Neuromorphic Compute
Neuromorphic & Alternative Compute
Computer chips that work differently from normal processors: 'neuromorphic' chips that fire like brain cells, chips that compute inside memory, or chips that use light. The aim is to do the same work with far less electricity, since power is becoming the limit on AI. The largest neuromorphic system has about 1.15 billion artificial neurons, but it is a research prototype. The chips work; the software to use them is the bottleneck, and almost nobody buys them yet.
Maturity
Where it feeds in
Where you have already met this
In 2023 BrainChip, an Australian-listed maker of neuromorphic chips, booked US$232,004 in revenue - down from US$5.1 million the year before - and lost US$28.9 million. The world's largest neuromorphic system, Intel's Hala Point, has 1.15 billion artificial neurons and sits at Sandia as a research prototype. The chips work. Almost nobody is paying for them yet.
If this moved
If ordinary AI models could run on these chips without being rebuilt by hand, the energy argument would finally be tested outside the lab. Automation of cognitive work would still sit at 2 - digital transformation, managerial capability, regulatory clarity and workforce trust hold it there too. The energy transition would stay at 1, held by consumption norms.
What to watch for
A mainstream AI framework compiling standard models to neuromorphic hardware without manual conversion. Neuromorphic or in-memory chips in a shipping product people buy, not a development kit.
Not to be confused with
Quantum computing. Quantum machines exploit quantum states for specific maths problems; neuromorphic chips are ordinary electronics arranged to compute the way neurons signal.
What it is
Processor architectures that do not follow the conventional model - neuromorphic chips, analogue and in-memory computing, photonic processing.
How it works
- 01Put memory next to the logicConventional chips shuttle data between processor and memory, and that shuttle costs much of the energy. Neuromorphic designs keep each 'neuron's' memory beside its logic.
- 02Stay quiet until something happensCircuits draw almost no power while idle and wake only when a signal arrives.
- 03Fire a spikeWhen the incoming signals cross a threshold, the circuit sends a short pulse to its neighbours.
- 04Weight the connectionsEach connection scales the pulse by a stored number, the way a synapse strengthens or weakens a signal.
- 05Rebuild the softwareStandard AI models are not written this way. They have to be converted or retrained - today, largely by hand.
State of play
- Intel's Hala Point packs 1,152 Loihi 2 processors, 1.15 billion neurons and 128 billion synapses into a six-rack-unit chassis drawing at most 2,600 W; it was deployed at Sandia National Laboratories as 'a research prototype' (Intel, 17 Apr 2024).
- Intel reports more than 15 trillion 8-bit operations per second per watt on the system; the figure was achieved without batching, on a proof-of-concept network (Intel, 17 Apr 2024; EE Times, 30 Apr 2024).
- Intel's neuromorphic lab director calls the scalability of software compilation 'a bottleneck' and says converting models is 'more of a manual process' than he would like (Mike Davies, in EE Times, 30 Apr 2024).
- BrainChip reported revenue of US$232,004 for 2023, down from US$5,071,252 in 2022, and a net loss of US$28.9 million (BrainChip Holdings, Annual Report 2023).
- Sony announced two event-based vision sensors developed with Prophesee, whose pixels report only changes in brightness rather than full frames (Sony Semiconductor Solutions, 9 Sep 2021).
Why it matters
It addresses the constraint that conventional architectures are hitting: energy per operation. If AI's growth is bounded by power rather than by silicon, the answer is architectural rather than incremental.
- It is one of the enablers holding automation of cognitive work at 2. It shares that job with four human and organisational enablers, so fixing it alone moves nothing.
- It is the architectural answer to an energy ceiling. The question is whether that ceiling binds before the software exists.
- It is a complementarity problem in its purest form: the device works, and everything around it assumes a different device.
What holds it back
Toolchains
There is no common way to compile a standard AI model to these chips. Each maker keeps its own software stack.
Model conversion
Models built for conventional hardware have to be converted or retrained, and today that is expert, manual work.
Fabrication access
Advanced chips are made in very few places, and experimental architectures queue behind mainstream orders.
Terms
- Spiking neural network (SNN)
- A network that passes information as short, timed pulses rather than continuous numbers.
- Von Neumann bottleneck
- The energy and speed cost of moving data back and forth between a processor and separate memory.
- In-memory computing
- Doing calculations inside the memory itself, so data does not have to move.
- Event-based sensor
- A camera sensor whose pixels report only when brightness changes, instead of sending whole frames.
- TOPS/W
- Trillions of operations per second per watt - a measure of how much computing a chip does for its power.
Steepens
| Megatrend | Weight | Role |
|---|---|---|
| Automation of Cognitive Work | moderate | shared weakest link (with 5 others) |
| Energy Transition and Electrification | moderate | |
| Precision Everything | marginal | marginal: not counted as a weakest link |
The gradient behaves like a sticky rubber band: it is pulled hardest toward the weakest enabler. Marginal links are shown but do not pull.
If this moved one step
Computed from the map as it stands: this score raised from 2 to 3, every other score unchanged. It is arithmetic, not a forecast.
Raised one step on its own, it lifts no megatrend's weakest link; each gradient it feeds still rises a little.
Tied as the weakest link
Here it shares the weakest spot with other enablers. Raising it alone leaves the weakest link where it is and lifts the gradient a little; the enablers it is tied with must move too.
- Automation of Cognitive Work2.7/5 , 2.8/5tied with Data & Datafication, Digital Transformation, Managerial Capability, Regulatory Clarity, Workforce Trust
The gradient also rises in
- Energy Transition and Electrification2.3/5 , 2.3/5
What would change the gradient
Conditional, not predictive. No dates, no probabilities.
- 01A software toolchain that does not require rewriting everything
- 02Energy cost per operation becoming the binding commercial constraint, which would force the switch
- 03Fabrication access, currently concentrated in very few places
Common misconception
That this is a niche research topic. It is the most plausible route around an energy ceiling that is already shaping where data centres can be built.
Talin, B. (2026). Neuromorphic & Alternative Compute. The Future Explorer, dataset v0.4.0 (2026-11-16). https://explorer.talin.digital/e/neuromorphic-and-alternative-compute/