Gemini is Cooked but GCP is Cooking

Max Kan · SemiAnalysis · August 07, 2026 at 02:32 · ⏱ 10 min read  | Read on Substack ↗
Summary
Google's DeepMind has effectively lost the frontier AI model race due to severe talent drain and a timid corporate culture, culminating in a massive leadership overhaul. However, this long-term AI failure is a massive short-term catalyst for Google Cloud (GCP), which is now aggressively monetizing its TPU infrastructure by selling compute to competitors, driving massive revenue acceleration and EPS upside.
  • •Google overhauled DeepMind leadership, with co-founder Demis Hassabis stepping back and Jeff Dean leaving to start a neolab called Discovery Loop.
  • •Gemini 3.5 Pro was silently canceled after Gemini 3.5 Flash flopped, leaving Gemini in 8th or 9th place globally behind models from OpenAI, Anthropic, and open-source competitors.
  • •GCP is prioritizing short-term financialization over DeepMind's frontier ambitions, selling over 20% of total TPU shipments from 3Q26 to 4Q27 directly to Anthropic.
  • •GCP has established a new revenue stream selling full TPU systems to external SPVs at approximately $35B per GW, generating an estimated $1.2B in 2Q26.
  • •Driven by an estimated $150B+ in TPU systems backlog, GCP's 2027 growth rate is projected to hit the mid-100s, significantly above the sellside consensus of 64%.
  • •The acceleration in GCP growth and high-margin TPU system sales is expected to add approximately $3 to Google's EPS in 2027.
Read time 10 min
Length 10,162 chars
Category finance
Ideas
Max Kan Substack author, SemiAnalysis
The author projects that GCP's pivot to selling TPU systems to external SPVs will drive 2027 cloud growth into the mid-100s (vs 64% consensus) and add ~$3 to Google's 2027 EPS.
The author projects that GCP's pivot to selling TPU systems to external SPVs will drive 2027 cloud growth into the mid-100s (vs 64% consensus) and add ~$3 to Google's 2027 EPS. Risk: Abandoning long-term competitiveness at the AI frontier could eventually threaten Google's core search and advertising moats despite the short-term cloud revenue boom.
GOOGL
Max Kan Substack author, SemiAnalysis
Meta is identified as a major beneficiary of GCP's compute sales, with Google committing to rent 'many hundreds of thousands more' TPUs to Meta over the next 6 quarters.
Meta is identified as a major beneficiary of GCP's compute sales, with Google committing to rent 'many hundreds of thousands more' TPUs to Meta over the next 6 quarters. Risk: Pre-paying for multiple GWs of compute is highly expensive and requires massive capital allocation to sustain open-source AI leadership.
META
More from SemiAnalysis
Kimi K3, The Manos, The Mythos, The Legendos
Aug 03, 19:42
The Wild Wild West Of LEGO Datacenters
Jul 29, 22:09
Can AMD break the CUDA Moat? AMD Advancing AI 2026
Jul 25, 00:33
Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis
Jul 23, 00:47
Meta’s Infrastructure Team Needs A Culture Reset
Jul 22, 02:41

This newsletter, published August 07, 2026, features Max Kan discussing GOOGL, META. 2 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Max Kan  · Tickers: GOOGL, META