The series changes altitude
For ten issues this series built downward into one problem: a reference condition-monitoring stack that ingests a Sparkplug B stream, scores vibration with an Isolation Forest, gates the false alarm, stores a long feature history, alerts, and shows the floor a dashboard, all on one Hetzner CX22 for a few dollars a month. That build is finished enough to stand on, and it is the reason this issue can change altitude and look up.
The question that has grown louder under the whole industry through 2026 is not how to build a monitoring stack. It is whether the AI being sold to manufacturing this year works at all, and what it costs when it leaves the demo. This issue, and the franchise from here, takes that question from the floor rather than the keynote. The credential for doing so is the build: a publication that has assembled the small version of machine-learning-on-a-machine has the standing to read the large pitch and say which parts survive contact with a plant.
The pitch, stated fairly
The pitch is not empty, and the honest broker states it at full strength before testing it. Adoption is real: by Rockwell's 2026 State of Smart Manufacturing survey about a third of manufacturing operations, 34 percent, are AI-augmented today, with manufacturers expecting more than half by 2030, and Deloitte's 2026 Manufacturing Industry Outlook expects agentic-AI adoption to roughly quadruple, from about 6 percent to 24 percent of manufacturers, even as only one in five reports being equipped to scale it. The capability being sold is a genuine step past the anomaly detector this series built. An agentic system is pitched not to flag a deviation but to act on it: read the alert, open the work order, adjust the schedule, trigger the supplier follow-up, across systems, without a human in every loop. The vendors at Hannover Messe and CES 2026 extend it into physical AI, robot cells trained in simulation and dropped onto the line, and into industrial foundation models like the multimodal one Siemens launched in April 2025, trained on CAD/CAE files, sensor time-series, and automation code rather than on the open web.
None of that is vaporware. The demos run. The question the honest broker asks is not whether the demo runs but whether the thing reaches the income statement of a plant that is not the demo, and three numbers answer it.
The first number: 95 percent reach nothing
MIT's NANDA study, The GenAI Divide, found that for 95 percent of companies, enterprise generative-AI pilots delivered little to no measurable return to the income statement, with only about five percent achieving rapid revenue acceleration. The pilots ran and demoed. What most of them did not do was move money the CFO could find. Gartner's projection runs parallel: 60 percent of AI projects will be abandoned through 2026 for want of AI-ready data, and at least 30 percent of generative-AI projects are dropped after the proof of concept.
The reason matters for this series specifically, because the missing data foundation is the thing the first ten issues were spent building. A pilot with no consistent, retained, queryable record of what the machine did cannot scale into anything, because the model at scale needs the history the pilot never kept. The series built that foundation cheaply and on purpose: the tiered InfluxDB layout, the retained feature bucket, the backup. The agentic pitch assumes the foundation is already there. On most plants it is not, which is most of why the 95 percent is 95 percent.
The second number: 17.7 percent
The sharper number is newer. FactoryBench, published in 2026 (Merzouki et al., arXiv:2605.07675), is the first benchmark built to test whether a large language model can read industrial machine time-series rather than retrieve a manual. It presents models with real robot telemetry, about 71,000 question-answer pairs across 27 fault types and more than 100 sensor channels from a UR3 cobot and a KUKA arm, and grades them up a ladder from reading a value to diagnosing a root cause.
The best current models did poorly in a way that is precisely diagnostic. On structured reasoning the strongest model, Claude Sonnet 4.6, scored in the mid-forties across all three levels, 46.8, 47.1, and 45.9 percent, never clearing fifty. On the level that matters on a floor, deciding the root cause and the action from the signal, the best model, GPT-5.1, scored 17.7 percent. The benchmark's own conclusion is that current language models fundamentally cannot understand the time-series signal of an industrial machine, and that retrieving a protocol document and interpreting sensor data are separate competencies these models conflate.
That is the empirical floor under the agentic pitch. An agent that opens the right work order from an alert is doing document retrieval, which the models are good at. An agent that decides from the vibration trace what is wrong with the bearing is doing the 17.7-percent task. The series reached the same conclusion the cheap way in issue 06: the thing that reads a vibration signature and flags the fault is not a language model at all, it is a small purpose-built anomaly detector in eighty lines of Python, and FactoryBench is the formal confirmation that this was the right tool and a frontier LLM is the wrong one.
The third number: ten thousand becomes five hundred thousand
The last number kills the business case quietly. A generative-AI pilot that costs ten thousand dollars a month becomes five hundred thousand a month at fifty times the scale, because the pilot was priced on a sliver of the plant and the token bill, the inference bill, and the integration cost all scale with it. A proof of concept that looked cheap at one cell is a different conversation across forty, and that conversation usually ends the project.
This is the failure mode the series has argued against since issue 02 in a different vocabulary. The whole point of the few-dollar stack was that the bill does not move when the plant grows, because the model runs locally on hardware the plant owns and there is no per-query meter. A cloud agentic platform inverts that: the demo is cheap and production is metered. Cost transparency is not an accounting footnote in industrial AI. It is the difference between a system a plant keeps and a pilot a plant cancels in the second budget cycle.
What works on the floor is the small model, with an asterisk the small-model camp omits
The constructive half of the issue is that there is a version of AI in manufacturing that does reach the P&L, and it is close to the opposite of the cloud agentic pitch. Dell's vice chair has called 2026 the year of the small language model, and Gartner projects organizations will use small, task-specific models three times as much as general-purpose LLMs by 2027. The floor profile of those models is specific: low single-digit to a few billion parameters rather than hundreds of billions, fifteen to forty watts rather than the draw of a datacenter GPU, and running offline on a ruggedized edge device with no cloud API bill at all.
Read that profile against the stack this series built and it is close to the same machine. The detector is a small, task-specific model. The whole thing runs on-premises, deterministic in its bill, owned. The series did not set out to build the 2026 edge-AI thesis, but it built one instance of it, because the constraints that produce a cheap working monitoring stack, local compute, a bounded bill, a narrow task, an owned data foundation, are the constraints that separate the AI that ships from the AI that demos.
The asterisk is that the small-model camp sells its own conflation, and an honest broker has to mark it on both sides. The headline edge figure, a model that flags bearing wear in eight milliseconds on a two-thousand-dollar box, measures a single inference pass, not autoregressive generation. The source calls it a three-billion-parameter model, but an eight-millisecond end-to-end latency is incompatible with token-by-token generation, which runs at the single-digit rates below and would need hundreds of milliseconds to seconds to produce a useful answer. Eight milliseconds is one forward pass of a classifier reading a signal, the kind of small model the i.MX 8M Plus and its 2.3-TOPS inference NPU were built for, and it is not the throughput number for a generative reasoning agent. On published 2026 edge-inference benchmarks a one-to-three-billion-parameter language model returns single-digit to low-double-digit tokens a second on edge hardware: an NVIDIA Jetson Orin Nano returns about six tokens a second on a 3B model, smaller one-to-1.5-billion-parameter models on the same board reach roughly nine to thirteen, and a Raspberry Pi 5 with an NPU hat runs one to six. That throughput is fine for drafting an offline work order in the background and nowhere near a real-time agent reasoning in the loop, and the hardware that does the eight-millisecond single-pass inference is often not the hardware that runs the generative model at all. The trick the agentic vendor plays at the cloud altitude, selling document retrieval as machine understanding, the edge vendor plays one altitude down, selling a single-pass detection latency as a reasoning model's throughput. Match the model class to the task: a small classifier for the signal, a small language model for the paperwork, and neither one doing the other's job.
The honest synthesis is not that AI in manufacturing is hype. The adoption is real and the small-model case is real. It is that the part being sold loudest, the cloud agentic platform that reasons over the machine, is the part the numbers do not yet support, and the part that works, a small model doing one scoped job on owned hardware, comes with a quieter limit the floor needs to plan around. The plant that wins in 2026 buys the small scoped model, refuses to fund the reasoning-agent platform until the FactoryBench number moves, and reads every latency figure to find out which model class it actually describes.
Where this leaves the bill, and the beat
The stack is unchanged this issue and so is its modeled bill: a few dollars a month for the VM and the offsite backup from issue 10's predecessor work. The change is editorial. The series spent ten issues proving a small, local, owned approach to one problem, and the rest of the industrial-AI market spent 2026 proving the same lesson the expensive way across many problems. From here, Setpoint covers that intersection directly and independently: the AI-in-manufacturing claim of the week, taken from the floor, modeled against the stack and the published evidence rather than the slide.
What lands in issue 12
The natural first test of the new beat is to put a language model where the pitch insists it belongs and model what it adds. Issue 12 takes the existing alert pipeline and asks a bounded question: when a gated flag fires, can a small local model turn the raw alert and the machine's manual into a useful first-line work order, the document-retrieval task the FactoryBench results suggest it can actually do, and where in that same loop does it cross into the diagnosis task it cannot. The work-order quality is assessed against published retrieval-augmented-generation results for maintenance text, and the latency and cost are modeled from the same edge-inference benchmarks cited above, so the question of what the model earns its place doing is answered in numbers a plant can check rather than a vendor's demo.
Method. Figures in this issue are modeled from published third-party benchmarks and vendor specifications, with sources linked inline. Where a number describes the reference stack, it is a modeled figure for that configuration, not a measurement of a production deployment.