Organoid intelligence (OI), which uses living human neurons cultured on multi-electrode arrays, has been proposed as an energy-efficient alternative to silicon-based computation. This scoping review examines the ecological and financial sustainability of OI across three dimensions: sample efficiency, energy consumption, and financial viability. Seventeen sources — 10 primary (7 peer-reviewed journal articles and 3 preprints) and 7 secondary — published between 2022 and 2025 are synthesised using the PRISMA-ScR framework. The review finds that biological cultures converge on simple learning tasks in far fewer training episodes than current deep-reinforcement-learning algorithms (Kagan et al. , 2024), though the headline "157× fewer episodes" figure compares two differently-shaped data series under asymmetric input conditions and is better read as a convergence-time result than a matched-input sample-efficiency comparison. Cloud access to the CL1 platform at 300 per week is 2–30× cheaper than equivalent GPU-cloud access depending on provider and commitment tier, not the 17× figure that appears in headline comparisons against on-demand hyperscaler pricing. The paper's original contribution is a first-order lifecycle energy estimate. Combining Cortical Labs' reported rack power draw (850–1, 000 W for a 30-unit rack) with a scaled-incubator estimate of life-support overhead (15–25 W per unit), total system draw is estimated at 1, 450–1, 600 W, a 60–71% increase over the headline figure. A sensitivity analysis identifies the thresholds at which this overhead would eliminate the biological advantage for small deployments. An independent lifecycle assessment remains outstanding. The Philippines, with its biomedical workforce, DOST grant infrastructure, and electricity rates among the highest in ASEAN, offers a concrete case for what thoughtful early engagement with biological computing looks like in a resource-constrained research economy. The review concludes that OI is a complementary technology suited to specialised niches (drug screening, neurological modelling, narrow sample-efficient learning) rather than a general-purpose replacement for silicon, and that governance frameworks for frontier AI, which currently assume silicon compute thresholds, will need to address biological substrates as the field matures.
Joaquin Ross (Tue,) studied this question.