By Adam Bridgman, Co-Founder & General Partner and Max Adelman, Associate
On November 30, 2022, OpenAI released the world’s first consumer-friendly Large Language Model (LLM) to the public. The apparent overnight success of ChatGPT was an iteration of research that Ilya Sutskever and others started at Google Brain a decade earlier, which was itself an iteration of neural network and deep learning research that began in the 1980s. The quality of ChatGPT’s output and its rapid ubiquity across different strata of society sent shockwaves across the world to everyone from Fortune 50 CEOs to middle school English teachers. AI has had implications beyond the balance sheet, with vertically integrated incumbents like Microsoft and Google effectively abandoning their net zero goals to participate in the arms race and utilities like Duke and FirstEnergy pulling back on their coal reduction commitments in the face of increased power demand. Although practical applications continue to need fine-tuning, we have seen a variety of use cases across healthcare, finance, media, tech, and, more recently, legacy heavy industrials.
As the initial ripples settled and the financial opportunities in developing and implementing AI became clear, one question was brought into focus: what is the underlying infrastructure needed to make AI work? The subsequent frenzy led to NVIDIA gaining nearly $3T in market cap to briefly become the world’s most valuable company, the biggest businesses in AI (Microsoft, Google, Amazon, Meta, Apple, NVIDIA, IBM, Oracle, and Intel) committing to $100B+ in annual capex to the sector, and the greatest shift in the venture capital funding landscape since the internet.
At the core of the AI revolution is a piece of infrastructure that has been around since the 1960s: the data center.
What is a data center? At its core, a data center is a facility with computing infrastructure to store, manage, and process large amounts of data. These facilities are the backbone of our modern digital ecosystem, enabling network connectivity, data security, website hosting, cloud services, LLM training, and more. Data centers are critical to our digital landscape, but ultimately, they are a physical asset made up of physical components: concrete, steel, insulation, and roofing make up the bulk of the shell, while the interior is filled with servers, chips, networking equipment, racks, cabling, and cooling infrastructure. Keeping the entire system running is a constant power supply, which is achieved with a combination of coal, gas, nuclear, solar, wind, storage, geothermal, and diesel generators. The explosion of AI has introduced significant new demands on the data center ecosystem, particularly in terms of computational power, storage, and efficiency. These developments have strained the existing infrastructure and necessitated a massive expansion—and rethinking—of data center development and construction.
What’s particularly interesting to Ironspring Ventures is the transformation of energy, industrial capacity, and global supply chains required to support this generative-AI boom. There’s so much attention on foundational LLMs and their applications, but the promise of AI is tethered to the real world and will only go as far as its supporting digital and physical infrastructure allows. This is the world that Ironspring lives in – industrial supply chain innovation and built world technology. We care about improving the way incumbent and next gen manufacturers design and create, and how those products are moved, distributed, assembled, constructed, and powered. We broadly categorize our digital industrial focus into four areas: manufacturing, transport & logistics, construction, and alternative energy. There are few places in the industrial supply chain that better overlap with all our areas of expertise than the data center. This piece primarily focuses on the data center supply chain lifecycle, and the associated “picks and shovels” within each component of the value chain. We are excited to actively engage and identify startups that are innovating across these infrastructure layers – both digital and physical – to ultimately enable this intelligent automation revolution.
Along with giving a primer on the current data center lifecycle, we will drill down into the four primary value chain components that represent the critical supply chain bottlenecks for data center development and potential areas of opportunity for venture-backable startups within them:
Do you want to build a data center? Here’s a crash course on how it’s done today.
Activities: project planning, land procurement, evaluating usage needs, securing power agreements
Key Considerations: power and water availability, fiber optic network reliability, cost of labor, proximity to population centers, geographic risk
Activities: design and distribution of chips, production of power and cooling equipment, coordination with downstream stakeholders
Key Considerations: supply chain bottlenecks, geopolitical risk, processing power, long lead
Activities: materials procurement, building, ongoing maintenance
Activities: power servers, maintain optimal internal temperature for peak performance
The data center industry is expected to undergo significant change as demand from more GPUs and the computations they perform become more complex.
While there will always be a place for co-located facilities, the future of data centers is likely to be characterized by larger, “campus-style” developments that emphasize scalability, energy efficiency, and compute power. These campuses will house hyperscale facilities alongside edge data centers, all interconnected through high-capacity networks. The integration of AI, automation, and advanced cooling technologies will drive operational efficiency, while modular and flexible designs will allow data centers to adapt to changing demands. This model lends itself to having on-site power and water resources that are distributed across multiple interconnected data center buildings within a single campus. This will allow for greater scalability, operational efficiency, and resource sharing. These facilities will dwarf current operations in terms of complexity, computing power, and resource demands.
In the rest of this piece, we will break down how the key components of the value chain also represent bottlenecks that venture-backable startups can address.
At the top of the funnel for data center deployment is the design and development stage. After the project owner lays out their goals, the actual design of the data center is mostly handled by the general contractor (GC) and specialty trade groups that have been hired for the project. Although there are varying degrees of vertical integration at the project owner level, the construction teams generally have a lot of autonomy to configure the internal “guts” of the structure. Occurring in tandem with the design of the facility is the complicated process of site selection, which represents a meaningful bottleneck in the project development timeline. Site selection must balance several key criteria:
So what? Successful site selection requires analysis of disparate data sources and collaboration of hyper-fragmented stakeholders with varying priorities. This planning and design phase is incredibly complex, yet the opportunity cost of securing these scarcest assets – potentially feasible land – is high enough to justify speed of acquisition at all costs. Last year Microsoft paid $76M for a 400 acre pumpkin patch in Wisconsin to build a facility. It’s a true land grab moment that is catalyzing a “picks and shovels” infrastructure overhaul. The design phase largely pre-determines the ROI of a project, so delivering tools that automate bottlenecks and ultimately support better, faster, and more dynamic decision making is critically important. There are opportunities for startups to improve multiplayer stakeholder collaboration, deliver flexible and future-proofed design optioneering, bring visibility to strained and nuanced supply chains, and facilitate real-time costing, scheduling, and forecasting analysis. This is not encompassing and many of these businesses will likely not be data center exclusive, but data centers could contribute meaningfully to their customer base. A few areas of interest to Ironspring include:
What does Ironspring think the future looks like?
The materials and components of data centers are wrapped up in a series of complex supply chains that create a sticky bottleneck for project development. Beyond the actual raw materials that go into construction (concrete, steel, copper, etc.), data centers are composed of an assortment of specialized equipment and componentry including: control systems, wiring, rack cable systems, control switches, power distribution units (PDUs), conduit, fiber optic cables, generators, and cooling systems.
Many of these components and materials, which were already constrained from broader global supply chain disruptions, have become further restricted due to the rapid expansion of data center deployment. Certain items can have lead times well over six months. For example, the fiber optic cable manufacturing market is controlled by approximately eight suppliers, only two of whom are in the U.S. (Corning and CommScope). The big tech companies have an unfair advantage over enterprise, co-locators, and edge developers when it comes to building because they can at least lay their own fiber (and have marginally more control over site location as a result). The hyperscalers have been working on channel partnerships to lock up supply of key components and materials such as fiber. From the OEM perspective, many ultimately do not want a small number of massive orders because there are downstream consequences to having your supply so locked up. However, while the hyperscalers remain price insensitive, they will continue to be able to monopolize the supply of long lead items.
Procurement lead times are a major constraint to project development, and while there are startups that are assisting with materials procurement and component manufacturing, most are too small to make a dent for the big players. The primary means through which developers are addressing manufacturing and supply chain challenges today is prefabrication and modular (PFM). PFM solutions enable parts of the construction process to take place off-site and can reduce construction times, costs, and labor constraints as well as improve safety and quality since the work is being done in a controlled setting. By using prefabricated components for the building as well as modular components for the electrical and cooling systems, developers can meaningfully reduce costs and (more importantly) speed up time to deployment. Developers are using PFM techniques for the prefabrication of structural and architectural components (concrete beams, walls, slabs, facades, etc.), skid-mounted and enclosed MEP equipment (cooling and power systems, water treatment, fire suppression, etc.), and holistic turnkey solutions for smaller facilities. Some sophisticated GCs and subcontractors can do manufacturing and integration in-house, whereas others do only one. For incumbent contractors who know the industry, this has been an enormous market opportunity.
So what? There are significant innovation opportunities in areas like next-gen advanced and sustainable materials, 3D printing and rapid production of critical parts and componentry, and modular and prefab solutions to streamline assembly processes and help consolidate supply chains. These are typically capital intensive initiatives by nature and are required to support the speed and flexibility demands of data centers but will also continue to be challenged by incumbents including sophisticated contractors and specialty trades that are vertically designing and bringing localized, micro-manufacturing capabilities in house. There are some exciting developments in robotics bringing intelligent automation capabilities off/near/on-site solving challenging manufacturing and assembly applications like circuit boards, wiring, and harnessing. The need for supply chain visibility and streamlined procurement is a must-have and large opportunity for new and existing players to deliver a secure, integrated, end-to-end platform (we have written about similar trends in building materials procurement here). AI will play a critical role here in optimizing manufacturing processes from procurement to production to quality control. Given the global arms race for compute capacity straining supply chains from every angle, there will be a continued push to onshore and localize as much as possible. As geopolitical tensions rise and data centers become increasingly critical to economies, the manufacture and sourcing of cyber-resilient systems and processes will be critical and highly regulated. To call out a few specific examples of opportunities and areas of interest to Ironspring, see below:
What does Ironspring think the future looks like?
At the end of the day, data centers are a physical asset for the digital ecosystem. For our virtual infrastructure to develop, our real-world infrastructure must come first. At the heart of the data center building process are groups of sophisticated general contractors and specialty trades such as mechanical, electrical, and plumbing (MEPs). Data centers require a vast amount of specialized labor, especially electricians, to handle the complex electrical systems needed to power and cool the facility. Data centers have relatively simple shells, with the real complexity of builds coming from controls, electricals, and power sources. These problems are all exacerbated by the scale of campuses and the rapid surge of these projects occurring in the same geographical areas, stretching labor thin.
The overall data center market is expected to grow ~7-10% annually (from a starting point of ~7,500 globally) through 2030. A significant portion of this growth is coming from the top of the market, with hyperscale data centers anticipated to grow at a 25%+ CAGR through 2032, expanding to nearly 2,000 by the end of the decade. What this ultimately means is that the data center market, while growing overall, is becoming concentrated as hyperscale facilities take over more market share. There will always be a place for co-located and edge data centers, but the shortage of available labor pools seems to largely be coming from the top of the market.
So what? As the data center function has changed (from storage and connectivity towards energy intensive AI training on GPU chips), several elements of the construction process have changed as well. The increased needs in compute power have necessitated adaptations in heating systems, cooling, electrical and mechanical equipment, thicker foundations and beams, and larger conductors to support additional amperage. These are all practical problems with practical solutions, but they continue to add strain to contractors and their supply chains. One of the most significant issues plaguing data center construction and installation is a severe shortage of experienced specialty labor. There are examples of a hyperscaler data center project of such scale that the GC was forced to contract nearly an entire state’s available electrician workforce at one time. As the skilled workforce continues to shrink, this supply and demand imbalance of qualified talent becomes untenable – the need for net new and novel labor capacity is critical. Project developers have already been addressing this with PFM design (see above), but we see several avenues where startups can address this bottleneck including existing skilled worker augmentation, labor marketplaces, knowledge management and training, worksite monitoring, and optimization workflows. To call out a few specific examples of opportunities and areas of interest to Ironspring, see below:
What does Ironspring think the future looks like?
Data centers consume an enormous amount of power. To put it into context, a single server in a machine learning data center (consisting of eight NVIDIA H100 GPUs) consumes as much power as five average American homes. A hyperscale facility that has tens of thousands of servers can require the energy input of a small city but in an area the size of a baseball field. The result of this much power in such a concentrated area is an enormous amount of heat that needs to be managed to maintain efficiency. Cooling and power needs have only grown as more energy-intensive chip designs have increased the power and heat density of facilities. Today, cooling accounts for ~40% of data centers’ energy consumption. New cooling technologies will continue to be developed over time to increase efficiency, but the real problem today is power. Upstream from construction, this is the true bottleneck for project development, to the point that many project developers discuss greenfield data center facilities in terms of power consumption rather than compute capabilities.
So how do you power a data center?
The most obvious answer is to connect it to the electrical grid, which is most often done. However, this presents a few problems: 1) the grid is becoming more strained, 2) lack of power reliability at the scale a data center needs, 3) lack of adequate redundancies in the event that the grid goes down, 4) inadequate power capacity, 5) cost of power, 6) environmental concerns.
The next step solution is independent power from renewables such as wind and solar, which makes sense especially given the carbon emission pledges that hyperscale developers have made. However, this presents a few problems as well: 1) existing infrastructure is often far from population centers and fiber networks, 2) supply chain and permitting for project development, 3) intermittency and variability, 4) cost of associated energy storage, 5) high development costs impacting project level returns, 6) geographical limitations, 7) energy density and capacity limitations. Geothermal energy is another option in this vein, but it has yet to be produced for reasonable cost at scale and is more of a novelty in the data center landscape at this stage.
It is clear that we need to rethink the way that we power data centers. Every industry expert that we have engaged with has spoken about the need for clean, reliable, base load power in the form of nuclear as the only real long-term solution for the data center power problem. Nuclear makes sense for several reasons: 1) high base load power with low interruption risk, 2) low greenhouse gas emissions allowing developers to meet their carbon goals, 3) high energy density, 4) power scalability, 5) long-term cost stability. Nuclear comes with its own challenges such as high upfront costs, safety risks, and permitting challenges. However, the high baseload power mixed with 24/7 reliability will ultimately be the best source of power for data centers. This is already happening today. To give a topical example, which was announced long after we wrote this section of the piece, Microsoft announced the largest power purchase agreement in history with Constellation Energy to use Three Mile Island to power their data centers in Pennsylvania. On a similar thread, Meta’s AI Chief recently announced that their data centers will be built next to nuclear plants, AWS is hiring for a data center principal nuclear engineer to evaluate SMRs and nuclear strategy, and Apple included nuclear power in their 2030 ESG goals.
So what? Unlike the construction and labor bottleneck, which can ultimately be solved with money and a little creativity, power is a problem that is difficult to brute force. Off grid behind-the-meter (BTM) power generation procurement and generation is becoming increasingly topical for project stakeholders, particularly as interconnection queues reach up to five years for new projects. BTM generation makes even more sense considering that data centers are not common use resources and would increase electricity costs for the entire rate base if plugged into the grid. The extremely long lead times for power deployment have created a situation where many developers are building around existing energy assets (such as the Microsoft / Constellation PPA) as their main criteria. There are opportunities for energy optimization platforms, battery storage and energy management systems, compliance and permitting automation specific to energy producers, and likely some very interesting software infrastructure to support the inevitable commercialization of nuclear energy. To call out a few specific examples of opportunities and areas of interest to Ironspring, see below:
What does Ironspring think the future looks like?
The rise of the data center has only exacerbated trends that Ironspring has been closely tracking for years including skilled labor shortages, power availability, supply chain visibility, and the return of advanced American manufacturing. While these problems are not news to us, the rate of change in the data center industry has created a sense of urgency for the key ecosystem players on the ground and new startups to take action.
What changes? Ironspring’s view is that data centers are going to move further away from urban centers, require more energy and water resources, be more modularized, and need more advanced labor, building materials, and electrical components. These facilities are going to be increasingly built around existing power and less so around existing fiber cables (as discussed above, the hyperscalers have an unfair advantage with this in their ability to lay their own fiber). Separately, but not discussed in detail in this piece, we see edge data centers as a major trend in potentially decentralizing the industry, with the market expected to grow at a ~23% CAGR over the next four years. Edge data centers are crucial to the ecosystem because they enable faster data processing and reduced latency by being closer to the data source. We chose to focus on hyperscalers and enterprise data centers in this piece because they pose the most immediate problems for heavy industrial supply chains.
We expect these trends will only get worse (or better, depending on your perspective), especially as companies like Google are planning to build 1GW data centers that will consume as much power as a mid-sized American city. From Ironspring’s perspective, this massive shift in our built environment will continue to manifest in opportunities for us to back world-class founders. What the rise of the data center has made us hopeful for is a future where our physical infrastructure has risen to its rightful place of prominence alongside the digital world that it facilitates. Hype cycles can be fleeting in the venture ecosystem, but AI will have massive impacts on global economies – the supporting picks and shovels opportunities are exciting and much of this is manifesting real time with data center backlog. We look forward to being an engaged ecosystem partner and backing founders who are taking novel approaches to address the bottlenecks around data center development.
Call for founders: at Ironspring, we’re committed to digitizing and innovating across legacy heavy industrial supply chains. If you’re an early stage founder building the next generation of data center tooling for construction, manufacturing, supply chain, or energy, we want to hear from you: max@ironspring.com
Special thanks to the fifteen experts spanning data center builders, startups, and developers we consulted to inform this piece. The perspectives of these ecosystem partners are incorporated throughout this Blueprint.