AI Is Breaking Operations

The operations stack at most data centers was built for a workload that no longer exists.

For twenty years, data center operations was an exercise in steady-state thermodynamics. Cloud and enterprise workloads aggregated millions of unrelated processes across thousands of servers. The law of large numbers smoothed the load curve. Building management systems, DCIM, and SCADA were built around that physics. Their job was telemetry: gather sensor readings, raise alarms, log compliance, occasionally tune setpoints. The control loop ran in human time. Operators read dashboards, called meetings, adjusted on a 15-minute cycle.

That world ended in 2023.

A 16,384-GPU training cluster does not behave like a cloud. Tens of thousands of accelerators move in lockstep through compute and communication phases. Power draw lurches across tens of megawatts in milliseconds. NVIDIA's roadmap is moving rack densities from roughly 40 kW on Hopper toward 1 MW on Feynman, with discussed 2 MW systems beyond that. The stack that worked for the prior workload was never asked to control anything in real time. It was asked to watch.

This piece is not a sustainability essay. It is an argument about P&L. AI workloads have introduced a structural mismatch between the physics of the load and the architecture of the control plane. The mismatch shows up as stranded capacity on the asset side, as preventable downtime on the OpEx side, and as a fast-growing capacity-market surcharge on the rate base. The vendors who built the existing stack cannot solve this with another dashboard. A new category is forming above the existing stack. It is being funded right now.

The cleanest production data on AI workload volatility comes from a paper jointly authored by Microsoft, OpenAI, and NVIDIA in August 2025, titled Power Stabilization for AI Training Datacenters. The paper describes the problem in physical terms most data center designers had never had to consider:

A second Microsoft Research paper, presented at ASPLOS 2024, quantifies the gap. Patel et al., Characterizing Power Management Opportunities for LLMs in the Cloud, measured production Azure clusters and reported three findings that should be familiar to any operator with AI tenants.

LLM training "incurs large swings in power consumption within short durations, up to 37.5% of the provisioned power capacity within 2 seconds, whereas inference only incurs a change of up to 9%." At the cluster level, "LLM training clusters incur massive and coordinated power peaks and offer a very small headroom (about 3%) to oversubscribe power. In contrast, LLM inference clusters offer substantial power headroom (about 21%)." Within a training iteration, model-specific behavior varies sharply. RoBERTa drops to 75 percent of GPU TDP at the iteration boundary. GPT-NeoX drops to 50 percent. Flan-T5 drops all the way to 20 percent, the idle power of the GPU itself.

In a conventional cloud or enterprise workload, no power signature looks like that. The aggregate load on a hyperscale facility a decade ago moved by single-digit percentages over minutes. AI training moves by 37 percent in two seconds, then back, then again, hundreds of times per training iteration, for weeks at a time.

Meta's Llama 3 405B paper acknowledged the same problem on a smaller cluster: "During training, tens of thousands of GPUs may increase or decrease power consumption at the same time. It can result in instant fluctuations of power consumption across the datacenter on the order of tens of megawatts, stretching the limits of the power grid." Meta's engineers, working on a 16,384-H100 cluster, recorded 419 unexpected component failures over 54 days, averaging one every three hours. The workaround was a command, pytorch_no_powerplant_blowup=1, that injects dummy workloads to smooth power draw. SemiAnalysis estimates that at gigawatt scale, the annualized energy cost of running such dummy compute exceeds tens of millions of dollars. That is real money being spent to keep the existing control plane from disconnecting the facility.

Oracle Cloud Infrastructure confirmed the pattern publicly in March 2026. In Behind the Scenes: GPU Power Smoothing for Large-Scale AI Training, OCI Senior Principal Engineer Arash Saifhashemi wrote that "at OCI scale, those swings can reach tens or even hundreds of megawatts. If the oscillation's frequency content lines up poorly with resonant characteristics of grid components, for example turbine generators or long transmission lines, it can create real stability concerns and even mechanical risk." OCI's own mitigation, called AET, is a millisecond-scale GPU heartbeat that injects low-priority GEMM kernels into power valleys with a roughly five-millisecond reaction time. Saifhashemi notes that the hardware-level mitigations in NVIDIA's newer racks "have a finite service life (one to two years, depending on usage) and are not a practical long-term answer."

Ed Ansett, founder of i3 Solutions Group, framed the operational consequence to Data Center Dynamics: "The problem is that you have, for example, 50MW of IT load that the utility is about to see, but it will see it very quickly, and the utility won't be able to respond that quickly. It will cause frequency problems, and the utility will almost certainly disconnect the data center, so there needs to be a way of buffering those workloads."

This is the gap incumbent operations is being asked to absorb without any of the tooling redesigned to absorb it. The market is responding through capital. Roland Berger's HYPERSPACE modeling forecasts data center battery demand growing from roughly 20 GWh today to over 70 GWh by 2030, a 26 percent CAGR, the highest of any battery segment globally. Form Energy and Crusoe announced a 12 GWh iron-air capacity agreement at CERAWeek on March 24, 2026 specifically for AI data center storage starting 2027. The investment thesis behind that storage and behind every behind-the-meter generation deal of the last 18 months is the same: when load cannot be controlled, the asset must be overbuilt.

McKinsey's base case has U.S. data center power capacity growing from roughly 30 GW in 2025 to more than 90 GW by 2030, a 22 percent CAGR. Training racks are heading toward 1 MW each. Inference racks settle at 30 to 150 kW. A 30 MW facility today is on a trajectory to host fewer than 100 racks within the decade, each of them a higher-amplitude, faster-moving load than the rack it replaces.

The pressure on the load side is meeting a control plane that, by every quantitative measure, is regressing under complexity. Uptime Institute's 7th Annual Outage Analysis, published May 2025, is the single most important operational document of the year for any CFO with data center exposure.

Three numbers from that report should be on every operations review. 40 percent of organizations have suffered a major outage caused by human error in the past three years. Of those, 85 percent stem from staff failing to follow procedures or from flaws in the procedures themselves. The share of human-error outages caused specifically by procedure failures rose by ten percentage points in a single year. Uptime now reports 58 percent of such incidents trace to procedure non-compliance, up from 48 percent the prior year. 80 percent of operators believe their most recent significant downtime incident was preventable with better management, processes, or configuration. Power remains the leading cause of impactful outages at 45 to 54 percent of incidents depending on year. Combined power-plus-cooling failures account for roughly 71 percent of all material disruptions.

The 15th Annual Global Data Center Survey, published July 2025, adds the staffing math. Uptime reports that 46 percent of operators struggle to find qualified candidates and 37 percent struggle to retain them. One in ten outages still cause serious or severe disruption. Roughly two-thirds of publicly reported outages over the past nine years involved third-party providers, colocation, cloud, telecom, meaning the failure mode increasingly crosses contractual boundaries and complicates root-cause analysis.

The cost numbers attached to those failures are higher than most boards realize. The Oxford Economics and Splunk 2024 study placed the average enterprise downtime cost at $9,000 per minute, $540,000 per hour, with Global 2000 enterprises losing roughly $400 billion annually to unplanned downtime in aggregate. Uptime's own survey: 54 percent of respondents say their most recent significant outage cost more than $100,000, and one in five say it cost more than $1 million.

The Uptime data is best read as a leading indicator. AI workloads are not yet the dominant cause of outages in the dataset because most AI capacity is still in its first operational year. They are, however, the dominant cause of operational complexity growth. Liquid cooling has no thermal buffer. GPU hardware can suffer damage in seconds, not the minutes that air-cooled systems tolerate. Rack densities have moved from 10 to 30 kW into a regime where a single procedural error in coolant valve sequencing can scrap a tray of $200,000-each GPUs. The existing stack does not close that loop. Humans do. And humans are scarce, exhausted, and increasingly unable to execute procedures complex enough to account for the new physics.

This is the operational debt that AI workloads are stacking on top of an already-stretched workforce. The dashboards report what happened. The operations team is asked to interpret it. The workload that produced the alert has already moved on.

The industrial software market makes a distinction that data center operators have largely ignored. It is the difference between a system of record and a system of action.

A system of record observes. It captures state, logs events, raises alarms, presents dashboards. Its time horizon is the operator's reaction window, minutes to hours. Its design center is documentation, compliance, and post-hoc analytics.

A system of action decides. It executes against an objective function under operational constraints. Its time horizon is the physics of the asset it controls, milliseconds to seconds. Its design center is closed-loop control with humans as supervisors rather than executors.

Every incumbent in data center operations is a system of record. Schneider EcoStruxure IT, Vertiv (which discontinued its Trellis DCIM platform in 2021), Nlyte (acquired by Carrier in 2021 and folded into Automated Logic Controls), Sunbird, Cormant, FNT Command, plus the BMS layer (Honeywell, Siemens, Johnson Controls) and the SCADA layer (Hitachi ABB, GE Digital, AVEVA), all observe assets and present that observation. They have added analytics, predictive maintenance, occasionally discrete automations. None sit inside the millisecond loop that an AI cluster requires.

A control layer does not need a dashboard to act. It acts and tells the operator what it did and why. It pre-empts thermal events instead of cooling them after the fact. It shapes workload distribution to match facility capacity instead of leaving capacity unallocated. It maintains the engineered safety envelope while operating closer to the physical limits of the asset.

Phaidra CEO Jim Gao, formerly head of DeepMind Energy at Google, makes the same point about the cooling stack in his company's product literature: "Traditional liquid cooling relies on Proportional-Integral-Derivative (PID) controllers, which wait for a sensor to register a coolant temperature change before taking action. Because coolant has high thermal inertia, this reactive feedback loop suffers from a 3-to-5-minute delay, resulting in rapid heat spikes that force GPUs to throttle performance to protect themselves."

A three to five minute reactive delay was a tolerable design choice when the load underneath it moved in hours. The same delay is now responsible for measurable revenue loss on every high-density GPU hall in production. The distinction the AI workload era has made unavoidable is this: monitoring tells you what happened. Control changes what is happening. They are different categories of software.

The temptation will be to read each incumbent's product roadmap and conclude that AI control is being addressed. It is not, for three structural reasons.

First, the business model. Schneider, Vertiv, Carrier, Honeywell, Siemens, and Johnson Controls sell hardware. Software is sold to drive hardware refresh cycles, not to extract performance from competitive hardware. A control layer that genuinely optimizes a heterogeneous facility is value-destructive for the dominant equipment vendor at that facility. Their incentives are to lock customers into a single-vendor stack, not to orchestrate across vendors.

Second, the architecture. DCIM platforms were designed against air-cooled, slow-changing thermal loads. Their data models, sensor cadence, and integration surfaces were not built for liquid cooling at sub-second resolution, for GPU-aware scheduling, or for grid signals. Industry analysis of the 2026 DCIM landscape observes that platforms "built around static air-cooled rack assumptions handle the shift to direct-to-chip liquid cooling, rear-door heat exchangers, and immersion cooling unevenly," and that current operator behavior is largely to bolt on a second platform to fill the thermal gap rather than replace the incumbent DCIM. The industry response is layering, not redesign.

Third, the trust gap. Uptime's 2025 Global Data Center Survey asked operators whether they would allow AI to control equipment. Most said yes for sensor analysis and predictive maintenance. Most said no for configuration changes, equipment control, or staff management decisions. Incumbents read this as a reason to slow down. The right reading is that the trust gap exists because the existing platforms cannot demonstrate closed-loop safety. A purpose-built control layer with verifiable behavior, engineered guardrails, and clear human override is the only way the trust gap closes. The incumbent dashboards cannot close it because they are not closed-loop systems.

The analog in adjacent industries is instructive. Palantir Foundry did not replace SAP or Oracle ERP. It sat above them and made decisions the legacy systems could not. C3 AI, Uptake, and Cognite built similar control layers above industrial historians and SCADA in oil and gas, manufacturing, and utilities. In each case, the incumbent system of record stayed in place. A system of action category formed above it. Investor multiples migrated. The same pattern is now beginning in data center operations.

Now the dollar math. Three independent industry sources converge on the same conclusion: AI workloads, layered onto the existing operations stack, are stranding 20 to 40 percent of the asset that operators have already paid to build.

Sunbird DCIM's customer data places the cost of bringing a single server-ready cabinet to operational state at approximately $15,000 to $20,000 per cabinet, with traditional capacity planning techniques stranding up to 50 percent of rack-level space capacity in a typical facility. Nordcad's 2026 stranded capacity analysis cites facilities running 3.9 times more cooling capacity than active IT load, an extreme illustrative case but directionally accurate.

This stranding is now monetized at unprecedented rates. Cushman & Wakefield's wholesale lease data shows that Northern Virginia wholesale rates, which sat at $103 per kW-month in 2020 after a multi-year decline, reversed and rose more than 30 percent in 2022 alone. Their H2 2025 Americas Data Center Update reports vacancy at 4.2 percent and pre-commitment rates approaching 89 percent across 25.3 GW under construction. Construction costs have moved from roughly $7 to $8 million per MW pre-COVID to a projected $11.3 million per MW global average in 2026.

Consider a representative 30 MW colocation facility in a tier-one U.S. market. At blended hyperscale wholesale rates around $180 to $220 per kW-month, the gross contracted revenue at full lease-up is on the order of $65 to $80 million per year. If 12 percent of the IT load capacity is stranded, well within the range every primary source describes as recoverable, the operator has effectively paid to build roughly 3.6 MW of capacity that will never be billable under the current control regime. At $200 per kW-month, that is $8.6 million of annualized contracted revenue forfeited against build costs north of $40 million. At a typical 25 to 35 percent EBITDA margin on a stabilized colo asset, recovering even half of that stranding moves the asset's EBITDA by single-digit millions per year on a single facility.

This is not theoretical. AI tenants are sophisticated about contracted-versus-actually-deliverable power. The operator who cannot prove, under workload-aware control, that the stranded portion of the asset is actually deliverable, leaves margin on the table at every renewal.

The story compounds at the grid layer. Monitoring Analytics, PJM's independent market monitor, calculated that forecast data center load drove $9.3 billion of the increase in PJM's 2025/2026 capacity auction, a 174 percent increase in the clearing price attributable to data center demand alone. PJM capacity prices rose from $28.92 per MW-day in the 2024/2025 delivery year to $329.17 per MW-day in the 2026/2027 auction. That is an order-of-magnitude move in two years.

On December 17, 2025, PJM's auction for the 2027/2028 delivery year fell short of the reliability requirement by 6,623 MW. It was the first time the auction failed to meet the reliability requirement since the capacity market was created in 2007.

CFOs at large data center tenants now face a capacity-market surcharge that is structurally larger than their pre-AI electricity bill. The operator who can demonstrably bid load flexibility into capacity markets, under closed-loop control rather than under a promise, captures rate relief the operator who cannot will pay for.

The capital has arrived. The category does not yet have a settled name, but its shape is clear.

Emerald AI, founded in 2024 by former U.S. climate envoy advisor Dr. Varun Sivaram, has raised $68 million in 16 months. The March 2026 $25 million strategic expansion round was led by Energy Impact Partners with participation from NVentures (NVIDIA's venture arm), Eaton, GE Vernova, In-Q-Tel, Lowercarbon Capital, Radical Ventures, Salesforce Ventures, Samsung Ventures, and Siemens. Emerald's Conductor platform orchestrates AI workload power consumption against grid signals. NVIDIA has named Emerald and Phaidra as orchestration partners in the Vera Rubin DSX reference design for new AI factories, including the 96 MW Aurora facility in Manassas, Virginia, described as "the world's first power-flexible AI factory at commercial scale."

Phaidra, founded by Jim Gao (formerly head of Google DeepMind Energy), Vedavyas Panneershelvam (formerly AlphaGo research engineer), and Katie Hoffman (formerly of Trane Technologies), has raised $120 million total. Phaidra applies reinforcement learning directly to industrial control, including cooling for CoreWeave and Applied Digital sites on NVIDIA hardware. Phaidra reports 10 to 15 percent energy savings in the first year of deployment with Fortune 100 pilot customers.

Around these two, the broader category includes Verdigris in real-time power monitoring, GridX in utility-side optimization, internal control tooling at CoreWeave and other GPU-cloud operators, and adjacent platform plays. NVIDIA itself acquired Israeli workload-orchestration startup Run:ai in April 2024 for approximately $700 million, with the deal closing December 30, 2024 after EU and DOJ review. The signal was unambiguous: workload-aware orchestration is now strategic infrastructure.

The DCIM incumbents are reacting, but their reactions are predictable. Schneider has rebranded EcoStruxure IT to "DCIM 3.0." Vertiv has built post-Trellis offerings around its UPS portfolio. Carrier is integrating Nlyte with Automated Logic Controls. None of these are systems of action. They are upgraded systems of record sold by hardware companies into the hardware refresh cycle.

Two adjacent capital flows matter. First, the battery thesis: the Form Energy and Crusoe 12 GWh iron-air agreement signed at CERAWeek on March 24, 2026 explicitly frames the deal as enabling "reliable, scalable, and cost-effective" power for the AI economy. Tesla's Megapack is now the default first-line buffer at xAI's Memphis Colossus site. Second, behind-the-meter generation: large hyperscalers are signing multi-gigawatt deals with natural gas turbine manufacturers as bridge power for sites where the grid cannot deliver under controlled load. Both flows are responses to the same control gap. When load cannot be modulated intelligently, the asset must be overbuilt.

That overbuild is the largest single line item in McKinsey's $6.7 trillion global data center capex forecast through 2030. The investor question is whether overbuild or orchestration captures the operating-margin uplift. The historical analog argues for orchestration. In every adjacent industrial vertical where a control layer formed above the legacy stack, the control layer captured more of the margin uplift than the upstream hardware vendors did.

The diagnostic is straightforward. A CFO with material data center exposure, owned, leased, hyperscale-contracted, or as a colo investor, should ask the following at the next operations review.

First, what is our stranded-capacity number, expressed in MW and in annualized contracted dollars at current market wholesale rates? If the operations team cannot produce a single number with confidence, the BMS and DCIM stack is operating as a system of record rather than a system of action.

Second, in the most recent 12 months, what fraction of our incidents trace to procedure failures or staff non-compliance with documented runbooks? Uptime's industry benchmark is 58 percent. Below that, the operation is performing better than peers. Above it, staffing and procedure complexity have outrun the existing control plane, and additional automation is the only economic answer.

Third, what is our capacity-market exposure in PJM, ERCOT, MISO, CAISO, or our local utility? At 2026 PJM clearing prices, every megawatt of contracted load that cannot demonstrate flexibility is paying full rate. Operators that can bid demonstrable closed-loop flexibility receive credits.

Fourth, what is our liquid-cooling thermal-response time on our highest-density halls? A three to five minute PID response is industry-standard. It is also long enough that a single coolant flow excursion can throttle a tray of $200,000-each GPUs for hours. The economics of one such event pay for a control-layer pilot.

Fifth, on our incumbent DCIM, BMS, and SCADA contracts, what is the closed-loop control surface? If the answer is dashboards, alarms, and discrete automations, the system is observational. Buyers should not pay for control on a software contract that does not provide control.

Sixth, what fraction of our recent outages crossed a third-party boundary? Uptime puts the public-incident figure at two-thirds. Above that, the operation is exposed to vendor failures it cannot diagnose without a unifying control layer.

Each of these questions produces a number. Each number is either improving or it is not. A CFO who runs this diagnostic quarterly will identify the operating-margin gap before the auditors do.

The pattern repeating in data center operations has played out three times in adjacent industrial software markets over the last fifteen years.

In each prior cycle, a workload changed faster than the incumbent control plane could adapt. Industrial historians could not handle continuous-process optimization in chemicals. AspenTech and AVEVA built layers above them. SCADA could not handle distributed energy resources at utility scale. Uplight, AutoGrid, and AspenTech AIoT formed above. ERP could not handle real-time supply chain decisioning under volatility. o9 Solutions and Kinaxis formed above. In every case, the incumbent kept its installed base and lost the operating leverage. The control layer captured the margin uplift.

Data center operations is now in cycle four. The workload is AI. The incumbent stack is BMS, DCIM, and SCADA. The failure mode is closed-loop control under fast physics. The capital is flowing. NVIDIA has named the partners in its reference architectures. Hyperscalers are buying the category through M&A (NVIDIA-Run:ai) and through co-development (the Microsoft-OpenAI-NVIDIA power stabilization paper). The mid-market of 10 to 50 MW facilities, the bulk of the operator universe, is exactly where the venture-backed entrants will land first, because hyperscalers can build internally and small enterprises cannot afford the integration cost.

Jensen Huang's framing of the new world at GTC 2025 was characteristically direct:

When revenue is power-limited, every megawatt of stranded capacity is foregone revenue. Every avoidable outage is foregone EBITDA. Every minute of unrecovered thermal headroom is GPU throughput left on the table. The capex race of 2022 through 2024 is conceding the stage in 2026 to an OpEx and operational-control race. CFOs at colo operators, hyperscalers, and tenant enterprises are converging on a single question: which control layer turns my asset into a revenue-producing factory under AI loads, rather than an overbuilt cost center?

The category that answers that question is the one being built right now. An orchestration layer above the existing BMS, DCIM, and SCADA stack. Vendor-neutral on hardware. Closed-loop on physics. Financially measurable in stranded-capacity recovery and capacity-market participation. The category is bigger than any one company. It is the most important new line item in data center operations for the next decade.

Arcestra.AI is the orchestration layer being described.