Nomura raises its 2027 AI data center construction forecast, but financing costs make its investment stance more neutral
Nomura has raised its forecast for new global data center deployments in 2027 from 32 gigawatts to 40 gigawatts and expects supply chain shortages to persist through 2027. At the same time, rising debt financing costs, project execution risks and limited visibility into 2028 demand constrain the optimistic view.
Original source：Nomura International (Hong Kong) Ltd., Taipei BranchFirst published on Unhidea：Original source date：
Key takeaway
Nomura believes demand for AI infrastructure construction still has room to rise, but growing financing pressure and substantial gains already recorded in share prices are making the risk-reward balance for technology stocks more neutral.
By how much was the data center construction forecast raised?
After accounting for project delays, cancellations and additions, Nomura tracked 51 gigawatt-scale projects. It expects annual new global data center deployments of 26, 40 and 28 gigawatts, respectively, in 2026–2028; the corresponding forecasts in June were 27, 32 and 23 gigawatts.
New deployments in 2027 are 8 gigawatts higher than in the previous forecast, but the 28 gigawatts projected for 2028 remain below the 2027 level. Nomura had hoped to see stronger demand in 2028, but has not yet reached that conclusion.
Under Nomura’s model, if all deployments in 2026 use GB300 and all deployments in 2027–2028 use Vera Rubin, the new capacity over the three years would correspond to demand for approximately 7.36 million, 7.66 million and 5.36 million chips, respectively. The model assumes that 70% of infrastructure power is used for computing.
Which plans could push demand higher?
SpaceX plans to add 10 gigawatts of computing capacity in 2027, bringing its cumulative total to approximately 13 gigawatts by year-end. Nomura’s baseline tracking currently assumes additions of approximately 7 gigawatts in 2027 and another 3 gigawatts in 2028, putting its 2027 estimate approximately 2–3 gigawatts below the company’s plan.
Broadcom says Anthropic plans to deploy 5 gigawatts of TPUs in 2027 and another 10 gigawatts in 2028. Among the large projects tracked by Nomura, deployments that Anthropic builds itself or may own amount to less than 1 gigawatt per year. Because of uncertainties over measurement methodology, double counting and project execution, Nomura has not included this gap in its baseline.
If the relevant statements by SpaceX and Broadcom and Nomura’s subjective judgment are incorporated, the demand scenarios for 2027 and 2028 could rise to 49 and 46 gigawatts, respectively.
As financing costs rise, where will the money come from?
Nomura believes the free cash flow of leading US cloud providers is insufficient to support further expansion in capital spending, making external financing necessary. The five largest cloud providers recorded a record net increase of 158 billion dollars in long-term debt in the first half of 2026; substantial bond issuance also pushed up interest rates and credit spreads.
Financing pressures vary among borrowers: the highest-risk tranche of the new debt SpaceX plans to issue in October has an expected yield of approximately 8.3%, above the 6% yield on its June bonds. CoreWeave’s June high-yield bond carries a coupon of 9.6%, and its secondary-market yield has risen to approximately 15%.
Chipmakers and cloud providers also help downstream companies obtain financing through backstop arrangements. Examples include commitments to lease back unsold computing capacity or guarantee the residual value of hardware. As of the second quarter of 2026, the major AI companies covered by the report had disclosed backstop exposure of 216.7 billion dollars, including 108.5 billion dollars for Nvidia; these arrangements also transfer downstream demand and asset-value risks back to the backstop providers.
What constrains project progress and the supply chain outlook?
More projects do not mean that every project is progressing as planned. OpenAI withdrew from Stargate projects in the UK and Norway, and Oracle-related projects also face regulatory and electricity supply obstacles. Nomura continues to use OpenAI’s 26-gigawatt chip agreement as the upper limit for its deployment estimates and has not adjusted the figure in response to these project changes.
Nomura expects earnings across Asia’s AI semiconductor and hardware supply chain to continue trending upward through 2027: demand forecasts continue to be revised higher, while new capacity built since the end of 2025 is not expected to begin production until 2028.
Why has Nomura become less aggressive on technology stocks?
Nomura believes demand and earnings remain supported, but AI technology stocks have already risen substantially, the introduction of new capacity in 2028 is drawing closer, and rising financing costs and limited visibility into 2028 demand are making the risk-reward balance more neutral.
Nomura also believes the recent valuation pullback was not driven by downward revisions to consensus earnings expectations, and further upward revisions to earnings forecasts could still lift share prices. It is particularly focused on the annualized recurring revenue of OpenAI and Anthropic: AI monetization performance will affect lender confidence, cloud providers’ willingness to build data centers and demand across the Asian supply chain.
