Devices & Robotics — W35: The agent moves to a box on your desk, and the buildout becomes a construction problem

Nobody shipped a humanoid this week. What shipped instead was the layer underneath every embodied system you'll deploy in 2027 — local inference hardware, an honest way to benchmark it, and the concrete-and-copper problem of building the rest.

Perplexity ships Portable Computer on NVIDIA DGX Spark. An agent runtime that executes on a desk-side DGX Spark box rather than a datacenter, with the sandbox enforced by the operating system instead of by prompt instructions, and zero metered token cost for any step that runs locally. That second detail is the product. Every enterprise agent pilot currently stalls on the same two objections — unbounded per-token spend, and the agent talking its way past its own guardrails — and this is the first credible hardware answer to both. No pricing or ship window established yet. Single-sourced this week — treat as reported, not confirmed. MarkTechPost

Liquid AI open-sources Pipette. A reproducible benchmarking suite that measures model, quantization scheme, runtime, and hardware as one combined system rather than varying one axis and freezing the rest. This is why vendor on-device numbers almost never survive contact with your actual target phone: the published figure was measured on a different runtime, at a different quantization, on different silicon. If you ship inference to handsets, laptops, or edge boxes, this is the week's most immediately usable thing. MarkTechPost

Evoke lands as an open-source world model with session persistence. It generates interactive environments and holds state across a session instead of regenerating from scratch each frame — which is the difference between a demo and a training substrate. Persistence is the whole requirement for simulation-based robot policy training: an environment that forgets what your manipulator just did teaches it nothing. Teams currently paying for proprietary environment generation now have a free floor. AI Search

AWS and Nvidia commit to 2 million additional GPUs. Plus a next-generation infrastructure tier, extending an already-enormous joint build. Read this as a hardware deployment schedule, not a market story: two million accelerators have to be physically racked, powered, and liquid-cooled somewhere inside roughly 18 months. The binding constraint has moved off the wafer and onto substations, land, and the specific contractors who can do medium-voltage and liquid cooling — a pool that is already booked. Corroborated. Nvidia newsroom

Nuclear-for-AI shrinks to fit the schedule. An investor writeup on Apollo Atomics laid out the arithmetic bluntly: gigawatt-scale nuclear runs 10+ years and roughly $20 billion, which is two model generations too slow. The bet is on smaller units sited next to the load. Practical consequence for anyone in the interconnection queue — expect behind-the-meter generation proposals rather than grid requests. Single-sourced, and it's an investor making the case for his own position. The AI Opportunities

Three independent pushes toward "run it on hardware you own" inside seven days — Perplexity's box, Pipette, and Nvidia's own local-AI developer materials — is no longer a coincidence, it's a category forming. Watch for a price and a ship date on the DGX Spark harness. Until local inference has a per-unit cost you can put in a budget line, the metered API keeps winning by default, however good the sandbox is.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.