What changes could data-center construction face after the elections?

  • Nomura believes a federal-level pause in data-center construction is unlikely, given divisions within the Democratic Party and the president’s veto power; constraints are more likely to come at the state level.
  • Candidates in Texas and Georgia who supported a “temporary pause” won, advocating a halt to new permits for AI data centers until rules are established; candidates from both parties in Ohio support approvals subject to conditions. No elected governor in a major data-center state has pledged a permanent construction ban.
  • Nomura’s baseline view is that policy is more likely to shift toward requiring companies to bear more electricity costs and meet stricter permitting conditions. Share prices of companies with large Texas exposure and high leverage could come under pressure until the state’s legislative session ends in May 2027.

Which years’ new computing capacity is most likely to be delayed?

  • Nomura expects most capacity scheduled to come online in 2027 to already be under construction, limiting the impact; many projects scheduled to come online in 2028 and later are still in the planning stage and are more vulnerable to permitting delays.
  • According to Nomura’s estimates, only about six months remain before projects must break ground to come online in 2028. Some delays could push commissioning to 2029, with the capacity involved amounting to about 7% of global capacity additions in 2028; in a bearish scenario, about 11%.
  • Several OpenAI Stargate sites scheduled for completion in 2028 are located in states where candidates supporting a temporary pause won, making the pace of local approvals especially important to construction timelines.

How can tech companies cushion delays, and what would it cost?

  • Nomura believes the main ways to cushion the impact in 2028 include increasing utilization of existing data centers, bringing projects under construction into compliance with new rules, accelerating GPU replacements, and securing already-permitted capacity. The first two measures could make the most of existing facilities by increasing rack density or adjusting project schedules.
  • These approaches involve costs or physical constraints: replacing GPUs could result in early-retirement losses and is constrained by rack power supply and cooling; already-permitted projects with grid connections may require a premium.
  • Relocating construction to more permissive jurisdictions or overseas is expected to have an impact mainly from 2029 onward. The former is constrained by power-equipment bottlenecks, while the latter is also affected by geopolitics and chip export licenses. Community compensation and bearing a greater share of electricity costs could also raise construction costs per gigawatt.

Why does Nomura think the impact on the memory market could be moderate?

  • Nomura believes projects will mainly be relocated or delayed, preserving cumulative demand; meanwhile, memory manufacturers can use long-term supply agreements to adjust supply two years in advance, limiting the risk of severe oversupply.
  • Construction delays could also push up cloud GPU rental prices, prompting customers to replace GPUs with higher-performance models earlier and thereby supporting semiconductor demand. However, this depends on how much tech companies’ responses can offset the delays.
  • The downside-price scenario shows that earnings remain sensitive to memory prices: assuming shipments and unit costs are unchanged and the share price is assumed to be 1.3 million Korean won, Nomura estimates that if SK Hynix’s DRAM and NAND prices in 2028 both fall by 30% relative to its current model assumptions, operating profit would decline from 606 trillion Korean won to 380 trillion Korean won.