AI and risk management in chip manufacturing

In 2022, Intel promised two chip factories in Ohio, running by 2025. Today, the first one isn't expected to open until 2030 or 2031—the company does not blame concrete, labor, or permits. It points to a variable most risk registers don't even list as a risk: whether customers still want what the plant was built to make.

In this article, we'll discuss why equipment, yield, and demand risks in chip manufacturing tend to move together instead of independently, how AI tools can build a full Monte Carlo model in minutes without ever flagging which assumption is actually driving the result, and why testing a correlation coefficient—not just picking one—is often the difference between a defensible risk model and a comforting one.

A case study: Intel's Ohio chip factory investment project

On Jan. 21, 2022, Intel announced an initial investment of more than $20 billion in "two new leading-edge chip factories in Ohio." The release stated that "Production is expected to come online in 2025" [1].

On Feb. 28, 2025 Naga Chandrasekaran, who runs Intel’s manufacturing operations, published a revised timeline. Intel would "complete construction of Mod 1 in 2030 and begin operations between 2030 and 2031." For Mod 2, it expected "to complete construction in 2031 and begin operations in 2032" [3].

The same statement explained how the work would proceed: "We will continue construction at a slower pace, while maintaining the flexibility to accelerate work and the start of operations if customer demand warrants" [3]. By that point the campus figure stood at $28 billion, with more than 6.4 million work hours completed and over 200,000 cubic yards of concrete poured [3].

Production expected in 2025, now expected between 2030 and 2031. And the company itself names the variable that governs the pace: customer demand.

That is worth sitting with, because it is close to the opposite of how most project risk registers are built.

The wider picture

In July 2026, construction spending on manufacturing structures across the United States ran at a seasonally adjusted annual rate of $169,795 million. In July 2025, the figure was $215,594 million [4]. That is a fall of about 21% in a year.

For anyone taking a plant construction proposal to a board, the effect is practical. The proposal must answer a question it could mostly avoid three years ago: what is the realistic spread of outcomes here, and who is prepared to put their name under it?

Where the real uncertainty sits

Ask most people what could go wrong with a factory and you get building answers. Weather, labour, materials, permits. In a chip plant those matter, but Intel’s own account says they are rarely what decides the outcome.

Three things usually do.

The first is the equipment. The machines can cost considerably more than the building that houses them, and they come from a small number of suppliers. When one supplier runs late, the others often do too because they are all under the same market pressure. So equipment cost and schedule length tend to move together rather than separately. This is my own reading from client work rather than a published finding, and I offer it as such.

The second is yield, and it needs stating precisely, because the risk is not in the building. Where a plant is ramping a new process generation, the learning curve dominates. Charles Weber’s study in IEEE Transactions on Semiconductor Manufacturing describes fault density dropping "by an order of magnitude every six months" early on, with the learning rate flattening "in the fifth year as random faults begin to dominate." His conclusion is the one that matters for a business case: "the yield-learning rate tends to be the most significant contributor to profitability in semiconductor manufacturing." In his model, moving the yield ramp forward by six months more than doubles the cumulative net profit of the venture, while delaying it by six months removes two thirds of that profit [5].

Individual plants do not publish their yield figures, so the curve for your own project must come from your history or your benchmark set. But six months either side of it is not a detail, and it does not belong in the cost estimate. It belongs in the revenue model.

The third is demand. Nothing in a risk register makes a plant profitable if the customers are buying something else by the time it opens.

Putting numbers on it

This is where simulation earns its place, and where the AI tools have changed the working day.

I start by getting the model built. Inside Excel, through the @RISK Agent MCP connection:

"Using the @RISK Agent MCP, turn the three cost lines in B6:B8 into PERT distributions from the low, likely and high figures in columns C to E, and set the build duration in B12 the same way. Propose alternative distributions with same parameters to compare"

That used to be two hours of typing. It now takes a couple of minutes, and the syntax comes out right.

Then comes the part the tool will not do for me. Left alone, the model treats equipment cost and schedule as unrelated. That gives a comfortable, narrow answer, because unrelated errors cancel each other out. In a chip plant they are not unrelated, for the reason explained. The figures here are illustrative, and you should set them from your own package data:

"Using the @RISK Agent MCP, correlate the three equipment packages at 0.6, and correlate equipment cost with installation duration at 0.5."

Worth saying how to do this properly, because correlation is where models go wrong in both directions. Correlate the driver, not the symptoms: if two cost lines move together because they share a supplier, the supplier is the variable, and a coefficient invented between the two lines is a shortcut you will struggle to defend.

Get the sign right before arguing about the size. Note that what @RISK takes is "a matrix of rank correlation coefficients" [6], so you are constraining the order in which values are drawn rather than fitting a straight line through them. Check the matrix is consistent; @RISK will warn you if it is not and offer to generate the closest valid matrix. It's important to check adjusted correlation matrix values to ensure they're appropriate for the model, as outlined in the @RISK documentation [7].

Take the coefficients from your own package history where you have it, and from a structured conversation with the people who buy the equipment where you do not, and record which of the two you did. Then test it: run at 0.3, 0.6 and 0.9 (you can use scenarios for this). If the decision flips somewhere in that range, the coefficient has stopped being an input and become the thing the project turns on.

I have never seen an AI tool suggest this on its own. AI will build whatever model you describe. It will not warn you that the independence buried in your description is why the answer looks reassuring.

Then I run 10,000 iterations manually from the ribbon and then ask the question that decides things:

"Using the @RISK Agent MCP, report the P50 and P80 cost and completion date, and rank the inputs by how much they drive the spread."

And separately, the question a cost model can never answer: what does a five-year delay do to the value of this plant? For a chip built to a particular technology generation, delay does not add cost. It removes the years when the product sells for the most. Treat that as a cost line, and you will be badly wrong about it.

Four things worth watching

1. Equipment risks move together, not separately.
   A model that assumes otherwise gives you a narrower answer than the situation deserves, and narrow answers get projects approved.
2. How long the equipment stays valuable is an assumption, not a fact.
   The Economist’s briefing on Nvidia of Sept. 3, 2026, sets out where that argument currently stands. It reports that Michael Burry has argued cloud providers inflate their profits by depreciating chips over five or six years rather than two or three, and that Jensen Huang maintains the chips stay useful and worth something for far longer [8]. The disagreement is the point either way: when informed parties differ by a factor of two on an asset’s useful life, that life is a range and not a number.
3. Yield belongs in the revenue model, not the cost estimate.
   Weber’s finding is that the timing of the ramp, not its eventual level, is what moves the return [5].
4. A supplier guarantee and the supplier’s ability to honor it usually depend on the same conditions.
   The event that triggers the guarantee is often the event that weakens the party giving it. Model the exposure and the capacity to absorb it as one thing, not two.

Where this sits in the AI and risk management series

The previous article looked at infrastructure, where the tendency to underestimate is well documented and the fix is largely discipline about comparing your project to similar ones. Chip plants are a different problem. There is less public data, the supplier base is smaller, and the risks are more tangled up with each other. Next month closes the series with cost estimation itself, and why Monte Carlo remains the most defensible method we have.

Catch up on the rest of the series:

• AI prompt engineering for risk managers: A practical field guide with @RISK Agent
• Monte Carlo simulation in the age of AI: What changes, what doesn’t, and what gets better
• AI and risk management in energy & utilities: How @RISK and AI are changing the game
• AI and risk management in infrastructure and civil engineering

Build with AI. Verify with @RISK. Defend with both.

Sources

[1] Intel Corporation, "Intel Announces Next US Site with Landmark Investment in Ohio", 21 January 2022. https://www.intc.com/news-events/press-releases/detail/1521/intel-announces-next-us-site-with-landmark-investment-in

[2] Intel Corporation, "Intel Breaks Ground in the Silicon Heartland", 9 September 2022. https://www.intc.com/news-events/press-releases/detail/1572/intel-breaks-ground-in-the-silicon-heartland

[3] Naga Chandrasekaran, "Ohio One Construction Timeline Update", Intel Newsroom, 28 February 2025. https://newsroom.intel.com/corporate/ohio-one-construction-timeline-update

[4] US Census Bureau, Monthly Construction Spending, July 2026, released 1 September 2026, Table 1. https://www.census.gov/construction/c30/pdf/release.pdf

[5] Charles Weber, "Yield Learning and the Sources of Profitability in Semiconductor Manufacturing and Process Development", IEEE Transactions on Semiconductor Manufacturing, vol. 17, no. 4, pp. 590-596, November 2004. https://web.pdx.edu/~webercm/documents/2004%20Weber%20Yield%20Learning.pdf

[6] Lumivero, @RISK Help, "RiskCorrmat". https://help-risk.lumivero.com/v8/en/@RISK/Function/5-Property/RiskCorrmat.htm

[7] Lumivero, @RISK Help, "Correlation Matrix Consistency". https://help-risk.lumivero.com/v8/en/@RISK/1-Define/5-Correlation/Correlation-Matrix-Consistency.htm

[8] "Nvidia is the central bank of AI", The Economist, 3 September 2026 (subscription required). https://www.economist.com/interactive/briefing/2026/09/03/nvidia-is-the-central-bank-of-ai