What If the Data Center Brought the Power With It?

At some point in nearly every conversation about a new data center, someone puts a very large number on a slide.

Usually it has MW after it.

That is when utility planners start thinking about substations, electrical engineers begin doing arithmetic in their heads, and everyone else starts wondering whether the neighborhood air conditioners are about to enter into a bidding war with artificial intelligence.

The concern is understandable.

Lawrence Berkeley National Laboratory's 2026 update estimates that U.S. data centers could consume roughly 649 TWh annually by 2030 in its reference case, with an uncertainty range of 521 to 843 TWh. That works out to roughly 9.5% to 15.3% of total U.S. electricity consumption, depending on how the industry develops.^[1]

That is real load.

It is also why I think the next generation of data centers needs to be designed around a different premise.

A sufficiently large compute campus should not simply arrive asking the local power system for several hundred megawatts.

It should arrive with a credible answer to a much better question:

What energy infrastructure are we bringing with us?

Done properly, the result begins to look less like an unusually large electrical customer and more like a privately financed energy campus that happens to contain a great deal of computing equipment.

And at the largest scales, that distinction may determine which projects get built, which get interconnected, and which communities actually want them.

Water deserves context - and local engineering

Water is one of the first objections raised around large data-center development, so the scale deserves some context.

Berkeley Lab estimates that U.S. data centers directly consumed about 66 billion liters of water in 2023, or roughly 17.4 billion gallons. Hyperscale and colocation facilities represented 84% of that direct consumption.^[2]

EPA estimates that Americans use nearly 1.2 trillion gallons of water every year just showering.^[3]

Put those numbers beside one another and the direct onsite water consumption of the entire U.S. data-center fleet in that benchmark year comes to roughly 1.5% of America's annual shower-water consumption.

Apparently the American shower is a surprisingly useful unit of national infrastructure planning.

That does not make the water question disappear.

Water is intensely local.

Berkeley Lab makes that point itself: evaporative cooling can create legitimate local water-availability concerns, and there are real tradeoffs between reducing onsite water consumption and increasing cooling-energy demand. Air cooling, liquid cooling, dry coolers, economizers, and reclaimed-water strategies all behave differently.^[2]

So the useful questions are not simply, "Does a data center use water?"

Nearly everything does.

The useful questions are:

• How much?
• From where?
• Under what seasonal conditions?
• With what cooling architecture?
• What is actually consumed rather than withdrawn?
• How much is reclaimed or reused?
• And what does that number mean relative to the local water budget?

Texas' current data-center audit is asking developers essentially those same questions, including projected water consumption, water sources, reuse, and cooling technology.^[22]

That is the right way to handle the issue.

Water is a siting and engineering constraint - not a particularly useful national shorthand for whether data centers should exist.

"Data center" is doing a lot of work

There is another bit of context worth establishing before we start casually throwing around 500 MW numbers.

A data center is not a size.

We use the same term for facilities whose electrical characteristics differ by orders of magnitude.

That is a little like discussing marine engineering while using "boat" interchangeably for a fishing skiff and an aircraft carrier.

They both float. After that, the terminology becomes less helpful.

There is no universally accepted megawatt boundary separating every category, because the terms describe ownership, workload, architecture, latency, and operating model as much as electrical capacity.

But three broad groups are useful from an energy-design perspective.

Edge data centers

Edge facilities put computing physically closer to the users, machines, or applications consuming it.

Latency is usually the reason.

A 2025 national-laboratory study uses less than 20 MW as a working definition for edge data centers and notes that many individual sites are below 5 MW. The same study points out an important grid issue: small edge facilities can accumulate on the same distribution feeder until a collection of individually modest loads becomes a very large local problem.^[5]

Electrically, however, this size class is also interesting.

A few megawatts of compute can plausibly be paired with enough BESS, local generation, and controls to create a genuinely resilient local microgrid without constructing something resembling an independent utility.

Enterprise data centers

Enterprise data centers are primarily defined by ownership rather than capacity.

DOE describes them as facilities owned and operated by a single company, critical to that organization's operations, often sizable but generally smaller than hyperscale facilities.^[6]

Banks, universities, manufacturers, healthcare systems, and large corporate campuses often fall into this world.

Their electrical architecture increasingly looks familiar to anyone who works around microgrids:

utility service, UPS, BESS, standby or prime generation, renewable generation where appropriate, prioritized loads, and an energy-management system sitting above the whole arrangement.

Hyperscale data centers

Hyperscale is where the power conversation begins changing character.

A recent Harvard-led study examined 403 operating U.S. hyperscale facilities, with an average electrical capacity of about 40 MW in its dataset. Meanwhile, national-laboratory planning work describes emerging hyperscale facilities in the 100 MW to multi-gigawatt range.^[16]

That apparent spread tells us something useful: the industry is moving quickly enough that the vocabulary is struggling to keep up.

The newest proposed campuses are dramatically larger than many operating hyperscale facilities.

And storage arithmetic becomes entertaining very quickly.

A 5 MW facility requiring four hours at full load implies about 20 MWh of usable stored energy.

At 50 MW, it is 200 MWh.

At 500 MW, it is 2 GWh.

Eight hours at 500 MW is 4 GWh.

A full day is 12 GWh.

Same multiplication.

Very different civil works package.

This is the same power-versus-energy distinction we discussed in the residential context in Whole-Home Backup Should Not Mean Every Load at Once: kW determines how much can run at once; kWh determines how long it can run. Hyperscale does not change the physics. It just adds rather more zeros.

This is why there cannot be one generic "data-center energy strategy."

For an edge facility, BESS may reasonably carry a substantial portion of outage endurance.

For an enterprise facility, storage may serve continuity, peak management, and generation transition.

At hyperscale, the BESS increasingly becomes part of the power system itself.

And that is where the opportunity gets interesting.

The old model has one very large arrow

The conventional electrical diagram for a data center is easy to draw.

Utility on the left.

Data center on the right.

Very large arrow between them.

That model becomes more difficult to defend as individual campuses move toward hundreds of megawatts and gigawatts.

The better diagram has arrows going both directions.

DOE's Office of Electricity is now explicitly examining microgrids as a way to integrate large loads such as data centers faster than traditional transmission and distribution expansion alone. DOE identifies onsite generation, storage, demand response, frequency regulation, and other ancillary services as ways these facilities can also support the surrounding electric system.^[4]

A future compute campus could combine:

• dedicated or electrically proximate firm generation;
• large-scale solar;
• modular BESS;
• controllable mechanical and computing loads;
• grid-aware microgrid controls;
• islanding capability;
• a defined resilience reserve;
• a flexible interconnection agreement; and
• the ability to reduce imports or export electricity when the surrounding grid needs help.

At that point we do not simply have a load.

We have an energy campus with a very large computer attached to it.

Bring the power with the compute

Early examples of this architecture are already appearing.

At DOE's Paducah site in Kentucky, a project announced in July 2026 proposes an approximately 1.8 GW AI and high-performance-computing campus, supported by 2 GW of new grid-connected natural-gas generation, transmission upgrades, and as much as 2.6 GW of battery storage.

The particularly interesting design choice is that DOE says the generation and storage are intended to exceed the campus's own requirements, allowing excess electricity to be delivered into the regional grid.^[18]

At the Savannah River Site, DOE/NNSA selected a proposal pairing a 1 GW data center with approximately 2 GW of onsite generation, initially using natural gas with an intended path toward nuclear generation. Again, the project is structured around dedicated generation rather than assuming the existing grid simply absorbs the new load.^[19]

There are obviously permits, emissions, transmission studies, fuel infrastructure, financing, and enough regulatory paperwork to keep several conference rooms occupied between an announcement and an operating plant.

Power projects have an impressive ability to turn a straight line into a flowchart.

But the architecture is important.

At gigawatt scale, bring-your-own-power is becoming a credible alternative to bring-us-a-substation.

Solar, batteries, and turbines are not competing for the same job

Energy conversations often improve considerably when we stop requiring every technology to prove that it can do everything.

It cannot.

Solar is exceptionally useful at producing low-marginal-cost energy while the resource is available.

Battery storage is exceptionally useful at moving that energy through time and reacting quickly to changes in grid and load conditions.

Firm generation is exceptionally useful when the outage, weather event, or grid constraint lasts longer than the stored-energy budget.

Those jobs overlap.

They are not the same.

Solar supplies cheap energy

A sufficiently large solar installation can offset daytime campus demand, charge storage, and reduce fuel consumption.

DOE notes that coupling solar with storage allows excess solar generation to be captured when production is high and dispatched later when load is high or sunlight is unavailable. Storage can also firm short-term solar variations and help provide resilient power.^[12]

There is an economic elegance to generating energy without ordering more fuel.

The sun's invoicing department remains remarkably understaffed.

BESS manages time

BESS handles the fast work.

It can:

• smooth ramps;
• clip peaks;
• absorb excess solar generation;
• bridge generator starts;
• provide spinning-equivalent reserve without necessarily keeping another combustion machine online;
• reduce campus import in seconds;
• support frequency and voltage;
• transition a microgrid into islanded operation; and
• hold a protected energy reserve for the data center itself.

A 2026 peer-reviewed national-laboratory demonstration modeled a 70 MW grid-interactive data center using BESS and microgrid controls. The system responded to utility demand-flexibility requests within about 10 seconds, supported voltage and frequency events, transitioned to islanded operation, and maintained compute service-level requirements.^[7]

That is a much more interesting job description than "big UPS."

Firm generation supplies endurance

Eventually, however, a multiday outage stops being a power problem and becomes an energy problem.

That is where dispatchable generation remains useful.

Natural-gas turbines are an obvious near-term option for some large campuses because they can provide substantial local firm capacity. DOE's Paducah and Savannah River projects demonstrate that this is already being considered at gigawatt scale.^[18]

That does not mean the turbines need to sit at full output twenty-four hours a day merely because they exist.

A well-controlled plant can use solar and grid energy when appropriate, charge batteries when energy is abundant, discharge storage through peaks, and maintain dispatchable generation as the endurance resource.

Longer term, the firm resource could be gas, nuclear, geothermal, or something else.

The function matters more than the logo on the machine.

Solar supplies energy.

BESS manages time.

Firm generation supplies endurance.

Give each technology the job it is actually good at.

The BESS should grow the way the data center grows

There is another reason storage fits data centers unusually well.

Data centers are already modular.

A 500 MW master plan does not generally mean somebody installs the final 500 MW of compute on Tuesday morning.

Buildings are commissioned.

Halls are populated.

Substations are added.

Cooling capacity grows.

Utility capacity may arrive in phases.

The BESS can follow the same development curve.

Oak Ridge National Laboratory has described modular systems and coordinated, autoconfiguring controls as an important architecture for utility-scale storage. Pacific Northwest National Laboratory has separately demonstrated modular, reconfigurable BESS concepts capable of independently controlling battery elements and isolating faults.^[8][9]

For a large data-center campus, that creates several advantages.

Storage can follow actual load

Build the electrical backbone and controls architecture for the ultimate campus.

Install the first BESS blocks with Phase One.

Add more as compute arrives.

There is little virtue in buying the year-ten battery on day one simply to demonstrate confidence in a forecasting model.

The AI hardware probably will not be the same in ten years.

The battery technology may not be either.

But conduit is still relatively predictable.

Power and energy can expand independently

A new compute phase may require additional MW of fast-response battery power without requiring another four hours of total-campus storage.

A new resilience standard may require more MWh without proportionally increasing inverter power.

A solar expansion may justify more stored-energy capacity simply to capture production.

A new grid-services agreement may increase required discharge power at the point of interconnection.

These are different design drivers.

The storage architecture should leave room for them to evolve independently where the selected technology permits.

Maintenance does not have to take down the fleet

A modular plant can also be designed around failure.

Because eventually something will fail.

This should not surprise anyone who owns equipment.

PNNL's modular research demonstrates one of the useful principles: failed elements can be isolated while the remaining system continues operating.^[9]

At campus scale, that philosophy can extend across battery enclosures, PCS blocks, transformers, protection zones, controls, and thermal systems.

A 2 GWh storage plant should not behave like one 2 GWh appliance.

It should behave like a fleet.

That also makes staged repowering and lifecycle replacement much more manageable.

At hyperscale, the energy system begins to look like an SLA

This is the piece I think becomes particularly important.

Data centers already live in a world of service-level agreements.

Availability.

Response time.

Capacity.

Recovery.

Performance.

The customer does not care that the electrical room had a difficult afternoon. The service either met its obligation or it did not.

At hyperscale, the relationship between the campus and the electric system may increasingly develop the same way.

DOE already uses the term "resilience uptime guarantees" for arrangements in which distributed energy resources are used to maintain a defined level of electrical service, specifically identifying data centers as the kind of critical infrastructure for which those guarantees are valuable.^[11]

That is getting awfully close to an energy SLA.

Imagine an interconnection or service agreement that defines things like:

Protected reserve: Maintain no less than a specified quantity of usable BESS energy for critical compute.

Grid-response capability: Reduce utility import by a defined number of megawatts within ten seconds of a qualifying event.

Curtailment capability: Remain below a specified grid-import threshold for two hours.

Ride-through: Maintain critical compute through defined frequency or voltage excursions.

Island transition: Transfer the designated campus loads into microgrid operation without violating compute availability requirements.

Recovery ramp: Return hundreds of megawatts of load to the utility in a controlled sequence rather than reconnecting everything in one enthusiastic step.

National-laboratory testing already demonstrates several of these behaviors in a data-center architecture.^[7]

That makes BESS more than backup equipment.

The battery becomes one of the machines enforcing the campus's service commitment to the grid.

Solar is part of the SLA - but differently

Solar needs slightly different treatment.

A photovoltaic array cannot promise a particular number of megawatts at 2 a.m. during a thunderstorm.

The weather has traditionally resisted contractual negotiation.

So solar by itself is not an uptime guarantee.

But it can absolutely form part of the campus's energy-performance obligation:

• installed renewable capacity;
• annual renewable-energy production;
• percentage of campus energy supplied by onsite generation;
• renewable charging requirements for BESS;
• annual maximum grid imports;
• emissions-intensity commitments; or
• export commitments during qualifying conditions.

DOE specifically notes that solar-plus-storage is more resilient than solar alone because the battery makes generated energy available later and can maintain stable power during changes in solar output or building load.^[13]

So I would frame it this way:

BESS supports the response and availability SLA.

Solar supports the energy-performance SLA.

Solar + BESS + firm generation + controls support the campus-level energy SLA.

At sufficiently large scale, these resources stop being sustainability accessories.

They are part of what allows the compute capacity to exist.

Flexible interconnection may make that mandatory in practice

This becomes especially interesting when utility service itself is conditional.

Princeton ZERO Lab's 2025 work on flexible data centers evaluates combinations of flexible interconnection and bring-your-own capacity, including batteries, solar, and firm resources. Its modeled configurations show how a data center could receive nearly continuous grid service while using onsite or contracted resources during the relatively small number of hours when transmission or generation capacity is constrained. The research was sponsored by Google, which is worth disclosing, but the work uses established power-system modeling approaches and makes the assumptions explicit.^[14]

Conceptually, a future agreement might look something like:

300 MW firm grid service.

Another 200 MW available provided that the campus can reduce its net demand when specified grid constraints occur.

Now the BESS is not optional.

The controls are not optional.

The generation plan is not optional.

They are the infrastructure supporting the additional 200 MW of commercially usable compute.

If one battery block is under maintenance, the controls need to know that.

If a turbine is unavailable, reserve requirements change.

If solar production is above forecast and the battery is full, export capability changes.

At that point, the useful question is no longer:

How big is the battery?

It is:

How much power and energy can the campus safely promise right now?

That is a controls problem.

And data centers happen to be quite comfortable with controls problems.

The battery should have two customers

There is another consequence of installing hundreds or thousands of megawatt-hours of storage beside a data center.

The battery does not have to be useful only to the data center.

Its first customer is absolutely the compute load.

The system must maintain the protected reserve necessary to satisfy its resilience obligations.

A facility that sells tomorrow's emergency reserve to improve today's market spread has confused energy trading with engineering.

Above that reserve, however, storage can potentially become economically active.

FERC Order No. 841 removed barriers to qualified electric-storage resources participating in organized wholesale capacity, energy, and ancillary-service markets.^[20]

Depending on the local market, tariff, and interconnection agreement, a campus BESS might provide value through:

• wholesale energy;
• frequency regulation;
• other ancillary services;
• capacity or resource-adequacy arrangements;
• demand response;
• peak reduction;
• congestion management;
• avoided demand charges; or
• bilateral utility agreements.

DOE and national-laboratory work similarly treat modern BESS as multi-service grid infrastructure rather than single-purpose emergency equipment.^[10]

The battery's first customer is the data center.

Its second customer can be the grid.

And I think there should be a third.

The community should be the third customer

If a data center finances generation, storage, and transmission improvements, there is no particular reason the public-facing benefit has to stop at:

"We promise not to increase your electric bill."

That is a rather modest sales pitch for billions of dollars of infrastructure.

Consider a campus that has established its protected resilience reserve.

The BESS is healthy.

Solar production is strong.

Firm generation is available.

The grid reaches its afternoon peak.

Instead of importing 500 MW, perhaps the campus imports 300.

Or 100.

Perhaps it becomes electrically neutral.

With sufficient generation and storage, perhaps the arrow reverses.

That creates measurable economic value.

Some belongs to the data-center operator.

Some may be monetized in power markets.

Some belongs to the utility or system operator that avoided a more expensive alternative.

Why should some of it not become a community energy dividend?

That might mean:

• peak-event credits for nearby customers;
• municipal energy-resilience funds;
• community BESS installations;
• infrastructure improvements;
• annual local energy credits;
• direct support for low-income ratepayers; or
• reduced local energy charges during defined export events.

And yes, I think a particularly good version could produce occasional zero-energy-charge community days.

Not necessarily a literal $0 utility bill.

Fixed customer charges, taxes, transmission, and distribution costs do not disappear because the battery had a particularly productive Tuesday.

But the energy portion?

That is at least technically imaginable.

Cornell researchers are currently examining a different but related concept - a Curtailment Credit Market through which data centers could pay other flexible loads to reduce consumption during constrained periods, moving some of the economic value of high-value compute back toward ratepayers.^[17]

Different mechanism.

Same interesting principle:

Compute generates extraordinary economic value. Some of that value can purchase flexibility elsewhere in the electrical system.

"Grid positive" is more useful than creative carbon accounting

This architecture also gives the industry a much better public story.

But I would resist trying to turn it into another creatively hyphenated version of "net zero."

If natural gas is burned onsite, there are direct emissions.

Count them.

If solar and batteries displace more expensive or higher-emitting marginal grid generation, count that too.

If the campus exports during a system peak, measure it.

If the campus imports heavily at 3 a.m. when the system has spare capacity and exports at 5 p.m. when everyone wants power, annual MWh alone does not describe its grid value particularly well.

A serious project should publish a scoreboard:

• annual grid energy imported;
• annual energy exported;
• maximum grid demand;
• maximum export;
• BESS power and usable energy;
• protected resilience reserve;
• dispatchable demand reduction;
• onsite renewable generation;
• firm-generation fuel use;
• water consumption;
• reclaimed-water percentage;
• direct combustion emissions;
• estimated avoided marginal emissions;
• transmission investment funded by the project; and
• community energy benefits delivered.

Then add one particularly useful metric:

What did the campus do during the grid's most constrained 50 or 100 hours of the year?

That is where a controllable data center can distinguish itself from an ordinary large load.

A facility can consume substantial annual energy and still provide meaningful grid value by reducing or reversing demand precisely when the power system is most stressed.

That is a much more useful claim than placing a green leaf beside a server rack.

EPA's AVERT tool provides one established way to estimate marginal grid-emissions impacts associated with changes in renewable generation, energy efficiency, and storage dispatch.^[27]

Texas has reached the limit of the extension-cord model

Texas is providing an unusually clear demonstration of the scale problem.

On August 3, 2026, Governor Greg Abbott directed the Public Utility Commission of Texas and ERCOT to audit data-center projects moving through ERCOT's interconnection process before they advance further.

ERCOT was considering more than 474 GW of connection requests - more than five times the system's record peak demand - and the governor's office says approximately 90% of those new power requests were associated with data centers.^[22]

Four hundred seventy-four gigawatts is enough to make even Texas say, "Hang on a minute."

ERCOT consequently paused the relevant Batch Zero advancement while the verification process is worked through.^[23]

Importantly, this is not a blanket prohibition on data centers.

ERCOT's large-load framework specifically recognizes customers proposing their own onsite generation and customers willing to accept curtailment obligations, while truly islanded facilities generally sit outside the ordinary grid-interconnection process.^[28]

That is exactly the distinction that matters.

A 500 MW inflexible load and a 500 MW energy campus capable of self-supply, curtailment, and grid support are not the same planning problem.

They should not be treated like one.

California has arrived at the same question from the other direction

California is handling the problem rather differently.

This will surprise no one familiar with either state.

CAISO launched its Large Loads initiative in 2026 to address how data centers and other major electrical customers connect to and interact with the transmission system. The effort explicitly includes reliability, demand flexibility, cost allocation, and prevention of cost shifting to other customers.^[24]

California's current planning materials forecast another 1.8 GW of data-center load by 2030 and 4.9 GW by 2040 within the CAISO balancing area.^[29]

The CPUC is simultaneously considering how new data-center and other large loads should be assigned the costs they cause, including the possibility of separate large-load customer treatment.^[25]

One California number also deserves clarification because it comes up surprisingly often in power discussions.

50 MW is not a generic threshold at which a data center suddenly "becomes a utility."

The California Energy Commission's 50 MW threshold is principally a power-plant siting and licensing threshold for thermal generating facilities. California also has separate opt-in permitting provisions for certain large clean-energy and storage projects.^[26]

Fifty megawatts matters.

It just does not perform every legal function occasionally assigned to it over lunch.

Texas and California have accidentally agreed on something

Politically, Texas and California agreeing on energy policy is uncommon enough that I would normally check the instrumentation.

To be clear, they have not adopted the same policy.

Texas has imposed an extraordinary audit and verification process on a massive interconnection queue.

California is working through transmission requirements, rate design, flexibility, and cost allocation.

FERC has now joined the same discussion nationally. In June 2026, the Commission ordered all six regional grid operators under its jurisdiction to justify or reform their rules for integrating large loads. The five areas FERC specifically identified include preventing cost shifting, accommodating behind-the-meter generation, flexible transmission service, and processes for generation serving electrically proximate large loads.^[21]

Different politics.

Different markets.

Different regulatory structures.

But the engineering principle underneath them is remarkably similar:

A very large new load should not compromise system reliability, and existing customers should not quietly inherit the infrastructure bill.

Harvard researchers studying AI and the U.S. grid have highlighted the same emerging questions around stranded infrastructure, behind-the-meter resources, flexibility, and determining who ultimately bears the cost of serving rapidly growing compute demand.^[15]

The electrons appear to have achieved bipartisanship.

The data center of the future may be a community energy asset

I suspect the most compelling hyperscale proposal ten years from now will not arrive with a slide saying:

"Here is how much electricity we need."

It will arrive saying:

Here is our compute load.

Here is the generation we are building.

Here is our modular BESS architecture.

Here is the storage capacity installed in Phase One and the expansion path through Phase Five.

Here is our solar production.

Here is our firm generation.

Here is the resilience reserve that never participates in economic dispatch.

Here is the amount of grid demand we can remove in ten seconds.

Here is what we can sustain for ten minutes.

Here is what we can sustain for four hours.

Here is what happens on day three of an outage.

Here is our water budget.

Here are our emissions.

Here is the infrastructure we are paying for.

Here is our maximum import.

Here is our maximum export.

Here is the energy-performance SLA.

Here is the grid-response SLA.

And here is what the host community receives in return.

At that point, calling the project merely a "data center" starts to understate the thing.

It is compute infrastructure embedded inside an energy campus.

It may operate a microgrid.

It may participate in wholesale power markets.

It may support the regional utility.

It may export power during peaks.

It may finance generation that would not otherwise exist.

It may eventually provide community energy credits.

From an electrical perspective, it begins behaving a little like a utility without necessarily being one legally.

And perhaps that is the useful direction for the industry.

The public conversation today is dominated by what data centers consume:

Power.

Water.

Land.

Transmission.

Infrastructure.

The better projects have an opportunity to add another column:

Generation.

Storage.

Grid flexibility.

Resilience.

Infrastructure investment.

Community energy.

A 500 MW data center will always be an enormous electrical load.

There is no clever controls package that changes arithmetic.

But there is a very large difference between a 500 MW facility that simply consumes electricity and a 500 MW facility surrounded by enough generation, BESS, and control capability to decide - second by second - whether it should be a grid load, nearly invisible to the grid, or part of the supply.

The latter is a considerably more interesting machine.

And if you're working through a project where the load has become large enough that generation, modular BESS, controls, utility requirements, and resilience all have to be treated as one system, feel free to reach out.

We've worked on electrical systems where the arrows point both directions.

Those tend to be the fun ones.