Nockchain's AI lane went live in August and capacity followed. The team says demand is now the constraint.

Nockchain’s AI lane is up and running. The live roadmap lists “AI Compute Network Launch” for August 2026, and it is marked completed.² Back in June, the roadmap listed an “AI Compute Market Launch” for June 2026.¹

The upgrade behind it, Logos, added a matrix-multiplication puzzle that Pearl miners also run.²¹ It sits alongside the ZK proof-of-work puzzle.²¹ This means miners now have two ways to win blocks on Nockchain.

The road from June to block 126,000

The first Logos pull request opened on July 18.²² It set activation at block 114,300, with a July 31 target.²² Logan Allen announced that height on July 20: “Logos and AI inference mining are proposed to activate at block 114,300, targeted for the end of this month.”³

Five days later, the @nockchain account moved the target: “Mainnet activation retargets to block 122,000 unless issues are found.”⁴ The Logos code on GitHub never carried that height, and it kept 114,300 until August 11.⁵¹⁰

On July 28, a reviewer described an attack on the new fork-choice rule in a comment on the Logos pull request.⁵ The reviewer wrote that activation should wait for a reviewed fix.⁵

The maintainers’ own write-up followed on July 29: “Activation must remain blocked pending a reviewed fork-choice fix.”²³ The document now lists the issue as resolved.⁶ It says the fix is a rule that credits each block with the expected work at its own puzzle’s target.⁶

The final date came on August 10 at 20:58 UTC: “The new AI Compute Network goes live on Friday (8/14) at Block 126,000. Code is being merged today.”⁷ Block 126,000 arrived at 02:05 UTC on August 14.⁸ The first AI-mined block, 126,859, followed at 16:27 UTC.⁹

The split between the two puzzles changed along the way. Until August 11, the code aimed for about 60% AI blocks.¹⁰ The commit that set the final height changed the target to about 30% AI and 70% ZK.¹⁰ A month later, the Anthropos upgrade at block 147,500 flipped it to 70% AI and 30% ZK.¹¹

Blocks are landing close to that target. Over the seven days to 12:06 UTC on October 10, AI puzzles won 69.6% of 3,971 blocks, counted by puzzle type in Nockscan’s prover statistics.¹²

Capacity followed

Nockscan measures the AI lane in multiply-accumulates per second. Over a trailing seven-day window, it read about 0.06 EMAC/s on August 16, two days after activation.¹³

Nockscan restarted the series at block 154,500, when a coordinated upgrade re-anchored both difficulty lanes.¹³¹¹ Readings from before that block can’t be compared with readings after it.¹³ Within the new series, the same seven-day measure rose from 12.5 EMAC/s at block 154,530 on September 26 to 23.0 EMAC/s at 12:45 UTC on October 10.¹³

Three weeks into the lane, the team described the next problem in a post on September 4: “Capacity is no longer the constraint. Demand is: turning it into inference someone paid for.”¹⁴

Routing and Lockstep

The roadmap’s plan for demand is the “Useful Inference Routing Protocol”.² It is listed for Q4 2026 with status current.² The roadmap describes it as “A major milestone that funnels compute demand into the AI Compute Network.”² It is absent from the Internet Archive’s June 16 capture of the roadmap and appears in the next capture, on September 30.¹²⁴

The team has also shared new research. On September 10, @nockchain published Lockstep, a paper and code release on verifiable inference.¹⁵

The 27-page paper defines a canonical inference contract for dense and mixture-of-experts transformer models.¹⁶ Under that contract, an optimized GPU engine and a portable CPU reference must return bit-identical outputs.¹⁶ The paper reports identical digests from the same source on Hopper, Ada and Blackwell GPUs.¹⁶

The serving implementation is a vLLM plugin that the paper describes as proprietary.¹⁶ The public release carries the specification, CPU references, a Lean formalization and conformance kits.¹⁵¹⁶ The paper also specifies a sampled-replay audit, with its primitives gated and integration disabled.¹⁶ It is open about its own limit: “Canonical execution is not a cryptographic proof”.¹⁶

Every use of “routing” in the paper refers to expert routing inside mixture-of-experts models.¹⁶ The paper never mentions Nockchain or proof-of-work.¹⁶ In Nocktane’s reading, the paper specifies the inference function itself and leaves the routing of paid requests to other work.

The announcement places Lockstep at the service layer, where “it can give customers and auditors a precise meaning for the inference work being routed among independent providers,” and adds that “the integration is not live and Lockstep is not currently a Nockchain consensus rule.”¹⁵ Nocktane found no published specification for the routing protocol as of October 10. That specification is the next piece to watch for.

The usefulness debate

In June, Abhinaba Basu posted “The Usefulness Gap in Proof-of-Useful-Work” to arXiv.¹⁷ The paper studies Pearl. Its headline finding is that Pearl’s network “produces zero useful AI computation.”¹⁷

The paper says Pearl’s verifier checks that a matrix multiplication was computed correctly.¹⁷ According to the paper, the verifier does not check whether the matrices came from an AI model, so miners can earn rewards on random matrices.¹⁷ An open-source miner Basu wrote had 44 shares accepted by a pool using random matrices.¹⁷

Logan Allen answered on September 2: “I didn’t find this paper’s methodology to be compelling.”¹⁸ The post argues that networks like this start out farmed by GPUs chasing the subsidy.¹⁸ It adds that the pieces needed for useful inference routing already exist and need a small amount of further work.¹⁸ On October 4, Allen described Pearl compatibility as a starting move: “Bootstrapping our AI puzzle with Pearl merge-mining compatibility was a tactic chosen to broaden mining distribution and improve initial network security.”¹⁹

Nockchain’s AI puzzle is Pearl-compatible by design.²¹ The Logos specification describes its matrices as “miner-chosen (arbitrary model, Pearl parity)”.²¹ Merge-mining is near universal in recent blocks: of the 7,043 AI blocks among the last 10,000 blocks, 7,042 declare a Pearl merge-mining commitment (Nockscan, 12:47 UTC on October 10).²⁰

The paper’s argument is therefore relevant to Nockchain’s lane too. That link is Nocktane’s inference. The paper itself never mentions Nockchain.

Next on the roadmap is the routing protocol. The roadmap describes it as the milestone that “funnels compute demand into the AI Compute Network” and lists it for Q4 2026.² Our tracker will record whether that date holds.