News Summary
TREE NEWS reports: Morgan Stanley’s September 4, 2026 research report (via 潮向研究) projects that hyperscaler data center capital expenditures will reach $1.5 trillion in 2027, up 60% year-over-year, but growth will plunge to just 12% in 2028. Total AI compute capacity is set to expand from 35 GW in 2025 to 145 GW in 2028 — a 4x increase. Custom chips (Google TPU, Amazon Trainium) will rise from 34% to 66% of incremental capacity. GenAI investment returns range between 25% and 50%, with model-layer API running on own infrastructure yielding the highest (~46%). About 25% of S&P 500 companies have now quantified GenAI benefits, up from 14% a year ago.
Industry Analysis
The report signals a pivotal shift: the AI build-out is moving from the hardware-intensive training phase to the inference-led deployment phase. The sharp deceleration in capex growth after 2027 implies that the massive spending on GPUs and data centers will gradually give way to software and enablement layers. Morgan Stanley argues this will trigger capital rotation from hardware and semiconductors toward enablement and software companies — with Amazon, Meta, Google, and Microsoft positioned as key beneficiaries of earnings upgrades and valuation expansion.
Custom silicon dominance (TPU, Trainium) underscores a trend toward cost-efficient, workload-specific compute — a direct challenge to NVIDIA’s hegemony. For crypto markets, this has nuanced implications: decentralized compute networks (e.g., Render, Akash) may find less demand for general-purpose GPUs, but the growth of inference workloads could open niches for edge or specialized compute. Meanwhile, the report’s estimate of a $20-30 trillion addressable market for digitalized knowledge work and ~$30 trillion for consumer AI suggests enormous long-term demand, potentially boosting AI-related tokens and projects bridging AI with blockchain.
Forward-Looking Perspective
As capex growth normalizes, investors should watch for: (1) margin expansion in software layers as AI monetization matures; (2) increased corporate AI spending, projected at $800 billion by 2027 — faster than public cloud adoption; (3) power infrastructure plays (BE, INIO, SEI, WULF, HUT, CIFR, RIOT) benefiting from sustained energy demand even as chip spending slows. For crypto, the intersection of AI and decentralized infrastructure remains nascent but promising, especially with AI agents transacting on-chain and models being tokenized. The next 24 months will likely see capital shift from ‘picks and shovels’ to ‘application layers’ — a pattern familiar to crypto veterans.
