When Will We Achieve AGI? AI Leaders Now Say It’s a Fuzzy Target

AI industry leaders are reportedly shifting the definition of artificial general intelligence away from a measurable point at which AI outperforms humans on most tasks, toward a more subjective or “spiritual” concept. Separately, China’s DeepSeek plans to deploy at least 160,000 Huawei accelerators at a large Inner Mongolia data center, underscoring efforts to substitute domestic AI chips for Nvidia hardware amid technology restrictions.

Analysis

The investable issue is not whether a single AGI milestone is reached, but whether AI buyers continue to underwrite infrastructure spending against open-ended capability claims. A less measurable end-state raises the risk that enterprise procurement shifts from "race to build" toward ROI-gated deployments, which would first pressure inference-adjacent hardware utilization and then hyperscaler capex guidance over the next 2-4 quarters. NVDA's premium multiple remains most exposed to any evidence that model progress no longer requires proportional training-compute growth; AMD and custom-silicon suppliers such as AVGO face the same cycle risk, albeit from lower valuation starting points.

Chinese accelerator substitution is a more durable competitive threat than an immediate earnings event for NVDA. Large domestic clusters create a software, systems-integration, and procurement feedback loop that can improve Huawei's ecosystem despite inferior chip-level performance, gradually reducing the addressable sanctioned-China opportunity and weakening NVDA's strategic scarcity premium over 6-18 months. The second-order beneficiary is Chinese data-center buildout demand for power equipment, networking, memory, and cooling, though investable exposure for U.S. portfolios is indirect and constrained.

Consensus may overread domestic Chinese deployments as proof of near-term Nvidia displacement. Hardware availability, interconnect reliability, compiler maturity, power efficiency, and model-training yield determine effective compute; a nominal accelerator count is not comparable to H100/H200 capacity. The bearish NVDA thesis is falsified if management demonstrates sustained ex-China demand plus stable gross margin despite China revenue loss, or if cloud-provider capex revisions remain upward through the next two earnings cycles.

Key Decisions for Investors

• Maintain a cautious tactical underweight in NVDA versus SMH over the next 1-3 months rather than an outright structural short; use a close above the post-earnings high or another upward hyperscaler capex revision as a stop, since the primary risk is continued demand visibility overwhelming China concerns.
• Pair trade: long AVGO / short NVDA in equal dollar amounts for 3-6 months. AVGO has greater custom-ASIC exposure if buyers prioritize workload-specific ROI, while the short leg hedges broad AI-semiconductor beta; exit if NVDA's data-center gross-margin outlook improves while AVGO's AI revenue trajectory disappoints.
• Set an earnings-season alert on MSFT, META, GOOGL, AMZN, and ORCL: any aggregate reduction in 2027 AI capex or explicit shift from training clusters to monetization discipline is a catalyst to add semiconductor downside via SMH puts. Absent that evidence, do not press a broad AI hardware short.
• Avoid treating Chinese domestic accelerator deployment as a standalone NVDA sell signal. Upgrade the China-substitution risk only if independent benchmarks show materially narrowing training-cost or software-tooling gaps, or if NVDA discloses incremental China-related revenue or inventory pressure.