The Next Big AI Trade May Not Be AI Everyone is spending fortunes to make AI compute cheaper, faster and more abundant. GPUs, data centers, inference clusters, custom silicon and massive power capacity are attracting extraordinary amounts of capital. But as more of that work is outsourced, distributed across providers and executed by increasingly autonomous software, another infrastructure problem becomes much more important: how do you prove that the inference you paid for is actually the inference you received? That is the question I think the market is still underestimating. An AI agent may purchase inference from an external provider. An enterprise may send proprietary workloads to third-party GPUs. A financial institution may need to prove that a computation satisfied a particular rule without exposing the underlying customer data. In each case, the customer does not just need an answer. It needs confidence that the computation was performed according to a defined specification, while keeping sensitive inputs and intermediate information private. The internet was largely built around trusting servers to return the right answer. The machine economy increasingly needs something stronger: computation that can be verified rather than merely trusted. That is why the base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722 thesis is much bigger than another AI and crypto narrative. @nockchain is building toward a distributed market for verifiable computation, with zero-knowledge proof generation embedded into the economic foundation of the network itself. Its whitepaper describes Nockchain as a Proof-of-Work blockchain in which miners generate STARK proofs over NockVM execution, with the same incentive structure potentially extending to useful workloads such as AI inference and distributed computation. And the part that makes the thesis especially interesting right now is not simply that Nock has ZK or AI. It is Lockstep. Lockstep has attacked one of the hardest practical problems in verifiable inference by demonstrating bit-exact transformer inference across heterogeneous hardware in its reported implementation. 1. The proof economy is already taking shape The demand for cryptographically verifiable claims is emerging independently of crypto. The European Commission's age-verification architecture allows users to prove their age without giving the online service their identity information. The Commission released a second version in October 2025 that incorporated zero-knowledge proof technology, and says the application was technically ready as of April 2026. Denmark, France, Greece, Italy and Spain are already taking up the technical solution with national customization efforts. @HesterPeirce made essentially the same argument in her September 23, 2026 SIFMA remarks. She criticized the enormous KYC and AML “haystacks” created by continually collecting personal and transactional data and argued that attribute-based credentials and zero-knowledge proofs could establish facts such as age, citizenship, accredited-investor status or sanctions status without exposing the underlying information. The BIS has already demonstrated the same basic architecture in cross-border finance. Project Mandala encoded jurisdiction-specific compliance requirements into transaction workflows and generated cryptographic proofs of compliance that could be verified without revealing underlying customer data. Its proof-of-concept architecture included a rules engine and proof engine, with the BIS explicitly highlighting zero-knowledge proofs and multi-party computation as privacy-preserving technologies used in the project. None of these projects depend on Nockchain. That is precisely why they matter. They show that the demand for privacy-preserving proofs is developing independently of crypto. The basic idea is simple: instead of handing someone all of the information, prove the specific fact that matters. Age verification is one example. Regulatory compliance is another. Computation is the much larger opportunity. 2. AI is creating the market for it The demand side gets much larger as inference becomes the dominant AI workload. Gartner expects global inference spending to reach $23.3 billion in 2026, ahead of the $19 billion spent on training, as enterprises move from model development toward continuous, real-time AI deployment. Every time an AI system analyzes information, calls an external tool, evaluates an opportunity, runs a workflow or makes a decision, it is consuming inference. Increasingly, that inference does not need to happen on the same machine that owns the model or the user's data. An enterprise can rent GPUs instead of buying them. An application can route requests across different providers. An autonomous agent can eventually select compute based on price, latency, availability and performance. McKinsey estimates that AI agents could mediate $3 trillion to $5 trillion of global B2C retail commerce by 2030 under moderate scenarios. The number itself is a forecast, not a certainty, but it illustrates the scale of the transition as software moves from recommending actions toward increasingly autonomous execution. NIST is already working on some of the infrastructure surrounding that transition. Its 2026 AI Agent Standards Initiative is focused on secure interoperability and agent security, while its work on software-agent identity and authorization specifically examines identification, authentication, authorization, auditing and non-repudiation for autonomous agents. And now @BlackRock is framing the broader economic shift explicitly. Its September 2026 The Machine-Native Economy describes AI as “machine-native intelligence” and digital assets as “machine-native money,” while arguing that compute is becoming an increasingly important economic resource. BlackRock says standardized claims on compute capacity could become a significant digital-asset use case for financing and programmable settlement, and cites consensus estimates putting the combined 2030 revenue of AWS, Microsoft Intelligent Cloud and Google Cloud at approximately $1.1 trillion. BlackRock is not saying Nockchain wins. It is not making a prediction about base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722. But it is describing the economic environment Nockchain is positioning itself around: machines increasingly accessing external resources, compute becoming an economic asset, and agents potentially discovering, provisioning and paying for compute programmatically. BlackRock's paper explicitly describes agents comparing compute providers on price, performance, latency, location and hardware specialization. Once machines start buying computation from other machines, the question of whether that computation was actually delivered becomes much more important. 3. The proof becomes the receipt Think about outsourced inference as a transaction between two machines. One machine requests a computation. Another machine performs it. The provider may not want to expose its execution trace, proprietary model or infrastructure. The customer may not want to expose its private inputs. Re-running the entire workload just to verify the answer defeats much of the point of outsourcing it. A cryptographic proof can sit between them. The customer receives the result together with evidence that the specified computation was performed correctly. The verifier does not need to blindly trust the provider, and the underlying private computation does not necessarily have to be revealed. That opens a market well beyond AI inference: confidential enterprise workloads, private financial calculations, compliance checks, model provenance, fraud detection, scientific computing, solvency proofs, private auctions and eventually marketplaces where a buyer pays for computation and receives cryptographic evidence that the job was performed. This is also how Nockchain's whitepaper frames the long-term opportunity. It describes Nockchain as a distributed market for verifiable computation and points to applications including private payments, sealed-bid auctions, solvency proofs, selective-disclosure identity and ML inference in which the model and input remain private. AI is simply one of the biggest markets where the problem becomes obvious. The underlying primitive is verifiable computation. 4. This is where @nockchain becomes interesting @nockchain is making a very different architectural bet from a conventional blockchain that adds ZK tooling after the fact. Nockchain's consensus is already based on Zero-Knowledge Proof-of-Work. Miners generate STARK proofs over NockVM execution, and the chain selects the heaviest chain according to accumulated proofpower rather than simply counting blocks. Today, the mining proof is tied to a block-bound ZK-PoW puzzle; the roadmap is to have the mining proof itself prove the candidate block transition. That matters because ZK is not merely an application feature sitting above the chain. It is part of the consensus mechanism. The stack underneath it matters just as much. Nockchain's published research includes a SNARK-friendly instruction set, a framework for compiling languages into efficiently verifiable virtual machines, the NockVM, and now Lockstep for bit-exact transformer inference. The important point is that the Nock stack is extremely difficult to replicate because the pieces were designed together from the ground up: the programming language, instruction set, VM, proving system, deterministic inference layer and consensus mechanism all reinforce one another. A competitor can pursue the same broad idea of verifiable AI, but reproducing the underlying stack, the research behind it and the years of optimization is a much harder engineering problem. And the fact that Nock started from the proving architecture rather than trying to bolt proofs onto a conventional execution environment is important because overhead ultimately determines whether the market exists. 5. Lockstep is the part people should be watching This is the section I would put much more emphasis on than the generic “AI compute” narrative. The hardest part of verifiable AI is not merely producing a proof. It is defining exactly what computation is being proved. Modern transformer inference does not automatically produce bit-identical outputs across GPUs. Different kernels, reduction orders, batch compositions, compilers and devices can legitimately return different floating-point strings. That creates a fundamental problem for verification: an auditor either has to reproduce the provider's exact environment or accept some tolerance around the result. Nockchain's September 10, 2026 Lockstep research attacks the problem directly by making the served inference function itself canonical. A versioned contract fixes the implementation-level floating-point choices, turning correctness into bit equality against a public computation that can be evaluated by both an optimized serving engine and a portable reference implementation. And the reported results are significant. Nockchain says Lockstep is implemented as a vLLM plugin, that five dense model families serve at or above stock throughput, and that Qwen3-30B-A3B reaches 1.32–1.37× stock throughput in its reported benchmarks. It also reports identical output digests across Hopper, Ada and Blackwell GPUs, with the reference implementation reproducing the results bit-for-bit on x86 and ARM. The contract core has been mechanized in Lean 4. That is the actual unlock. Different GPUs can run the same canonical inference and arrive at the same result. Once the computation is canonical, it can be proved. Once it can be proved, the customer can independently verify it. The key idea is not “Nock runs AI.” It is: make the inference deterministic, then prove it. 6. Private, deterministic, verifiable inference is the real target Put the pieces together and the product becomes much clearer. A customer has a private inference workload. It goes to an external provider. The provider runs the canonical computation on supported hardware. Lockstep's reported implementation demonstrates deterministic inference today. The proving layer can then be used to generate cryptographic evidence for the computation without requiring the customer to expose the entire underlying workload. This is the architecture Nock is trying to deliver: private, deterministic, verifiable inference. There is an important distinction here. The deterministic part exists today in the Lockstep implementation. The ZK-PoW proving layer exists today as the base consensus mechanism. But Nock has not yet shipped the final general-purpose ZKVM and complete verifiable-inference market described in the long-term thesis. That is exactly why the Q4 2026 and Q1 2027 milestones matter. The customer proposition is therefore a destination, not something that should be written as though it is already fully deployed. The goal is straightforward: the customer should not have to trust the GPU provider's word about what happened. The computation itself should provide the evidence. 7. The competitive landscape is broader than AI There are other teams working on useful computation, verifiable inference and ZK-based AI infrastructure. That is healthy competition, but I think the Nock differentiation is easy to miss when everyone is described using the same vocabulary. The important question is not simply who can say “AI compute,” “Proof of Useful Work” or “verifiable inference.” It is who can make the complete system practical. You need deterministic computation. You need an execution environment designed for efficient proving. You need low enough overhead that customers will actually use it. You need hardware that can perform the workload. You need a mechanism for generating and verifying the proof. You need routing that connects actual demand to that compute. And you need economic incentives that keep the system running. Nock is unusual because those pieces are being developed together instead of being treated as unrelated products. Its published research traces that progression from EDEN and the Nock instruction set, to the verifiable VM framework, to ZK-PoW consensus, and finally to Lockstep. That does not guarantee that the architecture wins. It does mean the competition should be evaluated at the stack level, not by comparing AI branding. 8. Why the TEE model does not solve everything Trusted execution environments will remain useful. But they make a different security tradeoff. A TEE asks you to trust a hardware-enforced execution boundary and its attestation mechanism. Cryptographic verification asks you to verify the computation itself. That distinction matters because TEEs have historically had vulnerabilities in exactly those trust boundaries. Foreshadow demonstrated extraction of SGX enclave secrets and the ability to forge local and remote attestation responses in vulnerable Intel SGX implementations. CacheWarp later demonstrated a software-based fault attack against AMD SEV-ES and SEV-SNP, including recovery of an RSA private key in one case study. None of this proves that TEEs are obsolete. It illustrates that their security model depends on assumptions about hardware, implementation and attestation. The Nock thesis is trying to reduce those assumptions for verifiable inference: rather than asking a customer to trust the execution environment, give the customer a cryptographic proof of the computation. That is a fundamentally different model. And it is why Nock does not need to win the argument that TEEs are “bad.” It only needs the economics of proving to become good enough that, for certain high-value workloads, customers prefer proof over trust. 9. The network is live. Q4 and Q1 are where the thesis gets tested Nockchain mainnet went live on May 21, 2025. In April 2026, Nockchain reported that the network had generated more than one billion zero-knowledge proofs, describing that first phase as the bootstrapping of a live proving market. Now the interesting part is phase two. The AI Compute Network was marked completed in August 2026. Nockchain describes it as an AI proof-of-work network that allows inference providers to merge-mine their existing workloads against Nockchain at negligible marginal cost. The next milestone is Q4 2026: Useful Inference Routing Protocol, which is explicitly designed to funnel compute demand into the AI Compute Network. Then comes Q1 2027: Full Nock ZKVM Completion and the ZK Verification Opcode. Nock describes the former as completing the full ZKVM for performant general-purpose zero-knowledge proofs of arbitrary computation, while the latter is designed to let fully verifiable applications run directly onchain. The roadmap then points toward generalized ZK Compute Markets in Q4 2027, extending verifiable workloads beyond AI to any provable computation traded on Nockchain. That sequencing is important. The proving network exists. The deterministic inference layer exists. The AI Compute Network exists. The next job is connecting that infrastructure to real demand and then completing the general-purpose proving environment. That is what Q4 and Q1 are really about. 10. This is where the BlackRock report becomes relevant BlackRock's September 2026 The Machine-Native Economy is unusually relevant to this thesis because it explicitly connects AI agents, programmable settlement and compute markets. BlackRock describes AI as “machine-native intelligence” and digital assets as “machine-native money.” It argues that compute is becoming a large and increasingly investable economic resource and says standardized claims on compute capacity could become a meaningful digital-asset use case for financing and programmable settlement. Its paper cites approximately $1.1 trillion of combined AWS, Microsoft Intelligent Cloud and Google Cloud revenue by 2030 as a rough proxy for the scale of the compute market. It also describes a future where agents discover, compare and provision compute based on price, performance, latency, location and hardware specialization, potentially using programmable payment rails. Again, BlackRock is not validating Nockchain. It is describing the macro environment. Machines will increasingly need to buy compute from other machines. Compute will increasingly become an economic resource that can be priced, allocated, financed and settled programmatically. Nock is approaching the next layer of that problem: how do you prove the computation was actually performed as specified? That is why the BlackRock report strengthens the thesis without having to mention base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722 at all. 11. The base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722 bull case is the convergence Look at the individual developments and the opportunity is easy to miss. Europe is building privacy-preserving age verification and has incorporated zero-knowledge proofs into its latest blueprint. The BIS has demonstrated cryptographic compliance proofs for cross-border finance. Hester Peirce is publicly arguing for attribute-based credentials and zero-knowledge verification to reduce unnecessary financial-data collection. NIST is working on identity, authorization, auditing and non-repudiation for autonomous software agents. Gartner expects inference spending to surpass training spending in 2026. McKinsey sees potentially $3 trillion to $5 trillion of global B2C retail commerce being mediated by AI agents by 2030. And BlackRock is now explicitly describing compute as a potentially large digital-asset market within a broader machine-native economy. Nockchain is approaching that convergence from the infrastructure side. The network is live. More than one billion proofs have been generated according to Nockchain's published materials. The AI Compute Network launched in August 2026. Lockstep has attacked one of the hardest practical problems in verifiable inference by demonstrating bit-exact transformer inference across heterogeneous hardware in its reported implementation. Useful Inference Routing is planned for Q4 2026, while Full Nock ZKVM Completion and native ZK verification are planned for Q1 2027. The roadmap ultimately points toward generalized ZK Compute Markets in Q4 2027. Then there is base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722. The current Nockchain research page specifies a hard cap of 2^32 units, minted through ZK-PoUW + AI-PoUW, with no premine and no founder allocation. The whitepaper also makes clear that phase two introduces a protocol-fund allocation from block emissions, so the accurate claim is not “all emissions go directly to miners forever”; it is that the coin remains capped and the allocation structure does not include a premine or founder allocation. That is where the economic thesis becomes interesting. If proving becomes a service customers actually pay for, if inference providers contribute hardware to the network, if useful computation creates demand for proof generation, and if the eventual ZKVM turns Nock into infrastructure for a much broader set of workloads, then base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722 sits underneath the network coordinating that proving economy. That is the potential flywheel. More useful computation creates more demand for proving. More demand for proving creates more incentive to contribute hardware. More proving capacity makes the network more useful. More useful applications create more demand for the underlying infrastructure. And the entire system becomes increasingly valuable as the machine economy becomes more dependent on outsourced computation. The end user may never know Nockchain exists. An agent could simply request an inference and receive a result plus proof. An enterprise could outsource a private workload and receive evidence that the specified computation was executed. A financial application could prove that a condition was satisfied without receiving all of the underlying customer information. A machine could purchase compute from another machine and eventually verify that the computation it paid for was actually delivered. That is why I would not frame the thesis as “AI + crypto.” The stronger thesis is private, deterministic, verifiable computation, with AI inference as one of the first enormous markets and the Nock zkVM underneath it. Lockstep is critical because deterministic inference is a prerequisite for clean verification. The NockVM and proof-oriented language stack matter because the proving system needs an execution environment designed around efficient verification. ZK-PoW matters because zero-knowledge proofs are already embedded into consensus rather than being bolted on afterward. Useful Inference Routing matters because the next step is connecting that proving capacity to real demand. And Q1 2027 matters because the full ZKVM and native verification layer are what start opening the door to the broader set of primitives beyond inference. The market is watching who builds the biggest AI compute stack. I think the more interesting question is who builds the infrastructure that can prove exactly what those machines computed, without requiring the customer to simply trust them. That is the market Nockchain is trying to build. And that is why base:0x9b5e262cf9bb04869ab40b19af91d2dc85761722 could ultimately be much bigger than “another crypto project tied to AI.” nockmilio.