Pivot: From Classroom to Server Hall
A 17-year-old high school student parsed a year of utility bills for his district and found $360,000 in gas overbilling, as reported by Y Combinator's Edviro page. That discovery, and the realization that no one on the facilities team had the tools to catch it, became Edviro.
Now the Y Combinator-backed platform is opening partnerships with data center operators. The shift is deliberate. Edviro's co-founders, Hursh Shah and Tanuj Siripurapu, both 19, built for the messy reality of physical infrastructure first: aging, mixed-vendor equipment, lean facilities teams, and almost no budget margin for waste. Schools operate large portfolios under those constraints. The company has deployed across 34 school facilities and verified more than $400,000 in avoidable energy costs, Y Combinator's Edviro page found.
The same architecture now targets a different pressure. Data centers are expected to drive roughly half of U.S. electricity-demand growth through 2030, according to Y Combinator's Edviro page. But the problem isn't only generation. Power must reach the right sites on data-center timelines with the reliability their workloads require. Once a facility has that power, operators still need to extract maximum useful compute from every available megawatt without compromising uptime. Inside a data center, power, cooling, controls, equipment, workloads, and maintenance are tightly coupled: a workload shift changes the thermal profile, a cooling-control change affects both energy use and equipment margins, a maintenance decision alters redundancy and available capacity. Yet the data and workflows used to manage these systems remain fragmented across BMS, DCIM, EPMS, meters, sensors, maintenance software, vendor portals, and spreadsheets.
Edviro's pitch: its "energy world models" (continuously updated representations of how a facility behaves) can bridge that fragmentation. The platform connects to a facility's BMS, utility, sensor, equipment, and maintenance systems to build a model that estimates the current state of the physical system, learns how components interact, and predicts how it will respond to a proposed action. It then uses that model to detect anomalies, simulate interventions before deployment, recommend the highest-value action while accounting for cost and reliability constraints, coordinate approvals and work orders through AI agents, and verify whether the expected impact actually occurred against meter and billing data. The company prices per connected site, not per user, and lets customers start with a single pain point before expanding.
The startup is explicit about what it lacks: full compatibility with every data center product, or autonomous control without operator approval. "We initially keep operators in control of every consequential action," the YC page states. "As the models are validated and earn trust, more of the operating loop can be automated." Edviro is recruiting data center operators as early design partners, particularly facilities dealing with power or cooling constraints, capacity-expansion planning, recurring manual investigations, fragmented controls, or reactive maintenance. The company is also hiring energy-modeling engineers across thermal systems, power systems, controls, scientific machine learning, and physical simulation. Edviro will appear at Yotta 2026 in Las Vegas, September 28–30, where it plans to show data-center-specific models and safety constraints built directly with operators.
The Crunch: Numbers That Stop Being Abstract
Lawrence Berkeley National Laboratory measured U.S. data center electricity use at 176 terawatt-hours in 2023 — 4.4 percent of the nation's total. That figure tripled from 58 TWh in 2014. The Department of Energy projects 325 to 580 TWh by 2028, a range that would push data centers to 7–12 percent of U.S. demand. Globally, the International Energy Agency counted 415 TWh in 2024, roughly 1.5 percent of world electricity.
| Forecast | 2030 Projection (TWh) | Notes |
|---|---|---|
| IEA base case | 945 | Global |
| Deloitte | 1,065 | Global |
| Goldman Sachs | 160–165% increase | Over 2023 levels |
| Berkeley Lab 2026 reference | 649 | U.S. only (range 521–843) |
The surge is not business as usual. For two decades U.S. electricity consumption barely grew. AI workloads broke that plateau. Training and inference now drive demand at several times the historical pace. NVIDIA's H100 GPU alone carries a 700-watt thermal design power, before the rest of the server is counted. Accelerated-server electricity use climbs roughly 30 percent annually through 2030; conventional servers crawl at 9 percent. In 2023 most AI server energy went to inference. Berkeley Lab models training taking a larger share by 2028 as more powerful chips deploy. Cooling and non-IT infrastructure will absorb about one-fifth of the net global increase through 2030.
Power density tells the operational story. Traditional racks draw 5–15 kilowatts. AI-optimized racks demand 30 to over 100 kW. That density overwhelms local grid capacity. Wholesale electricity prices near U.S. data center clusters have risen 267 percent. The IEA characterizes a conventional facility at 10–25 megawatts; a hyperscale AI campus hits 100 MW or more, and the largest planned projects run far larger. Nearly half of U.S. capacity sits in five regional clusters. The IEA estimates one in five planned projects faces delays without better grid integration.
Grid operators are blinking. Permitting for new generation and high-voltage transmission takes over a decade in the United States and Europe, with no guarantee sustained growth justifies the build. China added more than 400 gigawatts of new capacity in a single year. Utilities here struggle to string wire fast enough. The UN warns of voltage oscillations, unintended disconnections, and cascading failures, already appearing in renewable-heavy systems where volatile AI loads spike faster than grids can respond. Denmark's grid operator Energinet halted new connection agreements after an explosive demand surge. The country had 398 MW installed in 2026 with 208 MW under construction and 1.2 GW more projected by 2030. Ireland's data centers hit 21 percent of national electricity in 2023, on track for 30 percent by the early 2030s. The Netherlands imposed a nine-month moratorium on hyperscale permits. California now requires data centers to report water and electricity use, pay their own grid-upgrade costs, and meet clean-energy procurement rules, legislation designed to stop cost-shifting to ratepayers.
The economics leave operators no cover. AI servers consume six to eight times the electricity of traditional units. Anthropic projects a single frontier model will need five gigawatts by 2027; the U.S. AI sector overall will require 50 GW of new capacity by 2028. Former Google CEO Eric Schmidt told Congress to expect 29 GW more by 2027 and 67 GW by 2030. Deloitte sees AI data center demand jumping from 4 GW in 2025 to 123 GW by 2035. McKinsey tallies $6.7 trillion in global data center capital through 2030; NVIDIA estimates $1 trillion for AI-specific upgrades alone. The AI data center market itself may swell from $236 billion in 2025 to $934 billion by 2030, a 31.6 percent CAGR.
Hyperscalers are reacting. Microsoft, Google, and Amazon have committed to pay for power at any price. Microsoft's 2025 sustainability report shows emissions up 23.4 percent since 2020 alongside a 168 percent energy increase. In 2024, 40 percent of data center electricity came from natural gas, 24 percent from renewables, 20 percent from nuclear, 15 percent from coal. While 56 percent of new capacity through 2035 is slated as renewable, nearly two-thirds of incremental generation will still come from fossil fleets. Microsoft signed a 10.5 GW deal with Brookfield Renewable Partners; Google struck a 3 GW agreement with the same firm. The $5 billion Brookfield–Bloom Energy partnership deploys fuel cells behind the meter. Google works with NextEra on gigawatt-scale clean energy. FuelCell Energy taps captured coal-mine methane for on-site generation.
Operators are also learning to shape demand. Hyperscalers already shift non-urgent training and background jobs to hours when renewables flood the grid. At an Oracle cloud site, an AI workload manager dynamically throttled less time-sensitive jobs during a grid stress event, cutting the facility's draw 25 percent for three hours without degrading service. Duke University found that if data centers nationwide curtail only during the top few peak hours each year, the U.S. grid could absorb roughly 100 GW of additional load without a single new power plant. The Agora–Deloitte 2026 study puts the same figure at 4 GW of avoided fossil backup from just 120 flexible hours annually.
Water sharpens the constraint. U.S. data centers drank 17 billion gallons in 2023; 84 percent went to hyperscale and colocation. Direct hyperscale consumption could hit 16–33 billion gallons annually by 2028. One Oregon city reported nearly 30 percent of its water going to Google data centers, a tripling over five years. The IEA estimates a 100 MW facility averages 2 million liters per day, with 725,000 liters consumed on site.
The message is consistent across every forecast: supply cannot catch demand on the current timeline. The only lever left is using every megawatt more intelligently — in real time, across the fleet, with verification that the savings actually materialize. That is the operational gap Edviro is built to close.
Origin Story: A $360,000 Itch
Hursh Shah was a high school student when a teacher handed him a year's worth of utility bills for a class project. He wrote software to parse the data and found that PG&E had been overcharging the district on gas for the entire year. That insight, that the district lacked the tools or bandwidth to catch it, became the seed of Edviro.
Shah brought in Tanuj Siripurapu, a programmer since age nine, five-time hackathon winner, and former startup CTO who had done contract work for RTX and a major Bay Area networking company. Siripurapu, 19 and a Y Combinator alum, joined to build the platform. Together they founded Edviro to solve a problem that scales far beyond one district: schools across the country spend about $8 billion a year on utilities, and over 30 percent of that is wasted on poor scheduling, equipment failures, and basics like HVAC running in empty buildings.
The structural constraints are familiar to anyone who has walked a K-12 campus. Facilities teams are thin, often one or two people covering dozens of sites. Their software stack is fragmented: building management systems, utility portals, CMMS platforms, and paper work orders that don't talk to each other. Bills are opaque, meters are read monthly if at all, and verifying a savings claim means manual spreadsheet work that never gets done. "Facilities managers do not need more dashboards or alarms," the company says. "They need systems that continuously improve and take action."
Edviro's answer was an agentic platform that connects to existing meters, BMS, and utility data (no new hardware required) and runs a continuous loop: detect waste, simulate a fix, propose a work order for human approval, then verify the result against a learned baseline using IPMVP-standard measurement and verification. The first version caught boiler short-cycling at 14 starts an hour, after-hours runtime in empty gyms, ventilation drift in classrooms, and demand spikes before they hit the bill.
The team launched with free energy audits for early district partners. Palo Alto Unified and Los Gatos-Saratoga Union High School District signed on. Within the first year, 34 school sites were live and the platform had verified over $400,000 in savings for education customers.
That traction got them into Y Combinator's Summer 2026 batch, backed by YC and Reach Capital. The YC stamp matters less for the capital than for the network, access to partners who have scaled infrastructure software into regulated, risk-averse markets. It also signaled that the problem Edviro picked, operational energy waste in buildings with thin staff and messy data, was generalizable well beyond schools.
Shah's background in applied AI across climate, healthcare, and aviation (including a NeurIPS 2024 workshop paper and a patent-pending screening system) and Siripurapu's focus on scalable agentic systems gave them a technical vocabulary most cleantech founders lack. They weren't building a better dashboard; they were building an operating layer that could take a signal (a meter anomaly, a staff text, a sensor reading) and turn it into a reviewed, dispatched, verified fix.
The K-12 beachhead was deliberate. Schools have the pain, the data, and the public accountability to demand verification. But the architecture, meter-agnostic, BMS-agnostic, CMMS-replaceable or integrable, was built from day one to move into any facility class where energy waste is expensive and staff are scarce. Data centers were the obvious next vertical. The grid constraints, the cooling complexity, and the hyperscaler urgency to squeeze every megawatt made the fit unavoidable.
Edviro's founding insight — that the bottleneck isn't detection but the loop from detection to verified action — remains the core product thesis. The school districts proved the loop works. The YC batch proved the market believes it scales. The data center pivot is the test of whether it actually does.
Under the Hood: Three Technical Pillars
Edviro's platform sits on three technical pillars: a physics-informed world model that learns each facility's behavior, an agent layer that turns predictions into coordinated actions, and a measurement-and-verification engine that closes the loop with audit-grade evidence. The architecture was forged in K-12 districts — aging HVAC, mixed-vendor controls, lean staff, and utility bills no one had time to reconcile — and is now being stress-tested against the tighter coupling of power, cooling, and workload inside data centers.
World Model: More Than a Digital Twin
Most building analytics stop at visualization. Edviro's "energy world model" continuously estimates the facility's state, learns component interactions, and predicts responses to proposed actions. The model ingests interval meter data, BMS points, utility bills, sensor streams, equipment nameplates, maintenance logs, and occupancy schedules into a unified time-series context. That unification matters: an HVAC runtime anomaly that looks like a scheduling error on the BMS side may reveal itself as a demand-charge spike when correlated with the utility feed, or as a degraded coil when cross-referenced against a work-order history.
The company describes the model as continuously updated, not a static baseline but a living representation that adapts as equipment ages, seasons shift, or loads change. In a data center, where a workload migration can reshape the thermal profile in minutes, that adaptability is the difference between a useful prediction and a stale dashboard. Edviro's YC page frames it explicitly: the model "chooses the highest-value actions, and verifies the results against real operational, meter, and billing data."
Agent Layer: From Simulation to Work Order
Detection and simulation are table stakes. The differentiator is the agent orchestration that follows. Once the world model flags an anomaly — say, an air-handler running significantly above its learned baseline during unoccupied hours — the platform simulates candidate interventions: adjust the schedule, reset the supply-air temperature, stage a coil cleaning, or replace a faulty actuator. Each simulation is scored against cost, reliability impact, and physical constraints (redundancy requirements, ASHRAE guidelines, lease obligations).
The recommended action surfaces in the operator's queue as a reviewed work order with expected savings, payback, and CO₂ reduction pre-calculated. One-click approval dispatches the task to the facilities team or, where authorized, writes the setpoint change back to the BMS. Edviro's founders emphasize a graduated autonomy model: they initially retain operator authority over all such actions, then automate more of the operating loop once the models gain trust. That phrasing matters; it acknowledges that data center operators will not hand over chiller-plant control to a black box on day one.
The agent layer also coordinates multi-step workflows: assigning investigators, tracking parts, scheduling downtime, and closing the loop with post-action verification. The Projects tab manages the full lifecycle, baselining, assignees, work orders, asset inspections, while the Anomalies tab surfaces live issues: HVAC misconfigurations, scheduling drift, billing discrepancies.
Measurement & Verification: IPMVP-Grade, Automated
Energy savings claims have historically lived in spreadsheets that don't survive a board audit. Edviro bakes IPMVP (International Performance Measurement and Verification Protocol) compliance into the product. During onboarding, the platform establishes an independent energy baseline; for new data center builds, that baseline is set during fit-out, so variance appears while the contractor is still on site. Every subsequent intervention, whether a setpoint tweak or a capital retrofit, is measured against that baseline using the same meter and telemetry streams that feed the world model.
The Savings tab auto-generates reports: energy saved, cost avoided, carbon reduced, all exportable as PDF for finance or sustainability teams. No manual regression, no "trust me" emails. The company says it has already identified such savings across 34 live facilities, a figure that includes both operational waste and billing overcharges caught by cross-referencing utility tariffs against actual consumption.
Edge Integration: No Rip-and-Replace
Edviro does not require a controls overhaul. It integrates with existing BMS platforms, DCIM, EPMS, utility APIs, Modbus/BACnet sensors, and solar inverters, reading and, where permitted, writing back. The LinkedIn demo notes direct integration with BMS and sensor providers for "the most granular information possible," plus solar export visibility from the inverter vendor. That read-first, write-when-authorized posture is deliberate: it lets a data center operator start in read-only mode, validate the model against their telemetry, and expand control scope incrementally.
The technical hire roadmap signals where the architecture is heading next: Edviro is recruiting for similar roles in those same disciplines, a mix that suggests deeper first-principles modeling (CFD-informed thermal zones, electrochemical battery degradation, transformer thermal limits) rather than pure data-driven curve fitting.
Why the School District Crucible Mattered
The K-12 origin forced the team to handle fragmented, noisy, incomplete data, that messy reality, without the clean lab datasets that make for impressive demos but brittle deployments. Schools operate large portfolios of mixed-vendor equipment with lean teams and almost no budget for wasted energy. That constraint shaped a platform that works with what's already in the mechanical room, not what a sales deck wishes were there.
Data centers are the inverse on paper: modern, instrumented, capital-rich. But the coupling of power, cooling, and compute creates a new class of fragmentation: BMS, DCIM, EPMS, vendor portals, and spreadsheets that don't speak to each other. Edviro's bet is that the same architecture that stitched together a district's 34 facilities can stitch together a hyperscaler's availability zones, and that the verification discipline honed on school-board audits will survive a data center operator's uptime review.
On the Ground: What Engineers Gain
The shift from reactive to predictive management is not a slide-deck concept for data center teams; it is a daily operational rewrite. Edviro's platform, built originally for school districts with thin facilities staff and aging equipment, brings a unified layer that connects meters, HVAC systems, batteries, solar, and utility bills into a single evidence chain. For a data center operator, that means an anomaly in a cooling loop shows up alongside the demand charge it triggers and the work order it spawns, all before a technician walks the row.
When readings veer off baseline, the system alerts operators, giving them a chance to intervene before equipment fails, reducing routine inspections, increasing worker productivity, and extending equipment lifetime. In practice, this translates to predictive energy models that simulate load behavior and detect anomalies before they escalate, allowing operators to act early and improve uptime while cutting both operational and carbon costs.
The work-order loop closes automatically. Edviro turns building signals and staff requests into reviewed work orders, verified fixes, and capital plans, usable as a native CMMS or integrated with existing systems. A facilities engineer sees a flagged deviation, the system drafts a work order with the relevant telemetry attached, the fix is executed, and the platform measures the result against the learned baseline. Measurement and verification then generates audit-grade, board-ready proof of savings automatically. That unified context matters: when an HVAC runtime pattern affects both consumption and demand costs, or a billing anomaly needs checking against operations before a capital decision, the data sits in one place.
Remote diagnostics change the economics of expertise. Connected supervision platforms let experts resolve many issues without physical intervention, reducing downtime, limiting travel-related emissions, and accelerating recovery. These predictive and hybrid maintenance approaches combine human expertise with automation, ensuring high performance even under unpredictable workloads. For hyperscalers running GPU clusters at 40–100 kW per rack — where load variations can reach several hundred percent within milliseconds — that speed is not optional.
The role itself is shifting. Facility managers are evolving into strategic orchestrators, balancing resilience, energy efficiency, and sustainability while managing risk in real time. As monitoring becomes more integrated, teams transition from local, reactive management to fully connected, predictive environments; supervision, analytics, and automation become central to daily operations. Deloitte notes that agentic AI systems are already being piloted to autonomously manage complex scheduling, coordinate workflows, and mitigate risk, while IoT devices and 5G connectivity transform asset tracking and predictive maintenance.
Edviro's earliest data center deployments focus on a high-leverage niche: verifying cooling changes against telemetry and providing independent baselining for new builds. In a facility where a single rack can exceed 30–40 kW and liquid cooling is indispensable, the ability to simulate a setpoint change, dispatch the adjustment, and verify the PUE impact without a manual test cycle is the difference between a controlled optimization and a thermal event. The platform does not replace the engineer; it removes the forensic scramble that used to follow every alarm.
Talent War: Hyperscalers Buy the People Who Know Power
The talent war that once centered on AI researchers and machine learning engineers has pivoted. As of January 2026, energy-related hiring at Big Tech ran 34 percent above the prior year and remained 30 percent above pre-ChatGPT 2022 levels, per Workforce.ai data compiled for CNBC. The shift marks a step-change from traditional sustainability roles, which boomed during the Inflation Reduction Act era but lost steam amid ESG backlash that intensified after the 2024 election. Instead, recruiters report demand for operational veterans: energy procurement, markets, grid interface, and long-term strategy.
Microsoft has added more than 570 energy-focused roles since 2022, including Betsy Beck as director of energy markets in early 2025. Amazon leads with 605 hires, a figure that includes AWS. Google parent Alphabet has added 340 and moved to acquire data center operator Intersect in a $4.75 billion cash-and-debt deal. The four hyperscalers — Alphabet, Microsoft, Meta, and Amazon — are deploying nearly $700 billion in combined capex this year alone. "There are tech companies that are turning into energy companies," said Daniel Smart, group CEO of The Green Recruitment Company. "But they've never built one before. So they'll outsource the construction, and possibly even outsource the running of it, and just buy the energy."
That scramble for supply has pushed efficiency to "phase two," Smart said. For now, the priority is securing megawatts. The talent pool for that work is finite. Jeff Anderson, business development director at renewable energy recruiter Taylor Hopkinsons, said senior candidates from energy infrastructure are "becoming aware of the opportunities in data centres and the higher salaries on offer in tech, and are making enquiries and plans to move across long-term." Morningstar analyst Travis Miller argues utilities should see this as opportunity: tech firms need utility partnership more than acquisition. "That's the most efficient way to do it from a workforce perspective and from an infrastructure perspective," he told CNBC. "It's such a large amount of energy that they can't do it themselves."
The bottleneck extends well beyond white-collar strategists. Randstad's analysis of 50 million job postings between 2022 and 2026 found demand for robotic technicians up 107 percent, HVAC and cooling engineers up 67 percent, and industrial automation technicians up 51 percent. Traditional skilled trades — construction workers, electricians — rose 27 percent. With roughly 12,000 data centers operating globally and thousands more planned, mechanical, electrical, and plumbing systems need refreshing every four to six years, said Mike Mathews of Marsh. The U.S. faces a potential shortfall of 1.9 million manufacturing workers by 2033, per National Association of Manufacturers data. One in four workers globally is nearing retirement. "Unlike software developers who can often work remotely, skilled trades possess very low geographic mobility," said Randstad CEO Sander van't Noordende. "An equipment technician must be physically on-site. When a company builds a new AI data center, it is a region-changing event that can instantly exhaust local talent pools."
Startups are rushing into the gap. Google for Startups selected 28 companies across Europe and North America for its AI for Energy Class of 2026 in August. Elemental Impact launched a Data Center Innovation Initiative in May backed by Amazon, Google, Meta, and Microsoft to accelerate commercialization of energy storage, advanced electrical systems, and cooling tech. ABB's Startup Challenge in April named winners developing AI for grid resilience. The 2026 Industry Growth Forum drew more than 400 startups and nearly 200 investors in April around the theme "All Systems Go." DayOne committed RM28 billion-plus to Malaysia, targeting 5,000 local jobs and a training hub in Johor to produce 1,000 data center engineers.
The response from incumbents is mixed. BlackRock launched a $100 million trades-workforce initiative in March, with CEO Larry Fink stressing that capital alone cannot unlock the $10 trillion infrastructure need. Noordende said winners will combine traditional recruiting with apprenticeships, community-college partnerships, military-veteran pipelines, and internal academies. Mercer's William Self noted cross-industry poaching is accelerating because operational skills overlap across energy, defense, and tech. "Companies that are not prepared to move quickly with a competitive package are losing out," said Sea Change Talent. Nvidia CEO Jensen Huang has predicted six-figure salaries for the workers building AI factories; hazard pay may follow as geopolitical risk — Iranian drone strikes hit AWS facilities in the UAE this March — enters the calculus.
The hyperscalers are no longer just buying power. They are buying the people who know how to generate, transmit, and manage it.
Kicker
That loop works, as those districts showed. The data center operators will decide whether it scales. The meter doesn't care where it's mounted — only that someone is watching it, modeling it, and closing the loop before the bill comes due.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.