The Enterprise AI Paradox: Risking Sensitive Data in the Public Cloud

SANTA CLARA, Calif. – October 13, 2026 – In the frantic race to deploy generative artificial intelligence, the world's most heavily regulated industries are playing a dangerous game of digital roulette. According to a sweeping new survey of enterprise AI leaders, a staggering 61 percent of organizations in sectors like banking, healthcare, and government are transmitting sensitive data to public cloud AI services.

They are doing this despite a glaring, almost universally acknowledged risk: 93 percent of these same leaders admit they are deeply concerned about exposing proprietary prompts, intellectual property, and regulated data to third-party AI infrastructure providers.

This glaring contradiction sits at the heart of the "State of Enterprise AI Deployment" report released today by Fortanix Inc., a data and AI security vendor pioneering Confidential Computing. Conducted by Atomik Research, the study polled 300 AI leaders at major U.S. enterprises with over $500 million in revenue. The findings paint a picture of a corporate landscape caught in a paralyzing tug-of-war between the fear of missing out on a generational technological leap and the existential dread of a catastrophic compliance failure.

The AI Pragmatism Paradox

For business leaders, the pressure to integrate advanced machine learning models into daily operations is immense. Boards of directors are demanding AI strategies, and competitors are aggressively automating workflows. Yet, this rush to the cloud is hitting a massive regulatory wall.

The survey reveals that security friction is not just a theoretical concern; it is actively derailing innovation. Over the past year, 59 percent of surveyed organizations have delayed an AI project due to data security or privacy concerns, and 41 percent have blocked a deployment entirely. Regulatory compliance now ranks as the single largest obstacle to deploying AI on sensitive data, cited by 46 percent of respondents, outpacing even the fear of outright data exfiltration.

"It is a daily battle between the board's fear of missing out and our strict regulatory obligations," noted one chief information security officer at a leading U.S. healthcare provider, reflecting the broader industry sentiment. "We know the public cloud models are powerful, but the moment protected health information leaves our controlled environment and is processed on shared infrastructure, we are potentially violating HIPAA. The fines and reputational damage aren't worth the efficiency gains."

This is the AI pragmatism paradox: organizations are knowingly engaging in risky behavior because the perceived business value of public cloud AI is too massive to ignore, even as their own security teams hit the emergency brakes on nearly half of all new initiatives.

Beyond Encryption at Rest: The 'Data-in-Use' Vulnerability

To understand why 93 percent of AI leaders are losing sleep over public cloud deployments, one must look at the fundamental architecture of modern data security. For decades, enterprises have relied on robust encryption to protect data "at rest" (stored on servers) and "in transit" (moving across networks).

However, traditional encryption fails during the actual computation phase. For a large language model to analyze a financial portfolio or summarize a patient's medical history, that data must be decrypted in the server's memory. This creates a "data-in-use" vulnerability. During inference—the moment the AI is actively processing the prompt—the sensitive information is exposed to the underlying infrastructure, the cloud provider's hypervisor, and potentially privileged insiders.

While hyperscalers like Microsoft, AWS, and Google have implemented strict data governance policies, assuring customers that enterprise prompts are not used to train public models, these policy-based promises are increasingly insufficient for highly regulated entities. Regulators like the Financial Industry Regulatory Authority (FINRA) and the Department of Health and Human Services require cryptographic guarantees, not just terms of service agreements.

“It's clear that AI leaders in regulated industries have become educated about the fact that their data is at risk when a model starts processing it,” said Anand Kashyap, CEO and co-founder of the Santa Clara-based security firm that commissioned the study. “What they need is proof that the environments their models and data are processed in are secure and untampered."

The On-Premises and Sovereign AI Resurgence

Driven by these insurmountable compliance roadblocks and the fundamental flaw in data-in-use security, a massive architectural shift is underway. After a decade of relentless "cloud-first" migration strategies, regulated enterprises are pulling their most advanced AI workloads back behind the corporate firewall.

The survey indicates that 64 percent of organizations have already deployed closed-weight frontier AI models on-premises. Furthermore, a near-unanimous 97 percent of respondents stated they are at least somewhat likely to adopt leading AI models if they can be made securely available on their own local infrastructure.

This pivot is fueling the rise of "Sovereign AI"—the concept that an organization or nation must maintain absolute physical and cryptographic control over its AI infrastructure, models, and data. An overwhelming 95 percent of AI leaders now consider sovereign AI to be important or critically important to their overarching strategy.

Looking ahead, enterprise architects do not view the public hyperscaler cloud as their primary destination for AI inference. Instead, AI leaders expect hybrid environments to dominate over the next 24 months. By keeping sensitive inference tasks on-premises or in highly controlled private clouds, organizations can satisfy local data residency laws and shield their proprietary algorithms from the open internet.

Confidential Computing: The Missing Link

If hybrid and on-premises deployments represent the location of the future, Confidential Computing represents the underlying architecture that will make it secure. When asked what would increase their confidence in deploying AI on sensitive data, respondents pointed to advanced infrastructure security like Confidential Computing more than any other single solution.

Confidential Computing solves the data-in-use vulnerability by processing information within hardware-based Trusted Execution Environments (TEEs), often referred to as secure enclaves. Supported by major chip manufacturers like Intel and NVIDIA, these enclaves ensure that data and model weights remain encrypted even during active computation. The cloud provider, the system administrator, and the operating system itself are cryptographically barred from accessing the data inside the enclave.

For a global bank, this means they can run a proprietary fraud-detection model on a public cloud GPU without the cloud provider ever having the technical ability to see the transaction data or the model's inner workings. It provides the cryptographic proof of isolation that regulators demand.

As enterprises scale their AI factories from experimental sandboxes to mission-critical production environments, the focus is rapidly shifting from mere capability to verifiable trust. The organizations that will dominate the next decade of business will not necessarily be the ones with the largest AI models, but rather those that figure out how to deploy them without compromising the very data that gives them their competitive edge.