The rapid and unprecedented expansion of artificial intelligence is
driving an insatiable demand for global computing power, placing data centers squarely at the center of a growing policy and economic debate in the United States. While newly proposed facilities promise to unlock vital computing capacity for a rapidly digitizing economy, mounting concerns over surging electricity costs, heavy water consumption, and broader environmental impacts have sparked intense public scrutiny. These pressures have even prompted legislative proposals calling for a temporary pause on new construction.
However, beneath the surface of the environmental and infrastructural debate lies another critical, often overlooked challenge: securing the physical and digital infrastructure that powers the AI revolution. As modern enterprises increasingly rent graphics processing unit (GPU) capacity from specialized cloud providers—known widely in the industry as neoclouds—complex questions arise regarding hardware security, shared multi-tenant infrastructure, and the strict isolation of sensitive customer workloads.
To explore these pressing issues, we spoke with Yakir Kadkoda, Co-Founder and CTO of Lava Security, a cybersecurity firm dedicated to protecting neoclouds and advanced AI data centers. Kadkoda, who previously served as the Director of Security Research at Aqua Security and currently co-leads the OWASP Top 10 for Neocloud and AI Data Center Security Risks project, shared his insights on the legislative debate over data center expansion, hidden vulnerabilities in GPU cloud infrastructure, and the shared responsibilities of operators, customers, and policymakers.
Why the U.S. Needs More AI Data Centers and the Ripple Effects Beyond Big Tech
Artificial intelligence is no longer a localized feature or a passing tech trend; it has evolved into foundational infrastructure, comparable to power grids or modern roadways. Every modern hospital, bank, university, and manufacturing plant will soon run sophisticated AI models to remain competitive. Yet, the vast majority of these institutions lack the capital to purchase GPUs by the thousand. Instead, they must rent them. An expansion in data center capacity directly translates to greater market supply, more competitive pricing, and drastically shorter wait times for organizations outside the sphere of Big Tech hyperscalers.

Consider real-world examples outside the tech giants. A hospital network handling sensitive patient data cannot simply transmit that information to a public cloud application programming interface; it requires localized or securely rented GPUs, and today those institutions often wait months or pay steep premiums for access. Similarly, university research laboratories frequently find themselves queuing for weeks on shared academic clusters, where expanded capacity would directly accelerate scientific breakthroughs. Meanwhile, emerging startups operating on neoclouds without heavy capital expenditures feel the immediate sting of every market price hike.
If federal lawmakers were to enact a moratorium on new data center construction and expansion, startups, researchers, and smaller enterprises would feel the blow first. Big Tech corporations already command enormous internal compute capacities. Everyone else depends on renting. Restricting the construction of new facilities would artificially inflate prices, crush availability, and concentrate even more economic power into the hands of a few dominant players. Furthermore, the relentless global demand for AI development would not simply halt; it would migrate seamlessly to foreign nations continuing to invest heavily in foundational digital infrastructure.
Assessing Existing Capacity, Unused Hardware, and Environmental Concerns
A recent analysis by Lava Security highlights the complexities of how existing infrastructure is managed. Between March and May 2026, researchers identified more than 2,100 servers publicly exposing monitoring data from over 12,000 NVIDIA GPUs, representing roughly $100 million in hardware. Nearly 60 percent of those resources reported zero utilization at the exact moment of observation.
Kadkoda clarifies that this snapshot does not indicate that 60 percent of all industry GPUs sit permanently idle. Some hardware may have been transitioning between jobs, held in reserve, or undergoing routine maintenance. Rather, the more alarming discovery was not that GPUs went unused, but that thousands of expensive, high-performance AI systems were inadvertently leaking sensitive operational telemetry to anyone browsing the open internet. Addressing these gaps requires improving industry-wide visibility, tightening security protocols, and optimizing existing hardware utilization concurrently with the construction of new physical facilities.
While supporters of data center moratoriums raise valid concerns regarding electricity consumption, water usage, and environmental strain, stopping construction entirely is ultimately the wrong solution. Society does not stop building hospitals simply because they consume electricity; instead, engineers improve how those facilities are constructed, powered, and operated. The same logic applies to AI infrastructure. Establishing clear regulatory requirements, enforcing them strictly, and holding operators fully accountable provides a sustainable path forward. The definitive answer to infrastructure challenges is better infrastructure, not less of it.

When advising local municipalities considering the establishment of large data centers, communities should demand firm commitments in writing across three key pillars: measurable economic benefits, responsible resource consumption, and robust operational security. Developers should transparently detail projected electricity and water usage, outline who absorbs the cost of local grid upgrades, and commit to creating permanent jobs. Crucially, towns must require comprehensive security plans. Modern AI data centers are immensely complex environments featuring thousands of interconnected components, including servers, management controllers, networking gear, and industrial control systems.
The Hidden Security Risks of Rented GPU Infrastructure
Renting GPUs from specialized cloud providers fundamentally changes an organization’s risk profile. While an enterprise hands over direct control of physical hardware, firmware, networking, storage, and management systems to a third-party provider, it never hands away the underlying security risk. Customers routinely share physical infrastructure with other tenants, sometimes running on the exact same hardware nodes or crossing interconnected internal networks.
The primary obstacle for these customers is a stark lack of visibility. While clients can closely monitor their own running workloads, they typically possess zero insight into the underlying infrastructure supporting them. They cannot independently verify whether neighboring tenants are properly isolated, whether the foundational hardware is secure, or whether previous users left residual data behind. This creates an untenable paradigm where customers are forced to blindly trust infrastructure they cannot see or verify.
Lava Security’s own public research underscores these dangers. By renting GPU infrastructure like standard customers, researchers discovered that virtual machine and container-level security boundaries frequently failed to extend to the underlying infrastructure. In certain instances, researchers traversed internal management systems that should have been entirely inaccessible to standard tenants, establishing a potential pathway from a single rented machine directly into the provider’s broader network. Notably, these discoveries rarely required sophisticated exploits; often, the vulnerabilities stemmed simply from improper infrastructure configuration.
Navigating Tenant Handover and Overlooked Blind Spots Like Baseboard Management Controllers
When a high-performance GPU server transitions from one customer to the next, cloud operators must rigorously verify that the previous tenant has been entirely expunged—not just from the operating system, but from the underlying physical infrastructure. This necessitates securely sanitizing storage devices, stripping away lingering credentials and cryptographic keys, wiping persistent data, and thoroughly verifying the integrity of system firmware and management controllers. A simple system reboot or operating system reinstallation falls woefully short.

Economic pressures often complicate these security measures. GPU providers generate revenue when their hardware actively runs compute workloads, creating an inherent financial incentive to minimize the turnaround time between clients. However, speed must never compromise tenant isolation. One customer’s proprietary models, datasets, or credentials must never bleed into the environment of the next.
Compounding these challenges are overlooked infrastructural components like Baseboard Management Controllers (BMCs). A BMC functions as a dedicated, independent computer embedded within nearly every enterprise server, possessing advanced capabilities to monitor and manage hardware remotely even when the primary operating system is powered down. Traditional cloud security tools routinely fail to monitor this deeply embedded layer. Lava’s research identified more than 36,000 management interfaces directly exposed to the open internet, many burdened by critical authentication flaws. An attacker who compromises a BMC effectively gains control over the entire physical server, entirely bypassing standard operating system defenses.
Establishing Industry Standards and Accountability
To mitigate these risks before deploying valuable intellectual property onto rented hardware, organizations should demand three distinct pieces of evidentiary proof from cloud providers. First, they must require verifiable proof of tenant isolation to ensure neighboring workloads and management networks remain completely walled off. Second, companies should request proof of hardware and firmware integrity, confirming that the underlying physical components have not been subverted. Third, providers must supply proof of secure tenant handover, demonstrating that prior user remnants have been permanently erased. Relying solely on standard compliance paperwork, such as a SOC 2 certificate, is insufficient; true security requires verifiable technical evidence.
As lawmakers and industry leaders weigh the future of American technological competitiveness, the path forward relies on a balanced approach. By holding data center operators strictly accountable to local communities regarding resource usage while mandating independent, stack-wide security verifications, the nation can foster innovation without compromising public safety or data integrity. Securing the AI revolution requires embedding accountability into the very foundation of digital infrastructure from day one.

Comments (0)