The mathematics are straightforward, if a little dizzying. Supply chain constraints have led to delays or cancellations for approximately half of planned U.S. data center projects in 2026, according to Bloomberg and Wood Mackenzie. It's not a demand problem—AI's hunger for compute shows no signs of abating. Nor is it capital. The hyperscalers remain committed to spending hundreds of billions.
The real bottleneck? Power. Or more precisely, the electrical grid's inability to deliver it quickly enough, combined with the data center industry's surprisingly inefficient use of what it already has.
That inefficiency is where Niv-AI thinks it's found an opening. The Tel Aviv startup surfaced from stealth in mid-March with $12 million in seed funding and an eyebrow-raising assertion: somewhere in the neighborhood of 30% of contracted power capacity in AI data centers sits stranded, unusable because of the way operators handle GPU power spikes. The company's pitch centers on millisecond-scale monitoring of GPU power "fingerprints" and predictive orchestration aimed at unlocking that dormant capacity without sacrificing performance.
Whether the claim holds up remains to be seen. But the timing is undeniably sharp. Reporting from earlier this year, citing data from Bloomberg and Wood Mackenzie, indicated that supply-chain snarls around transformers and switchgear have contributed to delays or outright cancellations for roughly half of planned U.S. data center projects. Meanwhile, Goldman Sachs Research has projected that satisfying AI-driven electricity demand will require approximately $720 billion in grid investment by 2030—assuming the industry can build fast enough to stay ahead of the curve, which is far from guaranteed.
The Scale of the Challenge
The sheer scale has a numbing quality to it. Global data center electricity consumption was estimated to be around 415 terawatt-hours, accounting for about 1.5% of global electricity, according to the IEA 2025 report. The IEA is forecasting annual growth around 15% through 2030. In the United States, where the major cloud providers are racing to deploy AI infrastructure, the picture sharpens. Data centers accounted for approximately 4.4% of U.S. electricity in 2023, according to a Department of Energy analysis released late that year, with projections suggesting the share could climb to somewhere between 6.7% and 12% by 2028.
BloombergNEF, in a December 2025 analysis, estimated U.S. data center power demand could reach 106 gigawatts by 2035. PJM Interconnection, the grid operator overseeing much of the Mid-Atlantic and Midwest, is preparing for up to 30 gigawatts of new data center load between now and 2030. Goldman Sachs expects data center power demand to jump 165% by 2030 compared to 2023 levels.
These aren't theoretical exercises. In January, Georgia's Public Service Commission approved what amounts to a roughly 50% expansion in generation capacity, driven in large part by data center demand. The following month, the Department of Energy announced a $26.5 billion loan package to utilities in Georgia and Alabama for power expansion tied heavily to data centers. By April, the Associated Press was reporting that utilities nationwide are struggling to balance this load growth against clean-energy commitments—a tension that's unlikely to resolve itself anytime soon.
Yet even as grids strain under the weight, a different kind of constraint has emerged. Research published on arXiv in early April characterized "execution-idle" as a distinct energy state in GPU clusters—stretches where GPUs remain at high power despite minimal activity. AI training and inference workloads create wild power fluctuations, transients that challenge both facility infrastructure and grid connections. Data center operators, forced to plan for worst-case spikes, end up reserving capacity that sits idle most of the time.
Which brings us back to Niv-AI's central claim.
The Physics of Waste
NVIDIA's GB200 NVL72 rack configuration—a reference design for next-generation AI infrastructure—pulls approximately 120 kilowatts at full load, according to company documentation. That's roughly triple the power density of traditional enterprise racks. Vertiv and NVIDIA developed a 7-megawatt site blueprint incorporating direct liquid cooling to manage the thermal load, with designs supporting up to 132 kilowatts per rack.
The trouble isn't the average draw. It's the spikes. GPU workloads can surge from near-idle to peak in milliseconds, creating demand transients that electrical infrastructure—built for steadier, more predictable loads—wasn't really designed to handle. Operators respond conservatively: they contract for peak capacity, install oversized uninterruptible power supplies, implement cautious power budgets to avoid tripping breakers or destabilizing feeds.
Niv-AI's founders—CEO Tomer Timor and CTO Edward Kizis—frame this as fundamentally a data problem rather than a hardware one. Their view: the current approach, which sizes infrastructure for worst-case scenarios and throttles compute to stay within safe limits, leaves meaningful capacity stranded. The company's technology deploys high-resolution sensors to capture millisecond-scale GPU power signatures, then applies predictive algorithms to anticipate and synchronize loads across racks.
"We're building an intelligence layer between data centers and the grid," Timor told TechCrunch in March—a framing that's either prescient or wishful thinking, depending on whether the technology delivers.
The 30% stranded capacity figure, cited in both the company's announcement and subsequent reporting, hasn't been independently verified. But it does align with broader industry observations about the gap between contracted and utilized capacity. If even directionally accurate, the implications are substantial. At 120 kilowatts per rack, 30% stranded capacity translates to roughly 36 kilowatts per rack sitting idle. Scale that across hundreds or thousands of racks, and the math becomes meaningful even for hyperscalers accustomed to operating at enormous scale.
Carving Out a Niche
Industry experts consider Niv-AI to occupy a specific niche in the data center power efficiency market. Founded in May 2025 with a team of 10 employees as of March, the company is deploying rack-level sensors now and targeting operational systems in what it describes as "a handful of U.S. data centers" within six to eight months—putting us somewhere in the September-to-November window, per TechCrunch's reporting.
The $12 million seed round, led by Glilot Capital Partners with participation from Grove Ventures, Arc VC, Encoded VC, Leap Forward Ventures, and Aurora Capital Partners, positions the startup at the intersection of infrastructure software and energy management. A Glilot partner characterized Niv-AI as building a "foundational control plane for data center power." Grove Ventures board commentary, captured by TechCrunch in March, emphasized that current data center build approaches are unsustainable and that Niv-AI aims to improve utilization and grid responsibility through synchronized power and compute management.
The competitive landscape is heating up. Two days after Niv-AI's announcement, Claros emerged with a $30 million seed round targeting integrated voltage regulators and DC power distribution to reduce conversion losses. The approaches differ—Claros focuses on electrical architecture at the chip and rack level, while Niv-AI operates at the orchestration layer—but both reflect investor conviction that power efficiency represents a critical chokepoint.
The broader technology stack is converging around these constraints in ways that might create an opening for software-driven approaches. NVIDIA publicized an 800-volt DC architecture for megawatt-scale racks at GTC 2025; Texas Instruments unveiled an 800-volt-to-GPU core conversion system at the following year's conference; Infineon has been promoting 800-volt DC data center power paths since late 2025. Higher-voltage distribution reduces resistive losses and simplifies power delivery for ultra-dense racks.
At the same time, liquid cooling—once reserved for supercomputing niche cases—has gone mainstream. Schneider Electric acquired Motivair in early 2025 and now offers rear-door heat exchanger systems supporting up to 75 kilowatts per rack, along with cold-plate solutions for AI and high-performance computing workloads.
These developments create something of an enabling environment for predictive power management. High-voltage DC, liquid cooling, rack-scale integration—all provide the physical foundation. Software-driven orchestration, Niv-AI's territory, sits atop that foundation, potentially extracting value from infrastructure already deployed or under construction.
Policy Meets Physics
Grid reforms and policy shifts are moving in parallel, though not necessarily fast enough. FERC Orders 2023 and 2023-A, finalized over 2023-2024, reformed interconnection queues to address backlogs. PJM has cleared much of its queue and initiated a new cycle process starting this year, explicitly planning for large data center loads. States are adapting in fits and starts: Virginia approved new rate classes for loads exceeding 25 megawatts in 2025; Georgia expanded generation capacity by roughly half in January.
Yet the physical reality remains stubbornly challenging. Transformer lead times stretch months to years. Half of planned U.S. builds face delays or worse. Utilities, particularly in regions like Virginia's "Data Center Alley," are navigating fraught tensions between economic development, ratepayer impacts, and decarbonization goals. The European Union's Energy Efficiency Directive, updated in 2023, now mandates energy and performance reporting for data centers of 500 kilowatts or larger, with the European Commission publishing its first analysis of 2024 data last July.
Against this backdrop, the economics of stranded capacity start to look compelling. Hyperscaler capital expenditure signals underscore the stakes. Amazon indicated $100 billion in capex for 2025, per CNBC reporting in February. Microsoft spent $34.9 billion in the first quarter of fiscal 2026, with plans to increase AI capacity by more than 80% during the fiscal year, according to Data Center Dynamics reporting from October. Meta's 2026 capex, per investor coverage in February, could range from $115 billion to $135 billion, heavily concentrated on AI infrastructure.
When capital deployments reach that scale, marginal improvements in utilization generate outsized returns. If Niv-AI's approach unlocks even a fraction of the claimed 30% stranded capacity, the value proposition extends beyond simple cost savings to timeline acceleration—enabling operators to extract more compute from existing electrical connections while waiting for grid infrastructure to catch up.
Which is perhaps the most compelling argument for what Niv-AI is attempting, even if the company itself falls short.
The Road Ahead
Niv-AI's immediate milestones are straightforward enough: deploy sensors, validate the predictive models in live environments, demonstrate capacity unlocks without performance degradation or new failure modes. The company's pilots, expected to go operational later this year, will provide the first independent data points on whether this actually works at scale.
The broader industry challenge is considerably messier. Even aggressive efficiency gains won't eliminate the need for massive grid expansion. The IEA's 15% annual growth rate for data center electricity through 2030, Goldman Sachs' $720 billion infrastructure estimate, PJM's 30-gigawatt load projection—all assume substantial new generation and transmission capacity coming online. Efficiency buys time and reduces the magnitude of the build-out. It doesn't replace it.
But in an industry where half of planned builds are hitting delays, time has real value. The question was never whether AI would continue driving exponential compute demand—that trajectory appears more or less locked in. The question is whether physical infrastructure can keep pace, and whether software intelligence can extract more value from constraints that aren't disappearing anytime soon.
Niv-AI is wagering that the answer to the second question creates breathing room for the first. Whether that bet pays off—for the company, for the hyperscalers, for utilities scrambling to keep up—depends on execution in a market that doesn't forgive vaporware. The pilots will tell the story. Until then, it's just another startup with $12 million and a compelling theory about what's broken.
