Research analysis · Neuromorphic computing

A rigorous monograph maps the neuromorphic turn, and what it quietly asks of organoid intelligence

Volker Kempe's 88-page monograph, an English translation of his October 2026 German book, walks from the transformer architecture inside ChatGPT through biological neurons to spiking neural networks, the Neural Engineering Framework, and neuromorphic hardware from Intel's Loihi 2 to commercial edge processors. It contains no new experiments. It is nonetheless one of the more useful documents an organoid-intelligence researcher can read this year, because it states, carefully and without hype, exactly which biological tricks engineering has already absorbed, and at what power.

Source: AI at a Turning Point: From ChatGPT through Neural Networks to Neuromorphic AI, arXiv:2610.39796, 29 October 2026. CC BY-NC-ND 4.0; translation of the German edition, DOI 10.5281/zenodo.23278918. Primary source. Read: full PDF, all 88 pages; every number quoted here verified inline.

What the work claims

State the claim precisely, because the genre matters. This is a synthesis and expository monograph, not a primary result: every mechanism it describes is established science or shipped hardware, and the author's contribution is arrangement, rigor, and a point of view. The point of view is that spiking neural networks (SNNs) are not a replacement for connectionist AI but an extension of it, best deployed in hybrid systems where each approach contributes its strengths. The organizing tension is energy. Training and running large AI models consumes gigawatt- to terawatt-hours of electricity, while the human brain runs on a continuous 20 watts; SNNs attack that gap with event-driven computation, local learning rules instead of global backpropagation, and time itself as a carrier of information1.

The monograph is unusually honest about the limits of its own framing. The 20-watt juxtaposition is presented with an explicit caveat that it is "no scientific efficiency comparison", since biological and artificial systems have entirely different tasks, structures, and capabilities. That sentence is worth more than most numbers in the energy debate, and I will return to it.

How it works

The middle of the book is a technically serious tour. The transformer chapter derives self-attention and the training cost of large language models. The biology chapter covers the neuron, action-potential propagation, and spike-timing-dependent plasticity (STDP): a local learning rule in which the millisecond-scale ordering of pre- and postsynaptic spikes strengthens or weakens a synapse, letting networks extract causal temporal structure without any global error signal. The SNN chapter surveys rate, population, and temporal coding, learning methods that work around the discontinuity of spikes, and the Neural Engineering Framework (NEF), which treats neural populations as representational substrates whose dynamics can be specified analytically and then compiled onto hardware. The genre's founding classification, Wolfgang Maass's labeling of spiking networks as the third generation of neural network models, frames the whole chapter12.

The hardware chapter is where the numbers live. Intel's Loihi 2, fabricated in the Intel 4 process with roughly 2.3 billion transistors on about 31 square millimetres, offers 128 neuromorphic cores with 8,192 neurons each, up to about one million logical neurons, graded spikes that carry a numeric payload in addition to their timing, and programmable local learning rules built into the cores. The energy argument is architectural: synaptic weights live next to the computation that uses them, so no energy is spent shuttling weights from a distant memory, and work scales with the event rate rather than with the size of the network1.

Then the commercial edge, which is the part the organoid field should read most carefully. SynSense's Speck chip pairs an event-driven vision sensor with 328,000 spiking integrate-and-fire neurons in the sub-milliwatt range; its Xylo line processes weak sensor signals at a few hundred microwatts; Innatera's Pulsar, an SNN engine plus RISC-V microcontroller for always-on edge inference, is reported in the microwatt range. The monograph's own conclusion from this tier is that SNNs are arriving in practice not as general-purpose neuromorphic computers but as hybrid systems in which spiking circuits, classical microcontrollers, and conventional accelerators each do what they are good at1.

Where a skeptic should push

The most load-bearing assumption in the whole narrative is that the 20-watt brain and the megawatt data center belong in the same comparison at all. Kempe himself flags this, and the flag is warranted. A brain's 20 watts omits the energy of its construction: roughly 20 years of development, billions of years of evolution, and an entire organism's metabolism wrapped around the organ. A fair amortized energy accounting for biological intelligence has never been done, and the monograph does not attempt one. Equally, the microwatt figures for edge chips are vendor-reported numbers for narrowly scoped inference tasks, not for the open-ended generative workloads that drive the data-center demand; no neuromorphic chip today runs a language model at competitive quality, and the monograph does not claim otherwise.

There is also a calibration issue the reader must carry: as a synthesis with no new measurements, the book inherits the bias of its sources, which skew toward the success stories of neuromorphic engineering. What it documents well is existence: chips that exist, rules that train, frameworks that ship. What it cannot document is parity: no head-to-head joules-per-task comparison between a neuromorphic system and a conventional accelerator on the same workload appears anywhere in its 88 pages. That comparison, not the 20-watt juxtaposition, is the measurement the field actually needs.

Organoid intelligence and the 20-watt yardstick

Read as an organoid-intelligence document, the monograph's real content is a moving baseline. The classical rhetorical structure of a biocomputing pitch is: the brain computes at 20 watts, silicon does not, therefore living tissue wins. This book shows, chapter by chapter, engineering systematically strip-mining that gap. Event-driven computation: absorbed, and shipping at microwatt power. Local learning without backpropagation: absorbed into chip hardware as programmable core-local rules. Time as an information carrier: absorbed into graded spikes and temporal coding frameworks. Each of the brain's efficiency tricks that has been understood well enough to write down has been, or is being, written down in silicon. The efficiency-only case for organoid computing is therefore not standing still; it shrinks every time one of these mechanisms is productized.

The non-obvious implication cuts the other way as well. The monograph's own thesis is hybridization: classical and neuromorphic systems each deployed where their strengths lie, because neither dominates. That is exactly the right frame for organoid intelligence, and it is more honest than the substrate-war framing the field often uses. An organoid is not a competitor to Loihi 2; it is a candidate co-processor with properties no chip in the book has. The monograph's list of what engineering has absorbed is, by contraposition, a specification of what living tissue must still demonstrate: developmental self-organization rather than configured topology, adaptation after deployment without any retraining procedure, chemical self-repair, and a native interface to biological signals. None of those appears in the monograph's hardware chapter, and none is close to being absorbed.

The opportunity and the threat follow directly. Opportunity: the organoid field can borrow the monograph's analytical discipline. The NEF idea of specifying what a neural population computes before asking how, and the neuromorphic habit of quoting power, event rates, and task accuracy together, are exactly the reporting standards organoid papers most often skip. Threat: hype-correction. If a microwatt edge chip with 328,000 spiking neurons can already do the sensing task, an organoid that does the same sensing task at milliwatts with uncontrolled variability is a worse engineering artifact, whatever its biological interest. As the baseline tightens, "it computes on living tissue" stops being an answer to any engineering question by itself. The field needs claims of the form: this culture, this assay, this metric, beats this number. The monograph shows what such a claim looks like; it remains for organoid intelligence to produce one.

The bottom line

Established: nothing new; this is a synthesis, and a good one. Its transformer-to-Loihi 2 tour is mathematically serious, its hardware facts check out, and its energy caveats are more honest than the field's average. Established with caveats: the microwatt power figures are vendor-reported for narrow tasks, and the book contains no task-parity energy comparison. For organoid intelligence, the durable consequence is that the efficiency rationale is being eroded from below by silicon that absorbs one biological mechanism at a time, while the viable rationale, developmentally self-organizing and intrinsically adaptive compute, is precisely the one that has no silicon analog yet. What would confirm the monograph's optimistic frame: a published joules-per-task parity study between neuromorphic and conventional hardware on an open-ended workload. What would break it: evidence that hybrid SNN deployment stalls at the sensor edge and never reaches central workloads, which would narrow the whole neuromorphic program the book describes.

Frequently asked questions

What kind of document is this?

A synthesis monograph. Volker Kempe's 88-page book explains transformers, biological neurons, spiking neural networks, and neuromorphic hardware; it reports no new experiments. This article is an English translation of the German edition, DOI 10.5281/zenodo.23278918.

What is the 20-watt figure and is it a fair comparison?

The human brain runs on about 20 watts continuously, juxtaposed in the monograph against data centers drawing up to 500 MW, roughly 4.38 TWh per year at full utilization. Kempe himself states this is not a scientific efficiency comparison, since the systems differ in task, structure, and capability.

What does Loihi 2 actually provide?

About 2.3 billion transistors on roughly 31 square millimetres, 128 neuromorphic cores with 8,192 neurons each, up to about one million logical neurons, graded spikes that carry a numeric payload, and programmable local learning rules in each core.

How low is the commercial edge power, really?

Vendor-reported: SynSense Speck runs 328,000 spiking neurons in the sub-milliwatt range, Xylo processes weak sensor signals at a few hundred microwatts, and Innatera's Pulsar is reported in the microwatt range for always-on single inference tasks. These are narrow tasks, not general workloads.

What does this mean for computing on living tissue?

That the efficiency argument alone is shrinking as silicon absorbs each understood biological trick. The durable case for organoid intelligence rests on what tissue still has and chips do not: developmental self-organization, intrinsic adaptation without retraining, self-repair, and native biological interfaces.

References

  1. Kempe V. AI at a Turning Point: From ChatGPT through Neural Networks to Neuromorphic AI. arXiv:2610.39796. 2026. https://arxiv.org/abs/2610.39796. Translation of Kempe V. KI am Wendepunkt: Von ChatGPT über neuronale Netze zu neuromorpher KI, version 4, DOI 10.5281/zenodo.23278918. Accessed 2026-11-03.
  2. Maass W. Networks of spiking neurons: the third generation of neural network models. Neural Networks 10(9). 1997. (Origin of the "third generation" framing discussed in reference 1; full citation as given in the reference list of reference 1.)