NVIDIA dropped the Jetson T4000 module this week. Blackwell architecture, 1,200 FP4 TFLOPS, 64GB of unified memory, configurable 70-watt envelope, $1,999 at 1,000-unit volume. That is roughly four times the performance of the previous-generation Orin AGX with double the memory, for about $500 more. It is the new ceiling for "what fits at the edge," and for most builders working in physical AI it is the most important hardware announcement of the year.
I am not buying one. Not for the next basement. Not for the next community center. Not for the next sump pump. Here is the practitioner economics, and the portfolio architecture where the T4000 actually does fit.
The headline numbers
The T4000's spec sheet is doing two things at once. First, it raises the inference ceiling — a quantized 7-13B-parameter vision-language model fits on this module with room left over. Second, it does it inside a 70-watt power envelope that runs on a normal 15-amp circuit with no electrical work. That combination did not exist before this announcement.
| Module | Price (1Ku) | Memory | Power | Notes |
|---|---|---|---|---|
| RPi 5 + Hailo-8L | ~$200 | 8GB sys | ~10W | Sufficient for sensor fusion + small vision models |
| Jetson Orin Nano 8GB | $499 | 8GB | 15W | What we deploy per building today |
| Jetson Orin AGX 64GB | ~$1,500 | 64GB | 60W | Previous-gen ceiling for VLM at the edge |
| Jetson T4000 (Blackwell) | $1,999 | 64GB | 70W | 4x perf, runs modest VLA models comfortably |
For a single-asset deployment — one sump pump, one small commercial building, one custom industrial monitor — that price step is not justified by the workload. A Jetson Orin Nano at $499 runs the model that watches the basement camera, fuses the sensor packs, and runs the local language layer that explains what just happened. It does that today. It has been doing it for two years on the agentic sump pump.
The cost of the T4000 is not the $1,999. The cost is the $1,500 of payload it asks you to put on top of a $99-a-month service contract. With a 24-month customer lifetime, $1,500 of capex per building means the first 15 months are at zero margin. That is not a deal I will sign for a single building. It is also not a deal a small commercial owner will sign — they did not call you because they wanted to underwrite NVIDIA's R&D depreciation curve.
The Salesforce signal
The story NVIDIA buried under the hardware announcement is the bigger one. Salesforce integrated NVIDIA's new Cosmos Reason model and cut incident resolution time 2x on operational video. That is a production deployment of a video-understanding model on real operational data, by a customer that has nothing to do with robotics, that came in measurably under the previous baseline.
This is the architecture I have been running in basements for two years. A camera feed, a vision-language model that knows what it is looking at, a structured output, an alert. Salesforce calls it incident resolution. We call it "why the sump pump is making that noise." Same shape. Four orders of magnitude apart in price.
What changed this week is that Cosmos Transfer 2.5, Cosmos Predict 2.5, Cosmos Reason 2, Isaac GR00T N1.6, and Isaac Lab-Arena were all open-sourced. The model layer for physical AI just commoditized. A practitioner can now download the same Cosmos Reason that Salesforce is using and run it locally on a Jetson Orin against their own customer's camera feed, without paying per-inference cloud fees. That changes the operating cost of a small monitoring service, not the capex of the edge module.
Where the T4000 does fit
The T4000 is not the wrong module. It is the wrong module for one building. The right deployment shape is a "site brain" architecture: one T4000 serving a small portfolio.
The portfolio version looks like this. Each building runs an Orin Nano (or RPi5 + Hailo) doing the always-on sensor fusion, the camera triggers, and the immediate-action local control. A T4000 — on-prem in a maintenance office, or co-located at the largest building in the portfolio — runs the heavier reasoning passes: video summarization, weekly per-building model retraining, root-cause analysis across multiple buildings, the long-context conversational layer the customer talks to. The Orin Nodes report up to the T4000 over the customer's LAN or a small mesh; the T4000 keeps the data on premises and never sees the cloud.
The economics of that architecture make sense at three or more buildings under one ownership group, or at one building large enough to justify a dedicated maintenance office. A 50-unit apartment portfolio. A small school district with five buildings. A funeral home group with three locations. A property manager running 12 small commercial sites in one zip code. All of them have a budget for "the site brain" that one $199-a-month building does not.
And critically — none of those customers need the T4000 right now. The Orin Nano deployment we run today is already doing the work. The T4000 becomes interesting when one of those portfolio customers wants to add VLA-based root-cause analysis across the fleet, or wants to keep all video data on-premises for compliance reasons, or wants to run a local conversational agent over their entire building inventory. Those are upsell triggers, not opening offers.
The 80-percent threshold
The other physical-AI announcement this week was Boston Dynamics integrating Google DeepMind's Gemini Robotics-ER 1.6 into Spot. Spot now reads gauges and sight glasses autonomously, looks for spills and debris, and applies vision-language reasoning to industrial inspection routes. Boston Dynamics has several thousand Spot units deployed at paying customers. The reasoning upgrade is real, and a meaningful test of whether embodied AI is production-ready.
The line in the announcement that practitioners should pin to the wall is this: operators need at least 80 percent reliability before tasks stop feeling like "the robot is crying wolf." Below that threshold, alerts get ignored, trust collapses, and the deployment quietly dies whether or not the model is technically accurate.
I have seen exactly that failure mode in small buildings, where the supplier-side BMS shipped a generic alarm template and the maintenance team turned the email forwarding off in the third week. The model did not need to be more accurate. The alarm prioritization layer needed to exist. Last week's post on the construction loophole covered the same idea from the operations side — QuadReal runs a centralized alarm list standardized across 60 properties, and the standardization is the product, not the AI.
The T4000 will not move the alarm-trust dial. A one-page document that names which conditions trigger a phone call, which trigger an email, and which sit in a monthly-review queue will. That document costs nothing to produce. It is the part of the deployment that the supplier side is not selling, and the part that determines whether the customer renews.
What I'm actually doing this week
Three concrete changes coming out of this week's announcements:
- Pulling Cosmos Reason 2 onto a test Orin Nano against the basement camera feed. The model is open. The current local pipeline runs a smaller vision model. The honest question is whether Cosmos Reason on the Orin Nano (likely with quantization) materially improves the per-event explanations the customer sees. If yes, that is a software upgrade for the same hardware — zero capex change, better product. If no, the deployment story stays as it is.
- Building a portfolio-tier proposal template that names the T4000 as the optional site-brain upsell. The single-building proposal does not change: $99-199/month, Orin Nano per building, retrained on the customer's data, prioritized alarm list. The portfolio addendum names a one-time $2,499 site-brain install plus a $99/month managed software fee for cross-building analysis. That is a real product, priced honestly, available when the customer wants it.
- Adding the 80% reliability threshold to the landing page. Boston Dynamics is now the public benchmark. We can cite their number, and explicitly position the alarm-prioritization layer — the one-page document that says what triggers what — as the part of the service that gets us past the threshold and keeps us there. That is the part of the deal that the customer is actually paying for. The hardware is the substrate.
The bottom line
The Jetson T4000 is the most important edge-AI hardware announcement of 2026. It will absolutely show up in our deployments — but not in the first building. It shows up as the optional shared site-brain for the customer who already has three buildings on the service, who wants local-only video reasoning, who wants weekly per-building retraining, who wants to talk to a local conversational agent about the entire portfolio. That is a real product. It is not the opening offer.
The opening offer is still a $499 Orin Nano per building, $99-199/month, retrained on that building's own data, with a prioritized alarm list documented on one page. The Salesforce result this week — 2x incident resolution improvement using the same architecture at corporate scale — is the validation. Cosmos Reason being open-sourced is the operating-cost improvement. The T4000 is the headroom for what we sell next year, to the customer we earn this year.
The interesting move in physical AI right now is not buying the most expensive new edge module. It is recognizing which workload belongs on which tier, and pricing the tiers honestly. That is what small-building practitioners do well. The hyperscaler launch cycle just made it easier to do it cheaper.
We deploy physical AI on the right module for your building.
$499 edge node per building, retrained on your data, prioritized alarm list documented on one page. Optional shared site-brain for portfolios. Local inference, local action, local logs. Under $500 to start a pilot.
See What We BuildRead the case studies and related posts: How edge AI prevented a basement flood | 42% off energy costs on a community center | Closing the construction loophole on a small building | The edge AI hardware wave