TSMC weighs a Texas campus as Washington's chip tax credit runs out
TSMC is weighing a new Texas campus of several fabs, each costing at least $20bn, according to people familiar with its plans cited by Bloomberg. The 35% US manufacturing tax credit behind it requires construction to start by the end of December, and no extension has a path this year.

The plans are early. TSMC told the market it has no comment on rumours. The people familiar with its thinking spoke to Bloomberg, which The Next Web reported on 3 October. The arithmetic is not rumour: a single fab at $20bn or more would cost roughly twice TSMC's entire Dresden project, which is budgeted above EUR 10bn.
For anyone tracking US chip policy, the Texas story is the clearest test yet. Do Washington's incentives still pull in the most advanced manufacturing? Or has the money already done its job, leaving the next decision to cost alone? Neither TSMC nor the senators named in the reporting have said what happens if the credit lapses. That is the open question.
TSMC has not been idle in the US. It has committed $265bn in Arizona: ten fabs, two packaging plants and a research centre. North America brought in more than 75% of its wafer revenue this year, according to The Next Web. That makes the case for more American capacity look like a straight commercial calculation rather than a subsidy chase. The company keeps its most advanced research in Taiwan whatever it builds abroad, so the Texas campus would be a production bet, not a technology transfer.
The credit has a deadline, and Congress has not moved
The 35% manufacturing tax credit requires construction to start by the end of December. No extension has a path this year. Senators Mike Crapo and Ron Wyden say they are committed to one, according to the same report. But a commitment is not a statute, and the clock is the only thing in this story that is not negotiable.
That timing matters beyond TSMC. Samsung already runs a chip complex in Texas, Musk's Terafab is there, and Texas Instruments is expanding in the state. A new TSMC campus would land in a crowded market, competing for the same construction crews, the same power and the same suppliers. If the credit lapses, the incentive gap between the US and other locations widens. That happens at exactly the moment TSMC is choosing between Texas and Singapore, the other site it weighed.
Singapore is not Europe. That distinction is the part of this story Europe should read closely.
Europe has one plant, and it does not make AI chips
TSMC's only European plant is the Dresden joint venture, budgeted above EUR 10bn. Brussels approved EUR 5bn in state aid for it in August 2024. The plant will make 40,000 wafers a month, with equipment moving in this half, production starting in 2027 and full capacity arriving in 2029.
It will not make AI chips. Dresden runs 28 and 22nm CMOS alongside 16 and 12nm FinFET, several generations behind the nodes Nvidia and Apple buy. Europe's flagship TSMC investment targets automotive and industrial customers, not the accelerators that dominate the current spending boom. A single Texas fab, at $20bn or more, would cost about twice the whole Dresden project.
The EU's record is not empty. Infineon opened a EUR 5bn power chip fab in Dresden in July on EUR 1bn of EU support, the first win under the bloc's chip subsidy regime. But the target behind it will be missed. The EU wants 20% of the global market by 2030, held 9.8% in 2022, and its own forecast puts it at 11.7%.
The European Court of Auditors called the goal very unlikely in April last year. Industry has asked for a Chips Act 2.0, and the Commission is drafting one. The first act was meant to mobilise EUR 86bn, of which the Commission funds 5%, with member states and private investors covering the rest. Europe's problem is not a lack of announcements. It is that the money is mostly not Brussels' money, and the leading edge is not what it is buying.
The question for Europe is not the leading edge. It is whether it is in the room when the next campus is picked.
The demand side is the other half of the bet
Every fab decision rests on an assumption about how much compute the world will want. Epoch AI published an estimate on 3 October that puts the scale in perspective. It finds that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents. That supplies as many weekly working hours as about 140 to 720 million full-time employees, assuming full deployment and allocation to those workloads.
Memory shipped during 2025 and 2026 alone could support 16 to 56 million concurrent agents once deployed, Epoch says. Including shipments through 2027 raises the figure to about 30 to 170 million. More efficient models could push it further: applying DeepSeek V4 Pro serving benchmarks to the projected hardware supply yields approximately 1.9 billion concurrent agents.
Those numbers are capacity, not demand. Epoch is explicit that using even 20% of its central capacity estimate would imply $2.6tn to $5.3tn a year in API-equivalent spending. That compares with roughly $1tn in developer revenue by the end of 2027 at fivefold annual growth. Agent workloads are not cheap to run: its analysis of agent traces found Codex workloads averaged roughly $16 to $18 per hour of continuous agent activity, against $24 to $50 for Claude Code workloads. If demand falls behind supply, the overabundance of capacity lands on the same balance sheets that paid for the fabs.
The controls are not stopping the hardware
Export rules sit over all of this. Recent reporting has focused on enforcement rather than new restrictions. Ars Technica reported on 2 October that US authorities arrested a tech CEO accused of smuggling $300m in Nvidia chips into China. The Register, the Los Angeles Times and Business Insider picked up the case. ProductNation reported on 3 October that the US says two Malaysia-based firms sent $176m in a separate chip-smuggling case.
The pattern is consistent: controls on paper, prosecutions against the routes around them. That matters for TSMC's calculus because demand from restricted markets does not vanish. It moves. Every shipment that goes through a third country or a front company is a data point that the licensing system is leaky. It is also a reason for US policymakers to keep the domestic incentive carrot in place alongside the export-control stick.
Meanwhile the compute is being leased around the rules. Tencent has leased 100,000 AI chips from Oracle for $7bn amid US export curbs, ET Datacenters reported on 3 October. Neither company has said how much of that capacity sits in which jurisdiction.
What to watch before December
Three things decide how this ends. First, whether Congress extends the 35% credit before construction deadlines bite; Crapo and Wyden say they want to, and nothing has been filed. Second, whether TSMC picks Texas or Singapore. Brussels will read that choice as a verdict on Europe's ability to attract leading-edge capacity at all. Third, whether the demand Epoch describes shows up as revenue rather than inventory.
There is a sceptical case worth keeping in view. Ed Zitron wrote on 22 September, in a piece republished to his site and surfaced on 3 October, that Bloomberg had reported Microsoft holding about 12 gigawatts of capacity. Only about 2 gigawatts centred on AI-specific chips, he wrote, and he argued that NVIDIA's revenue growth rests largely on speculative purchasing. Zitron's piece is opinionated and his numbers come from a source he worked with on a Guardian investigation, which put Microsoft at approximately 2.2 million chips. Treat the framing as his. The underlying question is whether the hardware being bought is being plugged in. That is the same one TSMC's board will be asking before it commits $20bn to a Texas field.
For now, the most concrete fact is the simplest. TSMC has committed $265bn to Arizona and is weighing another multibillion-dollar campus in Texas. The tax credit that makes the maths work expires for new construction at the end of December. Everything else, including Europe's chances, follows from what Congress does next.
Sources
8- 01TSMC eyes a multibillion-dollar Texas campus for more AI chipsEN
- 02How many AI agents could run on the AI chips shipped through 2027?EN
- 03Where're All the AI Chips?EN
- 04Welcome to Opt Out October. Let's Take Control of Our Data and Our DevicesEN
- 05The Void: From Zero-Byte Responses to Continuation ControlEN
- 06Garmin watch as a controller for the Flipper ZeroEN
- 07Vx – One Language, Every ChipEN
- 08Enterprise Web App and API security under your controlEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.