# The AI boom has a bottleneck problem, and that could spell opportunity > Source: > Published: 2026-08-18 20:45:11+00:00 # The AI boom has a bottleneck problem, and that could spell opportunity **AI demand is outrunning infrastructure****Power, heat and data are bottlenecks****These ASX small caps are targeting solutions** For the first stage of the artificial intelligence boom, the investment story seemed fairly simple: the world needed more computing power, so back the companies[ making the most powerful chips](https://finance.yahoo.com/technology/ai/articles/jensen-huang-says-memory-now-042000029.html). That is still part of the story. But buying thousands of GPUs is one thing. Getting them to work together all day, every day, is another. Each chip needs electricity, produces enormous amounts of heat and must constantly exchange information with other chips. Data also needs to reach the chips quickly, all inside a data centre that can take years to approve, build and connect to the grid. Morgan Stanley US thematic strategist Michelle Weaver recently described computing capacity as “a constrained resource”. She said shortages of power and skilled workers, along with political resistance to new data centres, could limit supply for the next few years. Microsoft can already see that shortage in its own business. In its Q3 FY26 earnings call, CFO Amy Hood said demand for its Azure cloud services continued to exceed supply. The next stage of the boom may therefore belong not only to the chipmakers, but also to the companies clearing the roadblocks around them. ## First, the chips need electricity Every AI project eventually runs into the same basic question: where will the power come from? Without electricity, even the most advanced GPU is an expensive piece of metal and silicon. The International Energy Agency (IEA) said electricity use by data centres rose 17% in 2025. At AI-focused facilities, it jumped 50%. The IEA expects global data centre electricity demand to more than double to about 945 terawatt-hours by 2030, slightly more than the whole country of Japan uses today. Producing that much power is only part of the challenge. New power stations, transmission lines, substations and grid connections can take years to deliver. Developers may have their land, funding and chips ready, but still be left waiting for electricity. One ASX company targeting the power shortage more directly is [ 1414 Degrees (ASX:14D)](https://stockhead.com.au/company/1414-degrees-14d/). Its proposed [Aurora Energy Precinct in South Australia](https://stockhead.com.au/tech/1414-degrees-locks-in-ai-data-centre-partnership-path-at-aurora/) could host up to 1GW of AI data centre infrastructure, supported by renewable generation and battery storage. But it remains an early stage development, with due diligence, approvals and definitive agreements still required. ## The ASX contractors plugging AI in For investors looking for work already under way, the more direct small cap exposure may be the contractors connecting the data centres now being built. [ SKS Technologies (ASX:SKS)](https://stockhead.com.au/company/sks-technologies-sks/) installs the electrical and communications systems inside hyperscale data centres. Its data centre revenue reached $140.7m in FY25, while one major contract has since been expanded to $210m. [ Southern Cross Electrical Engineering (ASX:SXE)](https://stockhead.com.au/company/southern-cross-electrical-engineering-sxe/) provides electrical systems, switchboards, communications, fire protection and security for data centres. It generated about $120m of data centre turnover in FY25, and said in February it was bidding for more than $1bn of additional work. Neither company generates electricity or adds capacity to the grid. Once power is available, however, they install the systems needed to distribute it throughout the data centre. ## Then, someone has to build the data centre itself Even with power secured, a large data centre cannot appear overnight. Traditional facilities are huge, custom built projects. Customers may want capacity within months, but approvals, labour shortages and construction can stretch across years. [ DXN (ASX:DXN)](https://stockhead.com.au/company/dxn-dxn/) is trying to shorten that wait. It builds modular data centres in a factory and transports them to site. DXN says it has delivered nearly 100 modular facilities. In June, it secured an $8.8m contract to build a 1.36MW modular AI data centre for a US-based neo-cloud operator. ## Once it is running, heat is the next hurdle Switch on a powerful AI chip and most of the electricity entering it eventually becomes heat. Place dozens together and that becomes a major engineering problem. Nvidia’s GB200 NVL72, for instance, connects 72 Blackwell GPUs inside a liquid-cooled rack using about 120 kilowatts. If heat is not removed quickly enough, the chips slow down. [ GCM Corporation (ASX:GCM)](https://stockhead.com.au/company/gcm-corporation-gcm/) is trying to move that heat away from the chip more effectively. Its Very High Density, or VHD, graphite can be shaped into heat sinks, heat spreaders and cold plates[ that carry heat away from the chip.](https://stockhead.com.au/tech/gcm-brings-graphite-tech-to-the-fight-against-ais-heat-problem/) In like-for-like testing by UNSW, reported by GCM, VHD moved heat along its surface 1.3 times as effectively as copper and three times as effectively as aluminium. It also reacted to changing heat loads faster than both metals. ## The chips also need to talk to each other An AI cluster is not just a collection of processors. It has to behave like one enormous computer. Thousands of GPUs constantly exchange parts of the model and their calculations. If the memory and network cannot keep up, those GPUs sit idle. But another traffic jam is developing inside the chips. The tiny copper wires that connect billions of transistors (called the interconnects) are becoming narrower as chips shrink. At very small sizes, copper becomes less efficient. Electrical resistance rises, signals slow down and more heat is produced. [ Adisyn (ASX:AI1)](https://stockhead.com.au/company/adisyn-ai1/), through subsidiary 2D Generation, is developing a low-temperature process that grows graphene directly on semiconductor surfaces using equipment (ALD) already common in chip factories. Graphene could help tiny chip interconnects carry signals more efficiently while producing less heat. Adisyn has now scaled the process[ from small copper samples to a full 200mm wafer](https://stockhead.com.au/tech/adisyns-graphene-tech-makes-a-200mm-leap-as-us-patent-moat-widens/), with independent testing detecting graphene at all 12 sampled locations. It has also received a second US patent allowance. But the company must still prove the process is repeatable before progressing to 300mm wafers used in advanced chips. ## Getting data to the AI is another traffic jam Even a perfectly powered and cooled AI system is not useful if the right information cannot reach it quickly. That matters even more as the industry moves from chatbots to AI agents. A chatbot waits for a question and gives an answer. An agent can keep working and taking actions across several systems. That constant flow of information must be checked and filtered before reaching the model. [ Fortifai (ASX:FTI)](https://stockhead.com.au/company/fortifai-fti/) is developing its Nol8 AI Data Plane to manage that traffic. Using specialised FPGA hardware, Nol8 processes and filters information while it is moving, reducing the amount of unnecessary data sent on for AI workloads. In company-run [testing across Megaport locations](https://stockhead.com.au/tech/fortifais-nol8-proves-its-ai-mettle-across-megaports-global-network/) in Virginia, Tokyo and Sydney, FortifAI said Nol8 processed 34 terabytes of governed data a day – compared with 2.6 terabytes for its CPU-based benchmark. It also removed up to 64% of unwanted data, potentially allowing GPUs to spend more time on useful processing. ## Or avoid the data centre altogether There is one more way to reduce pressure on AI infrastructure: stop sending every task to the cloud. Edge AI lets a drone, camera, sensor or wearable process information locally – reducing network traffic and delivering a faster response. [ Nanoveu (ASX:NVU)](https://stockhead.com.au/company/nanoveu-nvu/) owns EMASS, developer of the low-power ECS-DoT edge-AI chip. Its existing 22nm chip is designed for always-on AI tasks, while a more capable 16nm version entered fabrication at TSMC in January. Nanoveu has demonstrated voice recognition using less than one milliwatt of computing power. In company-run drone tests, it reported an average 27.2% improvement in cruise energy efficiency, with peak gains of up to 51% on more complex routes. The key commercial test is whether manufacturers adopt the chip and place repeat volume orders. *At Stockhead we tell it like it is. While 1414 Degrees, GCM Corporation, Adisyn, and FortifAI are Stockhead advertisers, they did not sponsor this article.* *This story does not constitute financial product advice. You should consider obtaining independent advice before making any financial decision.* ## Related Topics ### UNLOCK INSIGHTS Discover the untold stories of emerging ASX stocks. Daily news and expert analysis, it's free to subscribe. By proceeding, you confirm you understand that we handle personal information in accordance with our [Privacy Policy](https://stockhead.com.au/privacy-policy/).