Huawei's latest report proposes 10 key directions to Intelligent World 2035

Huawei’s Intelligent World 2035 report forecasts global annual AI-token consumption will rise 100,000-fold by 2035, with autonomous-agent traffic accounting for more than 90% of total traffic. The company identifies infrastructure priorities including 100x larger computing clusters, 1,000x lower agent-task costs, SuperPoDs, intelligent connectivity, agent-native cloud platforms, and data-center power and thermal upgrades. Separately, Huawei and Tsinghua University project AI will generate more than $27 trillion of cumulative economic value over the next five years, while global digital and intelligent-infrastructure investment grows at a 19.14% CAGR to exceed $4 trillion by 2030.

Analysis

This is a vendor-led long-range framework rather than a measurable order, capex, or product event; it does not independently validate near-term revenue for listed AI infrastructure names. The useful signal is directional: agent workloads shift the bottleneck from training FLOPS toward inference throughput, network latency, memory bandwidth, power delivery, and cooling. That mix favors NVDA and AVGO at the silicon/interconnect layer, ANET and CIEN in high-speed fabrics, VRT/ETN/TT in data-center electrical and thermal infrastructure, and MU where memory-content growth persists.

The second-order implication is that token-cost compression could be more important than aggregate token-volume forecasts. If model, hardware, and systems optimization reduces cost faster than usage expands, hyperscaler capex may migrate from general GPU fleets toward proprietary accelerators and networking; this is relatively constructive for AVGO, MRVL, ANET and optical suppliers, while raising valuation risk for companies priced solely on scarce-GPU duration. Enterprise agent adoption also increases identity, observability, and machine-to-machine authorization spend, supporting PANW, CRWD, OKTA and NET only if agent deployments move from pilots into production workloads.

Over the next 1-3 months, the actionable catalyst remains hyperscaler capex guidance and evidence of power-constrained data-center delivery, not industry vision documents. Over 6-18 months, grid interconnection delays and transformer/switchgear availability can cap realized AI buildouts despite robust chip demand, making VRT, ETN and PWR more direct beneficiaries of the physical constraint. Falsify the infrastructure thesis if MSFT, AMZN, GOOGL and META collectively guide 2027 capex lower, lead times normalize materially, or inference unit economics improve without a corresponding increase in deployed capacity.

Key Decisions for Investors

• No event-driven trade on this release; treat it as a watch item rather than evidence of Huawei-linked revenue acceleration, particularly given limited transparent disclosure and geopolitical export-control constraints.
• Maintain a 6-12 month barbell: long VRT and ETN versus a short basket of high-multiple AI software names with limited production-agent revenue disclosure. The physical-power bottleneck has more durable pricing power than unproven application monetization; reassess if data-center backlog conversion or hyperscaler capex guidance weakens.
• For 1-3 month positioning around earnings, prefer long ANET versus short a broad semiconductor ETF (SMH) in modest size. AI cluster scale raises network-content intensity, while the pair reduces market-beta exposure; stop out if ANET backlog/order commentary indicates cloud digestion or if SMH outperforms by 10% after a negative networking read-through.
• Add MU only on evidence that HBM allocation and pricing remain tight through the next two quarters; memory is the cleanest beneficiary of agent inference memory intensity but carries the highest cycle risk. Avoid initiating if management signals conventional DRAM supply additions are catching up faster than HBM demand.
• Monitor PWR and utility/grid-capex indicators as a 6-18 month constraint trade. A sustained rise in interconnection queues and large-load contracts would support long PWR/ETN; falling transformer lead times or regulatory rejection of data-center load connections would invalidate the thesis.