DCD>Connect | London 2026 took place at the Business Design Centre on September 16–17, with more than 4,000 attendees, 250+ speakers, and 120+ exhibitors across DCD's Connect, Compute, and Investment Forum programs. The audience spanned operators, hyperscalers, colocation providers, utilities, engineering firms, investors, and a large number of suppliers new to the sector.
DCD is a mission-critical infrastructure event, focused on power, cooling, software, connectivity, and financing rather than compute itself. Most vendors on the floor were there because data centers are growing. Very few showed much interest in why, or in how the compute trends behind that growth will change what customers need from them. That matters, because the industry's requirements are increasingly set by NVIDIA, AMD, Broadcom, and the server Original Equipment Manufacturers (OEMs), on product cycles far shorter than the life of a data center.
Three compute trends frame our takeaways. Rack densities are climbing fast, from roughly 120–130 Kilowatts (kW) today toward around 600 kW on NVIDIA's roadmap. Training and inference have different location needs: training follows cheap power north, and while Central European operators expect inference to bring demand back, European inference appears to be driven more by sovereignty than latency, which Nordic sites can satisfy at around a tenth of Belgian power costs, according to figures discussed at the event. And because no single site can secure enough power, large deployments are being split across multiple facilities that must operate as one system. Here is what that means for each layer of the infrastructure stack.
1. Power Supply: Planning for a Moving Target
Grid access dominated. Independent network operators Advanced Electricity Networks and Eclipse Power both ran sessions on securing grid connections for U.K. data centers, and a main-stage panel with Telehouse and the government's digital infrastructure directorate asked what comes next after Critical National Infrastructure (CNI) designation. These are useful conversations, though grid reform mainly improves how existing capacity is allocated, so availability will remain a key planning factor.
On-site generation was the most crowded theme. Clarke Energy and INNIO Jenbacher presented gas engines paired with battery storage as bridging power for data centers; Baker Hughes ran both a showcase and a panel on diversifying the energy mix alongside NTT Global Data Centers and Pure Data Centres Group; Rehlko's Chief Executive Officer (CEO) spoke on energy resilience; and Baudouin addressed backup generators for Artificial Intelligence (AI), highlighting lead times of around 6 months. Piller presented a new whitepaper on independent conditioned power, built around a joint venture with Bergen Engines and Marelli Motori.
The more interesting shift is in Uninterruptible Power Supply (UPS) and batteries, which are seeing a resurgence after years as a replacement-driven segment. Large training clusters create rapid, synchronized swings in power draw that stress generators, equipment, and the grid, so UPS and storage are increasingly specified to smooth these swings, as well as provide backup. DCD reflected this: a debate on which batteries and backup systems will power AI brought together CSB Energy Technology, Uptime Institute, GreenScale, and AtkinsRéalis. CSB ran a separate showcase on battery backup, C&D Technologies hosted a power lunch, and Eaton hosted another with Siemens Energy. The prominence of established Valve-Regulated Lead-Acid (VRLA) battery makers alongside lithium-ion and kinetic storage suggests the technology mix is far from settled. The open question for all of them is how UPS evolves as NVIDIA's move toward 800 Volt (V) Direct Current (DC) changes where energy storage sits in the power chain. That question got less airtime than generation, and it is where we would encourage vendors to focus.
2. Cooling: Liquid for the Core, a Second Life for Air at the Edge
Liquid cooling for AI halls is now mainstream, and the major vendors used DCD to reinforce their positions. Schneider Electric hosted an invitation-only briefing on AI data centers and liquid cooling; Trane sponsored the main-stage panel on building the AI data center of the future with QTS and Prime Data Centers; and Munters led a panel on the new age of cooling with Start Campus, Kao Data, and Edge Centres.
The more useful sessions were at the facility level. Grundfos made the case for moving from centralized to distributed pumping; BAC presented high-density heat rejection; GF Piping Systems spoke on flow design; and Rentaload examined how direct liquid cooling challenges traditional commissioning. That last point deserves attention: commissioning and operating liquid-cooled halls requires skills the industry is still building. RISE Research Institutes of Sweden also presented work on emulating compute-dense servers to test power and cooling, a practical response to hardware that changes faster than facilities.
Edge sites are a different story. With limited space, no facility water, few staff, and tight budgets, liquid cooling is hard to justify for a handful of racks, and ABI Research expects a resurgence of air cooling at the edge driven by price. Daikin's presence illustrated the continuing strength of air-side technology: the company showcased glycol-free air-cooled chillers with integrated free cooling, operating from –20°C to +55°C. Air-cooled accelerators such as NVIDIA's L40S and RTX PRO 6000 Server Edition can handle many inference workloads at densities air still supports. A debate on whether Europe should bet on central cloud or edge AI, with nLighten, EdgeNebula, CityFibre and Galaxy Data Centers, showed the edge case is still being argued. If much European inference follows cheap power north, fewer edge sites will need high densities, and the edge cooling market may prove real but mostly air-cooled and price-sensitive.
3. Software: From Monitoring to Coordination
Software at DCD centered on operations within the facility. RiT Tech's Chief Technology Officer (CTO) spoke on Agentic AI and the self-aware data center, presenting its framework for moving beyond conventional Data Center Infrastructure Management (DCIM), and hosted an operations briefing with State Street; MCIM sponsored a main-stage panel on AI's impact on data center operations featuring NVIDIA, VIRTUS, JLL, and star.london. Daikin also presented its Intelligent Data Center Manager, claiming energy savings of up to 20% over conventional chiller plant control. Cadence exhibited, representing the growing role of digital twins in testing facilities against new hardware before capital is committed.
While these are valuable tools, software for the problems the compute trends are creating was largely missing: coordinating workloads, power, and network capacity across multiple sites, and managing the interface between AI loads and the grid. NVIDIA is moving into orchestration through Run:ai, and startups such as Emerald AI are working on power-aware scheduling, but vendor-neutral coordination across facilities, operators, and utilities remains an open opportunity for independent software vendors.
4. Networking: Scale-Across Opens a New Market
Connectivity featured from the first session. Zayo Europe's CEO joined the opening keynote on aligning capital with capacity, and a debate on whether connectivity can keep up with AI demand brought together euNetworks, Sumitomo Electric Lightwave, and Equinix.
The exhibition floor showed the fiber supply chain moving quickly to position for AI. AFL focused on high-density connectivity for Graphics Processing Unit (GPU) environments; its DENALI modular fiber platform, launched in July 2025, delivers up to 576 fibers in 4 Rack Units (RUs) and supports speeds up to 800G and beyond. CommScope promoted fiber infrastructure for Generative Artificial Intelligence (Gen AI) and high-speed networks, and Ekinops brought its optical transport and data center interconnect portfolio to the show. Together, they illustrate two distinct opportunities: dense fiber inside the AI hall, and high-capacity optical links between sites.
While there is a distinct opportunity for third-party vendors when it comes to scale-across, scale-up and scale-out (within the rack and across the cluster) are largely delivered by the major semiconductor companies through server OEMs such as Dell, HPE, and Supermicro. Scale-across, linking separate data centers into one compute fabric, is different. Driven by power constraints, it covers distributed training, model replication, data movement, and resilience across sites. NVIDIA's Spectrum-XGS and Broadcom's Jericho4 target it, but the wider opportunity sits with optical transport and Data Center Interconnect (DCI) vendors such as Ekinops, Ciena, Nokia, and Marvell, fiber and connectivity suppliers such as AFL, CommScope, and Sumitomo Electric Lightwave, and fiber operators such as euNetworks, Zayo Europe, and GlobalConnect. If capacity settles in the Nordics and Baltics, sovereign long-haul routes to central Europe become strategic infrastructure. Network planning still tends to follow site decisions; the fiber and optical companies at DCD are well placed to engage earlier and help shape them.
5. Building a Clearer View of Demand
The vendors at DCD brought strong products and deep engineering expertise. The opportunity is to connect that expertise to the compute trends shaping customer decisions: how AI loads behave on the power system, where inference will land, what edge sites will realistically need, and how distributed deployments will be coordinated and connected. Vendors don't need to become compute experts, but following chip roadmaps and tracking where hyperscalers and neoclouds such as CoreWeave, Nscale, Nebius, Lambda, Vultr, and TensorWave place capacity will be increasingly valuable.
These questions extend well beyond the United States and Europe. Grid access is constraining growth in established hubs from Northern Virginia to Frankfurt, Dublin, and Singapore, and capacity is moving to wherever power can be secured, whether that is the Nordics, the Baltics, the U.S. interior, or the Middle East. The same compute trends are reshaping power chains, cooling strategies, and network architectures in every market.
Over the next year, we will watch whether grid reforms in major markets translate into faster connections, how 800 V DC reshapes the UPS and battery market, whether sovereignty requirements pull inference into national markets or allow it to follow cheap power across borders, how quickly multi-site coordination software matures, and whether operators worldwide start buying scale-across capacity as a strategic commitment. The data center opportunity at DCD Connect | London 2026 was clear to everyone. The vendors that best understand where demand is heading will capture it.