In September, at SEMICON India, Ashwini Vaishnaw told an audience he keeps a nail empty on the wall of his office, waiting for the first wafer out of Dholera. Good line. It also tells you where things stand four months after I called this programme serious.
Four months is not long. In fab time it is an epoch.
The surface reading is encouraging. Applied Materials has committed $5 billion over a decade, five plants are in commercial production, and the mission has gone from ₹76,000 crore to ₹1,27,500 crore. On any reasonable accounting that is a good quarter.
None of it is a wafer.
Three slips, not one
The programme slipped three ways. Only one is the date everyone discusses.
The timeline. At the foundation stone on 13 March 2024, Vaishnaw said the first chip would come out in December 2026, at 28, 50 and 55 nanometres. The Cabinet's Semicon 2.0 release, PIB PRID 2284796 of 15 July 2026, says “With the first fab scheduled to be commissioned in 2028.” Eighteen months, and no public explanation from the government or from Tata. A third date is unreconciled: on 3 December 2025 MeitY told the Lok Sabha that operations were expected “by FY 2029-30.”
The node. On 17 July Bloomberg reported Dholera will start mainly on 90 nanometres, sourcing people who would not be named. A Tata Electronics spokesperson confirmed the fab spans 28 to 110 nanometres and that the plan “was always to start with 55 nm and 90 nm, followed by 28 nm.” Neither company confirmed 90 nanometres; both declined to contradict it. Two of the three nodes announced in 2024 are now a later phase, and Tata Sons' FY2024-25 report put 28nm at the start of that journey while conceding the cutting edge “will take several years of serious work.”
The counter-argument deserves a hearing: 90 and 55 serve automotive and embedded MCUs, where yields are quickest. MeitY Secretary S. Krishnan argued exactly that at Yashobhoomi on 24 September, reported the next day — India has chosen to subsidise the mature end rather than chase 2nm, because advanced manufacturing is “simply too expensive.” He was describing a band running from 28 to 180 nanometres, wider than the 28 to 110 the fab itself is licensed for. Rational. But it is not the position the government sold in June.
The measurement. No Tata, MeitY, PIB or Gujarat publication has ever given a construction completion percentage. The figure in circulation — roughly 50 per cent civil construction — traces to one unnamed-source paragraph in Tom's Hardware in May. When Gujarat's chief minister walked the site on 16 October, the state release carried no number and no milestone. The absence is the datapoint.
There is a fourth slip, and it ran the other way. The Cabinet approved Dholera at “up to ₹91,000 crore.” The financial support agreement was signed at ₹91,526 crore. The project exceeded its own ceiling before a single tool was bought, and nobody has explained why.
It was never a joint venture
Dholera is not a Tata-PSMC joint venture. The chain is Tata Sons, wholly owning Tata Electronics, wholly owning Tata Semiconductor Manufacturing. PSMC holds no disclosed equity. Its role is a technology licensor under a definitive agreement of 26 September 2024.
India bought know-how. It did not buy a partner, and PSMC carries no project risk.
The financing tells you how lenders price Indian fab risk. Tata Semiconductor Manufacturing signed a ₹6,835 crore facility with five foreign banks through GIFT City branches on 5 February 2026, reported by Mint on 23 March off the MCA filings. The 163.5-acre Gujarat plot leased to Tata Electronics was valued at zero, and pledged, against a company into which Tata Electronics had put about ₹690 crore since its 2023 incorporation. A Tata executive told Mint: “The banks' credit decision is based on Tata's global creditworthiness, and not the land valuations.” That is a lender pricing a greenfield Indian fab as parent-guaranteed corporate risk.
The better evidence that tools are coming is indirect. Kelington announced on 14 September a provisional $452 million letter of award, as main contractor, to hook up “up to 886 tools” across both modules of “a semiconductor wafer fabrication facility in Gujarat” — client unnamed, thirty months' work, February 2029 target. The May 2026 ASML agreement, by contrast, is only a memorandum, and ASML India was incorporated in Ahmedabad in June.
Committed is not disbursed
ISM 1.0 breaks down as ₹65,000 crore for chip production, ₹10,000 crore for the SCL Mohali laboratory and ₹1,000 crore for design — though Vaishnaw conceded on 28 November 2025 that the SCL line “is a fungible amount.” In August 2025 MeitY Secretary S. Krishnan said: “I think we have committed close to ₹62,900 crore or so already.” On 1 September he went further: the government has “nearly exhausted” the ₹76,000 crore.
Here is the number, and it is the government's own. The Expenditure Budget publishes scheme-wise actuals: ₹681 crore in FY2023-24, ₹638 crore in FY2024-25, ₹2,886 crore in the first half of FY2025-26. About ₹4,200 crore against a ₹76,000 crore headline, two years in. And the sub-scheme line that matters most — Semiconductor Fabs — reads ₹0.00 crore in FY2023-24, ₹0.00 crore in FY2024-25, and ₹0.33 crore to 31 December 2025.
Two years into a ₹76,000 crore scheme, the government's own accounting says nothing has been disbursed to the fab. Vaishnaw told the Standing Committee why: legal agreements must be executed before the fiscal support can be released. That is an answer. Whether it is an explanation is a separate question.
Even that is a budget line, not a project. Asked in the Lok Sabha on 5 August 2026 for “the funds sanctioned, released and utilised, project-wise,” the government replied with a project count and an operating-plant count. Dr Shashi Tharoor's July question on the design incentive scheme got a commitment figure and nothing else.
Beside it, PLI publishes its own: fourteen sectors, 892 applications, ₹2.40 lakh crore invested, ₹15.2 lakh crore of exports, ₹35,354 crore of incentive actually disbursed as of 31 March 2026. And PLI's employment series runs backwards — 12 lakh at March 2025, 14.39 lakh at December 2025, back to 14.15 lakh at March 2026 — while exports went from ₹6.5 lakh crore to ₹15.2 lakh crore over the same year. A series that rises and then falls inside twelve months is not a measurement. Scheme rules require claimed sales to be verified against GST returns, income-tax returns and bank realisation certificates, but no published reconciliation ties them to HSN-coded customs data, and a NITI Aayog database commissioned in 2022 to do exactly that was never announced as finished. The 14.15 lakh is thinner than it looks: the indirect half of it, 5.66 lakh, is self-reported and exists for only three of the fourteen sectors.
Bigger scope, lower price
Semicon 2.0 is the biggest thing that has happened, and it is good news for the wrong reason to celebrate.
The Cabinet approved ₹1,27,500 crore on 15 July, sixty-eight per cent above ISM 1.0, notified in the gazette on 31 August as CG-DL-E-31082026-275860. Alongside it went a ₹62,500 crore phone scheme replacing PLI-LSEM, and an increase in the electronics components scheme from ₹22,919 crore to ₹40,000 crore, with 75 applications worth ₹61,671 crore approved as of 30 March.
The re-scoping is right. ISM 1.0 bought assets; ISM 2.0 goes after the supply chain into them — gases, chemicals, equipment. India imports more than 90 per cent of its chipmaking equipment and 85 to 90 per cent of its specialty chemicals and electronic gases, on Equirus Securities' estimate. Vaishnaw declined to put a number on it at the Bloomberg New Economy Forum on 14 October, saying only that materials are “the biggest dependencies.”
Now the rates, from the notification itself rather than the trade press. Silicon fabs: 40 per cent of eligible capital expenditure, down from ISM 1.0's 50 per cent of project cost — the base changed as well as the rate. Advanced packaging, display and compound fabs, 35. Legacy packaging, 25. Equipment and materials, 30. R&D and talent, up to 75 — though those two are percentages of project cost including opex, not of the eligible capital expenditure the rest of the table sits on. The design incentive, which most of the coverage leaves out, is 9 per cent of net sales.
The scope widened and the price of a silicon fab fell. The subsidy now leans hardest toward what India has least of, which is defensible, and lightest toward what it has most of, which is a confession. And it does not touch Dholera, whose support was fixed under ISM 1.0: Bloomberg reported on 17 July that the new fund “will apply only to future, incremental investments and will not retroactively offset spending already committed.” The project that costs India most is the one the new scheme helps least.
By August the government had three plants producing — Micron from February, Kaynes from March, CG Semi from July. On 17 September PIB inaugurated CDIL's Mohali discrete line and Suchi Semicon's Surat packaging line in a single ceremony and put the total at five.
Suchi Semicon is the execution story nobody wrote up. Approved 5 May 2026, in commercial production 17 September 2026: four and a half months from Cabinet approval to a working line, on an ₹868 crore plant in Surat. That is the strongest execution fact in the whole file. It is also a packaging line — as are all five plants now producing. Nothing India actually fabricates yet.
Sanand is not what I said it was
I gave Micron's Sanand plant more credit in June than the evidence supports, and the company's own press release has been quietly correcting the record since.
On 28 February 2026, at the inauguration, Micron wrote: “The expansion of conventional assembly and test operations in India complements Micron's planned development of advanced manufacturing and packaging capabilities in the United States and strengthens the company's global assembly and test network.” The second half of that sentence is the part that gets cut. An unnamed Indian government official told the Economic Times in April that the plant “is not planned to host HBM manufacturing yet.”
The “10 per cent” figure is not Micron's. Where it can be pinned down it is attached to at least two different denominators — Vaishnaw's “global memory production,” and Business Today's “global packaging and test output,” the only one of them a single-floor assembly and test plant could plausibly deliver — and Micron's own release contains no percentage at all.
On ramp: tens of millions of chips this year, hundreds of millions next, said Mehrotra again on 17 September. A company at full ramp does not re-announce its ramp. Nor does it carry one 2027 number. It carries two, and neither is a filing. A company that will not put a plant in its revenue line will not put it in its guidance either.
In July, unnamed people aware of the developments told the Economic Times that Micron had handed Phase 2 to L&T, displacing Tata Projects after delays in Phase 1; Micron has confirmed neither the award nor the reason. Nor is that the only slip on the record: Phase 1, set for 2024, was by February 2025 still targeting early 2025 completion. And in Micron's record FQ4-26 results on 30 September, $54.23 billion in a quarter, Sanand does not appear — not in the release, and not in any document the company has ever filed with the SEC. Dell is the only customer Micron has named anywhere.
None of that makes Sanand a bad plant. It is probably the best-run asset in the programme: an IDM drawing wafers from its own global network and putting them through test in India, staffed by 700-odd fresh graduates who trained in Malaysia and Singapore first. Not an argument against Micron — an argument against importing their forward guidance as an Indian capability claim.
The binding constraint is not money
If you want to know what will actually stop an Indian fab, it will not be capital and it will not be the exchange rate. It will be the fact that nobody has ever priced one.
On 25 March 2026 the Gujarat Electricity Regulatory Commission, in its tariff order for Torrent Power's Dholera distribution licence, declined to determine a tariff for the Dholera special industrial region at all, and continued Dahej's rate card unchanged for FY 2026-27 under the saving provision in its own regulations. Dholera has been on somebody else's price since 2022. The card is not generous: the high-tension industrial energy charge runs ₹2.90 to ₹3.20 depending on billing demand, with fuel pass-through on top and time-of-day penalties, which GERC's own cross-subsidy table puts at an all-in ₹6.26 to ₹6.31 a unit. The semiconductor incentives run on different clocks again — a ₹2 a unit tariff subsidy for ten years from the start of commercial production under the Gujarat Semiconductor Policy, and a duty exemption that runs five years in a special investment region and ten only inside a notified SEZ, which Dholera is not. Nationally, Power Secretary Pankaj Agarwal told EDICON in New Delhi on 21 January that industry pays about $105 a megawatt-hour, of which about ten dollars is cross-subsidy.
Read that order again. Gujarat has spent a decade assembling every incentive a semiconductor cluster could ask for, and when a regulator finally had to price a fab, it reached for the saving provision and borrowed another licensee's rate card, because the Dholera licensee has no cost history to justify one of its own. A regulated tariff is built from audited load, sanctioned demand and cost of supply, and Dholera has none of the three. That is harder to argue with than any reliability average: it is a regulator saying out loud that it cannot yet tell what a fab costs.
I would not have reached for a reliability number even if one were available. In India the instrument is patchy and mostly self-reported: of 72 distribution companies, 36 carry no outage data on their websites at all, an audit by The Leap Journal found in September 2025. What does exist is REC's Consumer Service Rating, which graded 66 licensees last year. On industrial supply, Gujarat's DGVCL came in at 10.6 interruptions and Assam's APDCL at 13.9, which is respectable. It is also a frequency count with no duration attached, the definition is not published, and six of REC's fifteen special-category utilities reported no industrial figure at all. Assam's APDCL ranks first of those fifteen on REC's composite score, not on this measure. The worst numbers in the dataset belong to Punjab and Karnataka, not Bihar.
Water is better sourced, though still second-hand. T Kannan reported in The Statesman on 29 August that Dholera's own environmental-appraisal file carries a total requirement of about 17,549 kilolitres a day, 6,571 of them fresh and the rest recycled, piped from the Narmada the better part of a hundred kilometres down a canal built for farms and towns — a river he describes as already fully spoken for. He cites no clearance number and no allocation order, and I have not been able to read the clearance file. If the numbers are right, they point at something structural: project-by-project clearance never asks what happens to the basin when the next nine plants answer the same question the same way. India has a water application form, not a water budget.
None of the rest is failure either. The scheme supports projects for up to six years, NITI Aayog's roadmap runs ten, and Vaishnaw relays the Prime Minister's line that this is a twenty-year journey. Asked about timing at the Bloomberg New Economy Forum on 14 October, he offered three horizons rather than a deadline. On ITIF's February 2024 count, drawn from a handful of interviews, roughly one per cent of India's 600,000 annual electronics-stream engineering graduates can walk onto a factory floor within three months of graduating. The regulator's own data says the pool is a quarter of that — AICTE counted 160,450 electronics and communications seats filled in 2024-25. CSIS repeated the one per cent in September 2025, citing a source that does not contain it. Slow. On this evidence, not stupid.
Washington, which did nothing
Let me take the geopolitical claim by claim, because this is where the June framing was most wrong.
Count them. All twenty-three Bureau of Industry and Security items in the Federal Register between 1 June and 18 October were read, and not one of them tightens semiconductor, advanced-computing, SME, HBM or advanced-packaging export control. The only advanced-computing rule in that window is the one two paragraphs below, and it loosens rather than tightens. The last tightening was in January. Section 232 semiconductor tariffs are still Phase 1 only: a 25 per cent duty, effective 15 January, on what the proclamation calls a very narrow category, described in an annex the Federal Register publishes only as an image. The H200 and MI325X labels come from January's export-control rule, not from the tariff. Data centres, R&D, startups, repairs, consumer and civil-industrial use and the public sector are all carved out to zero. One item touches a chip input — a September stockpiling prohibition under Proclamation 11052 — and it is a Section 232 import measure, not an export control.
India's status is unremarkable, which is a different thing from unfavourable. India sits in Country Group B, outside the enforcement perimeter, so top-tier accelerators ship with no licence required at all — NLR, which is a stronger and different thing from a licence exception. It is also already in Country Group A:5, and therefore already has Strategic Trade Authorization eligibility under 740.20©(1)(i). None of that was granted to India. It follows from where India is listed.
The July rule was different in kind. On 10 July, effective two days before it appeared in print, BIS published at 91 FR 43034, “Enhanced Favorable Treatment for the United Arab Emirates Under the Export Administration Regulations.” The UAE came out of Country Groups D:3 and D:4, joined Country Group A:5 — the bracket India already sits in — and gained a new entity whitelist, Supplement No. 8 to Part 740, covering licence-free access to advanced computing items for G42, Core42, Amazon.com, Apple, Google, Meta, Microsoft, OpenAI, Oracle and xAI. MGX is not on that list, only flagged for favourable review, and the G42 carve-out expires on 6 April 2027 unless they become American companies. This is what a carve-out made for one country looks like. Nothing in it was made for India, and nothing in it needed to be.
And Section 232 is uncomfortable. Proclamation 11002 of 14 January set out a two-phase plan, and the clocks are what they are: a status report on the negotiations due to the President by 14 April, a data-centre market update due by 1 July, and a 180-day statutory fallback whose date, 13 July, has passed. Only Phase 1 was ever operative; Phase 2 is permissive and nothing has triggered it. What has been implemented is thinner than the headlines suggest. Taiwan's investment agreement took effect in May, and what it capped was 232 duties on Taiwanese automobile parts, timber and wood derivatives. The semiconductors were left out of it. India risks becoming the residual jurisdiction: the one that pays the tariff without receiving the carve-out.
What actually got harder
Not the chips. The Strait.
Between 6 and 9 August, the Guangdong Maritime Safety Administration announced traffic control on northbound vessels entering the Taiwan Strait through its southern approach, framed around a typhoon. Taiwan's Coast Guard Administration told shipping to disregard it. ISW's reading was that Beijing may have used typhoon cover to test expanding its law-enforcement jurisdiction over a waterway it does not control. Cheap. Deniable. Legally untested.
On 31 August, Taiwan's Ministry of National Defence told the Legislative Yuan that the PLA is stepping up blockade preparations — the annual military-power report describes the intent as building the capacity to construct a blockade, sever Taiwan's sea lanes and encircle it. The ministry's own answer that week was thinner than the question. The anti-blockade work is visible elsewhere: a first-ever joint navy and coast guard escort operation on 12 August, shepherding strategic cargo into Hualien and Suao. EIU's own verdict, published back in February, is that a direct purposeful invasion is unlikely and that a blockade — or even an invasion — cannot be fully ruled out.
And it has not stopped. Between 8 and 17 October, six maritime administrations closed large parts of the Bohai, Yellow, East and South China Seas, including a live-fire no-sail zone off the southern Yellow Sea that runs to 30 November. On 14 October the Southern Theater Command put four landing ships through a simultaneous live-fire salvo and the Liaoning carrier group moved into the western Pacific. The timing is not subtle: it sits either side of Keen Sword, the US–Japan exercise that runs 19 to 29 October and puts systems on Yonaguni, 110 kilometres from Taiwan.
What changed is that the cost of preparing a blockade has fallen to the point where it can be rehearsed under cover of a weather advisory. Every leg of this strategy is an external dependency. Coherent. Not a hedge.
The window
Eight-inch utilisation across the top ten foundries recovered to 88 per cent in 2026 and should reach 90 in the second half; prices rose 5 to 15 per cent between the first and second quarters, and a third round could extend into 2027. The tight window is real. It is why India's bet is rational. It also has an end date.
Start with the mechanism, because it is usually stated backwards. TrendForce's thesis is not that TSMC and Samsung are releasing mature-node capacity. It is the reverse: they have been cutting eight-inch lines since the second half of 2025, and TSMC is taking roughly 50,000 wafers a month — 15 to 20 per cent of Fab 14's twelve-inch mature capacity — out by 2028. TrendForce's May reading was that the full impact becomes more pronounced after the second half of 2027. The squeeze gets worse, not better.
Which means new mature-node capacity comes from everyone who is not TSMC or Samsung. That is Dholera's market. Counterpoint had 40 to 90 nanometre utilisation stuck near 80 per cent in January, and TrendForce noted in April that customers were negotiating to postpone planned price increases for the second half of 2026. A new entrant arriving in 2028 is the marginal wafer in a market that already has spare wafers, in a price cycle whose customers are asking for relief in the quarter it is supposed to extend.
Set that against the decision MeitY's secretary made out loud on 24 September. India has chosen to subsidise the legacy end of the market because the leading edge is “simply too expensive.” That is defensible. It is also a decision about which glut to walk into, and the glut is the Chinese one. The Semiconductor Industry Association, working from SEMI's fab forecast, put China's share of global mature-node capacity at nearly double — 19 to 33 per cent — between 2015 and 2023, measured on installed nameplate capacity at 28 nanometres and above. CSIS, separately, could not find an expert consensus that this becomes an oversupply by 2030. Note the trap in the number: TrendForce also has China's mature-process share reaching 33 per cent, in 2027.
Dholera's first wafers are scheduled for 2028.
I am not forecasting a collapse. I am pointing at something harder to argue with: a fab arriving in the last year of the price cycle it was planned against, into a market where the two best-run foundries on earth are deliberately shrinking and the ones adding capacity are subsidised to add it, while their customers are already trying to talk the price down. India does not lose by being late to a fab. India loses by being late to a fab that arrives into the end of a price cycle.
What I think
None of this is an argument against Dholera. A 2028 fab at 90 nanometres in India is worth more than no fab, and a fab that trains process engineers and drags a supplier base into existence around itself is worth more than the ₹91,526 crore on the invoice. The state has done several things right. It let Micron build. And it has built an OSAT cluster in Sanand that is real rather than a press release. Any honest assessment starts there.
What I don't believe is that the programme is being managed to a schedule that survives contact with the market.
Taking the claims one at a time. The timeline moved eighteen months and the government is carrying three live dates it has not squared. The node dropped two generations. The completion figure traces to one unnamed paragraph. No anchor customer has been named. By its own accounting the scheme has disbursed nothing to the fab. The capital went elsewhere — SEMI's 2026 construction-start ranking has no India row, and Reuters called India a late entrant. And the strategic instrument everyone was watching has done nothing for India, while the real escalation has moved to a mechanism nobody can hedge against.
The fair summary is this. India now has a genuinely good packaging industry, an underrated OSAT cluster, a real design workforce, a properly large subsidy programme, and one fab that will start producing two years from now at a node two generations behind the promise, into the part of the cycle where the industry stops paying for capacity.
That is not a failure. It is an expensive lesson with a very long fuse, and the fuse is already lit.