI revenue from Anthropic in 2026 and $42 billion in 2027. The filing on Monday did not contain a dollar amount.

Meanwhile, Broadcom is also collaborating with Anthropic rival OpenAI on custom silicon for AI. Both model builders currently rely heavily on graphics processing units from Nvidia through cloud providers such as Amazon, Google and Microsoft. OpenAI has also committed to drawing on six gigawatts of AMD’s GPUs, with the first gigawatt set to come in the second half of this year.

WATCH: Final Trades: Broadcom, Spotify, Applovin and Uber

Over a decade ago, Meta – then known as Facebook – hired social science researchers to analyze how the social network’s services were affecting users. It was a way for the company and its peers to show they were serious about understanding the benefits and potential risks of their innovations. 

But as Meta’s court losses this week illustrate, the researchers’ work can become a liability. Brian Boland, a former Facebook executive who testified in both trials — one in New Mexico and the other in Los Angeles — says the damning findings from Meta’s internal research and documents seemed to contradict the way the company portrayed itself publicly. Juries in the two trials determined that Meta inadequately policed its site, putting kids in harm’s way. 

Mark Zuckerberg’s company began clamping down on its research teams a few years ago after a Facebook researcher, Frances Haugen, became a prominent whistleblower. The newer crop of tech companies, like OpenAI and Anthropic, subsequently invested heavily in researchers and charged them with studying the impact of modern AI on users and publishing their findings. 

With AI now getting outsized attention for the harmful effects it’s having on some users, those companies must ask if it’s in their best interest to continue funding research or to suppress it. 

“There was a period of time when there were teams that were created internally who could start to look at things and, for a brief window, you had some absolutely outstanding researchers who were looking at what was happening on these products with a little bit more free rein than I understand they have today,” Boland said in an interview.

Meta’s two defeats this week centered on different cases but they had a common theme: The company didn’t share what it knew about its products’ harms with the general public.

Jury members had to evaluate millions of corporate documents, including executive emails, presentations and internal research conducted by Meta’s staff. The documents included internal surveys appearing to show a concerning percentage of teenage users receiving unwanted sexual advances on Instagram. There was also research, which Meta eventually halted, implying that people who curbed their use of Facebook became less depressed and anxious.

Plaintiffs’ attorneys in the cases didn’t rely solely on internal research to make their arguments, but those studies helped bolster their positions about Meta’s alleged culpability. Meta’s defense teams argued that certain research was old, taken out of context and misleading, presenting a flawed view of how the company operates and how it views safety.

“The jury got to hear both sides of the story and a very fair presentation of the facts, and they got to make a decision based on what they saw,” Boland said. “And both juries, with very different cases, came back with clear verdicts.”

Meta and Google’s YouTube, which was also a defendant in the L.A. trial, said they would appeal.

A federal judge in San Francisco granted Anthropic’s request for a preliminary injunction in its lawsuit against the Trump administration. 

Judge Rita Lin issued the ruling on Thursday, two days after lawyers for the artificial intelligence startup and the U.S. government appeared in court for a hearing. Anthropic sued the administration to try to reverse its blacklisting by the Pentagon and President Donald Trump’s directive banning federal agencies from using its Claude models.

Anthropic sought the injunction to pause those actions and prevent further monetary and reputational harm as the case unfolds. The order bars the Trump administration from implementing, applying or enforcing the president’s directive, and hampers the Pentagon’s efforts to designate Anthropic as a threat to U.S. national security. 

“Punishing Anthropic for bringing public scrutiny to the government’s contracting position is classic illegal First Amendment retaliation,” Lin wrote in the order. A final verdict in the case could still be months away. 

During Tuesday’s hearing, Lin pressed the government’s lawyers about why Anthropic was blacklisted. Her language in Thursday’s order was even sharper.

“Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government,” she wrote.

Following the ruling, Anthropic said it’s “grateful to the court for moving swiftly.”

“While this case was necessary to protect Anthropic, our customers, and our partners, our focus remains on working productively with the government to ensure all Americans benefit from safe, reliable AI,” the company said in a statement.  

Anthropic’s suit earlier this month followed a dramatic couple weeks in Washington D.C., between the Department of Defense and one of the most valuable private companies in the world.

In a post on X in late February, Defense Secretary Pete Hegseth declared Anthropic a so-called supply chain risk, meaning that use of the company’s technology purportedly threatens U.S. national security. In early March, the DOD officially notified Anthropic about the designation via a letter.

Anthropic is the first American company to publicly be named a supply chain risk, as the designation has historically been reserved for foreign adversaries. The label requires Defense contractors, including Amazon, Microsoft, and Palantir, to certify that they do not use Claude in their work with the military. 

The Trump administration relied on two distinct designations – 10 U.S.C. § 3252 and 41 U.S.C. § 4713 – to justify the action, and they have to be challenged in two separate courts. Because of that, Anthropic has filed another lawsuit for a formal review of the Defense Department’s determination in the U.S. Court of Appeals in Washington. 

Shortly before Hegseth declared Anthropic a supply chain risk, President Donald Trump wrote a Truth Social post ordering federal agencies to “immediately cease” all use of Anthropic’s technology. He said there would be a six-month phase-out period for agencies like the DOD.

“WE will decide the fate of our Country — NOT some out-of-control, Radical Left AI company run by people who have no idea what the real World is all about,” Trump wrote.

The Trump administration’s actions surprised many officials in Washington who had come to admire and rely on Anthropic’s technology. The company was the first to deploy its models across the DOD’s classified networks, and it was championed for its ability to integrate with existing Defense contractors like Palantir. 

Anthropic signed a $200 million contract with the Pentagon in July, but as the company began negotiating Claude’s deployment on the DOD’s GenAI.mil AI platform in September, talks stalled.

The DOD wanted Anthropic to grant the Pentagon unfettered access to its models across all lawful purposes, while Anthropic wanted assurance that its technology would not be used for fully autonomous weapons or domestic mass surveillance. 

The two failed to reach an agreement, and now, the dispute will be settled in court. 

“Everyone, including Anthropic, agrees that the Department of [Defense] is free to stop using Claude and look for a more permissive AI vendor,” Lin said during the hearing Tuesday. “I don’t see that as being what this case is about. I see the question in this case as being a very different one, which is whether the government violated the law.

WATCH: Anthropic vs. Pentagon hearing

U.S. District Judge Rita Lin said Tuesday that the decision by the Pentagon to blacklist Anthropic’s Claude artificial intelligence models “looks like an attempt to cripple” the company.

Anthropic appeared in San Francisco federal court on Tuesday to ask Lin to temporarily pause the Pentagon’s blacklisting and President Donald Trump’s directive banning federal government agencies from using that technology.

If the preliminary injunction is awarded, the AI startup will be able to continue doing business with government contractors and federal agencies as its lawsuit against the Trump administration plays out in court.

Without the injunction, the company has said, it could lose billions of dollars in business.

Earlier in March, the Department of Defense designated Anthropic a so-called supply chain risk, meaning that use of the company’s technology purportedly threatens U.S. national security. It was the first time an American company had been hit with that designation.

The label, if allowed to continue, will require defense contractors, including Amazon, Microsoft, and Palantir, to certify that they do not use Claude in their work with the military.

“This is something that has never been done with respect to American company,” Anthropic’s counsel Michael Mongan said during the hearing. “It is a very narrow authority. It doesn’t apply here, and it’s not a normal way to respond to the concerns that have been articulated by the other side.”

Palantir is continuing to use Claude in its work with the department as the legal battle plays out, CEO Alex Karp told CNBC on March 12. Anthropic’s model is also being used in the war with Iran.

Anthropic has argued that there is no basis to consider the company a supply chain risk. The company also said it is being unfairly retaliated against because it demanded that the DOD not use Claude for fully autonomous weapons or mass surveillance of Americans. The Pentagon insists it does not use the AI models for such purposes.

Lin said she expects to issue an order on Anthropic’s motion in the next few days.

On Monday, the judge gave lawyers for Anthropic and the government a list of questions she wants answered at the hearing.

One of those questions was: “What evidence in the record shows that Anthropic had ongoing access to or control over Claude after delivering it to the government, such that Anthropic could engage in acts of sabotage or subversion?”

In its motion seeking a preliminary injunction, Anthropic argued that such an order would prevent the company from incurring further economic and reputational harm.

“The government has infringed on Anthropic’s right to speak freely; it has disparaged the company’s good name by stigmatizing it with an unlawful designation as a national security risk; it has deprived Anthropic of government contracts and damaged its relationships with business partners in the private sector; and it has put millions, possibly billions, of dollars at risk,” the motion stated. “Absent immediate relief from this Court, those harms will continue to mount.”

The company also noted that an injunction would not require the U.S. government to use its models or prevent it from transitioning to another AI vendor. 

Before the conflict erupted in late February, Anthropic was one of the first AI companies to partner with many U.S. agencies as the government sought to rapidly upgrade its systems and capabilities with cutting-edge AI tech.

Anthropic signed a $200 million contract with the Pentagon in July and was the first AI lab to deploy its technology across the agency’s classified networks.

But as the company began negotiating Claude’s deployment on the DOD’s GenAI.mil AI platform in September, talks stalled over how the military could use the models.

The department has insisted on unfettered access to the company’s technology for all lawful purposes. 

During the hearing on Tuesday, Lin questioned if the DOD was punishing Anthropic for “acting stubbornly” in negotiations. The government’s lawyer Eric Hamilton said that the company was going beyond the normal scope of a contractor.

“Anthropic is not just acting stubbornly. It’s not just refusing to agree to contracting terms. Instead, it’s raising concerns to [DOD] about how [DOD] uses its technology in military missions,” Hamilton said. “What happens if anthropic installs a kill switch or functionality that changes how it functions? That is an unacceptable risk to [DOD].”

In February after Anthropic and the DOD failed to reach an agreement, Trump issued a Truth Social post ordering federal agencies to “immediately cease” all use of Anthropic’s technology.

“WE will decide the fate of our Country — NOT some out-of-control, Radical Left AI company run by people who have no idea what the real World is all about,” Trump wrote.

WATCH: Anthropic sues Trump administration over Pentagon blacklisting

Google said its fiber internet unit called GFiber is combining with Astound Broadband and forming an independent provider, with Google remaining as a minority shareholder.

The new company will be majority owned by investment firm Stonepeak and led by the existing GFiber executive team, “utilizing their expertise in high-speed fiber innovation to manage the combined network footprint,” Google said in a press release on Wednesday. The transaction is expected to close in the fourth quarter.

Google Fiber, launched in 2010, was an early effort by Google to build ultra-fast fiber-optic broadband networks in the U.S., starting with a gigabit-speed rollout in Kansas City in 2012. Google proposed building gigabit fiber connections to homes, far faster than typical U.S. internet speeds at the time.

Since then, some planned expansions were canceled and the company focused on select markets rather than a costly and time-intensive nationwide rollout.

The spinout comes at a time when demand is growing for high-capacity networks fueled by the increasing popularity of artificial intelligence services. The external capital will help the new entity expand across the country, the company said.

“This partnership with Astound and Stonepeak is the next step in our decade-long mission to redefine what customers can expect from their internet provider,” GFiber CEO Dinni Jain said in the release.

GFiber has been part of Google’s “Other Bets” segment, which includes non-core assets such as the Waymo robotaxi division and drug discovery business Isomorphic Labs. In 2025, the combined segment generated $1.54 billion in revenue, or less than 0.5% of Alphabet’s total sales, and recorded an operating loss of $16.8 billion.

The shift toward fiber infrastructure has become increasingly important as demand grows for networks that can support cloud computing, streaming and emerging AI services. U.S. tech giants are also rolling out a rapidly expanding network of transcontinental subsea cables, seeking to keep pace with growing bandwidth demand.

Astound is a major U.S. cable operator and broadband platform, which was acquired by Stonepeak in 2021 for $8.1 billion. Stonepeak specializes in infrastructure and real estate.

A Google spokesperson didn’t immediately respond to a request for comment.

WATCH: Google’s capacity advantage

Federal Communications Commission Chairman Brendan Carr lashed out at Amazon on Wednesday for opposing SpaceX’s orbital data center plans while it’s falling short of its own satellite “deployment milestone.”

“Amazon should focus on the fact that it will fall roughly 1,000 satellites short of meeting its upcoming deployment milestone, rather than spending their time and resources filing petitions against companies that are putting thousands of satellites in orbit,” Carr wrote in a post on X.

Amazon declined to comment.

Amazon last week urged the FCC to reject a SpaceX application for permission to launch a constellation of up to 1 million low Earth orbit satellites, which would function as a data center network in space to support artificial intelligence projects.

Amazon characterized the application as a “lofty ambition rather than a real plan,” noting SpaceX has provided scant details around how it will “deliver on these grand claims.”

SpaceX’s Starlink service currently dominates the internet-from-space market. Amazon has been vying to compete with Starlink via its Leo satellite service, previously branded as Kuiper. The company has invested more than $10 billion into the effort, and has sent up at least 200 satellites since last April via a variety of launch partners, including Elon Musk’s SpaceX.

In late January, Amazon asked the FCC for a waiver or 24-month extension, to July 2028, to meet a deadline that requires it to deploy roughly 1,600 internet satellites by July 2026. At the time, the company blamed delays beyond its control, including a “shortage in the near-term availability” of rockets and manufacturing disruptions.

Amazon noted in its request that the FCC has previously granted similar extensions. The FCC last month approved a separate petition from Amazon to deploy 4,500 internet satellites, which would more than double the size of its constellation.

Starlink has around 9,000 satellites in orbit today and roughly 9 million customers. It recently received authorization from the FCC to put another 7,500 satellites into orbit.

Scientists have decried the SpaceX proposal to launch one million satellites into orbit, citing a wide range of issues, including light pollution, orbital debris and other harms to the broader orbital environment, as well as increased risk of “Kessler syndrome,” a scenario in which debris and clutter in space can cause a chain reaction that makes low Earth orbit unusable.

Amazon pointed to these concerns from astronomers and environmental groups in its petition, and said SpaceX’s application “risks worsening international backlash” from regulators who are concerned about monopolization of space resources.

“Granting the application would worsen matters further, forcing every other operator in Low-Earth Orbit to plan around a constellation that may never exist, distorting international spectrum and orbital coordination proceedings, and lending regulatory legitimacy to what amounts to a publicity and narrative-shaping exercise,” Amazon wrote in its request to the FCC.

The FCC hasn’t yet approved SpaceX’s request, but in separate remarks to Reuters on Wednesday, Carr said he doesn’t expect Amazon’s petition to “get much traction.”

Carr is a longtime public fan of SpaceX who has mocked environmental concerns from those calling out Musk’s company for launches that harmed public lands and endangered species’ habitat.

He also accused the FCC, under former President Joe Biden, of “regulatory harassment” of SpaceX when the agency found the company’s Starlink WiFi service was not fit at the time to fulfill the program needs of a rural broadband initiative.

Defense Secretary Pete Hegseth’s decision to label Anthropic a “Supply-Chain Risk to National Security” on Friday resulted in more questions than answers.

“It’s all very puzzling,” Herbert Lin, a senior research scholar at Stanford University’s Center for International Security and Cooperation, told CNBC in an interview.

Anthropic is the only American company ever to be publicly named a supply chain risk, as the designation has traditionally been used against foreign adversaries. But the company hasn’t received any official declaration beyond social media posts.

A formal designation will require defense vendors and contractors to certify that they don’t use Anthropic’s models in their work with the Pentagon.

The dispute centered around how Anthropic’s artificial intelligence models could be used by the military. The Department of Defense wanted Anthropic to grant the agency unfettered access to its Claude models across all lawful purposes, while Anthropic wanted assurance that its technology would not be tapped for fully autonomous weapons or domestic mass surveillance.

With no agreement reached by Friday’s deadline, President Donald Trump directed federal agencies to “immediately cease” all use of Anthropic’s technology, and said there would be a six-month phaseout period for agencies like the DOD.

Experts told CNBC the supply chain risk designation is highly unusual, especially since the U.S. and Israel began carrying out strikes in Iran just hours later. A group of retired defense officials, policy leaders and executives wrote to Congress on Thursday, defending Anthropic and calling the Trump administration’s designation a “dangerous precedent.”

Anthropic’s models are still being used to support U.S. military operations in Iran, even after the company was blacklisted, as CNBC previously reported.

Talks between Anthropic and the DOD are now reportedly back on, according to the Financial Times, but there are still big questions hanging over the issue as of Thursday.

Why is the U.S. government still using Claude?

Stanford’s Lin doesn’t understand why the DOD is still using Anthropic’s models in sensitive settings if they pose such a threat. If the Trump administration really sees Anthropic as a risk to national security, he said, it wouldn’t make sense to phase out the models over an extended period of time.

“OK, wait a minute, they’re a really dangerous player for U.S. national security, so you’re going to use them for another six months? Huh?” Lin said. 

Michael Horowitz, a senior fellow for technology and innovation at the Council on Foreign Relations, said it’s “especially notable” that Anthropic’s models were used to support the U.S. military action in Iran. He said “there’s no clearer signal” of how much the Pentagon values the technology.

“Even in a situation where there is this intense feud between the company and the Pentagon, they are using their technology in the most important military operation that the United States is conducting,” he said. 

Transitioning away from Anthropic toward a new vendor takes time and comes at a significant cost in terms of efficiency, said Jacquelyn Schneider, a Hargrove Hoover fellow at Stanford University’s Hoover Institution.

Until recently, Anthropic was the only AI company approved to deploy its models across the agency’s classified networks. OpenAI and Elon Musk’s xAI received clearance, but their systems can’t be deployed or adopted overnight.

What’s the actual threat?

By designating Anthropic a supply chain risk, the DOD is suggesting that the company is “really bad” for U.S. national security, Lin said. But he stressed that the agency hasn’t clearly outlined what kind of threat the company poses. 

“They don’t point to any technical failing, they don’t point to any hack,” Lin said. “They say things like ‘They’re arrogant,’ and ‘We don’t want you telling the DoD what to do in some hypothetical situation that hasn’t happened yet.'”

Lin said the other punishment that Hegseth was threatening to impose on Anthropic, invoking the Defense Production Act, also contradicts the idea that the company threatens national security. 

The Defense Production Act allows the president to control domestic industries under emergency authority when it’s in the interest of national security. It could essentially compel Anthropic to let the Pentagon use its technology. 

Horowitz said he thinks the clash between Anthropic and the DOD is “masquerading” as a policy dispute. 

Months earlier, venture capitalist and White House AI and crypto czar David Sacks criticized the company for “running a sophisticated regulatory capture strategy based on fear-mongering,” after an essay published by an executive, and conservatives have repeatedly accused Anthropic of pushing “woke AI.”

Anthropic CEO Dario Amodei took a different approach than other tech executives, avoiding getting cozy with the Trump administration in its early days.

“This feels to me like a dispute that is about politics and personalities,” Horowitz said. 

Is an official designation on the way?

Anthropic hasn’t been designated a supply chain risk by any official measure, and there’s an open question as to if or when the company should expect one. Defense contractors have to decide whether they will follow Hegseth’s directive on social media or wait for more formal guidance. 

Several executives told CNBC that their companies are moving away from Anthropic’s models, and a venture capitalist said a number of portfolio companies are switching “out of an abundance of caution.” But others, including C3 AI Chairman Tom Siebel, said he doesn’t see a “need to mitigate” the technology “until it gets litigated.” 

Schneider said businesses are rational, and if they think it’s high risk to work with Anthropic, whether it’s formally declared a supply chain risk or not, they’re going to hedge and look for other partners.

“There’s all sorts of decisions that have been made within the Trump administration that, by law, require more codification,” Schneider said. “Even the example of moving from DoD to [Department of War]. That by law needs more codification, but all the contractors are using DoW.”

Even so, Samir Jain, vice president of policy at the Center for Democracy and Technology, said social media posts likely aren’t enough to actually cause a designation.

“There’s a process that the statute requires, including an actual finding that Anthropic presents national security risks if it’s part of the supply chain,” he said in an interview. “I don’t think, factually, that that predicate could possibly be met here.”

Anthropic said in a statement Friday that it will challenge “any supply chain risk designation in court.”

Does this have anything to do with the U.S. strikes on Iran?

For Schneider, the war in Iran now looms large over the spat between Anthropic and the DOD. She said she’s left wondering whether the two conflicts were happening in parallel, or if they were somehow related. 

“Obviously, you’re not going to walk away from technologies that are deeply embedded in your wartime processes right before you go to war,” Schneider said.

She said planning a military operation of that magnitude would have required “a lot of sleepless nights,” so she was surprised the DOD was willing to spend such a “remarkable amount of energy” on a public clash ahead of the initial attack.

What happens next?

As the war in Iran stretches into its sixth day, Anthropic’s path forward with the DOD remains a big mystery.  

Horowitz said he would bet that the six-month off-boarding period will become a “a locus for some re-examination” within the Pentagon, especially since members of Congress and broader public markets have shown so much interest in the dispute. 

Lin expressed a similar sentiment, and said he wouldn’t bet on Anthropic’s models being out of the DOD a year from now.

Schneider is less convinced. 

“I wish I had a more definitive thought about where this is all going to go, but everything is so unprecedented,” she said. When it comes to historical examples or analogous cases, Schneider said: “I don’t have those. It’s just super limited.”

The DOD declined to comment. Anthropic didn’t provide a comment.

WATCH: Anthropic tops $19 billion in annual revenue rate

Amazon‘s data center in Bahrain was targeted by Iran’s Islamic Revolutionary Guard Corps for the company’s support of the U.S. military, Iranian state media said Wednesday.

The company’s cloud computing unit said Monday that one of its facilities in Bahrain was damaged due to a nearby drone strike on Sunday. Two data centers in the United Arab Emirates were also damaged after they were “directly struck” by drones.

All of the facilities remain offline, according to the Amazon Web Services health dashboard.

The attack in Bahrain was launched “to identify the role of these centers in supporting the enemy’s military and intelligence activities,” Iran’s Fars News Agency said on Telegram.

The incidents came after joint U.S.-Israel strikes on Iran over the weekend. Iran has retaliated against Israeli and U.S. bases across the Gulf.

Amazon declined to comment.

In addition to structural damage, the data centers also experienced power disruptions and some water damage after firefighters worked to put out sparks and fire. Some popular AWS applications experienced “elevated error rates and degraded availability” due to the incident.

AWS advised cloud customers to back up their data, consider migrating their workloads to other regions and direct traffic away from Bahrain and the UAE.

AWS announced its Bahrain region in 2019, and it hosts significant workloads for governments there. The company also operates a corporate office in Bahrain that is primarily for AWS employees.

Earlier this week, Amazon instructed all of its corporate employees in the Middle East to work remotely and “follow local government guidelines” amid escalating instability in the region.