Theses›Datacentre & AI compute›Decentralised AI

Decentralised AI

By 2030, a top-3 large language model by number of users will be a decentralised AI.

  • Bear
  • Early
  • $35B by 2035

The callModels nobody can switch off are wanted, but training across the open internet costs 6x the compute and the value pools in tokens; nothing here for equity yet.

D1D2R1R2R3C1C2foragainst D1D2R1R2R3C1C2foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the restraints.
  1. D1Demand grows for models nobody can switch off, censor or price-gouge, which three or four labs cannot offer by construction.
  2. D2Millions of idle GPUs and commoditising open-source models give a network something to coordinate and something to serve.
  3. R1GPUs on the open internet cannot swap gradients fast enough, so training pays about 6x the compute at scale and every credible result is at 10-72B parameters.
  4. R2The value pools in tokens, and no equity exit above $50M is on record.
  5. R3The gradient-traffic problem is being solved in software, so there is no chip, photonics or materials company in it.
  6. C1A permissionless pre-training run of 100B parameters or more finishes over the open internet at near-centralised throughput. The largest so far are 40-72B, and the 40B run paused at about 6% of its data.end-2027
  7. C2Gensyn's mainnet, live since April 2026, or Nous's Psyche carries a paying commercial training job above 1,000 H100-equivalents, proving the network sells compute rather than tokens.end-2027
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
20272028202920302031today1234predictionswatch 20272028202920302031today1234predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1By end-2027 at least one decentralised AI model or app reaches >=10M monthly active users or a top-20 ranking by users.Dec 2027 · open
  2. 2A decentralised model or app at 10M monthly active users or top-20 by users.Dec 2027 · open
  3. 3By 2030, a top-3 large language model by number of users will be a decentralised AI.Dec 2030 · open
  4. 4The original 2023 call: a decentralised model in the top three by users.Dec 2030 · open
  • ·A scaling result that closes the multi-node compute penalty, or open training data alone matching the leading models.
Expressionslisted, private, technologies
Private · 8 private companies, 6 of them in the US; none in continental Europe.

Bittensor (Opentensor Foundation) · Fetch.ai · Akash Network (Overclock Labs) · io.net · Prime Intellect · Nous Research · Gensyn · Ritual

Technologies
Who has already bet on itprivate rounds, from public sources

$362M raised by 8 companies, $251M of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: Prime Intellect, Gensyn. Backed by a strategic investor: Prime Intellect.

Reading
Show the reading