Nomura Raises AI Data Center Outlook to 40 GW for 2027, but Turns Neutral

Nomura now expects 26.25 GW, 40.43 GW and 28.29 GW of new AI data center capacity in 2026, 2027 and 2028 — a sharp upward revision from its June estimates of 27, 32 and 23 GW. If guidance from SpaceX and Broadcom earnings calls is factored in, 2027 and 2028 deployments could reach 49 GW and 46 GW respectively, though Nomura has deliberately not fully included those numbers, researcher Aaron Jeng wrote in the 22-page "Asia AI Semi & Server" report published October 26, 2026.

At the same time, the bank's risk-reward call has shifted from bullish to neutral. After three years of a super bull market, volatility has picked up since the third quarter of 2026.

What is driving the upward revision?

Nomura has tracked global data center construction since December 2025, and finds data points from Asia's semiconductor and hardware supply chains "not just useful — they're overwhelmingly positive or very positive."

The pipeline is big and getting bigger. Of 51 gigawatt-scale projects worldwide, the U.S. hosts 24 — nearly half. Outside North America, projects cluster in South Korea, France, India and the Middle East. By player type:

• Traditional hyperscalers: 31%, led by Meta and Google
• Agile new-cloud and crypto-mining converts: 26%
• Sovereign AI initiatives: 30% (15 projects), often with commercial partners

More than 50 sub-gigawatt projects should add another 16 GW, plus over 110 undisclosed-capacity projects conservatively worth more than 1 GW. The 40 GW and 28 GW figures for 2027/2028 translate into demand for 5–8 million AI chips; if built on Vera Rubin technology, chip demand would hit 7.658 million and 5.357 million units, requiring 851,000 and 595,000 CoWoS substrates — or over 5 million Rubin GPUs and roughly 5.8 million CoWoS packages across the full pipeline.

Expansion continues. Meta has grown Hyperion to 5 GW in Louisiana with more than $50 billion invested and commissioned 1 GW in Canada. Google announced a €13 billion data center in Finland. Microsoft halted one small UK cloud deal, though it's under 100 MW.

Where does the demand sit in the chip stack?

Broadcom's September 2026 earnings call laid out deployment schedules Nomura calls significantly above its own tracking: Anthropic will deploy 1 GW of Ironwood in 2026, 5 GW of TPU v8i in 2027 and another 10 GW in 2028; OpenAI's Jalapeno is slated for 1.3 GW in 2027 with potential to exceed 5 GW by 2028; Meta's third-generation MTIA should total 3 GW before 2028.

SpaceX plans 10 GW of added capacity by end-2027, with 2.5–3 GW cumulative by end-2026 and 13 GW by end-2027 — not ruling out a 20 GW target. Nomura assumes conservatively over 7 GW new in 2027 and 3 GW more in 2028.

The chip shortage should persist through 2027, Nomura argues, because greenfield capacity coming online by end-2025 won't contribute meaningful supply until 2027–2028. That supports accumulating Asian tech stocks on pullbacks, with earnings in the AI supply chain trending up through 2027. The cycle won't end in 2027; upcycles persist while shortages last.

How serious is the credit risk?

Serious enough to change the stance. The five major CSPs added a record net $158 billion in long-term debt in the first half of 2026. Disclosed guarantee exposure across major AI companies totals $216.685 billion as of 2Q26: NVIDIA $108.5 billion (about 50%), Broadcom $29 billion, Google $43.785 billion, Meta $28 billion, AMD $4.1 billion, Oracle $3.3 billion.

Nomura flags the structure — suppliers lending to customers who buy the suppliers' products — as resembling Cisco and Lucent vendor financing before the 2000 bubble. Mitigating context: AI financing is 3.6% of GDP (Q1 2026, 3.5% in Q2), below subprime's 5.4% in early 2007 — though cloud capex as a share of revenue already exceeds the dot-com peak.

Yields tell the story. SpaceX's $40 billion October 2026 debt priced at roughly 8.3% at its riskiest tier, up from 6% in June. CoreWeave's June bond, issued at a 9.6% coupon, now trades around a 15% secondary-market yield. U.S. 10-year Treasuries have surged past 5%, a record.

Execution cracks are showing too: Oracle and OpenAI canceled their Texas flagship expansion; OpenAI suspended Stargate UK and exited UK and Norway projects; Oracle's Jupiter and Lighthouse face regulatory and power hurdles; OpenAI's 3.2 GW Project Camellia in Georgia was excluded for overlap.

Do valuations still work?

Cheap by history. The Philadelphia Semiconductor Index trades at 18 times earnings against a ten-year range of 13–27; Taiwan large caps ex-TSMC sit at 14 times versus a 10–20 range. SOX gained 19%, 42% and 52% year-to-date in 2024, 2025 and 2026; TAIEX returned 28%, 26% and 59%.

Even at a 10% financing cost, returns on roughly $30 billion ARR per gigawatt remain attractive given high ROIC, and major players keep insisting that "underinvestment in AI is riskier than overinvestment." xAI plans 20 GW of new capacity across 2027–2028.

Nomura expects AI-driven trading to return once credit markets stabilize or OpenAI and Anthropic's ARR resumes climbing — with ARR the single most critical variable to watch, ahead of funding costs, hyperscaler ROIC and capex guidance for 2027–2028 — though upside will be more muted than in 2023–2025.