Google Prioritizes Frontier AI Computing Power, Cloud Business Forced to Yield
nashnova research
Google CEO Sundar Pichai made a rare admission: the company is deliberately redirecting cloud compute to internal AI model development, even at the cost of near-term cloud revenue — this means the priority has shifted from 'sell to customers' to 'feed our own models first.'
Why is Google sacrificing cloud revenue to feed its own AI?
Google DeepMind and Google Cloud draw from the same compute pool. Pichai drew a hard line: frontier AI research gets its compute first, and external cloud capacity is planned around whatever remains.
This means → when compute is scarce, internal models eat first; external customers queue. Cloud revenue is being deliberately suppressed in the short term.
The immediate trigger: flagship model Gemini 4 Argon consumed 25 trillion tokens in its first 72 hours — three times Gemini 3.0's usage over the same window.
How severe is the compute shortage?
Enterprise demand for compute has broadly hit twice what companies planned for just two years ago. Traditional forecasting models have broken down.
Pichai confirmed that customers trying to book 2027 compute can barely secure enough capacity. He cited Amazon CEO Andy Jassy's assessment — cloud compute may already be pre-sold through 2028 — and called it "consistent with reality."
In plain terms = compute is now like sold-out concert tickets: the issue isn't price — it's that there's nothing left to buy.
Why can't supply keep up?
From investment decision to a data center going live takes two to three years. That construction cycle is nearly impossible to compress.
AI-agent-driven workloads are expanding exponentially, far outpacing original infrastructure plans.
This reflects an industry-wide bind: Google isn't the only one short on compute — every major cloud provider wishes its current reserves were double what they are.
Data centers in orbit — is Google serious?
Google has sent experimental TPU chips — tensor processing units, custom silicon designed for AI workloads — into sun-synchronous orbit, testing whether round-the-clock solar power can run a space-based data center.
Pichai drew a Waymo analogy: self-driving looked like science fiction early on, but Google's bet paid off. He believes orbital data centers could reach commercial scale within ten to twenty years.
This means → with ground-based expansion constrained by community opposition and grid-connection backlogs, the space route is Google's ultra-long-term hedge against the compute ceiling — but whether it delivers on a commercial timeline remains an open question.
市场有风险，内容仅供研究参考，不构成投资建议。