Manufacturing AI in 2026: downtime at $2.3M an hour and what to automate first

Summary. The most-quoted number in factory AI is Siemens' finding that an hour of unplanned downtime costs a large automotive plant $2.3 million, more than $600 a second, roughly double the 2019 figure. That number is real, and it is also survey data from 181 interviews, extrapolated. Read it as a direction, not a quote for your plant. The same Siemens report carries a figure almost nobody repeats: plants now suffer 25 unplanned downtime incidents a month, down from 42 in 2019, and lose 27 hours a month against 39 in 2019. Downtime is getting rarer and more expensive per hour, which changes what you should build first. Meanwhile the hardware stopped being the obstacle. An NVIDIA Jetson Orin Nano Super developer kit runs $249 and delivers 67 TOPS, up from 40 TOPS on the same board via a software update. For Indian plants there is one fact worth knowing before you build a business case: the Ministry of Heavy Industries told the Rajya Sabha on 13 December 2024 that under SAMARTH Udyog Bharat 4.0, "No financial assistance is given to any industry including MSME for adopting Industry 4.0 enabled technologies." The four SAMARTH centres give you assessment, training and consultancy. The capital is yours. Under the separate PLI schemes, 806 applications had been approved across 14 sectors against a ₹1.97 lakh crore outlay, with roughly ₹1.76 lakh crore of investment realised by March 2025. Here is what to build, what it costs, and which numbers to distrust.

The downtime number everyone quotes, and what it actually says

Every vendor deck opens with $2.3 million an hour. Almost none of them tell you where it comes from, so start there.

The figure is from the Siemens Senseye report The True Cost of Downtime 2024. Its methodology section is explicit: results cover April 2019 to March 2023 and are "based on 181 completed online interviews with maintenance, engineering and IT professionals at large industrial organizations" across automotive, FMCG, heavy industry and oil and gas. The Fortune Global 500 estimates were extrapolated from that sample using public information on plant counts and employee numbers.

So: 181 self-reported responses, four sectors, extrapolated to the world's 500 largest firms. That is a legitimate industry survey. It is not a measurement of your line, and a $2.3 million hour describes a large automotive plant, not a 200-person component supplier in Pune.

The report says so itself. For smaller manufacturers, it puts downtime costs "reaching $150,000 an hour at the top end" and notes their sharper problem is not the hourly cost at all. It is OTIF, delivery on time and in full, which their large customers monitor. Miss it often enough and you lose supplier status. For an Indian MSME supplying an OEM, that risk is worth more than any per-hour arithmetic.

One more thing about the source. Siemens sells predictive maintenance software, and the client-outcome numbers in the report come from live deployments of its own product. That does not make them false. It does mean they are the results the vendor chose to publish, and you should read a "50% reduction in unplanned machine downtime" as a best case from a self-selected set, not an expected value.

What the Siemens data actually shows

Sector	Cost per hour of downtime	Change since 2019	Annual cost, large plant
Automotive	$2.3 million	2x (up 113%)	$695 million
Heavy industry	not stated per hour	4x (up 319%)	$59 million
FMCG	$36,000	broadly stable	just over $10 million
Oil and gas	tracks the oil price	2023 similar to 2019	half of 2019 in 2023
All sectors, average large plant	doubled over five years	2x	$253 million

For context on the 113% and 319% rises, the report notes US price inflation totalled 19% across the same 2019-2023 window. Energy prices, not general inflation, drove most of the gap.

Now the number that should change your roadmap. Across the surveyed plants:

Incidents fell to 25 a month per facility, from 42 in 2019, a 41% drop. Hours lost fell to 27 a month, from 39. Annual hours lost per plant fell to 326, down nearly a third. In automotive and heavy industry, hours lost to unplanned downtime halved over five years.

But recovery got slower. Mean time to get production running again rose from 49 minutes in 2019 to 81 minutes. Siemens attributes that to three things: maintenance skills lost during the post-Covid hiring churn, longer lead times on emergency replacement parts, and a selection effect worth understanding. As the report puts it, Industry 4.0 going mainstream means "the things that cause downtime now are the more challenging problems that are harder to detect and take longer to fix."

Read those together and the strategy inverts. The easy failures are already caught. What remains is fewer, harder, longer, and more expensive per hour. Buying another vibration-sensor dashboard to catch the failures you already catch adds nothing. The value has moved to diagnosis and recovery speed, which is a different engineering problem than detection.

Where AI actually pays on a factory floor

Four workloads carry almost all the return. They are not equally ready.

Workload	What it needs	Realistic first result	Main failure mode
Visual defect inspection	Cameras, lighting rig, labelled defect images, edge box	Catch a known defect class the line already understands	Not enough defect examples; the rare defect is the expensive one
Predictive maintenance	Current, vibration and temperature data plus maintenance records	Ranked machine-health alerts on your worst 10 assets	Data exists but sits in a historian nobody can query
Process optimisation	Historical process data plus a controllable setpoint	Narrower variance on one parameter	No one will let a model touch the setpoint
Agent-assisted diagnosis	Manuals, past work orders, machine logs	Faster root-cause on the 81-minute recovery window	The manuals are PDFs and scans

Predictive maintenance is the most mature, and the Siemens data shows why it is also the most crowded: almost half of surveyed firms already have a PdM team, double the 2019 proportion, and 87% already gather data that makes PdM possible. If half your peers have this and downtime incidents have already dropped 41%, the marginal alert is worth less than it was in 2019.

That is why the fourth row is the interesting one in 2026. Recovery time went from 49 to 81 minutes while incidents halved. Nobody has a product for that. It is a retrieval problem over your manuals, work orders and logs, and it lands on exactly the 32 extra minutes per incident that got worse while everything else got better. At 25 incidents a month, 32 minutes each is roughly 13 hours a month of recovered production that current PdM tooling does not touch.

Note the constraint hiding in that row: the manuals are PDFs and scans. Retrieval over tables and scanned documents is where these systems break, which we cover in our guide to building a RAG knowledge assistant. The factory version is harder, not easier, because the documents are older.

What the hardware costs

Edge inference stopped being the budget line that kills projects.

Component	Price and spec (July 2026)	Note
NVIDIA Jetson Orin Nano Super developer kit	$249, 67 TOPS, 102 GB/s memory bandwidth	Down from a $499 list; 67 TOPS is up from 40 via software
Industrial camera and lighting, per station	Varies by vendor; specify before you budget	Lighting is usually the accuracy variable, not the sensor
Cloud inference, agent workloads	OpenAI gpt-5.6-luna, $1.00 / 1M input, $6.00 / 1M output	Cached input $0.10 / 1M
Cloud inference, heavier reasoning	OpenAI gpt-5.6-terra, $2.50 / 1M input, $15.00 / 1M output	Regional data residency adds a 10% uplift on eligible models

The Jetson number deserves a second look because of how it moved. NVIDIA raised the same board from 40 to 67 AI TOPS and memory bandwidth from 68 to 102 GB/s through a software update, and existing developer kit owners got the boost by upgrading rather than buying. A 1.7x performance improvement with no new hardware is not a normal event in industrial capex planning, and it is a reason to lease-or-wait less and pilot sooner.

At $249 a node, a ten-station vision pilot is $2,490 of compute. The plant will spend more than that on the lighting rig and the mounting brackets, and vastly more on the labelling effort. If your business case fails on edge hardware cost in 2026, the business case was never about hardware.

Where the money actually goes: labelled data, integration with the PLC and MES layer, and the people who keep it running after the pilot ends. The real cost is usually the integration, not the model.

India-specific considerations

Start with the fact that changes business cases, because most vendors will not tell you.

Under SAMARTH Udyog Bharat 4.0, the Ministry of Heavy Industries has established four Smart Advanced Manufacturing and Rapid Transformation Hub centres: the Centre for Industry 4.0 (C4i4) Lab in Pune, the IITD-AIA Foundation for Smart Manufacturing at IIT Delhi, I-4.0 India at IISc Bengaluru, and the Smart Manufacturing Demo and Development Cell at CMTI Bengaluru. Ten cluster experience centres were approved under a hub-and-spoke model, to be established by C4i4 Pune.

What they deliver is real. C4i4 Pune has compiled over 50 Industry 4.0 use cases, completed over 100 digital maturity assessments for auto companies, identified 500-plus improvement initiatives and trained 500-plus digital champions. CMTI Bengaluru has trained about 5,000 professionals. I-4.0 India at IISc has developed 6 smart technologies, 5 smart tools and 14 solutions. C4i4 also publishes the Industry 4.0 Maturity Model, an assessment tool built for Indian manufacturers, with a free online self-assessment.

What they do not deliver is money. In a written reply in the Rajya Sabha on 13 December 2024, Bhupathiraju Srinivasa Varma, Minister of State for Heavy Industries and Steel, stated plainly: "No financial assistance is given to any industry including MSME for adopting Industry 4.0 enabled technologies under SAMARTH centre initiative of the Scheme."

Plan accordingly. Use the free I4MM self-assessment to establish a baseline, because an external maturity number is useful in a board conversation and costs nothing. Then fund the build from your own capital or from a scheme that actually disburses.

The PLI schemes do move money, and at scale. Against a ₹1.97 lakh crore incentive outlay, 806 applications have been approved across 14 sectors, and realised investment reached about ₹1.76 lakh crore by March 2025, up from ₹1.61 lakh crore of committed investment in November 2024. Total sales by PLI participants have exceeded ₹16.5 lakh crore, and the schemes have generated over 12 lakh direct and indirect jobs.

For an auto or component maker specifically, the numbers to know are these: the automobile and auto components PLI was cleared in September 2021 with a ₹25,938 crore outlay, has attracted ₹67,690 crore in committed investment, and as of March 2024 had seen ₹14,043 crore actually invested, generating over 28,884 jobs. The scheme covers 19 categories of Advanced Automotive Technology vehicles and 103 categories of AAT components.

Read the mechanism carefully, because this is where business cases go wrong. PLI pays against incremental production and incremental sales, not against digitalisation line items. There is no box on the form for a vision system. The connection to an AI project is indirect: the AI spend is justified by the output it protects or the yield it lifts, which may feed the production metrics PLI rewards. It is not reimbursed.

Two more India constraints worth designing around.

Data residency. If your plant data includes personal data of workers, India's DPDP Rules 2025 apply, with substantive obligations effective 13 May 2027. Shop-floor camera systems see people. A vision system that records operators is processing personal data, and "we only look at the parts" is an architecture claim you have to actually implement, through on-device masking or by not retaining frames.

Connectivity. Edge inference is not a preference in most Indian plants, it is a requirement. Design for the line to keep running when the link to the cloud drops, and treat cloud as where you train and aggregate, not where you decide.

How we scope a build

eCorpIT is a Gurugram-based technology consulting organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with partnerships including AWS, Microsoft and Google. We are a software organisation, and we are specific about that boundary: we build the vision, data and agent layers and integrate with your existing automation. We do not sell you the PLCs.

We start with a measurement week, not a pilot. What does an hour of downtime actually cost on your line, computed from your output value, your labour cost and your penalty exposure, rather than from a survey of 181 firms in other countries? Most plants have never calculated this, and the number is usually smaller than the vendor slide and larger than the plant manager's guess. Everything downstream depends on it.

Then we pick one workload, based on where your data already is. If your machines already emit current, vibration and temperature, predictive maintenance is a short path. If they do not, instrumenting them is a six-month project you should decide on deliberately rather than discover halfway through a pilot. If your problem is the 81-minute recovery rather than the incident count, the answer is retrieval over your manuals and work orders, and that needs no new sensors at all.

The pilot is scoped to a single line and a single defect class or asset group, with an acceptance number agreed before we start, measured against a held-out set your team labels. If it does not clear the bar, we say so and you have spent a pilot budget, not a programme budget.

Integration is where these projects die, so we plan it first: how the model output reaches the PLC or the MES, who acknowledges an alert, what happens at 2am on a Sunday, and who retrains the model when the product changes. A vision model that nobody retrains after a die change is a model that gets switched off in month four.

Handover includes the retraining runbook. If your team cannot retrain it without us, we have sold you a dependency, not a capability.

For the governance layer around a multi-plant AI programme, see our enterprise generative AI strategy guide. More about how we work is on our about page.

FAQ

Does an hour of downtime really cost $2.3 million?

In a large automotive plant, according to Siemens' 2024 survey of 181 industrial professionals, yes. For most plants, no. The same report puts FMCG at $36,000 an hour and notes small and medium manufacturers reach $150,000 an hour at the top end. Calculate your own figure before quoting anyone else's.

Is unplanned downtime getting better or worse?

Both. Siemens found incidents fell to 25 a month per facility from 42 in 2019, and hours lost fell to 27 a month from 39. But the cost per hour doubled in automotive and quadrupled in heavy industry, and recovery time rose from 49 to 81 minutes. Rarer, harder, more expensive.

Does SAMARTH Udyog give MSMEs money for Industry 4.0?

No. The Ministry of Heavy Industries told the Rajya Sabha in December 2024 that no financial assistance is given to any industry, including MSMEs, for adopting Industry 4.0 technologies under the SAMARTH centre initiative. The four centres provide awareness, training, maturity assessment and consultancy. The capital investment remains yours.

What can I get from SAMARTH centres for free?

Assessment and training, mainly. C4i4 Pune publishes a free online Industry 4.0 Maturity Model self-assessment tool built for Indian manufacturers, has compiled over 50 use cases, and has trained 500-plus digital champions. CMTI Bengaluru has trained about 5,000 professionals in smart manufacturing. Use the baseline assessment before commissioning any vendor.

How much does edge AI hardware cost now?

Less than the mounting brackets, roughly. An NVIDIA Jetson Orin Nano Super developer kit is $249 and delivers 67 TOPS with 102 GB/s of memory bandwidth. NVIDIA raised that from 40 TOPS through a software update to existing boards. A ten-station vision pilot costs about $2,490 in compute.

Which AI workload should a plant build first?

Whichever one your existing data already supports. If machines emit current, vibration and temperature, predictive maintenance is short. If not, instrumenting them is a separate multi-month project. If your incidents are already rare but recovery is slow, retrieval over manuals and work orders addresses that without new sensors.

Does PLI cover AI and digitalisation spending?

Not directly. PLI pays against incremental production and sales, not digitalisation line items, so there is no box on the form for a vision system. Against a ₹1.97 lakh crore outlay, 806 applications are approved across 14 sectors. An AI project is justified by the output it protects or the yield it lifts, rather than being reimbursed itself.

Do shop-floor cameras trigger DPDP obligations?

If they capture workers, they process personal data, so yes. India's DPDP Rules 2025 carry substantive obligations effective 13 May 2027. A vision system that only inspects parts still sees operators, so on-device masking or not retaining frames is an architecture decision to make at design time rather than during an audit.

How eCorpIT can help

We build the software layer of factory AI: vision pipelines, machine-health models, retrieval over your manuals and work orders, and the integration that connects them to the automation you already run. Our senior engineering teams start by computing what an hour actually costs on your line, then scope one workload against an acceptance number agreed up front. eCorpIT has been building software from Gurugram since 2021, is assessed at CMMI Level 5, and partners with AWS, Microsoft and Google, and we design applications aligned with DPDP requirements where shop-floor data includes people. If you want a build scoped against your numbers rather than a survey of somebody else's plants, contact us.

References

1. The True Cost of Downtime 2024, Senseye Predictive Maintenance, Siemens
2. The True Cost of Downtime 2022, Senseye Predictive Maintenance, Siemens
3. SAMARTH Udyog Bharat 4.0 initiative, Press Information Bureau, 13 December 2024
4. SAMARTH Udyog Bharat 4.0, Ministry of Heavy Industries
5. SAMARTH Udyog portal
6. Jetson Orin Nano Super Developer Kit, NVIDIA
7. NVIDIA Jetson Orin Nano Developer Kit gets a Super boost, NVIDIA Technical Blog
8. PLI Scheme: powering India's industrial renaissance, Press Information Bureau
9. PLI Scheme for Large Scale Electronics Manufacturing, MeitY
10. PLI Scheme for Automobile and Auto Component Industry, Ministry of Heavy Industries
11. OpenAI API pricing
12. Enforcement of the DPDP Act and notification of the DPDP Rules, Shardul Amarchand Mangaldas
13. India's DPDP timeline: critical compliance deadlines for 2026-27, India Briefing

Last updated: 16 July 2026.