Najnovšia správa Huawei - 10 kľúčových smerov k inteligentnému svetu do roku 2035

Huawei projects that annual global AI-token consumption will rise 100,000-fold by 2035, with autonomous agents accounting for more than 90% of usage, driving substantial demand for compute, networking, cloud, storage and data-center infrastructure. Its 2026 digital-intelligence report forecasts AI will generate more than $27 trillion in cumulative economic value over the next five years, while global investment in digital and intelligent infrastructure grows at a 19.14% CAGR to exceed $4 trillion by 2030. The company outlined 10 strategic directions spanning AGI, SuperPoD computing clusters, AI connectivity, chips, autonomous driving, data-center energy constraints, and agent security.

Analysis

This is directionally supportive for the AI infrastructure complex, but the company-sourced 2035 projections are not a near-term demand datapoint. The investable implication is that agentic workloads shift bottlenecks from training accelerators toward inference economics: high-bandwidth memory, networking, storage, power delivery and cooling. That favors SK Hynix, MU, ANET, VRT, ETN and CEG more than broad software proxies if enterprise deployment moves from pilot projects to persistent, latency-sensitive workloads over the next 6-18 months.

Huawei's framing also underscores a fragmentation risk that consensus US AI-infrastructure estimates may underweight. Export controls force China to optimize heterogeneous domestic compute, networking and model stacks; this can accelerate share gains for SMIC, CXMT and Chinese networking vendors while limiting the addressable China revenue pool for NVDA, AMD, AVGO and ANET. The offset is a potentially larger non-China capex cycle as hyperscalers and sovereign clouds duplicate capacity for resiliency, security and regulatory sovereignty.

Near term, there is no incremental, independently verifiable order or capex commitment and therefore no reason to chase affected names on this release. The critical 1-3 month evidence is hyperscaler capex guidance, HBM lead-time/pricing, 800G/1.6T optical order trends, and utility interconnection queues; sustained inference monetization—not token-volume forecasts—is what converts infrastructure spend into a durable 6-18 month earnings cycle. A material slowdown in cloud capex, falling HBM pricing, or evidence that model efficiency reduces compute per useful task would falsify the capacity-scarcity thesis.

Contrarian view: token growth need not translate one-for-one into silicon demand. Distillation, quantization, routing to smaller models and on-device inference could push value toward edge semiconductors and power-efficient connectivity rather than centralized GPU clusters. Qualcomm, MRVL and NXP are higher-beta beneficiaries only if commercial agent workloads require local, real-time operation; absent confirmed design wins, this remains a watchlist rather than a position.

Key Decisions for Investors

• Maintain a 6-12 month overweight in VRT and ETN versus the semiconductor basket: power and thermal constraints are harder to optimize away than model-token demand. Add on 8-12% pullbacks; reassess if hyperscaler aggregate capex guidance falls more than 10% or backlog conversion decelerates for two quarters.
• Pair long MU / short a broad AI-software ETF (IGV) over 3-6 months if HBM contract pricing remains firm: agentic inference increases memory intensity while software multiples remain exposed to monetization disappointment. Risk is rapid HBM supply expansion or enterprise AI seat-revenue acceleration.
• Use ANET as a conditional long only after confirmation of 800G/1.6T backlog acceleration or raised cloud capex guidance. The reward is operating leverage from network-intensity; the key risk is proprietary hyperscaler networking and China revenue restrictions.
• Do not initiate a directional NVDA or AMD trade from this item. Set an alert around China revenue commentary and export-control changes: tighter enforcement is negative to direct sales but could be offset by incremental sovereign and non-China capacity duplication.