6. Emerging Trends & Outlook

Edge computing, modular builds, the inference revolution, and whether we're in a bubble.

Edge Computing

Edge DC market: $14.7B in 2025, projected $39.8B by 2030 (17.5% CAGR). North America holds 42% market share.

• How it complements hyperscale: Edge handles latency-sensitive workloads (real-time inference, AR/VR, autonomous driving) while hyperscale handles training and batch processing
• 5G driver: Multi-access edge computing (MEC) market projected to reach $259.5B by 2034
• Autonomous vehicles: 14.11% CAGR through 2031, requiring sub-millisecond local processing
• CDN evolution: Moving beyond static caching toward dynamic AI-powered content optimization at the edge

Modular / Prefab Data Centers

Market projected to reach $67.5B by 2030. Up to 50% faster than traditional builds.

Key Developments

• Compass + Schneider Electric: $3B multi-year agreement. Prefabricated Modular EcoStruxure Pod deployed fall 2025. Revenue generation starts 3 months sooner
• Eaton + Flexnode: Turnkey AI factory infrastructure 3.5 to 35 MW per hall, 800 VDC power (January 2026)
• Modern modular: scalable, liquid cooling integrated, AI-optimized densities (30-150 kW/rack)

Sustainability

Water-Free Cooling

• Microsoft: Piloting zero-water evaporated cooling in Phoenix and Mt. Pleasant (2026). All new designs use this since August 2024
• Evolution Data Centres (Singapore): Fully air-cooled, closed-loop, no evaporative water
• Immersion cooling (LiquidStack, Iceotope): 90%+ cooling energy reduction

Waste Heat Reuse

• NTT DATA Berlin 2: Heating 1,000+ buildings (2 MW now, scaling to 37 MW)
• Microsoft Helsinki: World's largest waste heat recycling scheme
• Google Hamina (Finland): 80% of local district heating demand
• Germany: Mandates 10% waste heat utilization from July 2026 (15% in 2027, 20% in 2028)

Inference vs Training

The defining infrastructure shift of 2025-2026.

Inference-optimized chip market: $50B+ in 2026. The 2026 AI story is "inference at the edge, not just scale in the cloud." AI DC capex for 2026: $400-450B globally. See also AI Chips for hardware implications.

Quantum Computing Readiness

Quantum machines are beginning to sit alongside classical HPC clusters and AI accelerators.

Hybrid Quantum-Classical Architecture

• Layer 1 (Control): Quantum system controllers, real-time orchestration in hundreds of nanoseconds
• Layer 2 (Accelerators): CPU-GPU servers for calibrations and QEC decoding in microseconds
• Layer 3 (Applications): HPC clusters scheduling hybrid applications in milliseconds
• Demonstrated roundtrip latency under 4 microseconds, bandwidth exceeding 64 Gb/s

Industry Moves

• NVIDIA NVQLink (October 2025): architecture combining classical and quantum systems
• IBM (March 2026): new hybrid quantum-classical architecture with shared storage
• Infrastructure challenges: cryogenic environments, vibration isolation, ultra-low-latency fiber

Autonomous Operations

Google DeepMind Cooling

Deep neural networks fed by thousands of sensors every 5 minutes. Autonomously controls cooling tower speeds, chillers, fan speeds. Delivers 30-40% cooling energy savings.

Robotics

• Autonomous robots with thermal imaging: 15% reduction in cooling-related energy consumption
• Cable management robots: 40% reduction in cable-related downtime
• Microsoft: Researching modular robotics at automation Levels 2-3 (Level 4 = fully self-maintaining)
• Google DeepMind + Agile Robots: Gemini Robotics foundation models for DC operations

True "lights-out" DCs remain aspirational but progressing. Key barriers: robustness, fault tolerance, self-healing capabilities.

Interconnection Trends

• Submarine cables: 98%+ of international traffic. 50+ new cables 2024-2026. Hyperscalers now largest developers
• Dark fiber: Lead times ballooned to 60 weeks, prices surged 70%+ since 2021
• Metro fiber: DCI for gigawatt AI campuses. "Scaling across" links 100K+ GPU clusters across buildings
• High-density fiber: AI clusters need 10x more fiber. SN-MT/MMC connectors: 3x density of traditional MPO/MTP
• Hollow-core fiber: Metro-scale deployments between AI DCs targeted first

Space Constraints

• Multi-story: Common in Asia (Singapore: 1-2% vacancy). Emerging in US/EU but US trend shifting to slab-on-grade for AI weight/power density
• Underground: Growing trend — natural cooling, physical security, reduced footprint
• Repurposed facilities: Telecom buildings, industrial sites. Require assessment for AI suitability
• Bottom line: Power scarcity, not physical space, is the primary constraint

The "Data Center Bubble" Debate

Bull Case Overweight

• Current absorption shows no signs of overbuilding in the US
• Fundamentally different from 1990s: underpinned by long-term contracts with the world's most profitable companies
• Microsoft DC exec: "more worried that we are underbuilding rather than overbuilding"
• Physical constraint (power) naturally throttles unconstrained overbuilds

Bear Case Material Risk

• Mid-2025-2026 may be the inflection where supply catches up to inflated orders
• 2026 is the year massive debt-funded capacity comes online demanding tangible ROI
• Many enterprise AI projects stuck in pilot phase (McKinsey)
• Semiconductor innovation could strand some investments
• $4T cumulative CapEx by 2030 vs ~$2T revenue — the macro math doesn't balance yet

Dot-Com Parallels

1990s telecoms: $500B+ in fiber, financed with debt, on projections of 1000% annual growth. Key differences today: infrastructure backed by most profitable companies in history (not speculative telecoms), real revenue being generated, physical constraints throttle overbuild.

2030 Outlook

Key Themes for 2030

• Power constraints remain #1 bottleneck, not capital
• Early "scramble" gives way to a more disciplined, selective phase focused on capital efficiency
• Liquid cooling becomes standard (thermal management: 37% CAGR)
• Inference dominates as the primary workload category
• Edge and near-edge proliferate for latency-sensitive AI
• Sustainability mandates (especially EU) increasingly stringent
• Modular/prefab becomes default for new capacity due to speed-to-market
• AI accounts for 50%+ of total DC capacity by 2031
• Hyperscale operators: 61% of all capacity (up from 44% today)