The Neuromorphic Commons: a silicon access model read for what it denies wetware

What the work claims

THOR is a roughly $4M NSF community cyberinfrastructure grant to build a shared, remotely accessible neuromorphic-computing commons.¹ The numbers, bound to the record: $3,151,918 obligated against an estimated total of $4,000,000, running 2024-09-01 to 2029-08-31, led by Dhireesha Kudithipudi at the University of Texas at San Antonio, with the University of Tennessee Knoxville, UC San Diego, and Harvard. It sits in NSF's community research-infrastructure line, which funds shared resources for a field rather than one lab's science.

Define the object. Neuromorphic computing here means silicon hardware whose architecture is inspired by the brain: spiking neurons that communicate in discrete events, event-driven rather than clock-driven operation, and in-memory compute that co-locates memory and processing to avoid the energy cost of shuttling data. This is emphatically not computation on living tissue. The contrast term, wetware or organoid intelligence, means computation performed on actual living neural cells. This grant is entirely on the silicon side of that line.

What makes it notable is not a chip but an access model. The award's own analogy is explicit and load-bearing: it aims for impact similar in scale to the impact seen when high-performance computing systems became accessible to the engineering research community. The bet is that exotic neuromorphic hardware is currently gated behind a few industrial and academic silos, and that turning it into a shared, remotely accessible, benchmarked commons will do for neuromorphic research what shared supercomputers did for computational science. As an infrastructure grant, every capability it names is funded intention, not demonstrated outcome.

How it works

The grant names four deliverables, and they are best read as the actual instrument rather than as slogans. First, remote access to large-scale neuromorphic systems, provided through close-knit partnership with industry and specifically through collaborations with two prominent neuromorphic companies, which the abstract does not name, so neither will I. Second, open-source hardware and software co-design frameworks and tools, meaning the software layer that lets a researcher map an algorithm onto unfamiliar spiking hardware. Third, common benchmarks and competitions, meaning shared tasks and leaderboards that let heterogeneous systems be compared on one yardstick. Fourth, rapid algorithm development supported by a collection of learning modules, network models, and example frameworks, in effect a model library and a teaching corpus. Around these sits a training pipeline the grant describes as reaching from K-12 students through graduate researchers.

Note the shape. Nothing here is a scientific finding. It is a set of shared-resource commitments: hardware you rent time on remotely, software you download, tasks you compete on, models you reuse. That shape is what matters for the reading below, because the shape itself is the thing that does or does not transfer to living tissue.

Where a skeptic should push

The load-bearing assumption for any reading of this grant against living-tissue computing is that neuromorphic and bioinspired name one competitive space in which silicon and wetware substitute for each other. At the application layer that is largely a category error. Silicon neuromorphic chips and living neural cultures operate on different tasks at different timescales with different economics; a spiking accelerator doing low-power event-based vision is not competing for the same job as a neuron culture in a closed-loop learning experiment. Treating them as rivals for the same workload flattens a real distinction into a false one. So I bound the substitution claim deliberately: the two substrates compete at the funding and definitional layer, where they draw on an overlapping pool of brain-inspired-computing money, talent, narrative, and terminology, not at the application layer, where they largely do not do the same work. Any argument that reads THOR as a head-to-head competitor to organoid platforms on the merits of the computation is overclaiming, and this analysis does not make it.

The second caution is about genre. This is an infrastructure grant, so verbs like will catalyze a transformation are aspiration written into a proposal, not results in the world. As of this award record there is no reported outcome to evaluate, only a plan and a budget. The correct posture toward every capability claim is funded intention, not achievement.

Why the silicon commons will not fit wetware

Now the section this title exists to write: what THOR changes for platform access, vendor capability, and the governance of computing on living neural tissue. The thesis is non-obvious and cuts against the naive reading, so it is developed as opportunity and threat together, grounded in the grant's actual mechanism.

On access, THOR institutionalizes a specific model for exotic brain-inspired compute: remote time-sharing of industry-partnered hardware, an open co-design toolchain, shared benchmarks, and a reusable model library. That is precisely the access model that organoid-intelligence and wetware platforms conspicuously lack. There is today no remote commons where a researcher rents time on someone else's living neural culture, runs a standard benchmark, and reuses a shared model. THOR builds one for silicon.

By building it, THOR captures three things at once, and this is the threat to wetware. It takes the bioinspired-processing narrative that wetware advocates also claim, the cyberinfrastructure dollars flowing to brain-inspired computing, and, most consequentially, the benchmarks and the definitional authority that come with them. If neuromorphic becomes operationally defined as this silicon commons and its leaderboards, then living-tissue biocomputing is quietly reframed as a fringe alternative that must justify itself on silicon's yardsticks. Whoever owns the benchmark owns the terms of comparison, and THOR is explicitly a benchmark-and-competition instrument. This is a narrower and more concrete claim than a general point about who controls a research vocabulary; it is specifically about the benchmark-and-access layer, tied to THOR's competitions mechanism.

Here is the part the naive reading misses, stated carefully because it is easy to overstate. THOR's access model works because of physical properties that silicon has in full and living tissue has only partially. A neuromorphic chip is copyable, shippable, and remotely time-shareable as an identical instance: you can fabricate many indistinguishable units, ship them, and let strangers rent slices of one over a network to run a fixed task repeatably. Living neural tissue has weaker versions of each. It is partly clonable, since human induced pluripotent stem cell lines can be clonally banked to give genetically identical starting material; it is partly shippable, since organoids are cryopreserved and distributed by living biobanks; and it is partly remotely accessible, since cloud labs and robotic biofoundries already run experiments on cells for remote users. What does not carry over is the identical-instance, real-time time-share at the heart of THOR's model. Each organoid follows its own developmental trajectory and degrades as it is used, so there is no indistinguishable second copy to benchmark against and no stable instance to share in real time. The result is a difference of degree large enough to matter: wetware access stays comparatively gated by physical co-location and per-instance fabrication, a more centralized and higher-friction regime than a silicon commons. That gap is grounded in the biology rather than chosen by policy, and while governance and engineering can narrow it, they cannot fully erase it.

The divergence is not total, which is the opportunity. THOR's open-benchmark and co-design governance is a partial blueprint a wetware consortium could adopt at the tooling and benchmark layer even though it cannot adopt it at the substrate layer. Shared task definitions, open interface standards, and common evaluation protocols are substrate-independent; a living-tissue field could borrow the commons governance for everything except the one thing that does not copy, the substrate itself. Finally, the substitution is not ethically neutral. A silicon commons carries no obligations around tissue provenance, donor consent, or the moral status of the computing substrate. A wetware commons would inherit exactly those obligations by construction, because the substrate is living cells of human origin. Even if the access engineering could somehow be matched, the governance load could not. The silicon commons is cheap to govern precisely because it computes on nothing anyone need have consented to.

The bottom line

Established from the source: THOR is a funded, roughly $4M NSF cyberinfrastructure grant running 2024 to 2029 to build a shared, remotely accessible, benchmarked commons for silicon neuromorphic hardware, in partnership with two unnamed industry hardware providers. That is a documented commitment, not a demonstrated capability, and the grant says nothing whatsoever about living tissue.

Labeled clearly as my reading: the commons access model THOR pioneers for silicon does not fully transfer to wetware, because living neural tissue supports only weak versions of the copyability, shippability, and identical-instance time-sharing that make the model work and lacks the indistinguishable second copy that benchmarking and real-time sharing require, while THOR's success may separately capture the bioinspired-computing narrative and benchmark authority in ways that structurally disadvantage wetware. What would confirm the reading: a wetware field that attempts a shared commons and hits exactly the copy, ship, and time-share walls, and neuromorphic benchmarks becoming the default yardstick applied to biocomputing. What would break it: the arrival of clonable, shippable, time-shareable standardized neural substrates, which would itself be a landmark result, making a wetware commons feasible on THOR's model after all.