Big Tech execs warn enterprises handing AI labs their crown jewels

Microsoft CEO Satya Nadella this week published a blog post raising an alarm about how enterprises using AI models like OpenAI's and Anthropic's are "paying twice." They pay once for the tokens that do the work, he argues, but they also pay a hidden cost by handing over the proprietary knowledge needed to make those models useful. Nadella explained that "models learn from 'exhaust'—the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong." Each correction, he wrote, is "distilled into institutional know-how." The problem is that enterprises are unsuspectingly giving frontier labs the ability to one day compete with them directly. Meanwhile, Nadella notes, while AI labs freely train on the open internet, they restrict enterprises from doing the same to their own models.

Nadella's proposed fix is for companies to retain ownership of their own data and build "proprietary learning environments" in the cloud paired with "orchestration layers" that let them switch between AI providers instead of getting locked into one. (Microsoft, it happens, sells exactly this kind of cloud environment and orchestration layer.) Despite his self-serving angle, Nadella's warning carries weight because Microsoft has poured billions into both OpenAI and Anthropic, making his critique surprising coming from a major stakeholder in those companies.

Nadella is not alone in raising these concerns. On CNBC earlier this month, Palantir CEO Alex Karp argued that frontier AI labs are quietly extracting the very thing that makes a business valuable—proprietary data, processes, and competitive edge—while charging premium token prices for tools that often do not deliver commensurate value. Karp said enterprise clients are privately "livid" about their AI vendors and accused frontier labs of being more interested in "tokenmaxxing" than solving real business problems. Palantir followed up with a nine-point "AI sovereignty" manifesto declaring that "data retention is your treasure." Even from outside Big Tech, former White House AI Czar David Sacks called Karp "exactly right" on his All In podcast, arguing that OpenAI and Anthropic have formed a duopoly leaving enterprises with too little leverage over their own data and infrastructure.

The rhetoric is already reshaping the market. Enterprises are shifting toward open-weight models—models whose underlying parameters are published publicly, letting anyone download, run, and modify them—and away from closed-source models. Open-weight models offer enterprises more control and transparency over their data, often at a fraction of the cost of frontier models. The most advanced open-weight models are coming from Chinese-based companies and are closing the performance gap with frontier offerings. Vercel reports that open models now account for 29% of traffic through its AI gateway. Amazon's CTO Werner Vogels recently told the publication that he was seeing companies shift from frontier model providers to open-source options both to control costs and gain more transparency. The U.S. government's recent decision to block access to Anthropic's Fable 5 models has added urgency; if companies are relying on AI models for crucial workloads, they want to ensure no one can pull the plug overnight.

The criticism from Nadella, Karp, and others reflects a growing tension within the AI ecosystem: enterprises are dependent on frontier models for capability, but that dependence comes with hidden costs they are only now beginning to recognize. Nadella's argument is particularly striking because Microsoft has invested billions into both OpenAI and Anthropic, yet he is warning against the very business model those companies employ. The irony is not lost on observers—Nadella's proposed solution (proprietary cloud environments and orchestration layers that let companies switch between providers) is exactly what Microsoft sells.

Karp's framing of the problem in terms of "data sovereignty" and his "AI sovereignty" manifesto align Palantir's existing pitch with a broader enterprise concern. When even competitors acknowledge the asymmetry, it suggests the issue has crossed from self-serving complaint into a genuine business risk that enterprises are beginning to price in. The fact that open-weight models—particularly those from Chinese companies—are closing the performance gap and gaining adoption is no coincidence; enterprises are voting with their inference calls, choosing models they can control and audit over black-box systems that extract their knowledge.

No comments yet. Be the first to share your thoughts!