The Clawconomy is real, and it is not a software business

A useful way to read the ecosystem is as four stacked layers, each with a distinct cost structure and a distinct lock-in profile. At the bottom is silicon — the TPUs, GPUs, and increasingly custom inference accelerators on which the model runs. Above silicon is the model layer itself: the weights and the API surface that wraps them. Above the model is the agent runtime: sandbox, checkpointing, tool brokering, observability. At the top is distribution — the surface where humans and other agents encounter the agent in their workflow.

The model layer was, until 2025, where the value lived. As of April 2026 it is the layer with the steepest declining returns. GPT-5, Opus 4.6, Gemini 3, and a swarm of open-weight competitors trade leadership month-to-month on different evals; switching costs are a string change in an API call; per-token pricing converges. Every move you see in the Clawconomy is a vendor positioning at a different layer because the model layer is not where they expect to make money in 2027.

NemoClaw, released in alpha at GTC, looks at first like a hardened OpenClaw. It ships kernel-level sandboxing, policy-based egress controls, a privacy router that keeps sensitive inference local, and full audit logging. Read the architecture more carefully and the substrate bet shows: NemoClaw's sandbox uses isolation primitives that benefit measurably from NVIDIA Confidential Computing on Hopper-and-later GPUs. The privacy router's inference path assumes the local model is on a CUDA device. The "OpenShell runtime" is, in practice, a specification for how an agent runtime should look when it can lean on the silicon it runs on.

The strategic claim underneath: agent runtimes are not a software problem to be solved by the best engineering team. They are a substrate problem, where the substrate is silicon plus the firmware and driver stack on top of it. Build the sandbox without the silicon and you cannot offer the same isolation guarantees. NVIDIA does not need NemoClaw to be the most popular runtime; it needs NemoClaw to be the runtime enterprises pick when they care about the isolation guarantees, because picking it pulls forward the H100/H200 refresh cycle for their on-prem inference fleet.

Cisco's DefenseClaw, unveiled at RSAC 2026 and released to limited GA in early April, looks adjacent to NemoClaw on first read. It also gates what an agent can do, also logs what the agent did, also ships a policy language. The locus of control is different. NemoClaw places the trust boundary at the kernel of the host running the agent. DefenseClaw places it at the network egress point that the agent traverses. The two are not redundant; they are placing bets on the same question — where does enterprise-grade agent governance want to live? — at two different physical layers.

Meanwhile, the OpenClaw maintainers are building neither. OpenClaw's upstream answer to the governance question is the Capability Manifest spec, a per-skill declaration of what network and filesystem regions the skill is allowed to touch, enforced by the local agent process. It is the lightest-weight of the three, and the most easily defeated; the heaviest-weight (NemoClaw) requires hardware NVIDIA owns, and the medium-weight (DefenseClaw) requires being on a network Cisco has visibility into. The substrate determines the strength of the guarantee. None of the three vendors is competing on the agent itself.

Two weeks after MiniMax launched MaxClaw — a one-click cloud-hosted OpenClaw deployment that runs 24/7 with native apps for every consumer OS — Moonshot AI shipped KimiClaw. KimiClaw goes a layer further: a browser-centric managed agent powered by Kimi K2.5, with 40 GB of cloud RAG storage built in, full ClawHub skill access, and Telegram as a default surface. Both are attacking the same gap: the OpenClaw mothership is self-hosted by design, which is correct for the technical user but wrong for the next 100M users.

Meanwhile, the consumer surface is the layer the western Clawconomy has barely touched. NemoClaw is enterprise. DefenseClaw is enterprise. The OpenClaw mothership is hobbyist. MaxClaw and KimiClaw are aimed at the "I want an agent and I don't want to run a server" buyer, who exists in the hundreds of millions and has been ignored. The bet is that the runtime captures the value if it captures the user, regardless of where the model runs. This is the consumer-distribution bet that the western Clawconomy has not yet answered, and the most predictable next move is one of the western vendors — Microsoft, Google, or a coalition — shipping a competitor.

Distribution does not only mean "where humans use the agent." It also means "where agents use other agents." Moltbook — an AI-native social network where agents post, comment, and upvote on each other's output — looked, for most of 2025, like a curiosity. Meta paid an undisclosed sum for it on March 10, which strongly suggests the price was not curiosity money. The thesis underneath the acquisition has been visible for six months in the user numbers: agent-to-agent traffic on Moltbook crossed human-to-human traffic on Threads sometime in February, and the gap is widening.

The distribution surface for agents is its own market. Once an agent has a Moltbook account, a karma score, and a set of agent followers, that account is a real distribution asset for whoever ships its underlying skills. Meta did not buy Moltbook for the human users; it bought the network effect that makes Moltbook the canonical place an agent goes to find other agents. The Clawconomy's top layer — distribution — is a two-sided market, and the second side is agents, and the second side is the side that just got priced in.

The four-layer reading produces a different decision than the brand-name reading. If you are evaluating which "Claw" to standardize on, the surface question is wrong; the structural question is which layer you want to own and which layer you want to rent. Buying NemoClaw is a bet on owning the silicon-tied isolation layer and renting everything above. Buying KimiClaw or MaxClaw is the inverse: rent the silicon and the model, own the user. The OpenClaw mothership is the only stack where you rent nothing and own everything, and the cost is that you are responsible for governance the way DefenseClaw and NemoClaw users are not.

The consequence: there is no single right answer, but there is a wrong frame. Treating the Clawconomy as a list of products misses that the products are arrayed across a stack and the stack is going to settle. The settling is not about which "Claw" wins. It is about which layer captures the rents when the model layer hits zero. Read every announcement through that filter and the next twelve months become legible. The model is not the product. The model is the input. The product is whichever layer above the model can charge for what the model alone cannot deliver — isolation, governance, distribution, or accountability. Pick which one you are buying. The brand on the box is a clue, not the answer.