Neurons already compute. The engineering problem is turning that biological fact into a system that is repeatable, addressable, maintainable and useful enough to beat an electronic alternative on a real workload.
Input and output bandwidth
Electrodes sample only part of a network. Three-dimensional tissue makes that access harder, and stimulation is coarse compared with the number of synapses inside the culture. Better interfaces are therefore as important as better tissue.
Reproducibility
Cell source, developmental state, network topology and culture conditions all change behaviour. A commercial processor needs enough standardisation that software written for one preparation transfers meaningfully to another.
Lifetime and state
Biological networks mature, adapt and eventually decline. CL1 advertises culture support for up to six months; FinalSpark reports organoid experiments lasting more than 100 days; 3D-MIND reports recording across six months. Long life helps, but it also means the substrate changes while it is being used.
Memory portability
A trained digital model can be copied. A learned biological network cannot currently be cloned with its acquired synaptic state. That makes backup, rollback, debugging and deployment fundamentally different.
Whole-system energy
Neural tissue is metabolically efficient, but the culture sits inside an engineered system containing temperature control, fluid handling, electronics and conventional compute. Claims about energy advantage need measurements at the boundary of the complete system.
Benchmarks
The field still lacks agreed benchmark suites that compare identical tasks, data and training budgets across biological and electronic systems. Better benchmarks may matter more than another dramatic game demonstration.