The core argument: Why this is not the peak

Nomura's central thesis is that the current AI infrastructure investment cycle has likely not yet reached its zenith . The firm asserts that supply constraints and hyperscaler spending trends point to further upside, not a downturn . This stands in direct opposition to the market's fear that massive capex commitments are building toward a supply glut.

The report's rationale rests on three pillars: a structural supply chain bottleneck, hyperscaler spending that "can't stop," and a forecast for AI server revenue that far exceeds consensus.

The 'unprecedented supply chain mismatch' of H2 2026

Nomura warns of a critical shift in the second half of 2026. Even if GPU availability improves, the bottleneck will migrate to other components, creating what the firm calls an "unprecedented supply chain mismatch" . Identified pinch points include:

• Substrates and packaging materials — BT resin and copper foil for flip-chip chip-scale packages (FC-CSPs) .
• Printed circuit boards (PCBs) — essential for data center infrastructure.
• Power management ICs (PMICs) — a growing constraint as server complexity rises .
• Electronic materials and semiconductor equipment — wafers, packaging materials, and driver ICs .

Nomura's analysis suggests these shifting bottlenecks will constrain supply and create persistent strain through at least 2027, even as GPU production expands .

AI server revenue growth: 78% in 2026, 76% in 2027

The report includes notably bullish projections: AI server revenue will grow 78% year-over-year in 2026 and 76% in 2027 . These figures are significantly above consensus at the time of publication. For context, IDC separately reported that the worldwide server market hit $122.6 billion in Q1 2026 alone, up 30.4% year-over-year, driven overwhelmingly by AI infrastructure demand .

Why hyperscalers 'can't stop spending'

Nomura argues that the five largest hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) have no choice but to maintain or increase capital expenditure. Key drivers include:

• Rising memory costs — Nomura raised its 2026 DRAM price growth forecast from 38% to 57% and NAND from 36% to 65%, forecasting a memory supercycle extending to 2027 .
• Expanding data center blueprints — over 6.3 gigawatts of data center capacity are under construction in North America alone, with power availability emerging as the primary bottleneck .
• Competitive dynamics — the big five are collectively projected to spend ~$750 billion on capex in 2026, with about 75% directed at AI-specific infrastructure .

Buy ratings on nine Asian AI companies

Nomura issued Buy ratings and raised target prices on nine Asian AI technology companies . Verified beneficiaries include:

• TSMC — the primary beneficiary of AI chip demand and CoWoS advanced packaging .
• ASE Technology — a leading semiconductor packaging and testing firm.
• MediaTek — cited as a key player in TSMC's CoWoS client mix .
• Samsung and SK Hynix — major memory suppliers benefiting from the AI-driven memory supercycle .

How the report counters the Meta Compute selloff

On July 1, 2026, Bloomberg reported that Meta Platforms was building a cloud business ("Meta Compute") to sell excess AI computing capacity to external customers . The news triggered a sharp market reaction:

• Meta's shares surged nearly 10% as investors welcomed a path to monetize its massive capex .
• But semiconductor stocks plunged globally, as investors interpreted the move as a signal that AI infrastructure supply may be exceeding demand . Korea's stock market was hit particularly hard .

Nomura's counter-argument — published just before the Meta news (June 30) and reinforced by its broader report — is that the market is misreading the situation . Specifically:

• The AI infrastructure investment cycle has likely not yet peaked despite the pullback in semiconductor shares .
• Supply constraints and hyperscaler spending trends point to further upside .
• Constraints will persist through 2027, straining both AI and non-AI segments — meaning the supply-demand imbalance is structural, not a sign of overbuild .
• Meta's decision to rent out compute is actually consistent with a market where supply is scarce — companies are trying to optimize utilization of hard-to-get capacity, not signaling a glut .

Yuanta Securities Korea echoed this view, stating that Meta Compute was not a sudden signal of capex cuts, but an existing AI infrastructure expansion strategy made concrete .

Verdict and key caveats

Verified as accurate. The core claims of the report — the 119-page length, the anti-peak-semiconductor stance, the specific bottleneck warnings on substrates/PCBs/PMICs, the ~78% and ~76% AI server growth forecasts, the "can't stop spending" hyperscaler thesis, the Buy ratings on TSMC, ASE, and MediaTek, and the rebuttal to the Meta compute-driven selloff — are all corroborated by multiple high-quality sources including Nomura's own materials , financial news coverage , IDC data , and Korean business media .

Caveats:

• The exact 78% and 76% AI server revenue growth figures are cited in financial news coverage of Nomura's report ; the original 119-page PDF was not directly accessible in full text, but the numbers are consistent with Nomura's broader published forecasts.
• Not all nine named companies were publicly confirmed in a single source; TSMC, ASE, MediaTek, Samsung, and SK Hynix are explicitly verified .
• The report is notably contrarian — Deloitte, for instance, projects the semiconductor industry reaching $975 billion in sales in 2026 but does not share Nomura's skepticism about a potential peak .
• Morgan Stanley's Brian Nowak estimates the top five hyperscalers could spend nearly $800 billion in 2026 and $1.1 trillion in 2027, which if accurate would further support Nomura's thesis but also introduces risk if spending disappoints .

The core argument: Why this is not the peak

Nomura's central thesis is that the current AI infrastructure investment cycle has likely not yet reached its zenith . The firm asserts that supply constraints and hyperscaler spending trends point to further upside, not a downturn . This stands in direct opposition to the market's fear that massive capex commitments are building toward a supply glut.

The report's rationale rests on three pillars: a structural supply chain bottleneck, hyperscaler spending that "can't stop," and a forecast for AI server revenue that far exceeds consensus.

The 'unprecedented supply chain mismatch' of H2 2026

Nomura warns of a critical shift in the second half of 2026. Even if GPU availability improves, the bottleneck will migrate to other components, creating what the firm calls an "unprecedented supply chain mismatch" . Identified pinch points include:

• Substrates and packaging materials — BT resin and copper foil for flip-chip chip-scale packages (FC-CSPs) .
• Printed circuit boards (PCBs) — essential for data center infrastructure.
• Power management ICs (PMICs) — a growing constraint as server complexity rises .
• Electronic materials and semiconductor equipment — wafers, packaging materials, and driver ICs .

Nomura's analysis suggests these shifting bottlenecks will constrain supply and create persistent strain through at least 2027, even as GPU production expands .

AI server revenue growth: 78% in 2026, 76% in 2027

The report includes notably bullish projections: AI server revenue will grow 78% year-over-year in 2026 and 76% in 2027 . These figures are significantly above consensus at the time of publication. For context, IDC separately reported that the worldwide server market hit $122.6 billion in Q1 2026 alone, up 30.4% year-over-year, driven overwhelmingly by AI infrastructure demand .

Why hyperscalers 'can't stop spending'

Nomura argues that the five largest hyperscalers (Amazon, Alphabet, Microsoft, Meta, Oracle) have no choice but to maintain or increase capital expenditure. Key drivers include:

• Rising memory costs — Nomura raised its 2026 DRAM price growth forecast from 38% to 57% and NAND from 36% to 65%, forecasting a memory supercycle extending to 2027 .
• Expanding data center blueprints — over 6.3 gigawatts of data center capacity are under construction in North America alone, with power availability emerging as the primary bottleneck .
• Competitive dynamics — the big five are collectively projected to spend ~$750 billion on capex in 2026, with about 75% directed at AI-specific infrastructure .

Buy ratings on nine Asian AI companies

Nomura issued Buy ratings and raised target prices on nine Asian AI technology companies . Verified beneficiaries include:

• TSMC — the primary beneficiary of AI chip demand and CoWoS advanced packaging .
• ASE Technology — a leading semiconductor packaging and testing firm.
• MediaTek — cited as a key player in TSMC's CoWoS client mix .
• Samsung and SK Hynix — major memory suppliers benefiting from the AI-driven memory supercycle .

How the report counters the Meta Compute selloff

On July 1, 2026, Bloomberg reported that Meta Platforms was building a cloud business ("Meta Compute") to sell excess AI computing capacity to external customers . The news triggered a sharp market reaction:

• Meta's shares surged nearly 10% as investors welcomed a path to monetize its massive capex .
• But semiconductor stocks plunged globally, as investors interpreted the move as a signal that AI infrastructure supply may be exceeding demand . Korea's stock market was hit particularly hard .

Nomura's counter-argument — published just before the Meta news (June 30) and reinforced by its broader report — is that the market is misreading the situation . Specifically:

• The AI infrastructure investment cycle has likely not yet peaked despite the pullback in semiconductor shares .
• Supply constraints and hyperscaler spending trends point to further upside .
• Constraints will persist through 2027, straining both AI and non-AI segments — meaning the supply-demand imbalance is structural, not a sign of overbuild .
• Meta's decision to rent out compute is actually consistent with a market where supply is scarce — companies are trying to optimize utilization of hard-to-get capacity, not signaling a glut .

Yuanta Securities Korea echoed this view, stating that Meta Compute was not a sudden signal of capex cuts, but an existing AI infrastructure expansion strategy made concrete .

Verdict and key caveats

Verified as accurate. The core claims of the report — the 119-page length, the anti-peak-semiconductor stance, the specific bottleneck warnings on substrates/PCBs/PMICs, the ~78% and ~76% AI server growth forecasts, the "can't stop spending" hyperscaler thesis, the Buy ratings on TSMC, ASE, and MediaTek, and the rebuttal to the Meta compute-driven selloff — are all corroborated by multiple high-quality sources including Nomura's own materials , financial news coverage , IDC data , and Korean business media .

Caveats:

• The exact 78% and 76% AI server revenue growth figures are cited in financial news coverage of Nomura's report ; the original 119-page PDF was not directly accessible in full text, but the numbers are consistent with Nomura's broader published forecasts.
• Not all nine named companies were publicly confirmed in a single source; TSMC, ASE, MediaTek, Samsung, and SK Hynix are explicitly verified .
• The report is notably contrarian — Deloitte, for instance, projects the semiconductor industry reaching $975 billion in sales in 2026 but does not share Nomura's skepticism about a potential peak .
• Morgan Stanley's Brian Nowak estimates the top five hyperscalers could spend nearly $800 billion in 2026 and $1.1 trillion in 2027, which if accurate would further support Nomura's thesis but also introduces risk if spending disappoints .