AI, CONVERGING TECHNOLOGIES (MANBRIC/BIOTECH) & ROBOTICS: The Capex Supercycle Faces the ROI Wall

The 2024–2026 AI cycle is the largest fixed-capital investment cycle in modern economic history. The market question is no longer “Will AI transform the economy?” but “Who bears the capex risk, and at what rate of return?” Humanoid robotics and converging technologies (MANBRIC: Metaverse-AI-Nano-Bio-Robotics-Info-Cognition, the extended NBIC family) are derivatives of this cycle: real bets on productivity that are still at the learning-curve stage.

Key data. – Data center capex in 2026: ~$850 billion (Morgan Stanley estimate; “we are only at 10–15% of projected demand”); projected to reach $1,500 billion in 2028. Goldman Sachs (supply-side model): ~$765 billion in 2026 and ~$7,600 billion cumulatively by 2031. Alphabet raised its 2026 forecast to $195–205 billion (August 2026). – Energy constraint: data center electricity consumption of 485 TWh (2025) → 950 TWh (2030) (IEA); demand from “AI” data centers +50% (total robotics market > $8.5 billion); verified deployments*: Figure 03 at BMW Spartanburg (1,250+ hours, 90,000+ parts, >99% [success rate?] by 2035), Morgan Stanley: $152 billion (2040), treat both figures with caution (4× gap = uncertainty around the adoption rate). – Biotech/convergence: the “bio” leg of MANBRIC is driven by generative AI applied to molecule discovery and sequencing (compressing R&D cycles); an early but growing venture-capital signal.

Quantitative framework. 1. Capex profitability test: to justify ~$850 billion/year, AI must eventually generate an annual incremental margin stream of around $250–400 billion (assumption: required ROIC of 12–15%, GPU depreciation over 4–5 years). As long as the “incremental AI revenue / capex” ratio remains below ~0.3, the cycle is driven by expectations, not cash. Monitoring this ratio quarterly is the only filter that matters. 2. Robotics unit economics: tipping point = total robot cost (amortized CAPEX of €130k over 5 years + OPEX) < fully loaded cost of a position operating two 8-hour shifts (€70–90k/year in Europe). At 2,000 documented productive hours/year (Figure/BMW), an 18–24-month payback is achievable for repetitive material-handling tasks; it has not yet been achieved for fine manipulation across multiple contexts. Adoption follows an S-curve, and we are at its lower inflection point: 2026 = proof of ROI; 2027–2028 = volume orders. 3. Energy as a limiting factor: +465 TWh by 2030 ≈ adding the electricity consumption of a country such as Japan. Any AI strategy without an energy component (PPAs, SMRs, grids) is incomplete, which explains the rationale behind our business idea No. 3.

Risks. Accelerated GPU depreciation (obsolescence before debt maturity); circular financing (suppliers financing customers); concentration risk: a handful of frontier labs (OpenAI, Anthropic) as the ultimate tenants of flagship projects; reliability of humanoid figures (massive gap between marketing claims and audited data, require documented operating hours); interest-rate risk: the cycle is financed at high long-term rates (highest levels since 2007).

Allocation implications. Play the cycle through cash-flow-generating picks and shovels (energy, cooling, networks, HBM chips) rather than growth-multiple applications; robotics: favor vendors with documented paying customers (Figure, Agility) over volume promises; in AI private credit: require DSCR covenants and collateral in real assets (not GPU goodwill), and stress-test at +300 bps.

KPIs to monitor. Incremental AI revenue/capex ratio of hyperscalers; documented humanoid operating hours; spot electricity prices at data center hubs (Virginia, Texas, Ireland); default rate of data center SPVs in private credit; average grid connection lead time (the real bottleneck).