AI Architecture Map

Non-Consensus Sub-Sectors

1. Onsite / Behind-the-Meter Power

Hypothesis: Data centers will increasingly bring their own generation because utility interconnection timelines are too slow.

Companies to track:

• Bloom Energy: onsite fuel cells for data centers.
• Vistra: power generation, especially where large-load customers need firm power.
• Constellation Energy: nuclear and clean firm power exposure.
• GE Vernova: turbines, grid equipment, electrification.
• Solaris Energy Infrastructure: distributed power and electrical infrastructure for hyperscale sites.
• Crusoe: energy-linked AI infrastructure and compute.
• Lancium: power-aware data center development.
• CoreWeave: not a power supplier, but useful as a demand-side signal for power procurement behavior.

Automation signals:

• New power purchase agreements with hyperscalers.
• Data center deals measured in MW or GW.
• Language like "behind-the-meter", "onsite generation", "time-to-power", "grid interconnection", "firm power".
• FERC, ERCOT, PJM, MISO, CAISO, SPP filings involving large-load interconnection.

2. Electrical Equipment: Transformers, Switchgear, PDUs, Busways

Hypothesis: AI data centers may be bottlenecked less by GPUs than by boring electrical components with multi-year manufacturing lead times.

Companies to track:

• Eaton: power distribution, switchgear, UPS, electrical systems.
• Schneider Electric: data center power, cooling, electrical management.
• Vertiv: power and cooling infrastructure for high-density AI data centers.
• Siemens Energy / Siemens: grid and electrification equipment.
• Powell Industries: electrical equipment for industrial and utility-scale projects.
• Hammond Power Solutions: transformers.
• nVent Electric: electrical connection and protection systems.
• Atkore: electrical raceway and infrastructure.

Automation signals:

• Backlog growth, book-to-bill, lead-time commentary.
• Mentions of transformer shortages, switchgear shortages, busways, PDUs, UPS, medium-voltage gear.
• Capacity expansions in North America.
• Hyperscaler preferred-vendor announcements.

3. High-Density Cooling and Water-Constrained Infrastructure

Hypothesis: Rack density and liquid cooling adoption create new winners in thermal management, especially where water, permitting, or reliability constraints matter.

Companies to track:

• Vertiv: liquid cooling and thermal management.
• Schneider Electric: data center thermal and power systems.
• Modine: thermal management.
• Watts Water Technologies: flow control and water infrastructure.
• Johnson Controls: building systems, chillers, controls.
• Carrier: cooling systems.
• CoolIT Systems: liquid cooling.
• Submer: immersion and liquid cooling.

Automation signals:

• "liquid cooling", "direct-to-chip", "immersion cooling", "high-density rack", "water positive", "heat reuse".
• Partnerships with Nvidia, hyperscalers, colocation providers, and AI cloud providers.
• Cooling capacity constraints in new data center markets.

4. Flexible Load / Grid-Interactive Data Centers

Hypothesis: The next unlock may be data centers that can modulate load, arbitrage power, and coordinate with grid operators.

Companies to track:

• Emerald AI: flexible AI data center load coordination.
• Nvidia: reference architecture and ecosystem influence.
• AES: utility and renewable power partner.
• NextEra Energy: power generation and large-load partnerships.
• Invenergy: power development.
• Constellation Energy: firm power and grid partnerships.
• AutoGrid / Schneider Electric: demand response and grid software.

Automation signals:

• "flexible load", "demand response", "curtailment", "grid-interactive", "load shifting".
• Utility pilots with AI data centers.
• Regulatory language allowing fast interconnection for flexible or self-powered data centers.

Secondary Radar: If The Automation Expands

Storage / Memory Bottleneck

Why: AI infrastructure is becoming memory- and storage-hungry beyond HBM. Training and inference create pressure across HBM, DDR5/RDIMM, NAND SSDs, nearline HDDs, CXL memory expansion, storage controllers, and advanced packaging.

Core non-consensus angle: HBM leaders are already widely discussed; the next opportunity may sit in the "boring memory stack" that lets AI factories feed GPUs, host agents, keep context/state, and store generated data cheaply.

US / global listed companies: Micron, Western Digital, Seagate, SanDisk, Pure Storage, NetApp, Astera Labs, Rambus, Marvell, Broadcom.

A-share / China listed companies: Montage Technology, GigaDevice, Longsys, Biwin Storage, DML, Ingenic Semiconductor, Shenzhen Kaifa Technology, Xiangnong Core-Tech, Sino-IC Technology, JCET Group, Tongfu Microelectronics.

Signals: DDR5 server penetration, HBM4 supply, NAND shortage, nearline HDD exabyte growth, CXL memory expansion, memory pooling, SSD controller demand, storage controller ASP, advanced packaging for HBM, domestic DRAM/NAND substitution.

Watch terms: "CXL MXC", "memory pooling", "tiered memory", "nearline HDD", "enterprise SSD shortage", "HBM4", "SOCAMM", "DDR5 RCD", "MRDIMM", "PCIe retimer", "NAND allocation to AI servers".

First deep-dive candidates: Montage Technology, Micron, Seagate, Western Digital, Astera Labs, Rambus, Longsys, Biwin Storage, Shenzhen Kaifa Technology.

CPU / Host Compute Bottleneck

Why: AI servers still need CPUs for orchestration, networking, storage, preprocessing, agent runtime, virtualization, retrieval, and serving workloads. As rack-scale systems pair more GPUs with tightly coupled CPUs, the CPU becomes an AI infrastructure attach-rate trade rather than a legacy PC trade.

Core non-consensus angle: The market talks about GPUs first, but AI data centers may pull through a second wave of host CPUs, Arm server CPUs, x86 share shifts, DPUs, SmartNICs, CXL fabrics, and memory-interface silicon.

US / global listed companies: AMD, Intel, Arm Holdings, Nvidia, Marvell, Broadcom, Astera Labs, Rambus, Qualcomm.

A-share / China listed companies: Hygon Information Technology, Loongson Technology, China Greatwall Technology, Sugon, Inspur Electronic Information, Foxconn Industrial Internet, Unisplendour.

Signals: GPU-to-CPU attach ratio, Grace/Vera adoption, AMD EPYC server share, Intel Xeon recovery, Arm server share, AI server motherboard wins, DPU / SmartNIC adoption, PCIe 6.0 / CXL 3.x design wins, domestic CPU procurement.

Watch terms: "AI server CPU attach", "Grace CPU", "Vera CPU", "EPYC Turin", "Xeon Clearwater Forest", "Arm Neoverse", "CXL 3.1", "PCIe 6.0 retimer", "DPU", "SmartNIC", "host CPU bottleneck".

First deep-dive candidates: AMD, Arm Holdings, Nvidia Grace/Vera ecosystem, Astera Labs, Marvell, Hygon Information Technology, Sugon, Inspur Electronic Information, Montage Technology.

Agent Runtime Security

Why: If AI agents gain tool access and operate at machine speed, enterprise security moves from prompt filtering to runtime control, identity, permissions, and audit.

Companies: HiddenLayer, Lakera, Protect AI, Noma Security, Lasso Security, General Analysis, SentinelOne / Prompt Security, Ent, Cisco AI Defense, Palo Alto Networks, CrowdStrike.

Signals: Agent runtime visibility, AI security posture management, prompt injection incidents, tool-call control, data exfiltration, autonomous remediation, enterprise agent governance.

Inference Factory / Serving Efficiency

Why: As usage shifts from training to inference, cost per useful token, memory bandwidth, routing, batching, and reliability become bottlenecks.

Companies: Groq, Cerebras, Fireworks AI, Together AI, Baseten, Modal, DeepInfra, OpenRouter, Etched, SambaNova, AMD, Broadcom, Marvell, SK hynix, Micron.

Signals: Tokens per second, cost per million tokens, HBM supply, custom inference silicon, reserved capacity, enterprise inference SLAs, open-source model hosting.