Invert the headline figure and it becomes easier to hold: 61.87 attojoules per operation is about 1.62 x 10^16 operations per joule, roughly 16 peta-ops per watt measured at system level [5][1]. How much that means depends on what counts as an operation, and the abstract does not say [11].
The retention spec is the part that is hard to get any other way. Gigahertz short-term dynamics alongside 7.45 seconds of long-term retention puts about ten orders of magnitude between the two timescales: one cycle of a 1 GHz process lasts a nanosecond, and 7.45 seconds is 7.45 billion of them [3][2]. The authors' own framing is that the architecture exists to attack the tradeoffs among speed, efficiency and scalability [8].
The construction described is a coupling of photonic and electronic mechanisms into reconfigurable neurons that carry both memory timescales [1]. The word "fully" is doing specific work there. It does not claim photons alone store everything; it claims the external digital memory that normally holds sensory inputs and intermediate activations, which the authors call throughput-limited and efficiency-constrained, is out of the loop [7].
What the drone result establishes is narrower than the number suggests. The abstract reports sensing, exploration and adaptation for autonomous racing-drone navigation [4] without saying whether a physical aircraft flew [12], and it does not state how the ten-core multilayer system relates to the 7,378-neuron chip, nor how many neurons sit in each core [13][2]. The full text costs $39.95, or $119 for a year of the journal [6], so for most readers the checkable record is one paragraph.
The reference list is candid about the setting. It cites Horowitz's 2014 accounting of computing's energy problem [10], which is where the case for in-memory anything starts: shifting data costs more than operating on it. It also cites a 2026 Nature Photonics commentary on the thin line optical neural networks must walk toward broad practical relevance [9]. A group that cites that commentary and then reports a navigation task with adaptation in it is answering the charge rather than dodging it.
So the attojoule figure is internally consistent rather than comparable. Nobody outside the author list can yet multiply it by an operation count and get the energy of one flight, which is the quantity a system designer would budget against. That is a limit of what has been published, not necessarily of the hardware.
