All-In panel on distributed compute: idle GPUs into shared inference & training pools
On the 27 June 2026 episode of the All-In Podcast , the panel turned to distributed (decentralised) computing as a way to convert idle and surplus GPUs — from data-centre overhang to home rigs — into usable AI capacity by contributing them to shared, lower-SLA training and inference pools. BittensorDealroom has a profile for this one. Try Dealroom →, Venice and PluralisDealroom has a profile for this one. Try Dealroom → were named as the projects building this, with Manifold's Targon cited as a permissionless market where anyone can rent or supply attested GPUs.
Training wants proximity; inference does not. The panel's core caveat: distributed training loses an order of magnitude or more in efficiency once nodes are not physically adjacent — even fibre-connected GPUs a couple of kilometres apart introduce crippling lag. Distributed inference , by contrast, tolerates higher latency, and the panel argued "distributed inference clouds are coming" as a real part of the American open-source AI ecosystem.
The hardware thesis. A floated model: sealed multi-GPU units with battery and Starlink placed in homes (e.g. bundled into a Tesla Power Wall) to serve neighbourhood-level inference, monetising otherwise-idle silicon. The unresolved question is how homeowners get paid — whoever solves that "pan-American" coordination layer captures the opportunity.
Bittensor. Open-source protocol (token: TAO) co-founded by Jacob Steeves and Ala Shaabana in 2019, running a peer-to-peer market where 100+ subnets reward miners in TAO for producing AI work — training, inference and other digital commodities — validated by a "proof of intelligence" mechanism. Network value sits around the $2B mark.
Targon / Manifold. Manifold Labs operates Targon, a decentralised "confidential compute" cloud and Bittensor subnet where operators stream attested, per-second GPU cycles into an open market; H200s rentable by the hour, permissionlessly, with validators verifying contributed compute. Its Targon Virtual Machine uses Intel TDX and NVIDIA confidential computing to keep workloads private on untrusted hardware. Backers include OSS Capital and DCG.
Pluralis. AI research lab (Melbourne / New York) building "Protocol Learning" for decentralised, model-parallel training in which models are never fully materialised by any single participant, keeping weights inside the protocol. Founded by CEO Alexander Long, the team came from Google, Anthropic and Amazon. Raised a $7.6M seed in March 2025 led by Union Square Ventures and CoinFund, with Variant, Topology, Eden Block, Bodhi Ventures, and angels Balaji Srinivasan and Clem Delangue.
Venice. Cited alongside as a community-oriented project in the same distributed/permissionless-AI space.
Read more: All-In Podcast (distributed-compute segment, 1:21:46) · FinSMEs · Manifold Labs · Wikipedia