Data Center Tech: what are startups building now?
In our data center market deck, you will find everything you need to understand the market
SUMMARY
Data Center Tech: what are startups building now? Data center tech startups are building the physical systems that let AI infrastructure exist at scale: power, powered land, liquid cooling, optical networking, modular construction and automation around the facility itself.
The biggest shift is that data centers are becoming an energy problem before they become a computing problem. Once projects reach hundreds of megawatts or gigawatts, access to firm power can decide whether a site gets built at all.
That is pushing startups behind the meter. Geothermal, dispatchable solar, batteries, fuel cells and eventually advanced nuclear are all being pulled into data center design because waiting for a conventional grid connection can take too long.
Cooling is the clearest near-term hardware transition. Air cooling still works across much of the installed base, but dense AI racks are forcing direct-to-chip liquid systems into mainstream deployment faster than previous data center cooling transitions.
The water question is also changing. The more useful goal is no longer simply “use less water,” but to separate rising compute density from rising local water consumption through closed-loop and non-evaporative designs.
Networking is becoming a second utilization problem. Expensive accelerators lose economic value when communication bottlenecks leave them waiting, which is why optical switching, transceivers and photonic interconnects are suddenly attracting much larger rounds.
Construction speed now has direct financial value. A data center that opens months earlier can put billions of dollars of rapidly depreciating GPU hardware to work sooner, making prefabrication and modular infrastructure more valuable than a narrow focus on construction cost alone.
The market is splitting in two directions at once. Gigawatt campuses are being designed for tightly coupled frontier training, while smaller distributed facilities are appearing around inference, sovereign AI and workloads that can be placed closer to users.
Startups do not need to replace Schneider Electric, Vertiv, Eaton or the rest of the established supply chain. The better opportunity is often to solve one fast-moving bottleneck well enough to become a critical component inside a much larger incumbent ecosystem.
The biggest source of hype is the gigawatt headline. Development pipelines, interconnection requests, leases, contracted capacity and operating megawatts describe very different levels of reality, so capacity claims need to be read with care.
Our clearest conclusion is that power currently organizes the whole market. Electricity determines location, rack density drives cooling, cluster scale drives networking, and long deployment timelines make modular construction and automation economically important.
This market map, featured in our data center market deck, highlights top companies and startups in the data center market
Why has data center tech suddenly become a startup market?
Data center tech has become a serious startup market because AI has turned previously boring infrastructure constraints into expensive bottlenecks. We are no longer dealing with a market where adding more servers mostly means finding more floor space. The largest AI installations increasingly need extraordinary quantities of electricity, much denser cooling, faster networking and infrastructure that can be delivered before the underlying chips become obsolete.
The scale change is difficult to overstate. McKinsey estimates that global data center capacity demand could rise from roughly 82 GW in 2025 to around 220 GW by 2030 under its continued-momentum scenario. AI capacity accounts for most of that increase. Its more recent US analysis estimates that data center IT power demand could grow about 27% annually through 2030 and reach roughly 121 GW.
That has changed what counts as valuable technology. A startup that saves a few percentage points of server cost is useful. A startup that makes 500 MW of otherwise inaccessible electricity deployable, allows 100-kW racks to operate reliably, or eliminates months from commissioning can determine whether billions of dollars of GPUs produce revenue this year or sit waiting for infrastructure.
We can see the shift in the companies attracting money and strategic partners. ZutaCore raised $100 million for waterless direct-to-chip cooling. iPronics raised $125 million for programmable optical switching, with Nvidia participating. Exowatt has raised $140 million around dispatchable solar systems aimed partly at AI data centers. Lancium has assembled more than 15 GW of powered-land development opportunities and brought Nvidia directly into its expansion.
What are data center startups actually building now?
Data center startups are currently building four things above everything else: new sources and architectures for power, liquid cooling systems, faster optical networks, and infrastructure that can be manufactured or commissioned much faster. Around those four layers, another group is building software that orchestrates the increasingly complicated physical system.
The previous data center startup cycle looked different. Cloud software, virtualization and server management captured much of the innovation. Today, the hard part has moved back into physical infrastructure.
At the power layer, companies such as Lancium, Exowatt, Fervo and nuclear developers are trying to create electricity that data centers can actually contract and use. At the thermal layer, ZutaCore, Iceotope and other cooling companies are redesigning how heat leaves increasingly dense GPU systems. At the network layer, iPronics, Mesh Optical Technologies and a growing collection of photonics companies are attacking the bandwidth and electrical-power cost of connecting enormous accelerator clusters.
Then there is construction itself. Prefabricated power rooms, cooling modules, containerized compute and standardized AI-factory designs are attempting to move work from unpredictable construction sites into factories. Software companies such as Netris are attacking another hidden delay: configuring the networks and infrastructure after all the physical equipment has arrived.
| Bottleneck | What startups are building | Representative companies |
|---|---|---|
| Power | Powered land, geothermal, modular generation, thermal storage | Lancium, Fervo, Exowatt |
| Cooling | Direct-to-chip, two-phase and precision liquid cooling | ZutaCore, Iceotope |
| Networking | Optical switching, transceivers and photonic interconnects | iPronics, Mesh Optical Technologies |
| Deployment | Modular infrastructure and infrastructure automation | Netris and modular-data-center developers |
As this chart shows, and as featured in our data center market deck, search interest in data centers has increased significantly
Has electricity become more important than the data center itself?
For many new AI projects, electricity is now more important than the building. The decisive asset increasingly comes before the server hall: a credible path to hundreds of megawatts of dependable power.
Traditional data center developers could choose attractive land and then arrange the electrical connection. That sequence becomes much harder when campuses are discussed in gigawatts and grid interconnections can take years.
Lancium illustrates the inversion particularly well. The company is essentially industrializing “powered land.” Its campuses combine land acquisition, transmission connections, generation, storage and energy orchestration. It recently said its development portfolio exceeds 15 GW, including about 4 GW of leased capacity. Crusoe and Lancium are developing a new 1 GW campus in Childress, Texas after already working together on the 1.2 GW Abilene development.
Those numbers put the new scale in perspective. A single 1 GW campus theoretically draws as much power as ten 100 MW facilities. Building several such campuses turns power procurement into energy infrastructure development.
Crusoe shows the same shift from the opposite direction. The company began with stranded-energy computing and evolved into a vertically integrated AI infrastructure developer. It recently disclosed 4.9 GW of contracted AI infrastructure capacity and a broader development pipeline exceeding 40 GW.
If you want more recent data on this point, please see our latest data center market report.
Are startups really building their own power plants for AI data centers, and why is geothermal getting so much attention?
Yes. Data center startups and adjacent energy companies are moving steadily behind the meter because waiting for the grid can be slower than building parts of the generating system themselves. Geothermal is one of the most credible options because it combines continuous low-carbon generation with the possibility of adding genuinely new supply.
Exowatt is developing modular systems that collect solar energy, store it as heat and dispatch electricity later, with the explicit goal of serving loads including AI data centers. The company said after its latest financing that it had raised $140 million in less than two years.
Fervo is pursuing enhanced geothermal at a dramatically larger scale. Google recently agreed to purchase 396 MW from Fervo's Cape Station development in Utah, with an option for about another 600 MW. That potentially puts one commercial relationship close to 1 GW. Fervo has also said it is examining behind-the-meter structures in which geothermal generation supplies a customer's load directly before or alongside eventual grid integration.
The scale matters because older geothermal projects were often discussed in tens of megawatts. AI campuses are increasingly discussed in hundreds of megawatts or gigawatts. Fervo is trying to make geothermal fit the second market by using horizontal drilling and other techniques refined by the oil-and-gas industry, while standardizing development around repeatable GeoBlocks.
Lancium takes a hybrid approach: grid interconnections remain valuable, but its campus designs add behind-the-meter generation, solar, storage and software that can adjust consumption in response to grid conditions.
Even fuel cells are resurfacing. Teragen Energy recently disclosed a $6 million pre-seed round to develop advanced fuel-cell systems for loads including data centers.
The common pattern is simple: electricity supply is becoming something the data center developer actively engineers instead of simply buying from a utility.
This chart, featured in our data center market deck, illustrates yearly venture capital funding for data center startups
Is nuclear power actually becoming a data center technology?
Nuclear power is becoming part of the data center technology roadmap, although small reactors are still well short of being a mainstream data center product. The commercial commitments have become substantial; the deployments remain largely ahead of us.
The strongest evidence is the size of the contracts. Meta has entered an agreement supporting Oklo's planned 1.2 GW advanced-nuclear campus in Ohio. The structure allows Meta funding to help Oklo secure fuel and advance the initial plants, with the first phase targeted for around 2030 and later expansion potentially reaching the full 1.2 GW.
TerraPower is also pursuing data center demand, while Last Energy explicitly markets a “nuclear-as-a-service” model for large industrial electricity users. The proposition is attractive: long-lived, high-capacity-factor generation with relatively small land requirements and potentially limited exposure to grid congestion.
But nuclear still carries a hard execution gap. A signed commercial agreement can be meaningful while operating power remains years away. Licensing, fuel supply, construction and commissioning are all considerably harder than deploying conventional generators or batteries.
The most plausible near-term architecture is mixed. Data center developers may use grid power, gas generation, batteries and renewables first, while geothermal or nuclear capacity arrives later. Oklo itself previously partnered with RPower around precisely this phased concept: natural-gas generation can supply electricity earlier, with advanced nuclear intended to replace or complement it later.
If you want more recent data on this point, please see our latest data center market report.
Is liquid cooling finally replacing air cooling?
Liquid cooling is becoming mandatory for the densest AI deployments, while air cooling will remain common across data centers generally. The important change is that liquid cooling has crossed from an efficiency option into an architectural requirement for a growing portion of new AI infrastructure.
Dell'Oro describes the threshold clearly. Traditional air-based thermal systems historically served rack densities around 20 kW, while liquid-cooling architectures can support 100 kW per rack and beyond. Its latest market work expects worldwide data center liquid-cooling manufacturer revenue to approach $7 billion by 2029 and says single-phase direct liquid cooling has established itself as the dominant architecture for large AI clusters.
Uptime Institute's latest global operator survey provides the other side of the evidence. Typical rack densities across the entire installed data center base are rising only gradually, but a growing share of operators now report peak densities of at least 30 kW. The industry is splitting by workload: normal enterprise racks can remain air cooled while extreme AI zones become radically denser.
The startup activity reflects that transition. ZutaCore recently raised $100 million to expand its two-phase, waterless direct-to-chip technology. Iceotope raised $26 million for precision liquid cooling. CAEPlus is developing active liquid-cooling technology and recently received strategic investment from Molex.
| Cooling architecture | Where it fits | Current direction |
|---|---|---|
| Air cooling | Lower-density conventional racks | Remains widespread |
| Direct-to-chip liquid | High-density GPU and AI racks | Becoming the mainstream AI approach |
| Two-phase direct-to-chip | Very high heat flux, water-conscious designs | Growing startup activity |
| Immersion cooling | Specialized high-density environments | Technically credible, adoption more selective |
This chart, featured in our data center market deck, shows how Equinix is capturing share in data centers
Are startups trying to eliminate data center water use?
Yes, although the more useful goal is to decouple higher compute density from higher water consumption. Some startups are explicitly designing cooling systems that avoid facility water use, while new data center projects increasingly advertise closed-loop or non-evaporative designs.
ZutaCore's proposition is particularly direct: its two-phase direct-to-chip system is designed as a waterless heat-removal architecture. Refrigerant moves through the system and changes phase close to the chips instead of relying on conventional evaporative water cooling.
Large campus designs show the same direction. Crusoe and Lancium's planned 1 GW Childress campus specifies closed-loop, non-evaporative liquid cooling. Circulating liquid inside a closed system does not imply continuously consuming comparable quantities of fresh water.
The pressure goes beyond environmental branding. Water availability can influence permitting and community acceptance, especially when enormous campuses compete with municipalities, agriculture and power generation for local resources. Recent research has also emphasized that a large part of the industry's broader water footprint can occur indirectly through electricity generation, so eliminating cooling-tower consumption still leaves part of the total footprint elsewhere.
Why are optical networking startups raising so much money, and could optics replace electrical links inside AI data centers?
Optical networking startups are attracting serious capital because connecting GPUs is becoming nearly as important as the GPUs themselves. Optics will not replace every electrical connection, but the optical boundary is moving progressively closer to processors as bandwidth and power requirements become harder for copper to satisfy.
iPronics gives us an unusually fresh example. The company raised $125 million in Series B financing, bringing disclosed funding to $177 million, and Nvidia participated. Its Optical Networking Engine is a programmable optical circuit switch designed to reconfigure AI-cluster connections dynamically.
Mesh Optical Technologies raised $50 million in Series A financing to mass-produce optical transceivers. Its founders came from SpaceX optical communications work, and the company is targeting the enormous volume of links required to make distributed accelerators behave like a coherent computing system.
The startup landscape now spans multiple layers: lasers, co-packaged optics, photonic integrated circuits, optical circuit switching, transceivers and alternative chip-to-chip links. Established suppliers are moving in the same direction. Molex recently introduced an optical-circuit-switch platform and expanded its co-packaged-optics components for large AI clusters.
The economics are straightforward. An expensive accelerator waiting for data or another GPU is underutilized capital. Across tens or hundreds of thousands of accelerators, small networking inefficiencies become a lot of stranded compute.
Electrical connectivity remains very efficient over short distances, while optics still carries packaging, laser, reliability, manufacturing and cost challenges. The likely architecture is hybrid: copper stays where it remains economical, while optics takes over increasingly demanding links.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, illustrates yearly funding for data center startups
Can modular construction and software make data centers much faster to deploy?
Yes. Data center deployment is becoming both a manufacturing problem and a software-automation problem because conventional project-by-project building and manual commissioning are too slow for the rate at which AI hardware changes.
A useful benchmark comes from Delta's new prefabricated AI data center architecture. Although Delta is not a startup, its product direction demonstrates where the market is heading: integrated power, cooling, piping and IT modules manufactured before they reach the site. Delta says its approach can reduce deployment time by up to 60%.
Canopy Wave recently entered the modular AI data center market with a factory-prefabricated architecture targeting roughly three-to-six-month core deployment cycles. Its logic is explicitly tied to the mismatch between GPU innovation cycles and conventional data center construction.
The financial argument is just as important as the operational one. Suppose a project contains several billion dollars of accelerators. Every additional month between delivery and productive use creates an enormous opportunity cost. Saving six months on the building can easily be worth more than shaving several percentage points from the construction budget.
Software attacks the same delay after the physical infrastructure arrives. Netris raised $15 million in Series A funding to automate networking for AI infrastructure providers and neoclouds. Switches, routing, tenant networks, security and connectivity still have to be configured after a facility has obtained its GPUs and networking hardware.
Energy orchestration is another part of the same problem. Lancium is using software to alter the power behavior of large compute campuses, combining grid interaction, local generation and batteries. Its Nvidia collaboration includes technology intended to adjust AI factory consumption against available power and increase the amount of compute achievable within a fixed electricity envelope.
Are startups building smaller data centers instead of gigawatt campuses?
Some are, and the contradiction is more interesting than it appears. The data center market is moving simultaneously toward gigantic gigawatt campuses and much smaller distributed AI facilities.
Gigawatt campuses optimize for enormous training clusters and hyperscale inference. Lancium's projects start around 1 GW, and Crusoe's contracted infrastructure already spans several gigawatts.
But not every workload requires one contiguous supercluster. Lektra, for example, is developing EdgeScale AI sites of 20 MW or less around locations where power infrastructure already exists. It says it has eight operational sites and is working with Penguin Solutions on AI systems for the facilities.
That is almost the opposite strategy: instead of waiting years for one 1 GW interconnection, find fifty places capable of supporting 20 MW.
The architecture works better for some inference workloads than frontier-model training, because training benefits heavily from enormous tightly coupled clusters. Inference is easier to distribute geographically and may actually benefit from proximity to customers.
The market is splitting into two useful architectures: huge synchronous clusters for frontier training, and smaller distributed facilities for regional inference, sovereign AI and enterprise workloads.
This chart, featured in our data center market deck, compares the main business model options for hyperscale data center operators
Are startups still trying to make data centers greener, or is speed winning?
Speed has become the dominant constraint, but sustainability still shapes which technologies can scale. The market is increasingly skeptical of solutions that are clean but cannot supply power soon enough, and equally skeptical of fast build-outs that create unacceptable local environmental costs.
We see this clearly in generation choices. Enhanced geothermal is attractive because it promises continuous low-carbon power rather than intermittent energy alone. Exowatt couples renewable generation with thermal storage because data centers cannot stop computing when the sun sets. Nuclear developers sell reliability and carbon-free output together.
At the same time, natural gas is clearly remaining part of the build-out. McKinsey's latest US power analysis found that among data center developers expecting to deploy on-site generation, 64% anticipated using natural gas. Google has also contracted large quantities of gas-backed generation for a Crusoe development while separately contracting geothermal power.
Cooling shows the same tension. Closed-loop liquid systems can support denser racks while reducing direct water consumption. Better thermal management can also reduce auxiliary electricity consumption.
| Requirement | Old optimization | Emerging optimization |
|---|---|---|
| Electricity | Low cost per kWh | Available, firm megawatts delivered on time |
| Cooling | Low facility PUE | Remove extreme chip heat with limited water |
| Sustainability | Annual renewable matching | Local power, water and carbon constraints |
| Construction | Lowest project cost | Fastest reliable time to productive compute |
If you want more recent data on this point, please see our latest data center market report.
Where is the biggest opportunity: power, cooling or networking?
Power is currently the largest strategic bottleneck, cooling is probably the clearest near-term equipment market, and networking may offer the deepest technology opportunity. Each captures value differently, so forcing one universal winner would hide more than it clarifies.
Power controls whether a project exists. A company that secures a credible 1 GW interconnection or supplies hundreds of megawatts of new generation can unlock an infrastructure project worth many billions of dollars. Lancium's scale demonstrates why powered land has become extraordinarily strategic.
Cooling sits one level downstream and benefits from a much more predictable transition. Accelerator power is rising, dense AI clusters already need liquid cooling, and Dell'Oro expects the liquid-cooling equipment market to approach $7 billion within a few years. That creates a clear commercial runway for companies with differentiated cold plates, coolant distribution, fluids or two-phase systems.
Networking is less obvious but potentially more transformative. If accelerator fleets continue scaling, bandwidth and communication power become fundamental limits on useful compute. The sudden financing of companies such as iPronics and Mesh Optical Technologies suggests investors increasingly understand this.
Measured by infrastructure value, power is the largest opportunity. Measured by an immediate equipment transition, cooling is cleaner. At the component level, optical networking may have the most interesting technology upside.
This chart, featured in our data center market deck, shows the revenue mix across customer segments in the data center market
Are incumbents too strong for data center startups to matter?
Incumbents dominate data center equipment, but the current market is unusually favorable to startups because AI is forcing customers to adopt technologies faster than traditional infrastructure cycles normally allow. Startups do not need to replace Schneider Electric, Vertiv, Eaton or established construction firms to build large businesses.
Cooling shows how this can work. Established suppliers control huge portions of power and thermal infrastructure, yet ZutaCore can raise $100 million because two-phase cooling addresses a problem whose technical requirements are moving quickly. Iceotope can develop precision cooling and work alongside larger ecosystem companies rather than recreate the entire cooling plant.
Optics follows the same pattern. iPronics does not need to manufacture every switch, cable and optical component inside a data center. It needs to solve a sufficiently expensive networking constraint and integrate into the broader Nvidia-led ecosystem. Nvidia's participation in its recent financing is significant precisely because it shows how component startups can become part of a much larger platform.
Large incumbents are also investing in or partnering with smaller companies. Molex's investment in CAEPlus around active liquid cooling is an example of the likely commercialization route.
What is mostly hype in data center tech right now?
The biggest source of hype is confusing announced capacity with functioning infrastructure. Gigawatts appear everywhere in data center announcements, but a development pipeline, an interconnection request, a signed lease, a financed project and an operational campus are completely different things.
Power startups deserve the same skepticism. A 1 GW nuclear agreement can be commercially meaningful without implying that 1 GW of nuclear electricity will arrive soon. A geothermal resource estimate is still several steps away from completed generation. A behind-the-meter generation plan still requires equipment, fuel, permits and transmission or distribution infrastructure.
Cooling has its own version of the problem. Very high theoretical rack densities make impressive demonstrations, but the commercial question is how many production systems are running reliably, how easily they integrate into existing facilities, and whether operators are willing to maintain them.
Optical networking can also be overinterpreted. The physical case for moving more communication into optics is strong, but many architectures are competing for the same transition. Category-level inevitability tells us very little about which startup eventually wins.
Finally, there is a temptation to label almost any company touching AI hardware as a “data center tech” startup. Generic GPU clouds belong in the category only when their differentiation comes from the physical or operational infrastructure itself.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, shows how hyperscale AI-ready campus technology has evolved over time
So what are data center tech startups really building now?
Data center tech startups are building the missing physical infrastructure of the AI boom. The center of gravity has shifted decisively away from generic server management and toward the technologies that turn scarce power into dense, operational compute: generation, powered land, liquid cooling, optical networking, modular construction and physical-infrastructure automation.
Our clearest conclusion is that power is the organizing problem.
Once campus requirements move into hundreds of megawatts and gigawatts, almost everything else follows from electricity. Power availability determines location. Chip power determines rack density. Rack density determines cooling architecture. Cluster size determines networking. Long interconnection queues make behind-the-meter generation attractive. Expensive GPUs make construction and commissioning speed financially critical.
That explains why companies that initially appear unrelated are converging on the same market. Lancium develops powered campuses. Fervo drills geothermal wells. Exowatt stores solar energy as heat. ZutaCore removes heat from processors. iPronics switches optical connections. Netris automates network deployment. They are all solving different stages of the same conversion process: electricity into useful AI computation.
Startups are effectively unbundling the data center and rebuilding the pieces whose old engineering assumptions no longer survive AI-scale density.
That is what data center startups are really building now: ways around the physical limits that have become the limiting factor on AI itself.
OUR METHODOLOGY
This analysis asks what data center tech startups are actually building now and which bottlenecks are pulling the most capital, customer commitments and infrastructure activity. We broke the market into power, cooling, networking, deployment and the software coordinating those physical systems, then compared evidence across those layers.
We prioritized evidence that moved beyond technical possibility. Funding rounds show where investors are placing bets; commercial agreements, leased or contracted capacity, product launches, deployments, operator surveys and market forecasts show whether those bets are connecting with real infrastructure demand.
Power claims were treated carefully because headline gigawatts can describe very different realities. Operational capacity, contracted capacity, leased capacity, commercial agreements and development pipelines were kept separate rather than treated as equivalent.
Large technology companies and established infrastructure suppliers were used mainly as validation points. Investments, agreements and partnerships involving Nvidia, Google, Meta, Molex and Delta help show where startup technologies are connecting with the broader data center ecosystem without turning those incumbents into the subject of the article.
We also avoided forcing power, cooling and networking into one ranking metric. Power was judged by its ability to determine whether a project can exist, cooling by the visibility of the near-term hardware transition, and networking by the depth of the technical constraint as AI clusters scale.
Key sources include McKinsey on global data center capacity demand, McKinsey on US data center power demand and on-site generation, Lancium on its 15+ GW portfolio and Nvidia partnership, Crusoe on contracted AI infrastructure capacity, Fervo on its 396 MW Google agreement, Oklo on the planned 1.2 GW Meta-backed nuclear development, Dell'Oro on liquid cooling, Uptime Institute on rack-density trends, iPronics on optical networking, and Netris on AI infrastructure network automation.
In our data center market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Data Centers: what’s changing now?
- Data Centers: what are the biggest unsolved problems?
- The latest funding news in the data center market
- What are the top startups in the data center market?
- What are the fundraising trends in the data center market?
- How strong is fundraising in the data center market right now?
