Infrastructure Lock-In Technology Choices Locking In the 2030s
January 2025. TSMC delayed the opening ceremony for its Arizona Fab 21 facility from December 2024, waiting until after the presidential inauguration to cut the ribbon. The optics were deliberate. The company had invested an estimated $65 billion in the project, faced repeated delays, and found itself at the exact intersection of US semiconductor policy, Taiwan security concerns, and the incoming Trump administration’s approach to tech tariffs. The story illustrates how infrastructure lock-in technology decisions often outlast the political and economic conditions that shaped them.
TrendForce, reporting on Common Wealth Magazine’s coverage, confirmed in May 2026 that TSMC still faces four unresolved operational constraints in Arizona: water supply in a desert environment, regulatory complexity, visa delays for Taiwanese engineers, and a labor shortage as the original three-year engineering assignments near expiration.
The ceremony went ahead. The fabs keep being built. The infrastructure lock-in technology decisions embedded in those choices will last longer than any trade policy.
Infrastructure lock-in technology starts when you pour concrete
Semiconductor fabs are perhaps the most durable infrastructure commitments in the technology industry. TSMC’s Arizona Fab 21 started as a 5nm facility, was upgraded to 4nm, and entered mass production in Q4 2024.
The second fab, targeting 3nm, has completed construction. Ground broke on a third fab in April 2025, targeting the 2nm and A16 nodes, with TrendForce reporting that production could begin as early as 2027, a year ahead of the original schedule.
The capital expenditure makes the infrastructure lock-in explicit. Digitimes reported in January 2026 that TSMC projects $40–42 billion in capex for 2025, rising further in 2026 and potentially exceeding $50 billion by 2028. You do not build fabs to serve current demand. You build them to capture a market position 10 to 15 years out. Every node commitment, every equipment installation, and every water rights agreement in the Sonoran Desert is a bet on what AI compute needs will look like in 2035.
The deeper issue: delays pushing back the 2nm and A16 nodes to Arizona could compel customers to rely on Taiwan-based facilities, leaving them vulnerable to geopolitical risks tied to Taiwan’s dominance.
The infrastructure lock-in technology works in multiple directions. If Arizona underdelivers, companies lock back to Taiwan. If Arizona succeeds, they lock into a desert water supply problem and a specialised talent shortage that does not resolve quickly.
Meta’s cable and the privatisation of ocean infrastructure
Subsea cables carry more than 95% of international data traffic, per Meta’s own announcement citing standard telecommunications data. There are currently over 570 cables in service globally, plus dozens more under construction, according to TeleGeography’s Submarine Cable Map.
What has changed in the past decade is not the number of cables. It is who owns them.
Oxford Internet Institute’s March 2025 analysis documented the shift: “Google has deployed several wholly owned private cables, such as the Curie, Dunant, Grace Hopper, and Equiano.”
Meta has also previously procured the private Transatlantic Anjana cable. Meta announced Waterworth, a 50,000-kilometer cable spanning five continents, primarily for its own internal consumption. These projects used to require consortia because no single company could justify the capital. The hyperscalers are now large enough that the business case works on its own, which means the world’s longest cable will be owned by a single Silicon Valley firm.
Insikt Group documented four incidents involving eight distinct cable damages in the Baltic Sea and five incidents involving five distinct cable damages around Taiwan in 2024 and 2025, with at least four involving Russia- or China-linked vessels. The vulnerability calculus looks different when the cable is a public consortium asset serving many stakeholders versus when it is a private corporate asset.
Repair prioritisation, rerouting decisions, and access to traffic data in a disruption event are all governed differently.
The AI architecture standard nobody is discussing
The table maps who makes each decision and what reversibility actually looks like. The AI architecture row is the least-discussed yet possibly the most consequential aspect of infrastructure lock-in technology for enterprise teams.
| Infrastructure Decision | Who Controls It | Lock-in Horizon | What Flexibility Exists |
|---|---|---|---|
| Semiconductor fab location | TSMC, Intel, Samsung, with state subsidies | 20+ years per facility | Minimal; capital invested, process nodes committed |
| Subsea cable routes | Google, Meta, and Amazon are increasingly private | 25 years per cable | Rerouting possible but expensive; ownership determines access |
| AI model architecture | OpenAI, Google, Anthropic, through API standards | 5–10 years via enterprise contracts | Open-weight models create some flexibility; most enterprise tooling locks to one vendor |
| Data centre location | Hyperscalers plus national programmes | 15–20 years per facility | Geographic diversification is growing, but slowly |
The Open Markets Institute flagged that the same concentration dynamic at play in physical cable infrastructure is also at work in the software layer: a small number of firms are becoming chokepoints through which AI capabilities flow, and the integration of control over physical and logical infrastructure by the same companies creates lock-in that spans layers.
The lock-in nobody announced
There is an argument that infrastructure concentration is not a problem. The hyperscalers are better at building and operating at scale than most governments.
TSMC’s Arizona yields reportedly match its Taiwan facilities, which is not a given when you relocate one of the most complex manufacturing processes ever developed. Meta’s cable may serve the internet more reliably than a consortium cable governed by competing national interests.
The counterargument is not that private infrastructure is bad. It is that infrastructure decisions made for single-company efficiency do not automatically optimise for systemic resilience. A cable routed for Meta’s data centre topology is not routed for everyone else’s. A fab built under the US CHIPS Act subsidies to serve specific customer commitments is not neutral capacity available to whoever needs advanced nodes.
The infrastructure lock-in technology choices being made in 2026 are locking in advantages for the organisations making them. What they are locking in for everyone else is access on those organisations’ terms.
Distilled
Infrastructure decisions often appear technical in the moment but become strategic constraints over time. Semiconductor fabs, subsea cables, AI architectures, and data centres are all examples of infrastructure that can shape market access, resilience, and competitive advantage for decades.
The common thread is concentration. A small number of organisations increasingly control the physical and digital systems that underpin global technology. Once those assets are built and standards become established, reversing course becomes difficult, expensive, and sometimes impossible.
The infrastructure lock-in technology choices being made in 2026 will influence where compute happens, how data moves, and who controls access well into the 2030s. For most organisations, the challenge is not making those decisions but operating within the constraints they create.
