The Morning Download: Companies Say the Risks of 'Open' AI Models Are Worth It

Good morning. As AI usage takes off, businesses have a growing need for smaller, customized and lower-cost models. Many of them are quite willing to find what they need among so-called open AI models that make their code publicly available and provide an alternative to proprietary giants such as OpenAI and Anthropic.

The Nemotron Coalition will advance development of frontier-level foundation models through shared expertise, data and compute, Nvidia Chief Executive Jensen Huang said Monday at the company’s annual GTC event underway this week.

Companies at GTC said open models play crucial roles in their AI initiatives. Isabelle Bousquette and I spoke to tech leaders from Capital One, ServiceNow and CrowdStrike, who said they appreciated open model strengths such as customizability and lower cost, even as they grappled with the challenges of making them secure. Read our coverage here.

Most AI architectures have room for both open and closed models and often assign them different functions.

“At the end of the day,” CrowdStrike Chief Technology Officer Elia Zaitsev said, closed models are “general purpose and very effective, but they are not customizable to specific use cases or niche domains.”

Closed models are generally trained on huge amounts of publicly available data from the internet, but struggle when confronted with tasks where training data might be less readily available, Zaitsev said. That includes CrowdStrike’s field of cybersecurity, where data for identifying threats is constantly changing as the threats and attackers themselves rapidly shift.

At Capital One, Milind Naphade, head of AI foundations, said using open models like Nvidia’s Nemotron and OpenAI’s GPT-OSS gives the company more control and better performance thanks to the ability to customize on its own data.

“It’s a level of customization that’s simply impossible with…fine-tuning that can be done on closed models,” Naphade said.

How do the risks and rewards of open models balance out for you? We’d love to know.

IBM closed its $11 billion deal for data-streaming company Confluent Tuesday. Chief Executive Arvind Krishna tells the WSJLI’s Belle Lin that Confluent’s technology will become the “backbone” of its own platform for helping business clients access their data for various AI uses. To make AI agents work, “you need to be able to get data wherever it is” and get it instantly, Krishna said.

The closure of the deal, first announced in December, capped a turbulent few months for the company, which saw its stock take a big dip in February—posting its worst decline in 25 years—triggered by an Anthropic Claude product announcement

Investors have been jittery so far this year about the threat AI tools could pose to some business models, especially in software.

Krishna told Belle that he is approaching the leadership challenge of steering IBM through that business risk in a nuanced way.

Keep calm and carry on. For employees and clients, Krishna said, the goal is very simple: “How calm can we keep everybody so they can stay focused on the task and not get distracted?”

Customers, if they are “delighted with our products and services,” will return the favor by writing checks, he added. “That is the simplest proof point that the naysayers are wrong.”

Focus on the customer. “For clients, we always talk about, ‘What value can we bring you for your business goals? Can we make you more productive? Can we help you eke out a bit more revenue with the same assets that you have?’”

If IBM can accomplish those things, then clients are happy, Krishna said. And, in turn, employees are motivated when customers are happy.

What Krishna said he can’t control are the markets and how they react. Supposing that IBM keeps its employees and clients satisfied, then “the markets will reward you over the medium to long-term,” he said.

Microsoft seeks more coherence in AI efforts. Microsoft is restructuring its Copilot teams, unifying the teams that work on its Microsoft 365 Copilot productivity offerings and the consumer version of Copilot, the Journal’s Sebastian Herrera reports. The move to simplify toward one product experience across business and consumer applications comes amid employee concerns about fragmentation and user confusion.

CoPilots face headwinds. Microsoft in February reported that it had sold 15 million Microsoft 365 Copilot “seats,” a fraction of Microsoft 365’s base of 450 million-plus paid seats. 

Simplicity is the new complexity.  Microsoft’s move to streamline its CoPilot efforts comes as OpenAI leaned into its own strategy shakeup, instructing employees last week to step back from a "do everything" approach and focus on the ‘core’ business, the Journal reported.

Nvidia says it's restarting production of AI chips for sale in China. Speaking at the company’s GTC event Tuesday, CEO Jensen Huang said that in recent weeks, demand signals out of China have strengthened. The U.S. in December approved sales of its H200 processor —a chip that is a generation behind its most powerful series of GPUs—in China, as long as the company shared 25% of its sales with the U.S. government. 

Nvidia also is working on a version of its Groq AI chips that can be sold in China, Reuters reports.

AI boom widening U.S. trade deficit. Tech companies imported record volumes of foreign chips and computers to build data centers, driving the goods deficit to a record $1.2 trillion in 2025, the New York Times reports. 

And now....back to the GTC

I arrived at a panel on AI agents a few moments before it was scheduled to begin and made my way to one of the few open seats, trying my best not to step on anyone’s toes but mine. Once seated, I met Ousmane Kabré, founder and CEO of Traack AI Inc. Traack is based in San Francisco and has a team of eight people. It builds AI workflow agents that manage high-volume communication and coordination for HR teams, universities, law firms, and other high-trust organizations, said Kabré, a native of Burkina Faso.

I asked Kabré, one of about 30,000 people at the annual Nvidia GTC event in San Jose, to share his thoughts on one of the dominant themes of the conference. He expects a lot from agentic AI “claws” that can work together with a high level of autonomy to get work done. “My view is that within the next year or two, most enterprise software will adopt claws at scale,” he said.

The key challenge is security, privacy, trust and governance, given that claws need access to files, email, internal systems and permissions, according to Kabré.

“At Traack AI, we’re already using this model in a controlled way and our team is of course continuing to learn and experiment,” he said.

— Steven Rosenbush

We want to know what you feel about AI. No, not what you think about AI, but what you really feel about it. Create an emoji that expresses your AI sentiment and send it to us. We’ll publish the best ones in future newsletters.

Russia has been expanding its intelligence sharing and military cooperation with Iran, providing satellite imagery and improved drone technology to aid Tehran’s targeting of U.S. forces in the region, people familiar with the matter said. (WSJ)

Uncertainty fueled by the war in Iran is expected to reinforce a broad agreement among most Federal Reserve officials to sit tight at their meeting this week. That will make any dissenting votes all the more notable as Jerome Powell nears the close of his tenure as Fed chair. (WSJ)

The president wants Congress to pass strict new voting eligibility rules before the midterm elections. But his crusade—which is seen as unlikely to succeed—is exposing deep fissures between Republicans as they clash over both tactics and policy. (WSJ)

Amazon.com  is planning to sharply cut the number of packages it ships through the U.S. Postal Service, a move that could cost the agency billions of dollars in much-needed revenue. (WSJ)