|
Like other software leaders these days, Palantir CEO Alex Karp is on a mission to convince businesses they need a trusted intermediary between their wallets and Anthropic or OpenAI, rather than dealing directly with the two AI firms. I spoke to Karp shortly after his fiery appearance on CNBC, where he appeared to accuse OpenAI and Anthropic of stealing data from and overcharging their customers, and positioning Palantir as a critical “application layer” to protect business and governments from those newbies and their bad intentions. Karp told me that, with Palantir’s help, some of his company’s U.S. government customers recently switched to using open-source models developed by Nvidia from proprietary models developed by the likes of Anthropic. He said he wasn’t permitted to disclose which agencies or units he was referring to, but his other comments implied the Department of Defense was at least testing Nvidia’s models. “There's just very deep frustration around…are they gonna optimize the models for me, or are they gonna take the alpha of my business, transfer in their weights, and compete against me?” Karp said about the proprietary model providers. Will this fear strategy work? So far, so good: Palantir’s stock has risen 12% since the comments while the broader tech stock-heavy Nasdaq Composite fell, though Palantir shares are down about 22% this year. While Anthropic has blindsided its business partners by introducing competing products, there’s no evidence it has used customers’ data to train models, which is prohibited by its terms of service. Colin Jarvis, OpenAI’s head of forward deployed engineering, said on X that, “At no point do we train on customer data.” Earlier this week, Palantir launched a product aimed at helping U.S. government agencies to use Nvidia open source models, Nemotron, effectively and securely through its software. Executives at Palantir told me the product helps customers tweak and continuously improve Nemotron models for their specific needs. Nemotron provides “equal or, in some cases, superior performance on the battlefield use cases, which are mostly highly classified,” Karp said. The U.S. military including the Army, Navy, Air Force and Space Force has long used Palantir software to identify human and non-human targets in conflict zones. Karp told me he expects every one of Palantir’s clients, including commercial ones, to leverage an open source model “as soon as they see it as being at parity” with proprietary ones. He said that, since the new “engine” from Palantir to improve Nemotron models was announced, his phone has been “blowing up” with calls from interested customers. Palantir makes digital copies of a company’s business data across various applications, and its software helps them to develop AI apps using that data and whichever AI model they want to use. Palantir also offers a tool called Evolve with a so-called model router to send AI tasks to different AI models based on whether the customer wants better performance or cost-savings. (See the five kinds of model routers.) A shift toward U.S.-made open models could make sense for the U.S. government, given the drama we saw between the Pentagon and Anthropic earlier this year over how the company’s models could be used. But it’s far from certain that open source ones will catch up in quality to the closed-source ones. For its part, Nvidia says it supports a diversity of models, both open and closed. “Proprietary versus open is not a thing. It’s proprietary and open,” Nvidia CEO Jensen Huang said at the company’s annual developer conference in March. According to benchmarking firm Artificial Analysis, Nvidia’s latest Nemotron model lags behind closed frontier models and Chinese open-source models in intelligence, but it’s the leading American open-source model. “We’re super happy with the performance [of Nemotron,” especially for how nascent of a push this was,” compared to other AI model developers that started much earlier, Nvidia researcher Venkat Srinivasan said at an Artificial Analysis event in San Francisco this week. “We want everyone in the open source community to use our data, use our methodologies, use our approaches and, indeed, use our models.” Palantir, Databricks Tout Different Kinds of Open Source Models Palantir and Databricks offer similar tools for businesses to use AI with their data, but their CEOs have taken different stances on Chinese open source models. Databricks CEO Ali Ghodsi struck an excited, optimistic tone when he discussed the popularity of Chinese models among his customers at a different event in San Francisco recently. “Chinese models, in particular, over the last year, over two, people were skeptical. They didn't want Chinese models until their costs showed up…Then they all want the Chinese [models].” He positioned open source models, particularly through the company’s new router tool, Unity AI Gateway, as a solution to companies “blowing through their budgets too fast. Our customers—their revenue is growing much slower than their cost of [using] large language models.” For all his talk about open source, Palantir’s Karp isn’t suggesting anyone switch to Chinese models, which he has said present potential security issues, though there’s no evidence such issues have occurred. Enterprises shouldn’t be tempted by the cost savings associated with such models, and software providers should champion Western ones. “We've got to fight for making sure the open models we have in America are usable,” he said of software and model providers including his own company. To be sure, Palantir allows customers to use models of their choice including Chinese open source ones through their own infrastructure, an engineer at the company told me. “The choice is going to be very quickly [an] open model from America or open model from China,” Karp said.
| 
