Why Is It So Hard to Make a Semiconductor Chip? Inside India's First Fab
A 300 mm silicon wafer. Photo: Peellden, CC BY-SA 3.0, via Wikimedia Commons.
On the salt flats south-west of Ahmedabad, in a planned city that for years existed mostly as a map, a building is going up that has almost nothing in common with the industrial sheds around it. Its foundations were poured to a tolerance measured in millimetres, because the machines that will sit on them cannot tolerate the vibration of a passing truck. Its air will be filtered until a cubic metre of it holds fewer particles than a cubic metre of air at the summit of a mountain. Its water will be scrubbed of minerals, gases, bacteria and ions until it is so aggressively pure that it is undrinkable.
This is the Tata Electronics–PSMC fab at Dholera, and by the government's own timeline it is meant to pattern India's first commercially manufactured semiconductor wafer around December 2026. India has been trying to reach this moment for forty-two years.
Which raises the question that almost nobody asks out loud, because the answer sounds like it should be obvious and isn't: why is making a chip so hard? India puts satellites in lunar orbit for less money than a single lithography machine costs. It builds the world's cheapest telecom networks. It writes the software that runs on chips designed in Bengaluru and made in Taiwan. Chips are just patterned sand. So why has this particular thing taken four decades, three national missions and the better part of ₹1.64 lakh crore in approved investment to get to the starting line?
The wrong mental model: a chip is not a product, it is a streak
Most of us imagine manufacturing as assembly. Parts arrive, they are joined, a thing emerges. If one join is imperfect you fix that join. Cars work like this. Phones work like this. Almost everything you have ever bought works like this.
A semiconductor does not. A finished wafer is the output of somewhere between several hundred and well over a thousand sequential process steps — deposit a film a few atoms thick, coat it in light-sensitive resist, expose it through a mask, develop it, etch what the light exposed, strip the resist, implant ions, anneal, polish flat, measure, repeat. The photolithography sub-process alone, about ten steps end to end, runs fifty or more times on a single wafer. Nothing can be reworked at the end. There is no join to go back and fix, because each layer is built on top of the one beneath it and both are invisible to the naked eye.
That single structural fact — sequential, unrepairable, invisible — is where all the difficulty comes from. And it produces an arithmetic that is worth naming.
The Thousand-Step Tax
Call it the Thousand-Step Tax: in a process where a thousand steps must each succeed for the product to exist, quality does not add. It multiplies.
Run the numbers, because they are brutal and they explain almost everything else in this article. Suppose your fab is excellent — every step works 99.9% of the time. That is one failure in a thousand, a standard most factories on earth would be proud of. Across 1,000 sequential steps your yield is 0.9991000, which is about 37%. Nearly two of every three wafers are scrap.
Now suppose you are merely very good: 99% per step. Your yield across 1,000 steps is 0.991000 — roughly 0.004%. Not a bad business. Not a business at all.
Go the other way. Push each step to 99.99% and yield climbs to about 90%. So the entire difference between a fab that prints money and a fab that burns it is the difference between one defect per hundred, per thousand, and per ten thousand — differences a human being cannot perceive and no single inspection can catch.
The Thousand-Step Tax is why you cannot buy your way into this industry the way you can buy your way into steel or cement or automobiles. Capital gets you the building and the tools. It does not get you the fourth nine. That has to be earned, step by step, over years, by people who have watched the same process drift and learned what it does when the humidity changes.
Why a speck of dust is a catastrophe
Once you accept the tax, the fab's famous obsessions stop looking like theatre and start looking like accounting.
Take dust. The features being printed on a modern logic chip are smaller than the wavelength of visible light. A particle of ordinary household dust, 10 micrometres across, landing on a wafer mid-process is not a smudge — it is a boulder dropped across a city, killing every circuit beneath it. So fabs run at ISO Class 3 to 4 cleanliness, which permits on the order of a thousand particles of half a micrometre or larger per cubic metre of air. Outdoor city air holds tens of millions.
Getting there is not a matter of mopping. It means the entire building is a filter: air pushed downward through HEPA banks in the ceiling at a controlled velocity, floors that are grilles so the air can leave, positive pressure so nothing infiltrates, humans sealed inside garments not to protect them but to protect the wafer from them. A person at rest sheds several hundred thousand skin flakes a minute. In a fab, you are the contamination.
The water problem nobody puts in the press release
Then there is water — and here Dholera's location becomes a strategic question rather than a real-estate one.
Chips are rinsed constantly, because every etch and every polish leaves residue that will become a defect if it stays. The rinse cannot use tap water; dissolved minerals would leave behind exactly the particles you spent a fortune excluding. It uses ultrapure water, produced by driving ordinary water through reverse osmosis, ion exchange, ultrafiltration, UV oxidation, degasification and polishing until essentially nothing but H₂O remains. A single 200 mm wafer can consume around 5,600 litres of it. A large fab running at capacity can draw up to ten million gallons of water a day.
India is siting fabs in Gujarat, a state that is water-stressed by any honest measure. That is not an oversight — Dholera's plan is built around a dedicated water pipeline and heavy recycling, and modern fabs reclaim a large share of what they use. But it is a constraint that will shape the industry here in ways it never shaped Taiwan's, and it belongs in any sober account of the project. India has a long habit of solving national coordination problems by standardising infrastructure rather than negotiating around it — the same instinct that gave the country a single time zone stretching across nearly 30 degrees of longitude.
Why the machines cost more than the building
The third obsession is the tooling, and this is where the numbers stop being intuitive.
The heart of a fab is the lithography scanner, the machine that projects the circuit pattern onto the wafer. Deep-ultraviolet scanners — the workhorses of mature nodes, and what Dholera will run — cost in the tens of millions of dollars each. Extreme-ultraviolet scanners, needed only at the leading edge, run from roughly $180 million to over $380 million per unit in 2026, and are made by exactly one company on earth, ASML in the Netherlands, which cannot build many per year.
A fab needs dozens of scanners and hundreds of other tools, each installed, levelled, calibrated and qualified. This is why the ₹91,000 crore figure attached to Dholera is not padding. It is also why technology has become an instrument of statecraft: when the world's most important machine has a single supplier, export permissions become foreign policy, and a chip fab becomes a sovereignty question rather than a commercial one.
So why is India starting at 90 nm and not 2 nm?
This is the objection that surfaces every time the Dholera project is discussed, usually as a complaint. The fab is approved for 28, 40, 55, 90 and 110 nanometre nodes, and Tata Electronics has confirmed that production will begin at 55 nm and 90 nm before 28 nm is added. Meanwhile TSMC is shipping at 3 nm and below. Isn't India starting a generation-and-a-half behind?
Yes. Deliberately, and correctly.
First, because of the Thousand-Step Tax. A mature node has a well-understood, forgiving process window. It is the node on which an organisation can learn to hold four nines — on which engineers who have never run a fab can become engineers who have. Attempting that education at the leading edge, where the process window is razor-thin and a single tool costs more than the rest of the line, is not ambition. It is a way to convert ₹91,000 crore into a very clean, very empty building.
Second, because the market is there. Nodes between 28 nm and 110 nm are where most of the world's actual silicon lives: power management, microcontrollers, automotive and industrial control, display drivers, sensors, defence and railway electronics. The 2021 chip shortage that idled car plants worldwide was not a shortage of 3 nm processors. It was a shortage of exactly these unglamorous parts.
The forty-two-year gap, and what actually caused it
India did not arrive late for lack of trying. The Semi-Conductor Laboratory at Mohali began production in 1984 on a 5-micron CMOS process, and by the late 1980s it had progressed to 800 nm — genuinely close to the global frontier, at a moment when Taiwan's TSMC was barely a year old.
On 7 February 1989, a fire gutted the facility. It reportedly started at multiple points at once, destroyed around ₹60 crore of imported equipment, and the inquiry was inconclusive. Suspicions of sabotage have never been settled and probably never will be.
But the fire is not really the explanation, and treating it as one is comforting in an unhelpful way. What the fire destroyed was recoverable capital. What India lost in the decade after it was the continuity — the accumulating institutional memory that the Thousand-Step Tax makes irreplaceable. TSMC's advantage today is not that it owns better machines; everyone can buy the same machines. It is that it has been running them without interruption since 1987, and every drift, contamination event and yield excursion in that time is written into its process recipes and its people. That is the asset a fire, a funding lapse or a policy discontinuity actually destroys, and it is the one you cannot re-import.
What ₹91,000 crore is really buying
Which reframes the whole project. The obvious reading is that India is buying import substitution — chips made here instead of bought abroad, in a market that has grown to roughly $45–50 billion and is projected to reach $100–110 billion by 2030.
The more accurate reading is that India is buying a resumption of the streak. As of July 2026, twelve units have been approved under the India Semiconductor Mission with cumulative investment above ₹1.64 lakh crore — one silicon fab, one silicon carbide fab, an integrated gallium nitride micro-LED display fab and nine packaging plants. In February 2026 Micron's assembly and test plant at Sanand was inaugurated. In April the Dholera site was notified as a special economic zone. In July the Cabinet cleared a second phase of the mission with an outlay of about ₹1.27 trillion.
Read as a list of announcements, that is politics. Read through the Thousand-Step Tax, it is something more specific: a deliberate attempt to guarantee that this time nothing interrupts the learning. Packaging plants first, because they are easier and they build a workforce. Mature nodes before advanced ones. Multi-year committed funding rather than annual appeals. The design is aimed less at the chips than at the continuity.
The human truth: a fab is an apprenticeship, not a purchase
There is a version of this story that is about money and machines, and it is the version that gets told. The truer version is about people, and it is slower and less quotable.
The two phases now under way at Dholera — cleanroom installation and equipment calibration — are the phases that decide whether the plant works. Not because they are technically hardest, but because they are where a few thousand engineers, most of whom have never worked in a volume fab, learn what “in spec” feels like. Yield is not commissioned. It is climbed, usually for two or three years after first silicon, and the climb is the actual product. December 2026 will produce a wafer and a photograph. Whether India has a semiconductor industry will be visible somewhere around 2029, in a yield number that nobody will put on a stage.
That is an unfamiliar shape of patience for a country that likes visible wins, though not an unprecedented one. India has been comfortable holding assets whose value compounds quietly and out of sight before — it is the same logic behind the world's largest private gold hoard. And it is the same logic driving the race to the Moon's south pole: get there first, imperfectly, and the capability compounds afterwards.
What to watch next
Three signals, in order of how much they will actually tell you.
First: does the December 2026 trial wafer arrive on schedule, or slip? A slip of a quarter or two is normal for a first fab and means very little. A slip of a year means the calibration phase found something structural.
Second — and this is the one to care about — does anyone publish yield? Fabs rarely disclose it, but customers know, and design wins are the tell. If Indian automotive and industrial customers begin qualifying Dholera silicon into shipping products through 2027 and 2028, the streak is real. If Dholera is still running engineering lots in 2029, it is not.
Third: does 28 nm actually arrive? It is the node that matters commercially, it was always planned to come after 55 and 90, and getting there will be the first proof that the organisation can move up a generation on its own rather than on a technology transfer.
None of that will trend. The Thousand-Step Tax is paid quietly, one nine at a time, and it is the only price in this entire story that cannot be negotiated.
Frequently asked questions
When will India's first made-in-India chip actually be produced?
The Tata Electronics–PSMC fab at Dholera, Gujarat is targeting its first patterned wafer, described as trial production, around December 2026. Commercial volume production typically follows first silicon by a year or more, once yields are stable.
Why is it so hard to make a semiconductor chip?
Because a chip is the output of several hundred to over a thousand sequential steps, none of which can be repaired afterwards. Since the success rates multiply, a fab running each step at 99% instead of 99.99% produces almost nothing usable. Cleanrooms, ultrapure water and extremely expensive lithography tools all exist to buy back those decimal places.
Why is India making 28 nm and 90 nm chips instead of the newest 3 nm ones?
Mature nodes have wider process windows, so they are where an organisation can realistically learn to hold very high per-step yields. They also cover most real-world demand — power management, automotive, industrial and display chips — and were the parts actually missing during the global chip shortage.
What happened to India's first semiconductor plant?
The Semi-Conductor Laboratory in Mohali started production in 1984 and had reached an 800 nm process by the late 1980s. A fire on 7 February 1989 destroyed much of the facility and roughly ₹60 crore of imported equipment. The cause was never conclusively established, and the decade of lost continuity mattered more than the lost hardware.
Sources
- Tata group — Tata Group to build the nation's first fab in Dholera
- Press Information Bureau — India Semiconductor Mission 2.0
- Business Standard — Not cutting-edge, but can Tata's chip fab reduce India's import bill?
- Semiconductor Engineering — How semiconductor fabs use water
- The Tribune — When India's dream of becoming a semiconductor powerhouse was shattered
