I have written some version of the same sentence in this blog all year: the hardware keeps getting cheaper, so putting a smart monitor on a small building keeps getting easier. Vibration sensors went from $600 to under $50. A teraflop of edge compute now costs less than a doorbell. Every month the frontier spent billions proving that the cheap end works, and every month my bill of materials got a little friendlier.
This week that sentence stopped being entirely true, and I would rather say so than write around it.
NVIDIA announced two new Jetson modules on Wednesday. That is normally a story about a new ceiling. It wasn't. The new modules are smaller. Less memory, less compute, half the physical size, half the power. And the reason NVIDIA gave, in its own positioning, is cost: memory has gotten expensive, and this is how you keep building without paying for it.
When the company that sets the ceiling for edge compute starts shipping downward and calls it a feature, that is worth ten minutes of your attention. Here is what actually happened and what it means if you own a building rather than a robot.
The spec sheet is the news
For years every Jetson generation moved the same direction: more compute, more memory, more money. The T5000 is the flagship, 2,070 teraflops and 128 gigabytes of memory. The T4000, which I wrote about in May, is 1,200 teraflops and 64 gigabytes at $1,999.
The new T3000 goes the other way.
| Module | AI compute | Memory | Power | Size |
|---|---|---|---|---|
| Jetson T5000 (flagship) | 2,070 TFLOPS | 128 GB | 40-130 W | 100 x 87 mm |
| Jetson T4000 | 1,200 TFLOPS | 64 GB | 70 W | 100 x 87 mm |
| Jetson T3000 (new) | 865 TFLOPS | 32 GB | ~70 W | 50 x 87 mm |
| Jetson T2000 (new) | 400 TFLOPS | 16 GB | ~40 W | 50 x 87 mm |
Look at the T3000 line next to the T5000. It has a quarter of the memory and about 40 percent of the compute, in half the space, on half the power. And here is the claim that makes it interesting: NVIDIA says the T3000 delivers inference performance similar to the T5000 on real multimodal workloads, the language models and vision models and world models people actually run on these things.
Read that again, because it is an admission. NVIDIA is telling you that three quarters of the memory on its flagship was not buying you performance on the workloads you care about. It only became worth removing when it started costing real money.
Why: the memory market broke
The reason is not in the Jetson press release. It is in the DRAM market, and the numbers are ugly.
DDR5 prices rose as much as 110 percent in the first quarter of this year. DRAM overall is up roughly 172 percent year over year. Server memory is on track to double year over year by late 2026. AI is projected to eat about 20 percent of the world's entire DRAM wafer capacity in 2026, and IDC has supply growth at 16 percent, below historical norms, with the shortage running into 2027.
The part that reaches my basement: this is not just server memory. The memory in an edge module or a single-board computer is LPDDR5X, the same class of memory in your phone, and it is being pulled into the same crunch, partly by NVIDIA's own appetite for it in the Grace and Vera chips. Every wafer that becomes a memory stack for a rack GPU is a wafer that does not become a module for a pump monitor. Analysts have mid-tier and premium device bills of materials rising about 25 percent as a result.
So the T3000 is not a product decision. It is a supply-chain decision wearing a product costume. NVIDIA looked at what 128 gigabytes of LPDDR5X costs now, looked at what it was actually delivering, and cut it.
What this changes for a small building
Here is where I have to update something I have been saying, and where the update turns out to be good news for the way I build.
When I put a monitor on a sump pump, it runs a small model on a small board. Eight gigabytes. I have always defended that on principle: you do not need a bigger brain to watch one machine. The model only has to learn what that pump normally does and notice when the pattern drifts. Everything above eight gigabytes was paying for headroom the job never used.
That was an argument about elegance and cost. This week it becomes an argument about exposure.
Memory is now the one input in my bill of materials that is scarce, appreciating fast, and impossible to substitute. You cannot engineer your way around needing DRAM. You can only need less of it. A design with an eight-gigabyte footprint is barely touched by a market that doubled. A design that needs thirty-two or sixty-four gigabytes is fully exposed, and its cost is repricing underneath it right now, this quarter, whether or not the vendor has told the customer yet.
The honest half: I am not immune, I am only less exposed. My boards and sensor packs are built out of the same market. If a $499 module becomes a $600 module, that is real. A monitor with hardware under $3,000 still works, and a service contract at $99 to $199 a month still works, but I would be lying if I told you my costs are frozen while everyone else's climb. What I can tell you is that a design needing eight gigabytes absorbs a memory shock that a design needing sixty-four does not survive without a price change. That is the whole difference, and it is the difference between a quote that holds and a quote that gets revised.
And don't wait for the new modules
One more practitioner note, because a splashy hardware week generates a lot of "should we hold off until" conversations.
No. The T2000 and T3000 ship in the first quarter of 2027. That is six months away, with no announced price, in a memory market nobody can forecast past next quarter. NVIDIA also announced Cosmos 3 Edge the same week, a four-billion-parameter model for on-device reasoning that you can adapt to a specific sensor in about a day. It sounds great. Its weights are not released. The larger Cosmos models are genuinely open and downloadable today, which I respect, but the edge variant that a practitioner would actually want is a "coming soon" aimed at modules that do not exist yet.
Neither announcement watches a pump this quarter. In May I turned down the $1,999 T4000 because $1,500 of capex on top of a $99-a-month contract means the first fifteen months run at zero margin, and a building owner did not call me because they wanted to underwrite NVIDIA's depreciation curve. The T2000 is honestly closer to the module I want: sixteen gigabytes, forty watts, half the size. I still will not buy a roadmap. I will buy it when it has a price and a ship date. Until then, the $499 board already in the basement runs the model, catches the drift, and has been doing it for two years.
What to take from this if you own the building
- The cost curve you have been waiting on has stalled. If your plan was to monitor that equipment "when the hardware gets cheaper," the compute got cheaper and the memory just got much more expensive. Waiting is no longer free.
- Ask any monitoring vendor what their box needs. If the answer is thirty-two or sixty-four gigabytes to watch a pump, that bill of materials is repricing under them, and either their margin or your renewal absorbs it. NVIDIA just cut its own flagship by three quarters for exactly that reason.
- Small footprint is now the durable design, not the budget one. One cheap sensor, a small model that learns one machine's normal, eight gigabytes, in your building. It was the cheap answer last year. This year it is also the stable one.
Last week I wrote that the frontier keeps proving, on its own dime, that a small model reading one cheap sensor beats an expensive rig. This week it proved something less flattering and more useful: the cheap-hardware tailwind that carried all of us is not a law of nature. It is a market, and the market just turned. The architecture that survives that is the one that never needed much to begin with.
A monitor that needs eight gigabytes doesn't care what DRAM costs.
Each critical asset, your compressor, your boiler, your pump, gets one off-the-shelf sensor and a small model running on a local board. It clamps on in an afternoon, learns that machine's normal, catches drift three to six weeks before a failure, and hands it to a person who knows the equipment. Nothing leaves the building. $99 to $199 per month, hardware under $3,000.
See how it worksSources: NVIDIA Jetson T3000 and T2000 module announcements, July 16 2026 (T3000: 865 FP4 TFLOPS, 1536-core Blackwell GPU, 8-core Arm Neoverse, 32GB LPDDR5X at 273GB/s, 25GbE, ~70W; T2000: 400 TFLOPS, 1024-core GPU, 6-core Neoverse, 16GB LPDDR5X at 137GB/s, 2x10GbE, ~40W; both ~50x87mm vs ~100x87mm for T4000/T5000; available Q1 2027; pricing undisclosed; NVIDIA claims T3000 inference performance similar to T5000 on multimodal workloads and positions migration as cost reduction amid elevated memory prices) via CNX Software, ServeTheHome, Embedded.com, IoT Tech News and NVIDIA. Jetson T5000 (2070 TFLOPS, 128GB LPDDR5X, 40-130W, $3,499 AGX Thor developer kit) and T4000 (1200 TFLOPS, 64GB, $1,999 at 1Ku) via NVIDIA and CNX Software. NVIDIA Cosmos 3 Edge (4B parameters on a dense 2B transformer, on-device vision reasoning and action generation, post-training for a specific embodiment in ~1 day; Cosmos 3 Nano and Super weights available on Hugging Face under OpenMDW 1.1 permitting commercial use, Edge variant not yet released) via NVIDIA Newsroom, NVIDIA Technical Blog and the Cosmos 3 technical report. Memory market data (DDR5 +110% in Q1 2026; DRAM +172% year over year; server DDR5 projected to double YoY by late 2026; AI consuming ~20% of global DRAM wafer capacity in 2026; IDC 2026 supply growth 16% DRAM / 17% NAND, below historical norms, shortage persisting into 2027; ~25% BOM increases on mid-tier and premium devices; LPDDR5X pulled into the crunch by NVIDIA Grace/Vera demand) via IDC, Tom's Hardware, Network World and Edge AI and Vision Alliance. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building.