• The US restricted Anthropic's new cybersecurity AI models, prompting the company to suspend all access.
• The restrictions are creating winners and losers in the red-hot AI race.
• Mistral and DeepSeek could benefit as sovereignty concerns boost the appeal of open-weight models.

One company's headache may be another's opportunity.

The White House's restrictions on access to Anthropic's new AI models have created winners and losers across the AI industry.

On Friday, US officials restricted access to Anthropic's cybersecurity-focused models, Mythos 5 and Fable 5, after concluding safeguards designed to prevent misuse of Fable 5 could be bypassed. The restrictions block foreign nationals from accessing the systems. In response, Anthropic shut down access for everyone.

The move has dealt a direct blow to Anthropic. But it may also strengthen the position of AI companies providing more open models that their customers can deploy and control themselves.

Here are the biggest winners and losers.

Mistral

The verdict: Winner.

Why: The French startup has spent more than a year making the case that Europe should not become dependent on American AI providers.

Unlike Anthropic, whose most advanced models are accessed through company-controlled systems, Mistral has championed open-weight models that customers can deploy on their own infrastructure and customize using their own data.

The Anthropic restrictions gave its CEO, Arthur Mensch, a real-world example of the risk he has been warning about.

In an X post on Tuesday, Mensch doubled down on Mistral's sovereignty pitch, saying the company's upcoming models would be open-weight because users should be able to "own, inspect, audit, or improve" the AI systems they use.

The timing couldn't be much better for Mistral.

France announced this week that its domestic intelligence agency would replace Palantir's AI data tools with those of a French provider, with Prime Minister Sébastien Lecornu warning against "strategic dependencies" on foreign technology.

DeepSeek

The verdict: Winner.

Why: Like Mistral, DeepSeek's open-weight approach may suddenly look more attractive.

Unlike Anthropic's Mythos 5 and Fable 5 models, which are controlled by the company, DeepSeek's models can be downloaded, modified, and deployed by customers themselves.

That makes DeepSeek a beneficiary if governments and businesses begin prioritizing control and sovereignty over access to the latest closed models.

The episode also gives China an opportunity to argue that reliance on US AI providers comes with geopolitical risks, giving it a boost in an increasingly narrowing AI race. During an Anthropic event last month, its CEO, Dario Amodei, said Chinese AI models were roughly 6 to 12 months behind leading US AI systems.

Anthropic

The verdict: Loser.

Why: Anthropic is the company directly affected by the restrictions.

The export controls block foreign nationals, including Anthropic's own employees, from accessing Mythos 5 and Fable 5, which limits the company's ability to distribute some of its most advanced systems internationally.

More importantly, the episode highlights a potential weakness of closed AI models. Because Anthropic controls access to its systems, governments can, in turn, exert greater influence over who can use them.

The episode is also the latest headache for Anthropic in a monthslong spat with the White House after the AI firm said its technology should not be used for domestic mass surveillance or fully autonomous weapons systems.

In response, the US government designated Anthropic a supply chain risk, then Anthropic challenged the move in court.

US AI companies with closed models

The verdict: Short-term winners, long-term losers.

Why: While the restrictions may give a short-term boost to Anthropic's US rivals, it also raises uncomfortable questions for them down the road.

Companies including OpenAI, Google, and xAI primarily distribute their most advanced models through platforms and services they control.

Meta is a partial exception. While some models in its Llama family are open-weight, the company has increasingly been moving towards closed models that it has tighter control over, such as Muse Spark.

For governments and businesses, the Anthropic episode serves as a reminder that access to AI can ultimately depend on decisions made by providers and the governments that oversee them.

That dynamic could strengthen the appeal of sovereign and open-weight alternatives in Europe and elsewhere.

• A neuroscientist worries some people are letting AI do too much of their thinking.
• Over time, she says, that could weaken cognitive reserve, a key defense against dementia.
• "How you use AI, not how often, will determine its impact," Vivienne Ming told Business Insider.

AI doesn't cause dementia, but how you use it could weaken one of the brain's core defenses against it.

That's the warning from Vivienne Ming, a theoretical neuroscientist, the chief scientist at the Possibility Institute, a metascience research group, and founder of Socos Labs, an AI and education firm.

"Your chatbot is not giving you Alzheimer's," Ming told Business Insider.

"My worry is the cumulative impact of chronic substitution: when you stop doing the cognitive work because something will do it for you, you stop building the reserve that protects you later," she said.

As AI has swiftly become an integral part of people's lives and careers, AI researchers and some tech leaders have been releasing warnings about its deskilling effect, the slow erosion of job skills, and the decline in independent thinking.

Ming went a step further, saying that repeatedly outsourcing mental effort to AI, especially among young people, could have real implications for long-term brain health.

"That's the group from whom I'm most concerned," she said. "How you use AI, not how often, will determine its impact."

Over the long term, Ming worries that routinely outsourcing thinking to AI could reduce cognitive engagement and make it harder to build cognitive reserve — the brain's ability to adapt and remain resilient in the face of damage or aging.

"The mechanism I'm describing is the classic 'use it or lose it,'" Ming said.

'GPT is the new GPS'

To drive her point home, Ming compared the effects of using GPS and an AI chatbot.

Researchers at McGill University in Montreal found in 2020 that people with greater lifetime GPS experience have worse spatial memory during self-guided navigation.

In a four-month small study conducted over four months last year, MIT's Media Lab found that people who used a large language model to help write essays showed weaker neural connectivity than participants who used search engines or no external tools, and often couldn't accurately quote passages from essays they had written minutes earlier.

These two examples, Ming said, are cases of cognitive offloading and surrender, or, as she put it, "delegating the effortful part of a task to an external system so your own networks never have to do it."

Her concern in both cases is that people may be engaging key brain functions less frequently, including the hippocampus, the part of your brain that is responsible for memory and learning, and the prefrontal brain networks that help with attention, self-control, and decision-making.

"The hippocampus and prefrontal networks doing that work are precisely the systems that matter for cognitive aging," she said.

"GPT is the new GPS," she added, referring to OpenAI's chatbot ChatGPT, which she said could erode cognitive skills if people increasingly rely on it to think for them.

A matter of cognitive reserve

Research has consistently linked mentally stimulating activities to higher levels of cognitive reserve and lower dementia risk.

One analysis conducted by the English Longitudinal Study of Ageing (ELSA) in 2020 on 12,280 adults aged 50 and older, found that older people with higher cognitive reserve can expect to have a 35% lower risk of developing dementia compared to those with lower levels.

"The principle that lifelong mental engagement delays cognitive decline is some of the most replicated research we have," Ming said.

Importantly, Ming said no biomarker study linking AI use to dementia pathology has been conducted yet. Most of the data right now is "correlational or short-term," she said.

However, she thinks now is the time to start analyzing this cohort, "while the behavior is still taking shape."

"By the time we have the dementia data, a generation will have already formed the habit," she added.

• Two pilots were killed after an Air Canada plane collided with a ground vehicle at LaGuardia Airport.
• Photos from the scene showed the plane on the ground, at an angle, its nose severely damaged.
• The airport will remain closed until at least 2 p.m. ET as federal investigators examine the incident.

An Air Canada aircraft collided with a ground vehicle at New York's LaGuardia Airport late Sunday, killing two pilots and forcing the airport to shut down as investigators examine the crash.

The Air Canada Express flight, a CRJ-900 operated by Jazz Aviation — a Canadian regional carrier that runs shorthaul flights on behalf of Air Canada — struck a Port Authority rescue and firefighting vehicle on the airfield shortly after landing, authorities said.

New York Port Authority Executive Director Kathryn Garcia told reporters early Monday morning that two pilots on board the aircraft were confirmed dead. New York Port Authority told Business Insider another 41 people were transported to the hospital, including 39 flight passengers and two officers.

Garcia said the fire truck involved in the collision was responding to a separate United Airlines aircraft that had reported an odor issue. Two officers in the truck were taken to hospital and are in stable condition with no life-threatening injuries, she added.

The airport will remain closed until at least 2 p.m. ET Monday to allow the National Transportation Safety Board to investigate, Garcia said, adding that federal investigators are already on-site.

Jazz Aviation said in a statement on its website that the plane was carrying 72 passengers and four crew members. The statement did not provide information about the number of injuries or deaths.

Air Canada has set up a helpline for friends and family of passengers on flight AC8646.

The flight left Montreal around 10:35 p.m. E.T., and touched down at LaGuardia at 11:37 p.m., per data from flight tracking website Flightradar24.

"The airport is currently closed to facilitate the response and allow for a thorough investigation," the Port Authority spokesperson said in the statement.

Per Flightradar24, 271 flights at LaGuardia were canceled on Monday.

An air traffic control recording from LiveATC.net appeared to capture the moments before the collision. In the recording, a controller urgently instructs the vehicle to stop. A few minutes later, the controller announces there was an incident on the airfield.

The Federal Aviation Administration issued a ground stop notice for all planes at LaGuardia Airport, per a notice by the agency.

A New York City Fire Department spokesperson told Business Insider that the department had responded to a call at 11:38 p.m., about an incident involving a plane and a vehicle on the runway.

LaGuardia is one of the three major commercial airports serving New York. It said in an X post earlier on Sunday that "weather conditions have caused LGA Airport flight disruptions," and advised passengers to "check with your airline to determine the status of your flight."

LaGuardia served over 30 million passengers in 2025, per the Port Authority.

• Mark Cuban said he bought a Mac Mini to fight a surge of AI-generated emails.
• He said he is training AI to auto-unsubscribe from spam flooding his inbox.
• Cuban believes AI outreach is a trial phase, and response rates will likely eventually drop.

Mark Cuban says the rise of AI-generated cold emails has gotten so overwhelming that he is now fighting back with AI of his own.

Speaking on the live-streamed tech show TBPN on Thursday, the billionaire investor said he recently bought a Mac Mini to help manage the growing flood of inbound messages.

"I do what everybody else does. I bought a Mac Mini," Cuban said.

Beyond AI-generated emails, he said the issue is unwanted email subscriptions.

"It's not even like the cold emails because that's pretty obvious," Cuban said. "It's people subscribing me to shit."

His fix, he said, is to use AI to automate the cleanup.

Cuban said he is training systems to take advantage of Gmail's built-in unsubscribe button, effectively creating a loop where AI filters out AI-generated noise.

"You just got to train it to hit the unsubscribe button," he said. "Then, I just review it and all that shit, so it's still a work in progress, but at least I have a path."

Cuban didn't immediately respond to Business Insider's request for comments.

A 'trial and error phase'

The approach reflects a broader shift in how executives are using AI to manage their inboxes.

LinkedIn CEO Ryan Roslansky has said he uses Microsoft's Copilot for "almost every" high-stakes message, and executives across industries, from tech to retail, recently told Business Insider's Ana Altchek that they rely on AI for day-to-day communications and reviewing documents.

Cuban framed the current moment as a trial-and-error phase, where people are testing what works and what doesn't.

"We're in that trial and error phase where people are like, 'We're going to try it, see what happens,'" he said, adding that response rates will likely fall as more AI-generated messages flood inboxes.

"Then they'll get bored, and then it'll drop off," he added.

• Sam Altman said AI would in the future be sold like electricity and water, metered by usage.
• "We see a future where intelligence is a utility," the OpenAI CEO said.
• Altman suggested demand is surging and compute will decide who gets access.

In the future, you could have another utility bill to pay for: artificial intelligence.

That's according to Sam Altman, who says AI will eventually be bought and sold as a basic utility like electricity and water that's metered by usage.

Speaking at the BlackRock Infrastructure Summit in Washington, DC, on Wednesday, the OpenAI CEO said tech companies like his are building toward a future where intelligence is delivered on demand.

"Fundamentally our business and I think the business of every other model provider is going to look like selling tokens," Altman said, referring to the units AI systems use to process and price input and output data.

"We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for," he added.

In that world, compute capacity determines who gets access — and demand for AI is only going up, Altman said. Compute capacity is the processing power required to train and run AI models, determined by infrastructure such as chips and data centers.

If OpenAI doesn't build enough compute capacity to meet demand, Altman said, it either "can't sell it or the price gets really high." That would push AI access toward the wealthy, or force governments to decide how limited compute should be distributed, he said.

The infrastructure sprint

Major tech companies are set to spend hundreds of billions of dollars this year on compute to meet soaring demand for AI.

In her keynote at CES 2026 in January, AMD CEO Lisa Su said the world will need more than "10 yottaflops" of compute — a scale 10,000 times larger than global AI capacity in 2022 — over the next five years to keep up with growth.

Powering that expansion is a significant infrastructure challenge.

AI data centers can consume as much electricity as small cities, and the strain on the US power grid — along with transformer shortages and slow permitting for transmission lines — could become a bottleneck.

In an episode of the "Moonshots with Peter Diamandis" podcast in January, Elon Musk said that electricity generation is now the limiting factor in scaling AI, predicting China could outpace the US in total AI compute because of its faster energy build-out.

Inside tech companies, compute is a valuable but sometimes scarce resource. Engineers are competing for access to GPUs, and some job candidates now ask about their AI compute budget alongside salary and equity.

Last December, OpenAI President Greg Brockman said the company, which has committed roughly $1.4 trillion on data center projects over the next eight years, wants "to be ahead of the curve," but said, "I don't think we will be, no matter how ambitious we can dream of being right now."

At the BlackRock Infrastructure Summit, Altman said the goal is to move away from a world of being "capacity constrained."