We concur with Haueis and Colaço that metabolism should be treated as a central constraint on cognitive modelling. We also support their constructive proposal that metabolic considerations should delimit what is biologically feasible (evaluative role) and guide the development of new models (generative role), thereby establishing an iterative process in which metabolism shapes and constrains the space of possible cognitive accounts. However, their review primarily emphasises the cellular and synaptic scale and pays comparatively little attention to system-level principles that have given rise to some of the most influential theoretical frameworks in cognitive neuroscience. These include sparse coding, predictive coding, and, more broadly, anticipatory schemes such as active inference – all global computational strategies that reflect and impose constraints on local neuronal activity.

Sparse coding builds on the efficient coding ideas of Attneave and Barlow by formalising a trade-off between information and metabolic cost in the encoding of relevant stimuli (Attneave, Reference Attneave1954; Barlow, Reference Barlow1961). This principle enforces that only a limited number of neurons are active at any given time (Olshausen & Field, Reference Olshausen and Field1996). Sparse coding is thus framed as a generative, iterative modelling scheme in which representations are successively evaluated at each stage using metrics that penalise excessive activity – serving as a proxy for the metabolic cost of spiking – while preserving representational accuracy (Olshausen & Field, Reference Olshausen and Field1996). Beyond its success in accounting for the receptive fields of sensory neurons in multiple modalities (Lewicki, Reference Lewicki2002; Olshausen & Field, Reference Olshausen and Field1996; Perez-Orive et al., Reference Perez-Orive, Mazor, Turner, Cassenaer, Wilson and Laurent2002), sparse coding has been observed in response to ecologically relevant stimuli in sensory (Machens et al., Reference Machens, Gollisch, Kolesnikova and Herz2005; Vinje & Gallant, Reference Vinje and Gallant2000), memory, and associative systems, suggesting a domain-general organisational rule (Rolls & Treves, Reference Rolls and Treves2011; Tallman, Steinmetz, & Wixted, Reference Tallman, Steinmetz and Wixted2025; Waydo et al., Reference Waydo, Kraskov, Quian Quiroga, Fried and Koch2006). Sparse coding thus provides a clear illustration of how global constraints induce specific local patterns of neural activity, with direct implications for ATP consumption. More broadly, the empirical support for sparse coding indicates that system-level metabolic constraints directly shape cognitive representations, here giving rise to parsimonious selective codes that balance explanatory power against metabolic cost.

Another influential systems-level framework with direct metabolic implications is predictive coding, which is considered in the target article primarily from a local perspective. Predictive coding draws on Helmholtz’s (von Helmholtz, Reference Von Helmholtz1866) view of perception as inference – where sensory input is interpreted against prior expectations – and on engineering approaches to efficient communication such as linear predictive coding (Makhoul, Reference Makhoul1975), in which signals are predicted from past inputs and only residual errors, the difference between prediction and input, are transmitted. Although not originally formulated in metabolic terms, its core mechanism – the local, distributed computation of prediction errors – minimises redundant signalling and reduces the need for sustained or long-range neural activity, thereby potentially lowering metabolic costs within a global computing scheme (Friston, Reference Friston2010; Rao & Ballard, Reference Rao and Ballard1999), and suggesting a canonical organisational principle (Keller & Mrsic-Flogel, Reference Keller and Mrsic-Flogel2018). Supported by a substantial – albeit partial – body of empirical evidence (Bastos et al., Reference Bastos, Usrey, Adams, Mangun, Fries and Friston2012; Caucheteux et al., Reference Caucheteux, Gramfort and King2023) and central to many modelling efforts (Hodson, Mehta, & Smith, Reference Hodson, Mehta and Smith2024), predictive coding has become a guiding principle in contemporary models of perception and cognition, offering a framework in which improved representational efficiency is achieved alongside reduced metabolic demand.

Building on predictive coding, active inference, a prominent framework in modelling cognition (Hodson, Mehta, & Smith, Reference Hodson, Mehta and Smith2024), combines the advantage of prediction – which reduces unnecessary computations by encoding only deviations from expected sensory input – with a frugal inference engine that performs tractable, locally computable belief updates. Indeed, at its core, the free energy principle employs variational inference, replacing exact but intractable inference with a simplified optimisation based on principled approximations (Friston, Reference Friston2010). Crucially, this notion of tractability – formulated as a system-level computational requirement and embedded within the architecture of active inference models – translates into metabolic constraints on neural systems, ensuring that inference remains feasible under limited energetic resources. In this sense, the free energy principle constrains the space of viable cognitive architectures: by relying on variational inference, it restricts cognition to tractable approximations, thereby placing principled bounds on cognitive complexity under metabolic constraints.

The examples above illustrate how global principles impose structured bounds on local neural activity and, consequently, on metabolic expenditure. These system-level restrictions complement the cellular and synaptic factors emphasised by Haueis and Colaço, together delineating a hierarchical organisation of metabolic requirements governing the viability of cognitive systems. These considerations naturally extend to even higher-level ecological constraints on cognitive modelling, which encompass and contextualise metabolic ones. In particular, they point to the necessity of grounding cognitive models in embodiment: cognition must be understood, and modelled, in relation to the sensory, motor, and environmental contingencies under which it operates. This perspective is essential both for explaining biological cognition and for guiding the development of more efficient artificial systems. In this respect, proposals such as the embodied Turing test (Zador et al., Reference Zador, Escola, Richards, Ölveczky, Bengio, Boahen, Botvinick, Chklovskii, Churchland, Clopath, DiCarlo, Ganguli, Hawkins, Körding, Koulakov, LeCun, Lillicrap, Marblestone, Olshausen, Pouget, Savin, Sejnowski, Simoncelli, Solla, Sussillo, Tolias and Tsao2023) emphasise that models of intelligence should be evaluated not only in terms of internal performance, but in terms of their ability to operate under the physical, ecological, and energetic constraints that characterise real-world behaviour. In this hierarchy, neuronal metabolism should be understood as one crucial constraint among many that jointly determine viable cognitive architectures.