|> Scope note (2026-08-10): This page absorbed the former "Artificial Intelligence" theme, which tracked the same hyperscaler capex cycle and largely the same beneficiaries from the model/monetization side. Rather than maintain two near-duplicate pages, AI Infrastructure now covers the full stack — capex/hardware buildout through model deployment and monetization.

AI Infrastructure refers to the physical and software layers required to train and deploy large-scale artificial intelligence models — from compute (GPUs, ASICs), networking (InfiniBand, Ethernet), and storage (HBM, NAND) through the power systems (grids, data centers) and specialized electrical EPC engineering that support them, up to the models and applications running on top of that stack.

We are in the midst of an unprecedented capital expenditure expansion cycle. The global AI infrastructure buildout is backed by a massive forward contractual ledger. Post-Q2 2026 reporting, combined cloud/AI infrastructure backlogs across the hyperscalers and neocloud players have crossed $2.5T (tripling from ~$750B in Q3 2025). Rather than purely aspirational guidance, this represents take-or-pay long-term agreements, prepayments, and multi-year enterprise commitments that provide extraordinary visibility across the hardware and physical supply chains.

Key drivers:

• Contractual Purchase Commitments ($2T+): Big tech non-cancelable contractual obligations have surged (e.g. Alphabet's obligations up ~9x Y/Y to $811B), locking in multi-year GPU/accelerator, networking, power, and memory capacity.
• Hyperscaler Capex Step-Up: Annual data center capex spending is projected by CFRA to reach $1.3T in CY 2026 and $2.0T by CY 2028. NVIDIA's purchase order backlog across Blackwell and Vera Rubin platforms has crossed $1.0T through CY 2027.
• Next-Gen Accelerator Inflection (Vera Rubin & Custom XPUs): Inflection confirmed for 2H 2026 / Oct-Q as NVIDIA ramps the Vera Rubin platform (carrying a ~$3.5M–$4.0M price tag per NVL72 rack, +25% vs Blackwell; CFRA puts Vera Rubin content at ~$42B per gigawatt vs ~$25B for Grace Blackwell) and Broadcom scales custom hyperscaler XPU revenues to $56B in FY26 and >$100B in FY27. NVIDIA's FQ3 guide of ~$108B ±2% and preliminary ~70% FY2028 growth indication put a number on the ramp.
• Standalone AI CPU & Agentic Orchestration: Entry of purpose-built agentic AI CPUs (NVIDIA Vera targeting $20B in FY27 revenue within a $200B TAM) to handle token orchestration and tool use, shifting compute economics toward token-throughput efficiency.
• Ecosystem Financing Platforms: Introduction of non-balance-sheet AI infrastructure financing vehicles (NVIDIA's $500B MOU with 6 major financial institutions; Broadcom/Apollo/Blackstone AI XPU platform) de-risking balance sheets and accelerating adoption among neoclouds, enterprise, and sovereign AI projects.
• High-Bandwidth Memory (HBM) Deficit — now the sector's binding constraint: Extreme supply tightness in HBM3e/HBM4 and rising component costs are compressing hyperscaler lease margins, providing immense pricing power to Samsung, SK Hynix, and Micron Technology through 2027. NVIDIA's FQ2 FY2027 call made this explicit: memory supply, not customer demand, is what caps its FY2028 growth at ~70% rather than ~100%. The cycle's ceiling is now set upstream of the accelerator vendors.
• Physical, Electrical & Optical Bottlenecks: High-speed clusters require advanced liquid cooling architectures, complex high-voltage data center substation engineering (Vinci SA), and high-density Ethernet networking silicon (Broadcom Tomahawk 6 @ 100 Tbps).

• Compute: NVIDIA Corporation, Broadcom Inc., Marvell, AMD.
• Cloud Platforms: Microsoft Corp., Amazon.com Inc., Alphabet Inc., Oracle Corp., Meta Platforms Inc., CoreWeave Inc. (neocloud leasing).
• Memory: Samsung Electronics Co. Ltd. (HBM4/HBM4E leadership, DS margin ~70%), SK Hynix Inc., Micron Technology, Inc..
• Power, Cooling & Data Center EPC: Eaton Corporation plc, Schneider Electric SE, GE Vernova Inc., NextEra Energy, Inc., Vertiv Holdings Co, Vinci SA (high-density electrical & cooling engineering).
• Manufacturing Equipment: Applied Materials, Lam Research, KLA Corporation, ASML Holding N.V..
• Models & Applications: Microsoft Corp. (Copilot, Azure AI, Agent 365), Salesforce (Agentforce), ServiceNow Inc., Alphabet Inc., Meta Platforms Inc..

• Power Constraints: Grid capacity and data center power availability are becoming the primary bottlenecks for new deployments.
• Capex Fatigue / ROI Scrutiny: Risk that hyperscalers may slow spending if AI monetization does not meet expectations — CFRA notes a ~40% probability of an AI capex bubble pop within three years if monetization lags supply catch-up.
• Technological & Execution Risk: Complex rack-scale manufacturing transitions (e.g. Rubin Ultra 78-layer PCB midplane / Kyber NVL144 execution risk) and co-packaged optics scaling.
• Model Gap / Regulation: Meta AI acknowledged behind Gemini/GPT, though reaccelerating; AI safety and DMA/antitrust scrutiny remain overhangs on the applications layer.

• 2026-08-26: NVIDIA's FQ2 FY2027 print reset the capex debate. Revenue of $96.22B (+105.8% Y/Y) beat by $3.84B, Data Center hit $89.02B (+116.6% Y/Y), and the ACIE sub-segment — AI clouds, industrial and enterprise, i.e. the non-hyperscaler half — grew +138% Y/Y and +25% Q/Q. The material disclosure for this theme is the preliminary FY2028 revenue growth indication of ~70% against a ~45% consensus, given as explicitly supply-constrained: management stated customer demand implies closer to +100%, with memory supply the binding bottleneck. Supply and capacity commitments were raised to $279B. NVIDIA also disclosed guarantees to SB Energy Corp. covering a land, power and shell buildout of ~4.25 GW of IT load — the vendor-financing perimeter widening from GPUs into the physical estate. See NVIDIA Corporation.
• 2026-08-24: Broadcom detailed custom XPU roadmap with Meta (1 GW in 1H CY27 to 3 GW by CY28), Alphabet (TPUs through CY31), Anthropic (1 GW to +5 GW in CY27), and OpenAI (1.3 GW in CY27). Broadcom AI chip revenue projected at $56B in FY26 (+175%) and >$100B in FY27.
• 2026-08-20: Oracle Corp. reported Q4/FY26 RPO surging to $638B (+363% Y/Y) driven by massive AI infrastructure contracts (including the OpenAI Stargate partnership), with OCI cloud revenue up 39% to $34.0B.
• 2026-08-18: Ingested Q2 2026 updates highlighting the cloud backlog inflection to $2.5T and Alphabet's contractual purchase obligations reaching $811B.

• Semiconductors
• Electrification
• Infrastructure
• Cloud Computing
• Agentic Commerce
• Software & SaaS