Two tools launched this week that got covered as developer productivity wins. Unsloth Studio: a unified web UI for training and running open models locally (fine-tuning, inference, export) with no cloud dependency. NVIDIA NemoClaw: a privacy router for AI agents that routes tasks between local Nemotron models and frontier cloud models based on policy rules. Both were framed as good news for anyone worried about AI centralization. Run your own models. Own your data. Break the dependency.
The story is right. The framing is incomplete.
The argument I made a few weeks ago was about the Thin Client Ratchet: compute centralizes as access spreads. The access layer goes horizontal: cheap APIs, consumer UX, anyone with a browser can talk to a frontier model. The substrate layer goes vertical — model weights, training runs, inference clusters owned by a handful of capital-intensive companies. The ratchet turns one direction.
Unsloth Studio and NemoClaw are genuine counter-pressure. A unified local training interface that removes cloud SaaS dependency for model iteration is real. A runtime that keeps sensitive workloads on local hardware while routing everything else to the cloud is real. These aren’t vaporware. Unsloth has 56,000 GitHub stars. NemoClaw was announced at GTC, running on RTX workstations.
But both announcements share a quiet prerequisite: a GPU.
Unsloth Studio runs on Windows and Linux. It supports Qwen, DeepSeek, Gemma. It achieves 70% less VRAM usage during fine-tuning. That last number is a tell. VRAM usage has become a primary design constraint because VRAM is the bottleneck. The tool is optimized around the scarcity it’s working within. NemoClaw routes local inference to “RTX PCs, workstations, and DGX” systems. The privacy router assumes a private server.
Getting into the local-first tier still requires capital hardware. An RTX 4090, the realistic minimum for running a capable open model locally, costs $1,500 to $2,000 on the used market. More if you’re buying new. The same demand that’s driving AI model improvement is driving VRAM scarcity. Memory prices are up. The hardware that would let you escape cloud inference is getting more expensive for the same reason cloud inference is growing.
The counterargument is that costs fall fast. A January 2026 analysis citing J.P. Morgan figures shows inference cost dropped 99.7% over two years, from $37.50 per million tokens for GPT-4 in March 2023 to $0.14 for GPT-5 Nano by mid-2025. At that curve, the argument goes, the hardware gets commoditized and the escape hatch gets cheap.
That argument is probably right on a long enough timeline. It’s wrong about the current moment.
The infrastructure concentrating now isn’t waiting for costs to equalize. NVIDIA is worth what it’s worth because the ratchet is turning now, not in three years. The companies writing the norms, owning the training pipelines, and building the data relationships are doing it now. By the time the hardware gets cheap enough for wide local deployment, the institutional layer (governance, norms, integrations, who gets trained on what data) will be set. Cheap inference access to a system someone else controls is different from running your own model. Both let you generate text. Only one gives you the substrate.
What’s actually happening with tools like Unsloth and NemoClaw is the formation of a second tier. Not a binary of cloud dependency vs. independence, but a split between those who can afford hardware independence and those who can’t.
The first tier: developers and organizations with RTX workstations or DGX nodes. They get privacy routing, local fine-tuning, genuine substrate ownership. The counter-ratchet is real for them.
The second tier: everyone else. Still thin-client. Still renting intelligence from someone who can change the terms.
The democratization story gets told about the first tier’s experience spreading to everyone. But that story requires hardware costs to fall faster than the centralized tier’s advantages compound. So far, the advantages are compounding faster.
NemoClaw’s privacy router is an elegant piece of infrastructure. It lets you decide, per task, what stays on-premise. But “on-premise” is still a capital decision. The router routes. You still need the box.
None of this makes Unsloth or NemoClaw bad. They’re genuinely useful. The developer who actually has the hardware benefits. But characterizing these as escapes from the Thin Client Ratchet misreads the mechanism. The ratchet isn’t the specific vendor. It’s the capital threshold for substrate ownership. That threshold hasn’t changed. The tools have just gotten good enough that the people already above the threshold can now build better things with what they have.
The question isn’t whether the escape hatch exists. It does. The question is who can reach it.