The ledger ate the GPU: full-stack verifiable compute hits mainnet

In the year of our algorithm, the phrase “trust the compute” has gone from a research prayer to a deliverable—the same way ironclads replaced wooden warships, but with cryptographic proof where the rivets should be. Three production-grade stacks now run the same play: off-chain model execution, cryptographically verifiable proofs, on-chain settlement or attestation; and Microsoft’s Agent Framework shows up like a stage manager to force the orchestration layer that agent-driven wallets will need.[^claim_908][^claim_912][^claim_915][^claim_911]

Cysic is the sector-spanning version, a full-stack verifiable compute network that welds zero-knowledge proofs, AI computation, decentralized hardware, and blockchain infrastructure into one architecture. CYS and CGT tokens went live at launch, and the network claims 10 million ZK proofs delivered with 260,000+ nodes onboarded, from mobile devices to professional operators.[^claim_909] CYS does double duty: it pays for ZK proofs, AI inference, and compute services, and it stakes in the Proof-of-Compute consensus. Splitting utility from governance in the token model keeps compute pricing in a separate ledger from protocol politics—good risk hygiene before you take a position.[^claim_910]

The tighter case comes from Verathos, operating as Bittensor Subnet 96.[^claim_912] It generates sumcheck proofs over Merkle-committed weights, and validators check those proofs on CPU in milliseconds. The production vLLM path generates the proofs in parallel during CUDA graph execution, no challenge-response round trip, and the overhead stays in single digits—latency near zero, hitting the CPU cache like a drill round.[^claim_912] That overhead profile is the trade: a verifier no longer needs a GPU—just an ordinary CPU—while miners still need NVIDIA GPUs with 24GB+ VRAM, the RTX 4090, A100, or H100 class.[^claim_914] The same proof system extends to training, verifying forward pass, backward pass (gradient GEMM), and optimizer step for both full fine-tuning and LoRA with AdamW, SGD, or Muon.[^claim_917]

The network exposes an OpenAI-compatible API with score-weighted routing across miners, and payments arrive as TAO, USDC on Base, or x402 pay-per-request.[^claim_913] For crypto, that’s the missing piece: an agent can call a model endpoint with the same interface as a centralized API, and settle per request without pre-funding—a repo trade without the margin call.

0G’s GLM-5.2 deployment stretches the pattern to long-context and privacy. The Mixture-of-Experts model carries roughly 744B total parameters, 40B active, in a 1M-token context window, and every inference runs inside 0G’s privacy stack with prompts and outputs encrypted, computation verifiable rather than “trust the provider.”[^claim_915][^claim_916] Long-context verifiable inference is what coding agents and audit agents need to reason over entire repos; encrypted compute keeps proprietary prompts off public ledgers—secrecy, like carry, is a paid premium.

Microsoft’s Agent Framework sits outside crypto, but its documented primitives—workflows, Agent Harness, human-in-the-loop checkpoints, resumable orchestrations, self-hosting, and Azure Functions/durable hosting—describe the execution model crypto agents need when a transaction or policy checkpoint interrupts an autonomous run.[^claim_911] Checkpoints and resuming are the price of admission before an agent wallet signs multi-step transactions safely.

The pattern that ships today is hybrid: miners run GPUs, validators verify on CPUs, model weights stay Merkle-committed, and payments clear as tokens or x402 requests.[^claim_912][^claim_914][^claim_913] The next battleground is training—Verathos already proves fine-tuning and LoRA—and the first network that makes verifiable training a paid, production market will set the standard for the rest.[^claim_917] Watch for third-party benchmarks on that overhead, and for whether 0G’s 1M-token encrypted context holds under real agent workloads.[^claim_915]