The AI compute trade is undergoing a regime change, and the consensus framing is now too narrow to price it. As of April 2026, the standard “is the B200 better than the Ascend 950PR?” question has stopped being the right question. Compute is decoupling into two partially isolated national ecosystems — silicon, memory, networking, energy, software, industrial capacity, regulation and adoption all moving together — and the asset at stake is the standard, not the next quarter’s revenue.

The signals are now concrete enough to act on. NVIDIA closed FY26 with $215.9B in revenue ($193.7B Data Center, +68% YoY) and guided Q1 FY27 at $78.0B ±2% assuming zero China Data Center compute — the company’s own guidance now treats China as effectively foreclosed. The May 2025 $4.5B H20 charge crystallized that foreclosure. On the other side, Chinese domestic accelerator makers captured ~41% of China’s 2025 AI accelerator server market; Huawei’s Ascend 950PR has a 2026 shipment ambition of ~750,000 units; and DeepSeek’s V4 model is being optimized to run on Huawei silicon, with Chinese tech giants placing large advance orders. The software–hardware layer is aligning with the domestic stack.

What the market is underpricing is the durability of the two-stack world. Most sell-side coverage still treats Chinese self-sufficiency as a headline risk to NVIDIA’s TAM. The harder, more important question is what a sovereign Chinese compute ecosystem — capable of serving its domestic market and exporting an alternative to non-aligned buyers — does to global pricing power, hyperscaler ASIC roadmaps, HBM allocation, CoWoS bottlenecks and the value capture across every layer of the stack.

And the binding constraint is no longer silicon. It’s energy and packaging. Global data-center electricity demand is on track from 415 TWh in 2024 to 945 TWh by 2030 (IEA). Time-to-power at AI campuses is now thesis-defining; turbine OEMs (GE Vernova, Siemens Energy) are quoting backlogs into 2029. CoWoS capacity at TSMC is the silent constraint on NVIDIA’s growth. HBM3E/HBM4 contract economics are setting the ceiling on what every accelerator vendor can ship. Investors who track only the front-end silicon comparison miss where the marginal dollar of value is actually accruing.

This brief is calibrated for a specific kind of decision. It is not a buy/sell rating, not a price target, not a tactical positioning call. It is a decision-support architecture for allocators, board members, corporate strategists and infrastructure leaders who must price multi-year capacity, capex and supply-chain decisions across a stack that no single equity-research note covers end-to-end. The investor indicator dashboard ships with quantified thresholds — each confirmed crossing (two consecutive observation periods) maps to a documented ±3–5 percentage-point adjustment across the five 2026–2028 scenarios, including the new oversupply scenario (S5). The watchlist names tickers organized by stack layer so the framework is operational, not theoretical. The methodology appendix documents every probability assumption — so you can audit, challenge or override the desk’s judgement rather than inherit it.

The 90-day signals checklist alone is worth the price of admission. It names the events, the triggers that would move the thesis, and the analytical response the desk will execute — turning the next earnings season, BIS action or HBM contract print from noise into signal.