Inside the NVIDIA Rubin Order Rush — AI Datacenter Capex: Bubble or Inevitable?

NVIDIA Rubin is shipping, and hyperscaler order books are exploding. Microsoft, Google, Meta, Amazon, OpenAI, and xAI placed a combined ~$30B+ in one quarter, per reporting. Let's lay out the Blackwell delta, the power constraint, and whether this is a bubble.

Rubin and Rubin Ultra at a glance

• Rubin GPU: 288GB HBM4, 2.5x training / 3.3x inference vs Blackwell
• Rubin Ultra: 4-die package, 1TB HBM4 per unit, 100 PFLOPS (FP4) inference
• NVL576 rack: 72×Rubin Ultra → 4.6TB GPU memory per rack
• Power: 150kW per rack (up from Blackwell NVL72's 120kW)

A single rack burns the power of 200 apartments. A hyperscale building approaches a nuclear plant unit's output.

Order book

Rough order ranges (sourced):

• Microsoft: ~$20B/year (Azure/OpenAI)
• Google: ~$10B alongside in-house TPU v6
• Meta: ~$8B (Llama 4/5 training + Reality Labs)
• xAI: ~$6B (Memphis Supercluster expansion)
• OpenAI: ~$10B via the Stargate project

Quarterly total >$30B isn't a stretch.

Power and cooling become the bottleneck

Rubin's real constraint isn't GPU supply — it's power and cooling.

• US datacenter approvals routinely take 3 years
• Ireland, Netherlands, Singapore have paused new permits
• In Japan, Tokyo Bay and Inzai are at capacity, shifting to Hokkaido/Tohoku
• Liquid cooling gives way to immersion; facility costs jump

Jensen Huang said it plainly: "Power is the new currency of AI." Amazon, Microsoft, and Google are now investing in small modular reactors (SMRs).

On the bubble debate

Silicon Valley is increasingly split — "Rubin orders are peak bubble" / "demand is real." Real risks:

• LLM improvement curves plateauing would crater training GPU demand
• Inference moves to dedicated ASICs (Groq, Cerebras, Etched)
• China's domestic semi self-sufficiency could cut 30% of global demand

But all of these going wrong simultaneously is a low-probability scenario. My base case is "correction in 2027–28 after over-investment" — similar to post-dot-com. Survivors absorb cheap infrastructure.

What this means in Japan

• SoftBank, KDDI, Rakuten, Sakura Internet's GPU cloud businesses may be sitting on excess capacity by 2028. Long-term anchor customers locked in now matters
• Enterprise AI adoption stays "rent someone else's cloud" — buying Rubin doesn't make sense
• Semi equipment and materials (Tokyo Electron, SCREEN, SUMCO, Shin-Etsu) stay in tailwind through 2028

FAQ

How much faster is Rubin than the Blackwell generation?

Rubin GPUs carry 288GB of HBM4 and deliver roughly 2.5x training / 3.3x inference over Blackwell. The higher-end Rubin Ultra integrates 4 dies for 1TB HBM4 per unit and 100 PFLOPS (FP4) of inference. At rack scale, NVL576 bundles 72 Rubin Ultra units for 4.6TB of GPU memory per rack.

Is AI data center investment a bubble?

Orders keep running at tens of billions per quarter, yet some argue capacity is "2x real demand." Plateauing LLM gains, inference shifting to dedicated ASICs (Groq, Cerebras, Etched), and rising Chinese self-sufficiency are all risks — but all failing at once is low-probability. My base case is a "2027–28 correction after over-investment," where, like post-dot-com, survivors absorb cheap infrastructure.

What's the power draw per rack, and is the grid a problem?

Rubin Ultra's NVL576 pulls 150kW per rack (up from Blackwell NVL72's 120kW). The real constraint isn't GPU supply but power and cooling — US datacenter approvals routinely take three years, and the shift from liquid to immersion cooling spikes facility costs. As Jensen Huang put it, "power is the new currency of AI," and securing it becomes the bottleneck on investment.

Should Japanese companies buy Rubin in-house?

For most manufacturers, "rent someone else's cloud" is enough — buying Rubin in-house rarely makes sense. GPU cloud providers (SoftBank, KDDI, Rakuten, Sakura) risk excess capacity by 2028, so locking in long-term anchor customers now is the right move. Semi equipment and materials (Tokyo Electron, SCREEN, SUMCO, Shin-Etsu) stay in tailwind through 2028.

Related

• Geopolitics of TSMC's 1nm era