Notes on Software Engineering Daily: The Hardware Bottleneck AI Can’t Fix, with Jason Hoch (co-founder of Nominal) talking to Kevin Ball.
The 30-Second Reality Check
AI agents are fast at software because code can be tested in seconds, for free. A hardware test means building the thing, driving it to a desert and fetching a hard drive back. No model speeds that up. Better data plumbing might.
Explain With Pictures
Software feedback loop Hardware feedback loop
(seconds, basically free) (days to months, $$$$$$)
+-------------+ +---------------------------+
| agent edits |<---------+ | design |
| code | | +-------------+-------------+
+------+------+ | |
| | +-------------v-------------+
+------v------+ | | supply chain (months) |
| unit tests | | +-------------+-------------+
| (2 sec) | | |
+------+------+ | +-------------v-------------+
| | | build + wire the test bed |
+------v------+ | | ($1M table, you have two) |
| pass/fail +----------+ +-------------+-------------+
+-------------+ |
+-------------v-------------+
Repeat 10,000 times | fly it / fire it (8 sec) |
before lunch. +-------------+-------------+
|
+-------------v-------------+
| pull the hard drive, |
| hope the config matched |
+-------------+-------------+
|
+-------------v-------------+
| "what happened?" |
| (a day or two later) |
+---------------------------+You can give an agent a test suite. You can’t give it a wind tunnel.
3 Hot Takes & Quotes
1. Software observability is built to forget. Hardware can’t afford to.
I always joked that as a software engineer, I was trained to think of my infrastructure like cattle, not pets. But for our customers, everything is a pet.
A rocket startup might spend years and tens of millions of dollars to get to an eight-second engine test, and that data is “your entire company”. Meanwhile, standard software observability is:
super different from software observability, where you tend to forget everything a week later.
We built a whole industry around P99s, 15-second buckets and retention policies, and then act surprised when that playbook falls over for people who can’t drop a single sample. Downsampling is fine when you can just redeploy. Less fine when the test took four years to get to.
2. LLMs know people writing about physics, which is a different thing from knowing physics.
it’s not thinking in terms of physics. It’s thinking in terms of humans who have translated physics into English
Code is the native language of a coding model, so it can check its own work. Ask it how a jet will behave and you get a well-written summary of other people’s summaries, which you then have to go and physically test anyway. And even a perfect design still waits on the real-world queue:
A lot of times, the supply chain is the supply chain in the world of hardware and you can’t shave off these months
No amount of tokens makes a part arrive faster. You’re still waiting on atoms, same as before.
3. Zoomed-out charts lie, and AI-built dashboards could lie with confidence.
How do you make sure that those zoomed out plots don’t tell a lie and don’t hide a data point that could be critical?
A multi-hour flight squeezed into one chart can hide the split second that matters. Nominal’s founder worries that AI-generated UIs lack that paranoia, leading to “compounding misunderstandings”. His safety bar is refreshingly clear:
I just think we’re willing to be pretty YOLO with OpenClaw, but we not want to with a rocket fire, or especially any flight that has a person inside the vehicle.
Vibe-code your to-do app. Leave the flight test dashboards to people who lose sleep over a missing data point.
The Post-Mortem / Verdict
Strip away the AI talk and the pitch is stuff software solved years ago. Catalogs, tagging, data lineage, hot and cold storage that tell the same story, and config that updates itself so a tired technician doesn’t garble a test flight. One customer thought all its data landed in network storage; Nominal found “only 10% of it even got to that point.”
My guess is hardware’s AI moment shows up after it has clean, labelled data to learn from. Until then the most useful thing a software engineer can bring to a rocket company is a decent data pipeline. Less exciting on LinkedIn, I know. As for letting agents design the rocket itself:
Not yet. Not yet. But let’s start with video games and then we can move on to physical systems. (Jason Hoch)
Links & References
- The episode: The Hardware Bottleneck AI Can’t Fix, Software Engineering Daily, 2 Jun 2026 (transcript)
- Jason Hoch (guest), co-founder of Nominal: X
- Kevin Ball (host), VP of Engineering at Mento: Website · X · LinkedIn
- Nominal, the hardware test data platform discussed in the episode