A Playbook for Evergy: Curing AI’s Power Bottleneck With Home Compute and Flexible Load
Evergy already pays 75 MW data centers to flex. A homeowner on its grid who wants to host a 12.5 kW compute node has no path at all. Here is the math for closing that gap, and why it beats another peaker plant.
I’m an Evergy customer, and I’m writing this for the people who run Evergy and the commissions that regulate it in Kansas and Missouri. Evergy is betting its growth on data centers. It has signed about 2.5 GW of them, carries a $21.6 billion five-year capital plan, and has more than 5 GW of new generation in its 2026 resource plan, most of it natural gas. That is a reasonable plan for firm load. It is an incomplete plan for the load that is coming.
AI compute is stuck between two bottlenecks. The power bottleneck is years-long grid connections and transformer shortages. The capital bottleneck is the build-big, borrow-big model, now showing strain at CoreWeave. The cure is a portfolio rather than one more giant campus: compute that backs off for a few hours a year at the grid’s peak, spread across homes, small businesses and new subdivisions, and grouped into a single flexible load the utility can call on.
Evergy has already written half of that policy. Its Missouri large-load tariff pays customers above 75 MW to cut load at peak, and its demand response programs in both states earn its shareholders money for reducing peak. Everything below 75 MW that happens to be compute, including every home node on its system, is left out. The reason is structural and fixable: Evergy earns on what it builds, and the rules pay far less for what it avoids building. Below I lay out the numbers, a proposed rider that would close the gap, a model you can run, why Evergy hasn’t moved, the objections I would raise if I worked there, and a 12-month plan.
The bottleneck has two halves
The power half is well documented. Lawrence Berkeley National Laboratory’s Queued Up report found a median of more than five years from interconnection request to commercial operation, with 2,061 GW still waiting at the end of 2025. Large substation transformers now take more than 160 weeks to arrive, per Wood Mackenzie. The smaller equipment is short too: NEMA says a new distribution transformer can take up to three years to arrive, against four to six weeks five years ago, and NREL estimates that about 55% of residential transformers are near the end of their lives.
The capital half arrived this month. CoreWeave, the largest independent AI cloud, carries about $35.6 billion of debt and $16.3 billion of lease obligations and lost $1.4 billion in the first half of 2026, according to Fortune. Its unsecured bonds now trade at yields around 15% to 16%, and its stock fell about 30% in October. CoreWeave says its roughly $104 billion contracted backlog more than covers its obligations, and nobody has filed for anything. But the market is telling builders that the model of borrowing billions against GPUs to fill giant campuses that wait years for power is fragile.
Evergy sits on both halves. On the power side, it has signed energy service agreements for about 2.5 GW of data center load, which it expects to drive 7% to 8% annual retail load growth through 2030, and it reports a pipeline of more than 10 GW. To serve it, the 2026 resource plan adds more than 5 GW through 2032: about 3.9 GW of natural gas, nearly 800 MW of solar and 450 MW of storage, including a 440 MW gas peaker at Mullin Creek in Nodaway County, Missouri. On the capital side, Evergy’s tariffs protect ratepayers with minimum bills of 80% of contracted demand, 12-plus-year terms and exit fees. Those protections are good. They are also a concentrated bet on a handful of very large customers whose financing, as CoreWeave shows, can turn fast.
The pushback is local, too. In Platte County, the presiding commissioner said publicly in October 2025 that he could not support a proposed $100 billion data center campus in Kansas City’s Northland, arguing that households would end up paying for the power plants and grid upgrades. Evergy disputed his link between its rates and data centers, and it reports that data center revenue reduced its pending Missouri Metro rate request by $25 million. Both things can be true. Customers are skeptical, and the utility needs an answer that does not ask them to underwrite more peak capacity.
The capacity Evergy already owns and rarely uses
Utilities build wires and power plants for the peak, the handful of summer evenings when air conditioners, ovens and EV chargers all run at once. The rest of the year, much of that capacity sits partly idle. Ratepayers pay for it either way.
Duke University’s Nicholas Institute put a number on that headroom in February 2025. Across the 22 largest US balancing areas, which cover 95% of US load, the grid could absorb about 76 GW of new load if that load agreed to curtail 0.25% of the time, and about 126 GW at 1%. In hours, 0.25% is about 85 hours a year, 0.5% is about 177 and 1% is about 366. And curtailment rarely means a full shutdown: during 88% of curtailment hours, half of the new load could keep running.
Evergy’s region gets its own line in that study. The Southwest Power Pool, which Evergy belongs to, could integrate about 10 GW of new load at 0.5% curtailment, tied with ERCOT for the third-largest figure in the country. That is four times Evergy’s entire signed data center load, available to loads willing to step back for about a week’s worth of hours spread across a year.
The point for Evergy is simple. Every megawatt of compute that agrees to step back at peak is a megawatt Evergy does not have to build peak capacity for. AI inference is unusually good at stepping back: requests can be routed to another region, batch jobs can wait an hour, and a home node can drop to idle in seconds. Google has already signed demand response agreements with Indiana Michigan Power and the Tennessee Valley Authority to cut machine learning workloads during grid stress. Compute can flex. What remains open is whether Evergy will let the small, distributed kind do it.
The toolkit: five ways to add compute without adding peak
Home and small-business compute nodes. SPAN, NVIDIA and PulteGroup are putting liquid-cooled 16-GPU inference nodes on the outside walls of new homes, with a 100-home pilot in 2026 and a target of more than 1 GW in 2027. Sunrun added an installation partnership for new Texas subdivisions in September and is piloting nodes in existing homes that already have its solar and batteries. SPAN plans a commercial node for offices and small businesses in early 2027. Each home node draws about 12.5 kW at full load and can drop to about 1 kW in seconds. No transmission study, no substation, no 160-week transformer order for the campus.
Flexible large loads. For sites big enough to sign a large-load contract, the tool already exists. Evergy’s Missouri Large Load Power Service tariff includes a Demand Response and Local Generation Rider that pays verified load reductions and lets Evergy curtail at peak, with bigger bill credits for customers who accept fewer constraints. Other utilities go further: Colorado Springs Utilities offers interruption of up to 100 hours a year with energy and demand credits, and East Kentucky Power Cooperative pays credits that rise with interruption hours and penalizes non-performance at five times the firm demand charge.
Batteries at the node. Every SPAN install comes with a home battery, and Sunrun’s pilot runs on homes with its storage already in place. A 13.5 kWh battery can carry a 12.5 kW node at full power for about an hour, which smooths short appliance spikes and the edge of an evening peak. It will not carry a three-hour peak alone, which is why the node itself must be able to throttle.
Headroom-first siting. Not every street can take a node. In the simulation from my previous piece, a 50 kVA transformer serving eight homes delivered 87% of full-power hours with one node, 72% with two and 38% with four. The cheapest capacity is the transformer that already has room. New subdivisions, where the builder sizes transformers for the compute load from day one, come next. Retrofitting crowded streets comes last.
Routing inference to where power is. Most AI inference does not care whether it runs in Overland Park or Phoenix. When SPP is tight, a fleet scheduler can move requests to nodes on other grids, and hold batch work such as embeddings, evaluations and synthetic data until the peak passes. That is what makes a curtailment promise credible: the operator keeps serving customers while Evergy’s peak drops.
The proposal: an Aggregated Flexible Compute Rider
Evergy can build this on a policy it already has: extend the flexibility it already pays 75 MW customers for to an aggregator that controls thousands of small nodes and behaves, from the grid’s point of view, like one large curtailable load. Here is a draft term sheet. The numbers are my proposals, meant to start the conversation, not settle it.
- Eligibility. An aggregator qualifies once it controls at least 10 MW of nodes across Evergy territory, with no single node above 80 amps (19.2 kW at 240 volts).
- Curtailment. Evergy may call up to 200 hours a year, enough to cover the Duke study’s 0.5% case, with 15 minutes’ notice. Each node must drop to 1.5 kW or less, verified by panel-level telemetry. Non-performance is penalized at a multiple of avoided capacity cost, as East Kentucky Power Cooperative does with its five-times charge.
- Interconnection. Headroom first. The aggregator may site nodes only where a distribution transformer has room, using Evergy’s own loading data. Any upgrade that serves compute is paid by the aggregator as a contribution in aid of construction, never added to everyone’s rates.
- Pricing. Energy at a rate that recovers embedded distribution and energy costs, and no peak capacity charge as long as the node meets its curtailment obligation. The bill reflects what the load actually costs the system.
- Hosts. Homeowners and small businesses receive bill credits through their Evergy bill, with plain-language disclosure of noise, heat, equipment ownership and removal terms.
- Where to start. New subdivisions with participating builders, where transformers are sized for compute from the start. A 10 MW pilot in each state, about 800 homes apiece.
- Data. Evergy receives real-time node telemetry, which doubles as the best distribution-planning data it has ever had at the street level.
One more term belongs to the regulators rather than the rider: count the rider’s curtailable megawatts toward Evergy’s existing demand response goals in Missouri and Kansas, so Evergy earns on the peak it avoids. Without that, the rider asks Evergy to give up rate base for nothing.
The rider turns home compute from something that happens to Evergy into something Evergy runs. It sets the terms, it gets the data, it keeps the peak flat, and it sells energy in the hours its system is underused.
The economics: give up 2% of hours, avoid 92% of the peak
Take a 100 MW program: 8,000 home nodes at 12.5 kW each, about 128,000 GPUs. Served as firm load, it adds 100 MW to Evergy’s peak. Served under the rider, every node drops to 1 kW when called, so it adds 8 MW. The program gives up 177 hours a year, about 2.0% of the year, and removes 92 MW of peak capacity Evergy would otherwise have to build or buy.
peak_firm = nodes × node_kW peak_flexible = nodes × idle_kW value_to_grid = (peak_firm − peak_flexible) × capacity_cost_per_kW_year cost_to_fleet = GPUs × curtailed_hours × utilization × price_per_GPU_hour
SPP has no capacity market, so I use PJM’s 2026 estimate of the gross cost of new entry for a combustion turbine in the 2030/31 delivery year, $642 per MW-day or about $234 per kW-year, as a proxy for what a new peaker costs. At that price the 92 MW of avoided peak is worth about $21.5 million a year. On the other side, at $1.80 per GPU-hour and 60% utilization, the fleet gives up about $24.5 million of compute revenue in those 177 hours. Per node, that is about $2,690 of avoided capacity against about $3,060 of forgone revenue.
That near tie is the honest result, and it matters. Curtailment is not free money for the operator, and Evergy should not expect to pay for it twice. What makes the trade work is speed. In my previous piece I estimated that getting capacity energized in six months instead of three years is worth about $8.9 million per MW in present value at today’s prices. For a 100 MW program that is roughly $890 million, which dwarfs $24.5 million a year of curtailed hours. The operator accepts curtailment as the price of connecting now. Evergy avoids building for a peak the load will not add. And the program still delivers about 542 GWh a year of energy sales, almost all of it in hours when Evergy’s system has room.
The curtailment budget is the lever to negotiate. At 85 hours a year, the Duke 0.25% case, forgone revenue falls to about $11.8 million. At 366 hours, the 1% case, it rises to about $50.6 million and the operator would want real compensation. The rider’s 200-hour cap sits near the middle.
"""Firm versus flexible: a proposed model for an aggregated home compute program."""
from dataclasses import dataclass
@dataclass(frozen=True)
class Program:
nameplate_mw: float = 100.0 # aggregated node nameplate
node_kw: float = 12.5 # node at full load
idle_kw: float = 1.0 # node when curtailed
gpus_per_node: int = 16
util: float = 0.60 # share of available hours sold
curtail_hours: float = 177.0 # Duke study, 0.5% curtailment case
gpu_price: float = 1.80 # $ per GPU-hour
capacity_cost_kw_yr: float = 234.0 # proxy cost of new peaking capacity
@property
def nodes(self):
return round(self.nameplate_mw * 1000 / self.node_kw)
def peak_mw(self, flexible: bool):
per_node = self.idle_kw if flexible else self.node_kw
return self.nodes * per_node / 1000
def capacity_avoided_musd(self):
avoided_kw = (self.peak_mw(False) - self.peak_mw(True)) * 1000
return avoided_kw * self.capacity_cost_kw_yr / 1e6
def hours_lost_share(self):
return self.curtail_hours / 8760
def revenue_lost_musd(self):
gpus = self.nodes * self.gpus_per_node
return gpus * self.curtail_hours * self.util * self.gpu_price / 1e6
def energy_gwh(self):
avg_kw = self.idle_kw + self.util * (self.node_kw - self.idle_kw)
return self.nodes * avg_kw * (8760 - self.curtail_hours) / 1e6
if __name__ == "__main__":
p = Program()
print(f"nodes: {p.nodes:,} GPUs: {p.nodes * p.gpus_per_node:,}")
print(f"peak demand, firm: {p.peak_mw(False):.0f} MW flexible: {p.peak_mw(True):.0f} MW")
print(f"hours given up: {p.hours_lost_share():.1%}")
print(f"peak capacity avoided: ${p.capacity_avoided_musd():.1f}M per year")
print(f"compute revenue forgone: ${p.revenue_lost_musd():.1f}M per year")
print(f"energy delivered: {p.energy_gwh():.0f} GWh per year")The script prints 8,000 nodes, 100 MW firm against 8 MW flexible, 2.0% of hours, $21.5 million of avoided capacity, $24.5 million of forgone revenue and 542 GWh of energy. Change capacity_cost_kw_yr to Evergy’s own avoided-cost figure and the conclusion is easy to check against the numbers Evergy files with its regulators.
Why Evergy hasn’t done this yet
Start with what Evergy is. It is a regulated monopoly: the only electric utility serving its territory, with the Kansas Corporation Commission and the Missouri Public Service Commission setting what it may charge and earn. That is a description, not an insult. But it explains the gap.
A regulated utility earns a return on the assets it builds. Evergy told investors in August that it expects its rate base, the asset base it earns on, to grow about 12% a year through 2030, on a $21.6 billion five-year capital plan plus about $1 billion of new generation tied to signed customers. A 440 MW gas peaker goes into that rate base. A rider that makes part of the peaker unnecessary does not. Economists have a name for this tilt toward capital-heavy answers, the Averch-Johnson effect, first described in 1962. Nobody at Evergy has to act in bad faith for it to work. The rules pay for building and pay much less for avoiding.
What makes this hard to excuse is that Evergy already knows how to earn on avoided peak. Under Missouri’s Energy Efficiency Investment Act, Evergy’s current business demand response program pays participants between $2 and $75 per kW-year, and Evergy’s shareholders collect an earnings opportunity once the program reaches 65% of its goal, according to Missouri PSC staff testimony from March 2026. In Kansas, the Corporation Commission approved Evergy’s efficiency and peak-reduction programs in September 2023 with an earnings opportunity that one dissenting commissioner called “unnecessary, excessive and flawed.” Evergy earns on peak reduction today. It simply does not count compute.
The same staff testimony shows how new load gets walled off. Nucor’s steel mill in Sedalia could not join Evergy’s Missouri demand response program because its special incremental-load tariff barred it from those programs. That barrier is a rule, not physics, and rules can be rewritten.
The fix is to make the flexible option good business for Evergy. Regulators in both states could count an aggregator’s curtailable megawatts toward Evergy’s demand response goals, so its shareholders earn on the peak it avoids. Or they could follow Hawaii, whose commission adopted performance-based regulation in December 2020 with a grid services incentive and shared-savings mechanisms that reward the utility for buying flexibility from customer-sited resources. Either way, the choice in front of Evergy’s leadership becomes whether to earn on the peak it builds for or on the peak it avoids.
The objections I would raise if I worked at Evergy
Neighborhood transformers are short too. True, and it cuts against any plan that starts by replacing equipment. NEMA’s three-year lead time and NREL’s estimate that 55% of residential transformers are near the end of their lives are exactly why the rider says headroom first, new subdivisions second and retrofits last, with the aggregator paying for every upgrade that serves compute.
Ratepayers will end up paying. Only if the rider is written badly. Upgrades are paid up front by the aggregator, the capacity charge disappears only for nodes that actually curtail, and the energy rate covers embedded costs. Written that way, a program sells hundreds of gigawatt-hours in off-peak hours over wires and plants customers already pay for, which helps spread fixed costs rather than adding them.
The nodes will not really shut off. Then penalize them. Panel-level telemetry makes every curtailment verifiable within seconds, and the East Kentucky Power Cooperative model shows a utility can price non-performance at five times the firm demand charge. Monthly drills during the pilot will show quickly whether an aggregator’s promise holds.
Neighbors will object. Some will. Coverage of Sunrun’s pilot already describes pushback, and SPAN has not published a noise figure for its cabinet. That is a reason to start in new subdivisions where buyers opt in with full disclosure, not a reason to leave the whole category to third parties operating without Evergy’s rules.
The aggregator could fail. That is the CoreWeave lesson, and the rider’s structure is the answer. Capacity arrives in 12.5 kW pieces rather than 500 MW contracts, aggregators post deposits as Evergy’s large-load customers already do, and a removal bond covers stranded equipment. A failed aggregator leaves a few hundred cabinets to collect, not a stranded campus.
It is more energy use, not less. Correct. The program above adds about 542 GWh a year. The rider does not make that energy clean. It makes it cheap to serve by keeping it off the peak, and pairing nodes with rooftop solar and storage, as Sunrun’s pilot does, is the obvious next step. Utah State University’s Robert Davies warns that efficiency gains tend to raise total energy use, and he is right to.
It is too small to matter. Evergy serves about 1.7 million customers. One percent of them hosting a node is 17,000 homes and about 212 MW of compute, connected in months with almost nothing added at peak. And SPP’s flexible headroom in the Duke study is about 10 GW.
A 12-month plan for Evergy
Months 1 to 3: open the data
- Publish street-level transformer loading data for compute hosting.
- Issue a request for information to aggregators, including SPAN and Sunrun.
Months 4 to 6: file the pilot
- File the Aggregated Flexible Compute Rider as a pilot with the Kansas Corporation Commission and the Missouri Public Service Commission.
- Ask both commissions to count the pilot’s curtailable megawatts toward Evergy’s demand response earnings opportunity.
- Sign two new-subdivision partners.
Months 7 to 12: prove it
- Energize 10 MW in each state.
- Run monthly curtailment drills.
- Publish the results: actual peak contribution, curtailment performance and energy sold.
None of this requires Evergy to abandon its large-load strategy. It requires Evergy to stop treating flexibility as a privilege of 75 MW customers, and its regulators to let Evergy earn on avoided peak from compute the way it already earns on avoided peak from thermostats and factories. The utility that writes the first good rule for small, flexible compute will set the template other utilities copy. Evergy has the tariff language, the demand response machinery, the headroom in SPP and the customers. It should move first.
If Evergy’s team wants to pressure-test any of these numbers, I would be glad to walk through the model with them.
Sources
- MarketBeat, “Evergy Q2 earnings call highlights,” August 2026. Link
- Evergy, Company overview. Link
- Lawrence Berkeley National Laboratory, “Electricity Rate Designs for Large Loads,” technical brief, August 2026. Link
- Utility Dive, “Kansas, Michigan regulators approve large load rules for Evergy, Consumers Energy,” November 2025. Link
- Missouri Public Service Commission, Staff rebuttal testimony of Mark Kiesling, Case No. EO-2026–0129, March 2026. Link
- Kansas Reflector, “Long-awaited Evergy energy efficiency programs approved by Kansas regulators,” September 2023. Link
- Hawaii Public Utilities Commission, Performance-Based Regulation Phase 2 decision summary, December 2020. Link
- Averch, H. and Johnson, L., “Behavior of the Firm Under Regulatory Constraint,” American Economic Review, 1962, summarized here. Link
- Utility Dive, “US grid headroom for flexible load,” on Duke University Nicholas Institute, Rethinking Load Growth, February 2025. Link
- Duke University Nicholas Institute, Rethinking Load Growth (report). Link
- Latitude Media, “Google expands demand response to target machine learning workloads,” August 2025. Link
- Fortune, “CoreWeave: A look inside the $52 billion debt and lease hole,” October 2026. Link
- Disruption Banking, “Who holds CoreWeave’s secured debt, and who holds the notes,” October 2026. Link
- Crypto Briefing, “CoreWeave shares drop 30% in October,” October 2026. Link
- LBNL, Queued Up, as summarized by the American Public Power Association, 2026. Link
- Data Center Knowledge, “Why AI data center projects face years of delays after approval,” 2026. Link
- Utility Dive, “Electric transformer shortage,” citing NREL and NEMA. Link
- Maryville Forum, on Evergy’s Mullin Creek peaking plant. Link
- FOX4 Kansas City, “Platte County commissioner can’t support $100 billion Northland data center,” October 2025. Link
- PJM, “Periodic review of default CONE and ACR values,” August 2026. Link
- SPAN, “What is XFRA?” support article. Link
- Latitude Media, “Span to launch distributed AI data centers for edge compute,” 2026. Link
- Sunrun, “Sunrun Launches Distributed AI Data Center Pilot,” GlobeNewswire, July 2026. Link
- IndexBox, “Span and Sunrun partner to install XFRA in Texas homes,” September 2026. Link
- Forkast, “Sunrun wants to turn your home into a mini data center,” 2026. Link
- Fortune, “Startups are installing tiny data centers in people’s homes,” 2026. Link
- Brian James Curry, “The Data Center in the Basement,” the previous piece in this series.
