Terafab, Fordlandia, and the Decision That Took Seventeen Years

How does Elon Musk's Terafab relate to Henry Ford's production?

In 1913 at Highland Park, Henry Ford and his engineers stopped building automobiles in place and started moving them past the worker instead. A chassis that had taken more than twelve hours to assemble eventually came down to roughly ninety minutes once the moving assembly line had been refined.

Every business school tells that story. It is usually a story about flow and scale, and it is the Ford story people naturally reach for when they look at what Elon Musk is building in Texas. But I think it is the wrong Ford story for this.

On August 6, Tesla and SpaceX confirmed the site for Terafab, a vertically integrated semiconductor campus in Grimes County, Texas, about an hour northwest of Houston. The first phase carries a $16.8 billion price tag, and the finished campus is planned at more than 100 million square feet. Musk has talked about a long-term ambition of reaching one million wafer starts a month, with logic, memory, advanced packaging and test all sitting on one site. The chips are intended for Tesla’s robots and self-driving vehicles, and for the data centres SpaceX wants to operate in orbit.

Those are enormous numbers, so the Highland Park comparison almost writes itself. The trouble is that Highland Park was really about how work moved through a system, not how much of the system Ford chose to own. For that, the more interesting comparison came later.

Ford Ran Two Experiments, Not One

Highland Park changed the production process. Ford stopped treating assembly as a collection of separate jobs and designed it as one continuous flow. The result was speed and volume the old craft method could never have supported, at a cost per car it could never have reached.

The River Rouge complex asked a different question. At the Rouge, iron ore, coal, sand and timber came in at one end and finished automobiles came out the other, with Ford’s own steel mill, glass plant, sawmill and power house in between. By then, the question was no longer simply how to make the line faster. It was which parts of the value chain needed to sit inside the company at all.

That is much closer to the question Terafab is asking.

Modern semiconductor manufacturing is built on specialization, and that specialization has worked extraordinarily well. One company designs, another fabricates, others supply memory, equipment, materials, packaging and test. The system works because different companies become exceptionally good at one part of the chain, but the tradeoff is dependency.

For most buyers, that dependency is manageable. Chips are an input, suppliers are suppliers, and the model works. Once the volumes become large enough and the chips become critical enough to your product roadmap, though, the relationship changes. Lead times begin driving product dates, allocation decisions affect what you can ship, and somebody else’s capital plan starts becoming part of your own operating constraint.

At that point, vertical integration begins to look less like empire-building and more like a decision about what you cannot afford to leave outside your control.

The Intel Detail Matters

One detail in the Terafab story matters more than it first appears. Intel joined the project earlier this year, and Musk has said Terafab plans to use Intel’s 14A process technology.

That tells us something about the boundary Tesla and SpaceX are drawing. Terafab is being built to bring a very large amount of semiconductor capability inside their orbit, but that does not mean every part of the technology stack has to be recreated there. The project can own enormous amounts of capacity, packaging, test and production coordination while still relying on outside expertise where that makes sense.

That is much closer to what Ford did at the Rouge than the popular version of the story suggests. Ford did not insist on making every machine tool himself. He relied on specialists where specialization worked, and he brought things inside when dependence on an outside supplier threatened the flow of the larger system. The Rouge was less a monument to owning everything than a series of judgments about which constraints were important enough to own.

Terafab appears to be making the same kind of judgment. The interesting question is not whether Musk wants greater control over the semiconductor supply chain; he clearly does. The interesting question is where that control stops.

Ford ran that experiment too, and the second version did not end nearly as well.

Ford Already Ran the Overreach Version

In 1927, Ford secured roughly 2.5 million acres in the Amazon with the intention of producing his own rubber. From his perspective, the logic was not absurd. Rubber was essential to automobile production, and Ford disliked depending on a market dominated by plantations in Asia. So he did what had worked elsewhere and tried to bring the constraint inside the system.

Development began the following year. Ford built a town, imported American-style work routines and tried to grow rubber at industrial scale. The project struggled almost immediately as disease, insects, poor agricultural assumptions and management problems undermined the original site.

By 1933, Ford had effectively reset the project at Belterra, about eighty miles away, where the soil and drainage were better. It was not an irrational move. If the first site was the problem, moving to better ground was exactly what a competent operator might do.

That is what makes Fordlandia useful as a story. The easy version is that Ford made one terrible decision in the jungle and eventually admitted defeat. What actually happened was harder to see while it was happening: a series of locally reasonable decisions kept the larger objective alive long after the original assumptions had started to fail.

Ford did eventually produce rubber. By 1942 there were nearly 3.6 million trees planted at Fordlandia, and the first harvest yielded about 750 tons. The original ambition had been enough rubber to supply two million tires a year. By 1945, the plantations were sold back to the Brazilian government for $250,000, within months of Henry Ford II taking control of the company.

Seventeen years had passed.

The Hard Decision Was Never the First One

With hindsight, it is easy to look at Fordlandia and ask how anybody could have approved it. I don’t think that is the most interesting question.

Ford had immense resources, a formidable record as an industrialist and a legitimate business reason for wanting more control over rubber supply. Even the move to Belterra, after the original site struggled, had a reasonable technical argument behind it. The harder question is why the objective itself survived so long.

People were warning that the project was in trouble years before it ended, so the problem was not simply a shortage of information. What the organization lacked was a reliable mechanism for turning that information into a decision to stop. That distinction matters because it is much closer to the way large programs fail now.

You have probably sat in some version of the meeting. The program is still being reported green, and every input to that status is defensible. The delivery lead is reporting on delivery. The vendor is reporting against the vendor’s scope. Finance is accurate for the month it covers. The dates have moved twice, and each move had a reasonable explanation on the day it happened. There is a mitigation plan, and the mitigation plan is real.

Every piece can be correct on its own while the whole thing is wrong.

That is the part that makes these situations so difficult. The people with the best information are often the people with the least room to challenge the premise. If you have spent six months leading the recovery plan, saying the recovery plan will not work also says something about your own last six months. If you sponsored the investment, reopening the original case means reopening your own judgment.

That is not primarily a character problem. It is a structural one. Better reporting does not solve it, and more dashboards do not solve it. Eventually somebody needs enough distance from the original decision, enough visibility to see the whole system, enough authority to reopen the case, and no personal stake in which way it goes.

Fordlandia had warnings. What it lacked was a dependable way to turn those warnings into an exit. Seventeen years is what that can cost.

The Boundary Is the Decision

Seen through that lens, the Intel relationship may be one of the more interesting parts of Terafab. It suggests that somebody has already made at least one boundary decision: there are parts of the semiconductor stack Tesla and SpaceX want to control, and there are parts they are willing to rely on others for.

That is sensible operating design. The harder moment comes later, when one of those external dependencies starts to hurt. Maybe it is supply. Maybe it is pricing, a technology delay or a disagreement over priorities. At that point, the natural response will be to bring one more capability inside, and it may well be the right response.

But this is also how vertical integration expands. Usually not through one grand decision to own everything, but through a series of decisions that each make sense on their own. One dependency becomes painful, so you internalize it. Then the next dependency becomes the constraint, and you internalize that too. Over time, the organization can find itself operating businesses it never originally intended to enter.

Terafab is plainly not Fordlandia, and the point is not that Musk is repeating Ford’s mistake. The point is that the same kind of boundary decisions are being made, and the quality of those decisions depends partly on whether they can later be reversed.

Three Positions, Not Two

Large projects tend to get discussed as though there are only two choices: continue or cancel. That framing strongly favours continuing, because cancelling is a career event while continuing is usually just another meeting.

There is a third position, of course, and Ford took it in 1933 when the project moved to Belterra: reset. Same objective, different ground, better conditions. It was a reasonable attempt to separate a failed method from a still-valid goal.

The problem was not the reset itself. The problem was that the reset was not followed by a defined point at which somebody independent had to ask whether the objective still made sense. A reset without a defined checkpoint is just a slower version of continuing.

That checkpoint is a design problem. It either gets designed when the case is written or improvised under pressure four years later. What evidence would force the original case to be reopened? Who is positioned to judge that evidence? Who has the authority to stop the work if the case no longer holds?

Those questions are worth asking when a large capital case is approved, but they matter even more eighteen months later, once real money has been spent, careers are attached to the outcome and every decision comes with history. By then, the question I care about most is simpler:

Who in your organization can see the whole program, and what would it cost them personally to say that the case no longer holds?

The expensive part of a bad decision is rarely the day it gets made. It is the time between the point when the evidence changes and the point when the organization becomes capable of acting on it.

Henry Ford II ended Fordlandia within months of taking control in 1945. The warning signs had been visible for years.

That is the part of the story worth remembering.

Derrick Greenwood writes about operating models, delivery and the decisions that get made in the room before the go-live. He gets called in when big projects have big problems.