Every time someone asks an AI chatbot a question, a building somewhere gets a little warmer. Multiply that by billions of questions a day, add the training runs behind the scenes, and you get one of the fastest-growing energy stories on the planet.

The International Energy Agency says electricity used by data centers worldwide will more than double to around 945 terawatt-hours by 2030, with AI as the biggest reason. That is roughly as much electricity as Japan uses today. Water is part of the story too. One industry estimate puts global data center water use at 5.6 billion gallons a day in 2026, and community pushback over water has already delayed or blocked more than $150 billion worth of projects.

So it is fair to ask: is there a way to run AI without draining the grid and the river?

There is, and it is not what most people expect. The biggest new idea is not a better solar panel or a bigger battery. It is swapping copper wires inside the data center for light traveling through glass. That sounds like a small engineering detail. It is not. Moving data around burns a surprising share of a data center's power, and light does it far more cheaply than copper.

This blog walks through why data centers are so hungry, what "optical" really means, how Cisco turned itself into a major player in that world, what a startup called Ayar Labs is doing inside the chips themselves, and what all of this means for the people who own, operate, and report on data center buildings.

Think of a data center as a giant hair dryer that never turns off. Inside are rows of computer chips doing math. Every bit of electricity those chips use turns into heat. That heat has to go somewhere, or the chips cook themselves.

For years the answer was air. Big fans blew cold air across the servers, and the warm air went outside. To make that cold air, data centers often used cooling towers. Those towers work like sweating: water evaporates, and evaporation carries heat away. It works well, but roughly 80% of the water that goes into a cooling tower is lost to the sky. That is why a single large data center can drink as much water as a small town.

AI made all of this worse, fast: a normal server rack used to draw 5 to 10 kilowatts. An AI rack packed with GPUs can draw ten times that or more. Too much heat is packed into too small a space for air to handle, which is why the industry is moving to liquid cooling, where a fluid runs right up to the chip. We will come back to water later, because liquid cooling helps with some problems and can make others worse.

Here is the part most people miss. Only some of a data center's electricity goes into actual computing. A big slice goes into cooling, and a surprisingly large slice goes into moving data between chips. A GPU is useless on its own. It has to talk to other GPUs, memory, storage, and the network thousands of times a second. For decades that talking happened over copper wires, and copper has a problem that gets worse every year.

Copper carries electricity, and electricity fights back. Push a signal down a copper wire, and it weakens every centimeter. The faster you send data, the worse the loss gets. Today's AI chips want to send data at 200 gigabits per second on a single lane. At that speed, Nvidia says a copper link can lose up to 22 decibels of signal, which is a polite way of saying most of it disappears.

To fix that, engineers add chips called retimers and signal processors that clean up and boost the signal. Those chips burn power. A single high-speed network port can draw around 30 watts to keep the signal readable. Multiply that across tens of thousands of ports, and you have megawatts of electricity doing nothing but shouting over the noise in a copper wire. Every one of those watts becomes heat, and every bit of heat needs cooling; cooling needs power, and often water.

There is a second cost. Because copper loses signal so quickly at high speed, it only works over short distances, a meter or two at most. That forces AI designers to cram GPUs as close together as possible, which packs even more heat into even less space. The chips are hot because they are close together, and they are close together because copper cannot reach any further.

Mark Wade, the CEO of Ayar Labs, puts it bluntly: AI is hitting a power wall driven by interconnect inefficiency. The wires, not the chips, are becoming the limit. One analyst who spent several days at the industry's biggest optical conference this year came home with a simple prediction. All AI data center interconnects will be optical within five years.

If you have fiber internet at home, you already use the trick. Instead of pushing electricity through metal, a tiny laser flashes light down a strand of glass. Light does not weaken the way electricity does. It can travel meters or kilometers with almost no loss; it does not heat the cable, and you can send many colors of light down the same strand at once, each carrying its own stream of data.

Data centers have used fiber for years to connect buildings and racks. What is new is how close to the chip the light now gets. There are three steps on that ladder, and each one saves more power.

Pluggable optics. This is the old way. A small module plugs into the front of a network switch, converts electricity to light, and sends it down fiber. It works, but each module has its own signal-cleaning chip inside, and those chips are power-hungry.

Linear pluggable optics (LPO). Same plug-in idea, but the signal-cleaning chip is removed, and the switch does the work instead. Cisco says this cuts the module's power use by about half and can trim total switch power by around 30%.

Co-packaged optics (CPO). This is the big one. Instead of a module at the front of the box, the tiny laser and light converter sit right next to the switch chip, inside the same package. The copper path shrinks from centimeters to millimeters. Nvidia says its CPO switches are 3.5 times more power-efficient than switches with pluggable modules, with ten times fewer failures because there are fewer parts to break. One early report puts the saving at around 3 kilowatts per switch. In a cluster with hundreds of switches, that freed-up power can run more GPUs instead of more fans.

The simplest way to picture it: copper is like passing a note hand to hand across a crowded room, with people re-writing it every few steps so it stays readable. Light is like saying it once, clearly, across the whole room: less effort, less heat, fewer mistakes.

Most people think of Cisco as the company that makes the routers in office closets. That is still true, but over the last two years Cisco has quietly become one of the three or four companies that matter most in AI networking, and optics is a big reason why.

The foundation is a chip family called Silicon One. In February 2026, Cisco announced the Silicon One G300, a switch chip that moves 102.4 terabits of data per second. Cisco says it is built for AI clusters measured in gigawatts and improves how fast AI jobs finish by about 28% because it handles the bursty, all-at-once traffic AI creates without dropping data.

The chip only matters because of what plugs into it. Alongside the G300, Cisco launched 1.6-terabit optical modules and 800-gigabit linear pluggable optics, the LPO design that strips out the power-hungry signal chip. That is the practical, ship-it-this-year version of the optical shift. Cisco's optics leader, Bill Gartner, has been publicly cautious about co-packaged optics, noting that a CPO package may have a thousand or more fiber connections glued to a piece of silicon and that reliability must be proven. Cisco's bet is to win on the step everyone can take now, LPO, while CPO matures.

The second move changed Cisco's position in the market. In 2025, Cisco and Nvidia agreed to combine their networking portfolios, making Cisco the only outside chip supported inside Nvidia's Spectrum-X AI networking platform. The first product of that deal, the Cisco N9100 switch, runs on Nvidia's silicon with Cisco's management software on top. For a company building an AI data center, that means you can buy Nvidia's GPUs and Cisco's network and have them designed to work together, which is exactly what large enterprises, governments, and sovereign cloud builders want.

Why does this matter for an ESG story? Because Cisco is the vendor most big companies already trust, and it is now selling the energy-saving optics as a standard part of its AI data center kit, not as an exotic option. When the default product from the default vendor cuts switch power by 30%, that saving spreads across thousands of buildings without anyone having to make a brave decision.

Cisco and Nvidia are bringing light to the network switch. A smaller group of companies wants to go one step further and bring it right onto the processor itself. The best-known is Ayar Labs, a Silicon Valley startup that has become one of the most heavily funded hardware companies of the AI era.

In March 2026, Ayar Labs raised $500 million at a $3.75 billion valuation, with Nvidia, AMD, and MediaTek among the investors. It added another $150 million in late September, and Wiwynn, a major server builder, joined as a backer. That is close to a billion dollars raised for a company most people outside the industry have never heard of.

The product is two parts. The first is a tiny chip called TeraPHY. It sits inside the same package as a GPU or AI accelerator and turns the chip's electrical signals into light. A single TeraPHY can move up to 8 terabits per second using 16 colors of light at once. The second part, called SuperNova, is a laser that sits outside the hot chip package and feeds light into it, a bit like a power supply, except it supplies light instead of electricity. Keeping the laser away from the heat is a clever move, because lasers hate being hot.

The energy numbers are the point. Ayar Labs says its optical links use as little as 5 picojoules per bit, which is a tiny fraction of what a long copper link needs at the same speed, and it claims four to eight times better power efficiency than today's pluggable optics or copper. Because light does not care about distance the way copper does, GPUs no longer have to be jammed into one rack. They can be spread across a room and still act like one giant computer. Less crowding means less concentrated heat, and less concentrated heat means a cooling system that can be smaller, cheaper, and thirstier for fewer gallons.

Ayar Labs is not alone. Startups such as Lightmatter and Celestial AI are chasing the same prize with different designs, and in March 2026 AMD, Broadcom, Meta, Microsoft, Nvidia and OpenAI formed an Optical Scale-up Consortium to agree on common standards. Nvidia has also put roughly $4 billion into optical component makers to ensure supply. When buyers, chip makers, and standards bodies all move at once, a technology stops being a lab demo and becomes a line item in next year's budget.

This is where a lot of blogs get it wrong, so it is worth slowing down.

Water use in a data center comes from one thing: removing heat. Every watt that goes into a chip, a copper wire, or a fan comes out as heat, and the heat has to leave the building. If it leaves by evaporating water in a cooling tower, you use a lot of water. If it leaves through a closed loop and a dry cooler, like a car radiator, you use almost no water on site, but you use more electricity to run the fans. There is no free lunch, only a choice about which bill you pay.

That is why liquid cooling gets confusing. Running fluid straight to the chip is far more efficient than blowing air, and it can sharply cut cooling energy. But it doesn't decide where the heat goes afterward. One industry survey found direct liquid-cooled AI facilities using 1.8 to 3.2 liters of water per kilowatt-hour when paired with evaporative heat rejection, more than some air-cooled sites. At the same time, closed-loop and immersion designs have been reported to cut direct water use by 70% or more. Same chips, very different water bills, depending on the last step.

So where do optical cables fit? They attacked the problem one step earlier. Light does not make the cooling system better; it makes the building need less cooling in the first place. Fewer retimers and signal chips mean fewer watts turning into heat. Co-packaged switches drawing kilowatts less each means a smaller load on the chillers. Spreading GPUs across a room instead of cramming them into one rack means heat that is easier and cheaper to remove. Every watt you do not create is a watt you do not have to cool, and roughly a liter or two of water you do not have to evaporate.

Put the three together, and you get the realistic picture of a low-water AI data center in 2027 or 2028: optical links so the chips waste less power talking to each other, direct-to-chip liquid cooling to pull heat away efficiently, and closed-loop or free-air heat rejection so the heat leaves without evaporating a river. Hybrid sites in cooler climates already report under 1.2 liters per kilowatt-hour by using outside air most of the year.

One last honest caveat. The water used to generate electricity at the power plant is usually larger than the water used for cooling at the data center. Cutting power with optics helps that number too, but a data center on a coal grid with perfect cooling still has a big water footprint; it never sees where the electricity comes from matters as much as what happens inside the building.

Most of the companies reading about optics are chip and network buyers. But a growing share of AI data centers are owned by real estate investors, REITs, and infrastructure funds who lease the shell and the power to a tenant. For them, the optical shift changes three things.

The numbers you report. Two metrics drive data center ESG reporting. PUE (power usage effectiveness) measures how much electricity goes to overhead like cooling versus actual computing. WUE (water usage effectiveness) measures liters of water per kilowatt-hour of IT power. Optics and better heat rejection move both in the right direction, and they do it inside the tenant's equipment and the building's cooling plant at the same time. That matters because investors, lenders, and regulators are starting to ask for site-level PUE and WUE, not company averages. Community opposition over water has already delayed or blocked more than $150 billion of projects. A credible water story is now part of getting a permit.

The lease. The lease determines who pays for power and water, who gets credit for savings, and who decides on cooling design. A tenant that upgrades to co-packaged optics and liquid cooling may cut its power draw per rack while asking the landlord for a very different cooling plant. Leases written for air-cooled 10-kilowatt racks do not cover that. Owners are starting to add clauses on liquid-cooling readiness, water budgets, and how they share efficiency gains.

The data. None of the reporting works if meter data, cooling plant logs, and the tenant's IT load are stuck in three different systems. Owners running portfolios on platforms like Yardi or MRI are finding that the ESG module is only as good as the utility and building-systems data feeding it. The firms doing this well treat energy and water data like rent data: metered, reconciled monthly, and traceable to a source.

There is a bigger point underneath all this. Over the next five years, the data center boom will be judged on whether it ate the grid and the water table or found a way not to. Optical cables are one of the few pieces of good news in that story, because they cut the problem at the source rather than cleaning up after it. Owners who understand that will be the ones tenants, lenders and towns want to deal with.

You do not need to understand photonics to act on this. Here is what to ask, whether you are building, buying or leasing.

If you are building or expanding

• Ask the network designer which optics they plan to use: standard pluggable, LPO, or co-packaged. Each step down that list cuts switch power

• Design the cooling plant for liquid cooling from day one, even if the first tenant is air-cooled. Retrofits cost far more

• Decide how heat leaves the building. Evaporative towers use water; closed-loop and free-air designs use power. Pick based on the local water situation, not habit

• Get a WUE target into the design brief, not just a PUE target

If you are leasing to an AI tenant

• Add liquid-cooling readiness, rack power density and a water budget to the lease

• Agree up front how efficiency savings are shared when the tenant upgrades its optics or cooling

• Require the tenant's IT load data in a form that feeds your ESG reporting

If you are investing or lending

• Ask for site-level PUE and WUE, not portfolio averages

• Check whether the local community or regulator has pushed back on water use nearby

• Look at the electricity source. A clean grid shrinks the water footprint you cannot see

If you are running the reporting

• Make sure meter data, cooling logs and tenant IT load land in one system, reconciled monthly

• Treat ESG data with the same rigor as rent and expense data, because lenders soon will

None of this is exotic. It is the same discipline good owners already apply to rent rolls and capital budgets, pointed at kilowatts and liters instead of dollars.