The value chain underpinning the latest wave of artificial intelligence (AI) is burdened with risks around cross-border trade of geopolitically sensitive materials and equipment. Large incumbents dominate supply chains. There is little to reference in terms of challenger suppliers generating returns for private equity backers. Yet investor appetite is clearly on the rise.

Evidence came earlier this month with a first close for the debut blind pool fund Celadon Partners. The firm, which has gradually pivoted into a data centre value chain specialist, headquartered in the US and investing mostly in Asia, raised approximately USD 200m against an overall target of USD 500m.

The thesis observes that only 30% of the total spend on building compute capacity in the age of AI is infrastructure – the rest is IT and power equipment, often undervalued and experiencing generational shifts in demand. Historically, businesses built around these inputs plodded along under the radar. Now, they are critical to a megatrend.

Geopolitical tensions are seen as most severe in semiconductor production, though still manageable. Other segments are much less vulnerable. Across the board, there is technology risk on the cutting edge, but it can be diffused by understanding the preferences, problems and behaviours of hyperscaler computing companies.

“Beyond semiconductors, areas like networking, power, circuit boards, and liquid cooling are becoming high value and important,” said Celadon CEO Donald Tang. “Many businesses that were lower-margin or cyclical are now critical to infrastructure and riding trends of long-term structural growth.”

Target areas

Central to the opportunity set is the idea that advanced AI requires better chips and consumes more electricity. In the past five years, the power density of data centre racks has increased 10x to around 50 kilowatts. It is expected to hit 100kW in the near term.

Big-ticket land and infrastructure investments and hyperscaler contracts define the value chain upstream. Private equity investors are expected to be more active in supporting functions around construction, project design, and component manufacturing.

Power and cooling systems are attracting the most attention in terms of equipment, although there are opportunities in specialised chips. Further downstream, investment will focus on operations and maintenance services, including software and specialised subcontractors.

Project owners and general contractors prioritise product reliability and will favour established suppliers, minimizing investor inroads in core equipment such as generators, transformers, and chillers.

Small to mid-size GPs will therefore see more opportunity in sub-components and potentially disruptive technologies, according to Joongshik Wang, EY-Parthenon’s ASEAN and Singapore strategy and execution leader.

Initial PE activity is likely to focus on conventional engineering, procurement, and construction (EPC) service providers transitioning their business models to more directly address the AI theme in their local markets. The key diligence question will be how to value such companies and weigh the amount of revenue coming from traditional operations versus AI.

Overall, there is a sense that the sector is becoming more fragmented as technical hurdles mount, and that as a result, everything is up for grabs in different bite sizes. This can be seen in the wide spectrum of investors chasing assets that are comparable in operations if not scale.

For example, A91 Partners, which typically writes cheques of USD 50m or less, recently acquired a significant minority stake in India’s Sasmos, a company that provides optical fibre assemblies to data centres. Bain Capital accessed the same niche in 2021 by participating in an acquisition of Coherent Corp at an implied enterprise valuation of USD 6.8bn.

“We continue to be bullish about AI-driven digital infrastructure investment opportunities,” Drew Chen, a partner of Bain, said, highlighting his firm’s investments in memory chips, power systems, and data centres themselves. “We will expect to see more deals in this space.”

Asia angles

Annual AI infrastructure spend is typically estimated in a range of USD 400bn-USD 500bn globally. Underlying equipment and services investments account for 50% to 70% of this, depending on the market, according to several investors contacted for this story.

Asia is the primary components production market in most categories and is expected to grow the fastest in terms of data processing for AI. The region is second only to North America with 20,300 megawatts of data centre capacity as of 2025, according to KPMG. This figure is expected to grow 13.1% through 2030, outpacing North America at 9.2% and Europe at 5.3%.

“There’s enough demand – but can you really capture the opportunity? That’s a different question,” said EY-Parthenon’s Wang.

“There are a lot of good local-level opportunities for private equity with traditional players that are willing to transform to address AI. But it’s really about the corporate level. Do they have a proven track record? Are they gaining a lot of IP and technology and engineering capability? You have to be careful.”

Wang expects private equity to be active primarily in power and electronics, including related software, analytics, and sustainability offerings. He is most optimistic about concepts that combine product and services delivery such as portable modular data centres (PMDCs).

In this model, power generators, power distribution units, cooling systems, and a data hall for housing server racks are sourced from original suppliers before being containerized and shipped as a single unit to the data centre site. It is considered relatively asset light and capable of speeding up project delivery timeframes from 24-36 months to 6-12 months.

Essentially, this is an offering in complex supply chain management. It will, in theory, appeal to global hyperscalers and data centre developers without the ability to coordinate Asian counterparties and Asian labour. Wang said EY is currently advising multiple PMDC acquisitions for private equity in the region. These companies’ annual sales can range around USD 500m.

“How do you organise relationships for cooling and power? How do you actually bring them together into one single location and build it into containers?” he said. “It is manufacturing capacity, but it’s not the expensive capex you would imagine.”

Talking shop

Private equity inroads will come into clearer focus as the industry expands its touchpoints with AI companies and data centre operators. Large firms with existing experience on the infrastructure side will have an early advantage.

Early investors in AirTrunk offer a uniquely Asian case in point. The Australia-founded data centre operator was acquired by a Blackstone-led consortium in 2024 for about USD 16bn. Key takeaways from the initial buildout included an education in hyperscaler uncertainty about entering Asia and market-by-market challenges around project management.

India can be an expensive market for land with significant uncertainty around development approvals. Korea can be subject to civil complaints and restrictions on power availability. Japan is considered a particularly tight market in terms of contractor labour.

Furthermore, best practices evolve rapidly. In some areas, such as optical networking, technology changes from generation to generation every 18 to 24 months. Depreciation schedules for IT equipment are therefore getting shorter. Some data centre tenants are pursuing more financing for equipment.

Nikhil Reddy, head of real estate for Asia Pacific at Goldman Sachs Alternatives, one of AirTrunk’s first investors, said his firm is tracking a dynamically evolving set of preferences, requirements and considerations from tenants in terms of equipment. That makes it more difficult to cater to what they want.

“Timing is probably the biggest challenge for tenants today,” he said. “Sometimes the building has been delivered but they may not be able to get contractors or subcontractors to install equipment because of capacity constraints. The sheer amount of financing needed also continues to grow, creating a new set of challenges.  A big part of our value in the process is to resolve these issues for clients.”

Reddy added that Goldman works closely with hyperscalers to understand what kind of liquid cooling they will adopt. No one approach has come to the fore. It has been an exercise in monitoring a still nascent market, punctuated by the emergence of start-up AI users as less predictable data centre tenants.

“We’re seeing more neoclouds in the market that cater to intensive computing needs, which require liquid cooling to run efficiently. This is driving changes in the design and location of data centres,” he said.

Smaller PE firms needn’t have direct data centre exposure to glean insights about serving the demand of AI companies. One senior executive with a middle market GP said their firm was in communication with hyperscaler engineers and procurement professionals daily, describing the reward as “instant profitability.”

“You don’t have to bet on a company growing over time and becoming profitable and then valuable. Everything is accelerated and de-risked,” they said.

“Data centre operators don’t necessarily have discussions at the same technical level as the engineers of the architecture of the network. The important thing is not owning data centres – it’s dealing with the key engineers who are designing the whole system.”

Which standard?

Rapidly evolving performance requirements and customer preferences are precisely the reason power and cooling have emerged as the most popular verticals for private equity.

Hong Kong-based and mainland China-focused Albamen Capital is leaning into the power side of the equation. Its debut 2024-vintage CNY 7bn (USD 1bn) fund is now fully deployed, 75% in renewable energy and 25% in data centres. Fund II is being organised.

Jason Wong, a partner at the firm, observed that the strategy, although well supported by macro tailwinds, entails some nuanced equipment-level challenges.

Many older data centres, for example, are outfitted with converters to channel AC power from the grid into computers that run on DC power. When renewable energy is installed on site, DC power is provided, requiring a second converter. The outcome is electrical current on site going from DC to AC, then back to DC, resulting in significant power loss.

Meanwhile, the broader industry shift toward high-voltage DC power distribution in new build data centres has come to a crossroads. On one side, Nvidia is pressing for 800-volt architecture, while Open Compute Project (OCP) has advocated a +/-400 volt feed.

“There are two different standards and no conclusion which will win yet. So, there is risk,” said Wong, who has led the development of more than 15 data centres in a more than 30-year career. “Once that’s resolved, I think there will be big opportunities in that area.”

Cooling as a category likewise encompasses challenges and opportunities at the facility, power plant, and server levels. There is an emerging consensus that liquid cooling is the industry’s future, although opinions are divided on delivery methods. These range from relatively conventional heat sinks and cold plates to immersion cooling, which submerges whole computers in specialised fluid.

Traditional air conditioning remains in widespread use but is considered difficult to crack due to a clutch of global suppliers dominating the market for core equipment.

Actis indirectly accessed the segment last July by acquiring a 100% stake in Singapore-based Barghest Building Performance (BBP), an efficiency software provider for the Asian heating, ventilation and air conditioning (HVAC) industry.

BBP began operations in 2012 with a focus on shopping malls and small industrial sites. In the past four years, data centres and semiconductor manufacturers have grown to represent about 50% of revenue. The company is paid a share of the energy savings it achieves for customers, so the larger the energy usage, the greater the revenue. Semiconductors is the biggest earner.

For Actis, the transaction fit squarely into its energy efficiency agenda rather than as part of an AI infrastructure thesis. But AI-related demand for cooling efficiency services is hoped to support the valuation on exit.

“BBP is the only company I know of looking holistically at the entire plant,” said Asanka Rodrigo, a managing director for energy infrastructure at Actis.

“We aren’t optimising individual pieces of equipment as an OEM [original equipment manufacturer] would. We optimise energy in and refrigeration tonnes per unit of energy out in real time as required by customers. On average, we reduce customers’ cooling energy consumption by 20%-23%.”

Greater rack density is also the driver behind optical networking, which can accommodate a greater complexity of chip-to-chip connections than copper wire. The essential fibre optics technology was proved out in the 1990s by the telecoms industry but is now seen as rerouted on a new growth path with AI.

In addition to optical receivers and fibre cable, critical components include light-to-digital converters and various types of lasers. Celadon has made this space a priority.

Chipping away

Semiconductors is where the bar for innovation could be highest. New-age circuit boards that can replace traditional cable-based connections are projected to be high-margin products when commercialised.

EY’s Wang also sees a small but worthwhile opening for private equity in fabless designers of application specific integrated circuit (ASIC). “They get just 2%-3% market share, but the markets are big enough,” he said.

Sporadically, there will also be investment in direct challengers to the nearly monopolized GPU industry. These are the only chips that can practically handle AI workloads. Nvidia and its manufacturing partner Taiwan Semiconductor Manufacturing (TSMC) have a more than 90% market share.

Singapore’s Silicon Box claims to have shaken up this space by interconnecting tiny integrated circuits called chiplets into packaged systems. It has positioned itself as the second operator globally to provide advanced packaging of GPUs with the high bandwidth memory necessary to run generative AI.

The company raised a USD 202m Series B round in 2024 featuring BRV Capital, Event Horizon Capital, Hillhouse Investment, Prasedium Capital, Tata Electronics, TDK Ventures, and UMC Capital. It was reportedly courting investment from Indonesia Investment Authority as part of plans to raise an additional USD 150m as of last September.

Still, as the AI value chain opportunity set matures, private investment appears likely to be concentrated in lower risk plays, especially around established middle-market businesses leaning into the theme. Interestingly, corporate pivots in this vein could play out in the high risk-high reward end of the spectrum as well.

Archana Hingorani, a managing partner at Indian deep tech investor Siana Capital, said she is not yet seeing start-ups directly targeting data centres as a business proposition but that start-ups with agnostic efficiency solutions are catering more to data centres.

“We have a few portfolio companies that are now being approached by some operators about adapting their technology for data centres. For example, can you convert an efficient EV [electric vehicle] battery into a large energy storage device?” she said.

“As soon as you start getting those inbound requests, there will be me-too start-ups trying to set up those business models. That may take a few years, but it won’t take too long.”