Actual Computer is a Bittensor subnet (SN95) building inference software for the hardware people already own. The product runs large language models on Macs, PCs, consumer GPUs, home servers, and edge boxes, coordinating them as a mixed device fleet rather than forcing a clean datacenter setup.
The bet is that inference moves out of datacenters as open-source models reach near-frontier quality and as the energy demand on datacenters becomes unworkable. SN95 is the Bittensor surface that turns the resulting fleet of consumer machines into coordinated AI supply.
The Thesis that Makes Actual (SN95) Work
The case for local inference rests on a few current realities of the AI industry.
1. Datacenter energy demand is at a building-capacity bottleneck. Based on current projections, max energy buildout looks pinned for a decade-plus. The supply side cannot keep pace with demand.
2. Massive idle compute exists outside datacenters. An estimate from Epoch AI’s director of research suggests there is roughly 4x more compute sitting in homes, offices, and consumer devices than inside datacenter racks. Most of it is unused.
3. Open-source models are now near state-of-the-art. GLM 4.7 benches close to top closed-source models. Kimi K2 and other 1T-class models are deep in capability. The gap between datacenter and home models is shrinking fast.
4. Consumer hardware is finally ready to run them. Apple ships Macs with 512GB of unified RAM. Nvidia’s GB10/Spark uses unified RAM on Blackwell. 5090 clusters run models at high speed. M5, M6, and post-Grace-Blackwell are the next inflection.
The team calls the underlying opportunity TEA: technological energy arbitrage. Turning underused compute and abundant residential energy into AI output instead of renting intelligence forever through subscriptions.
How the Software Works
Actual Computer’s software coordinates real hardware as it exists, not as it would look inside a clean datacenter rack.
What the product does:
1. Heterogeneous-first. Runs inference across mixed device fleets. Macs, PCs, GPUs, home servers, edge boxes. Different devices have different strengths and different failure modes, and the software is built to see that rather than flatten everything into a fake homogeneous cluster.
2. (GPT-Generated Unified Format) GGUF-based inference engine. First built on LlamaCPP and the GGUF ecosystem. A custom engine is in development to push record-setting performance across device configurations.
3. One-line install. Sign up, install on Mac, Linux, or PC with a single command, load models through the interface, and access OpenResponses-compatible endpoints. Actual never sees or stores inference requests.
4. Hardware floor: 2017 and later. Anything from a 3060 GPU and up can contribute. The bar is consumer hardware, not datacenter silicon.
5. Always-on home compute. Users can fire up a home cluster running a powerful model and access it from anywhere. Local, global, or both.
The team positions the software as infrastructure, not a developer tool: “the structure supporting the future of machine intelligence should look like breaker boxes and transformers, less like apps.”
The SN95 Mechanism
Actual Computer chose Bittensor as the distribution layer because no other chain currently houses or incentivizes digital commodities at the scale SN95 is targeting.
How the subnet works:
1. Participants run inference passively. A consumer machine sleeping eight hours a night can be running model inference instead. The operator pockets the spread between residential power costs and what the network pays.
2. The network is a compute fabric, not a rented GPU farm. Supply can be directionally controlled by geography, cost, and time of day. Closer to a CDN (Content Delivery Network) than a datacenter.
3. SN95 sits at the inference and execution layer. Not the model layer, not the training layer. Competition is on serving efficiency, latency, and request handling, not on changing the model itself.
4. The consumer surface is unclaimed. Other inference groups on Bittensor are largely datacenter-oriented. SN95 explicitly targets consumer and edge hardware.
In simple terms: SN95 is about using the brain (running models), not building it (training).
SN95 vs IOTA (SN9): Building Models vs Running Them
Actual Computer (SN95) and IOTA (SN9) often get grouped together because both deal with large language model infrastructure on Bittensor, but they sit at different points in the model lifecycle.
IOTA is a distributed pre-training network coordinating GPUs to train a large language model from scratch (the team has been working on a model in the ~100B parameter range). Actual Computer, on the other hand, is the inference and serving network that takes finished models and runs them efficiently on consumer hardware.
| Dimension | IOTA (SN9) | Actual Computer (SN95) |
| Pipeline stage | Pre-training: building a model from scratch | Post-training: running finished models live |
| Core problem | Coordinating distributed GPUs on a single training run | Coordinating heterogeneous consumer devices on inference |
| What miners do | Contribute training compute toward a shared large model | Serve inference requests across mixed device fleets |
| Hardware target | GPUs capable of contributing to training | Consumer machines from 2017 and later (3060 and up) |
| Output | A larger, more capable open-weight model | Lower-cost, lower-latency inference on existing hardware |
| Time horizon | Months to years (training cycles) | Real-time per request |
| Plain English | Producing the intelligence | Delivering the intelligence |
The cleanest read is that SN9 is a distributed lab building a model from scratch, and SN95 is a distributed network serving finished models to end users. Strip out one and the system loses a layer.
Without pre-training networks like SN9, open-source models stop improving relative to closed-source frontiers. Without inference networks like SN95, those models stay locked inside datacenters. The two cover adjacent ground in the same stack, not competing approaches.
The Actual Models Service
Alongside the core inference software, Actual runs a curated models service. This service offers:
1. Quality imatrix quants of popular open models. The team makes and publishes its own quants rather than relying on whatever happens to be available.
2. 8 models with 24 variants. Three quant levels per model: Q8_0, Q6_K, and Q4_K_M.
3. Featured lineup includes Qwen 3.6 35B A3B, Nemotron 3 Nano Omni 30B A3B Reasoning, Gemma 4 (31B IT, 26B A4B IT, 12B IT, E4B IT, E2B IT), and Qwen 3.6 27B.
4. Sizes range from 4.1GB to 35.2GB download. Memory asks from 4.7GB to 40.5GB.
5. Public checksums per quant. The snapshot attached was last updated on June 9, 2026.
The service removes one of the friction points for users running models locally, which is finding a quality quant in the right size for their hardware.
Setting Up SN95: The Bittensor Wallet Connection
Actual Computer integrates Bittensor at the account level through a wallet connection feature. From User > Settings, users can link a Bittensor wallet to their Actual account, giving the product an on-chain identity for each operator.
The setup is built around a signature-based proof of ownership rather than a transaction:
1. Polkadot-compatible browser wallet required. The user needs an active Actual account, a wallet extension installed and enabled, and a wallet account that supports raw-message signing.
2. Signature proves ownership. When the user clicks Connect, Actual reads accounts from the extension, opens a picker if multiple are present, generates a short-lived verification challenge, asks the extension to sign the raw bytes, and verifies the signature against the chosen address.
3. Challenge rules. Five-minute TTL (Time to Live) on each challenge, five-attempt-per-minute rate limit, address normalized to the Bittensor SS58 prefix.
4. One wallet per account, one account per wallet. A given Actual account can have one active Bittensor wallet at a time. A given wallet address can only be the owner of one Actual account. Already-connected addresses are rejected on collision.
5. Reversible. Disconnecting from the settings page revokes the active connection record. Users can reconnect later through the same verification flow.
If the wallet extension exposes multiple accounts, the picker shows the wallet provider, account name where available, and the full SS58 address so the user can pick the intended one before approving the signature.
Common errors mostly fall into a handful of buckets: no compatible extension installed, the extension exposes no accounts, the wallet does not support raw-message signing, or the five-minute challenge window has elapsed.
The integration is the bridge between Actual’s consumer-facing product and the on-chain side of the network. It proves the user controls the address (It does not move funds.)
The Team and the Vision
Actual Computer is based in New York. Lead designer Jack Vodka handles art direction and UI design, which the team has put real effort into. The public face of the company is Tom Lynch, who runs SN95.
The team’s two stated predictions about where this is going:
1. Actual Computer becomes the next Microsoft. The fundamental software layer that sits between consumer hardware and how people interact with their computers daily.
2. Large language models as they exist today are effectively irrelevant within five years. The transformer architecture may still be core, but the LLM as a category has a limited lifespan in its current form.
Both are bold by design, but while the first is the bet behind Actual, the second is the gap that whatever sits after LLMs has to fill.
The Heterogeneous Bet
Actual Computer (SN95) is one of the more specific bets on Bittensor right now. The energy argument (datacenters bottlenecked, consumer power abundant) and the hardware argument (open-source models near frontier, consumer chips finally ready) compound into the same conclusion: inference moves to where the hardware and electricity already live.
The software layer to coordinate heterogeneous consumer compute has been the missing piece. Actual is building that layer, and SN95 turns the resulting fleet into useful supply. Whether the team becomes the next Microsoft is a separate question. The substrate underneath the bet is what they call technological energy arbitrage: four times more compute outside the datacenters than inside, energy supply that cannot be cut off, and a network designed to route demand to where the compute actually lives.
➛ Join Actual Computer (SN95) Beta By Requesting Access Here
Enjoyed this article? Join our newsletter
Get the latest TAO & Bittensor news straight to your inbox.
We respect your privacy. Unsubscribe anytime.
Be the first to comment