Nomura: 75% Probability AI Arms Race Continues, Memory Chip Shortage by 2027 Virtually Locked In
nashnova research
Nomura sees a 75% probability that the AI capex race persists and 2027 memory supply is already locked via long-term contracts; Big Tech free cash flow has turned negative across the board, yet no one dares blink first — because falling behind costs more.
Why won't anyone hit the brakes?
Nomura frames the competition as a "game of chicken" — whichever company cuts AI spending first risks falling behind in the next round. This means → sustaining the burn is the rational choice, even as cash runs out.
TSMC CEO C.C. Wei says AI demand still exceeds supply capacity. Amazon CEO Andy Jassy says even with 2026 capex rising to $220 billion, the company cannot meet all 2026–2027 demand — and some 2028 demand has already been pre-booked.
Microsoft and Alphabet management confirm the same: cloud businesses are capacity-constrained. In plain terms = the bottleneck is not demand — it is the inability to build fast enough.
Why is 2027 memory supply "already locked in"?
Nomura notes that 2027 memory volumes and prices were locked through long-term agreements in October 2026, with prices rebounding.
Cancellation clauses are punitive: the defaulting party may owe 50%–100% of annual demand value as a penalty. This means → buyers cannot easily walk away, and sellers hold unusually firm order books.
This reflects a rare level of demand visibility and margin protection for memory suppliers such as Samsung, SK Hynix, and Micron in this cycle.
Can the giants still afford this?
Alphabet posted negative $5.9 billion free cash flow in Q2 2026. Meta: negative $2.1 billion in Q3. Amazon: negative $7.6 billion over the twelve months to June 2026.
Microsoft still showed $19.6 billion in positive FCF over the same period, but Nomura expects pressure ahead. In plain terms = three of the four are already spending more than they earn; the fourth is next.
Planned 2026–2027 capex remains enormous: Alphabet $195–205 billion, Meta $135–145 billion, Amazon roughly $220 billion, Microsoft about $175 billion for FY2027.
Not enough cash? Borrow.
Alphabet raised $49.6 billion in June 2026 through common stock and mandatory convertible preferred shares.
Meta issued $24.9 billion in bonds in Q2, paused buybacks, and authorized non-voting Class C shares to fund AI investment. This means → Big Tech has shifted from "invest from profits" to "invest on borrowed money."
CEOs are publicly justifying the spend: Google's Sundar Pichai says frontier-model training takes priority over cloud revenue; Mark Zuckerberg argues that even if superintelligence arrives later than expected, added compute can be redirected to ads and recommendation systems — giving shareholders a "heads I win, tails I don't lose" narrative.
What could stop the race?
Nomura outlines three scenarios: race continues at 75%, capex slowdown at 20%, financing failure at 5%. This means → the biggest risk is not demand evaporating — it is capital markets closing the door.
Power is the other hard constraint: a Sightline Climate survey shows 30%–50% of planned 2026 data-center additions face delays, with over 11 GW of capacity showing no sign of breaking ground.
Yet Nomura argues power shortages are more likely to delay and reshape demand than to destroy it. In plain terms = less power → higher premium on new chips and HBM — high-bandwidth memory, a type of ultra-fast DRAM designed for AI workloads — that deliver more performance per watt → older GPUs face replacement pressure.
What should investors watch?
2027 memory orders are largely locked in — near-term supplier margins have a floor.
The real tracking question is whether two forces can tighten simultaneously before 2028: deteriorating financing conditions (rates, equity markets, bond markets) and persistent power bottlenecks.
This reflects the market's core tension: demand certainty is high, but the funding chain underwriting that demand is stretching longer and growing more fragile — if it snaps, the correction will be sharp.
市场有风险，内容仅供研究参考，不构成投资建议。