DCD Connect | London 2026 Takeaway: Where’s the Compute Roadmap in Data Center Infrastructure Planning?

At DCD Connect | London 2026, more than 5,000 senior leaders and 150-plus exhibitors filled the Business Design Centre for 2 days dominated by the mission-critical stack. Power led the agenda, with Equinix, Bulk Infrastructure, Digital Realty, and LCL debating whether data centers are becoming Europe's new energy partners as utilities move from connection provider to gatekeeper, and demand-response markets and on-site microgrids begin to reshape how capacity is planned and traded. The public costs of the build-out got real stage time, too. Schneider Electric sponsored the net-zero reality check panel with Heim Datacenter, Tillion and the Net Zero Innovation Hub, confronting mandatory Public Usage Effectiveness (PUE), water, energy, and waste-heat reporting under the tightened European Union (EU) Energy Efficiency Directive and asking whether climate goals survive Gigawatt (GW)-scale expansion. STACK Infrastructure, Scandinavian Data Centers, Apto, and Google mapped the build-out beyond Frankfurt, London, Amsterdam, Paris, and Dublin (FLAP-D) into the Nordics, the Mediterranean, and Eastern Europe. Johnson Controls, Iceotope, and ZincFive featured in the technology showcase, and Ekinops exhibited for the first time.

What barely featured anywhere on the agenda was compute, even though the decisions that set the requirements for every vendor in the room had already been taken upstream: roughly 120–130 Kilowatts (kW) per rack today moving toward around 600 kW later in the decade, and a shift to 800 Volt (V) Direct Current (DC) distribution inside the facility.

That absence is the gap that matters, and it is the one the event does the least to close. On energy, by contrast, Data Center Dynamics (DCD) bridges it comprehensively. The industry now understands its own bottleneck in considerable detail, down to which grids can deliver, on what timeline, under which allocation model, and what regulators will demand in return. That is genuine progress, and it was hard won. But energy is the constraint everyone can see from where they already stand, whereas the compute roadmap runs the other way, from the top of the stack down. The value chain runs one way, from Graphics Processing Unit (GPU) to rack to facility, and the suppliers at DCD sit at the end of it, planning on 15-to-20-year horizons against a compute cycle that turns over every 1 to 2 years.

The roadmap they cannot see is already redistributing value within their markets rather than simply expanding them. The shift to 800 V DC distribution pushes conversion out of the facility and into the rack, changing the mix of switchgear, busway, and protection, and leaving suppliers built around Alternating Current (AC) architectures at risk of being specified out of reference architectures they never saw. Uninterruptible Power Supply (UPS) vendors face growth and displacement at once, as the synchronized load swings of Artificial Intelligence (AI) clusters turn the UPS into a load-shaping device while NVIDIA moves energy storage into rack power supplies. In thermal, the binding constraint has migrated from the rack to the facility side, which is why hydraulic design and pumping strategy have become main-stage topics; in software, computer vendors are addressing multi-site orchestration before the Data Center Infrastructure Management (DCIM) incumbents have a product; and in networking, scale-up and scale-out have been absorbed by the semiconductor vendors, leaving scale-across as the layer where optical and fiber suppliers retain real room to compete.

The commercial implication is uncomfortable for a market that has spent 2 years rebuilding itself around liquid. A growing share of new cooling demand will be air-cooled, spread across many small sites, price-sensitive, and operationally simple, leaving vendors optimized entirely for hyperscale liquid well positioned for the phase that is ending and considerably less so for the one that follows. None of this can be inferred from customer requirements, because customers are specifying today's training halls. It can only be read from the compute side, from which accelerators are moving in volume, at what power envelopes, for which workloads, and into which markets. That information exists, but it sits with the semiconductor vendors and their Original Equipment Manufacturer (OEM) partners, and currently reaches only the narrow group of suppliers closest to them.

For infrastructure innovators, the practical response is to treat the compute roadmap as a planning input of equal standing to the customer requirement, to build portfolios that serve both phases of the AI lifecycle rather than only the visible one, and to buy visibility by getting into the upstream ecosystem, because value in this chain is allocated through reference designs and partner programs long before it reaches procurement. The most defensible positions will be in the layers where the compute vendors are least likely to absorb coordination across facilities and utilities, grid-interface software, inter-site interconnect, and the deployment and operating skills that liquid cooling, DC architectures and load-shaping storage all outrun.

For the semiconductor vendors, the obligation runs the other way. Rack-scale systems cannot be deployed in facilities that cannot power or cool them, and delays at the facility end eventually surface as deferred revenue upstream, which makes educating the broader thermal, electrical, and fiber base a commercial safeguard rather than a courtesy. Published power and thermal envelopes and inference-class reference designs, not just training-class ones, would do more for deployment velocity than another partner announcement.

The infrastructure industry is not short on engineering capability. What it lacks is visibility into the part of the value chain that now writes its requirements, and closing that gap is what separates anticipating demand from being handed a specification once the decisions have already been made.