← AM·PM Brief

Gemini 4 Usage Triples in 72 Hours, Launch Halted by Compute Crisis | AI Daily Brief

· Evening brief · 10 news · 9:02

Audio in Mandarin Chinese · English transcript below

⚡ Gemini 4 Argon burned 25T tokens in 72h—3x Gemini 3.0's launch—Pichai signals Google is winning the inference war.

Compute is emerging as the primary hard constraint on the commercialization of top-tier LLMs. The launch of Gemini 4 Argon alone consumed nearly 25 trillion tokens, confirming the exponential surge in demand—yet Google has been forced to deploy distilled models and staggered rollouts to ease its compute hunger, with infrastructure now becoming the competitive bottleneck. What merits particular attention is that the sandboxing and testing required for Agent deployment further slows the pace; the core of the AI race is shifting from parameter scale to engineering delivery and the efficiency of compute orchestration.

Gemini 4 Usage Triples in 72 Hours, Launch Halted by Compute Crisis | AI Daily Brief

Today's Top 3 Headlines

  1. Opinion

    📊 Google CEO Sundar Pichai: Gemini 4 usage tripled in first 72 hours, launch constrained by compute

    Gemini 4 Argon shows strong agentic and coding capabilities, burning ~25T tokens in first 72h—3x Gemini 3.0's pace; compute constraints limit rollout to API and Pro/Ultra subscribers, with distilled smaller models expanding reach. Real-time monitoring, sandboxing and external testing for Agent deployment slow the rhythm; Pichai admits coding still has room to improve.

    Source ↗
  2. Opinion

    Arvind Krishna: Quantum computers may crack current encryption around 2030, migrate stacks now

    Krishna says DARPA's quantum benchmark program shrank from hundreds to ~13 participants; taking the midpoint of 2028–2033, existing cryptography will be broken by 2030—"not hypothetical." Post-quantum crypto is ready but unimplemented; stolen encrypted drives may be decrypted in 5 years, so replace IT stacks now with post-quantum keys as bridge. He admits this security rests on complexity assumptions, not proofs, so timing remains a range.

    Source ↗
  3. Opinion

    TBPN: SpaceX Q3 revenue up 161%, Musk targets $1T annual revenue run rate

    TBPN recaps SpaceX Q3: revenue up 161% to $14.6B, profit $4.4B; one partial Starship launch carries ~10x Falcon 9's Starlink payload. Call focused on ground compute: Musk targets 2.5–3GW by year-end, 10GW next year, predicts $500B–$1T recurring revenue by end-2027; TBPN flags this mirrors Oracle-style forward revenue claims from Sept 2025.

    Source ↗

+7 more headlines

  • Sameer: Look at 12-month revenue multiple, not current ARR
  • Claire Vo: Claude Duet best polished, but passive, no desktop, still can't enter workflow
  • Nando de Freitas: Local coding model Flint up to 2x faster than peers
  • Greg Brockman: GPT-Rosalind Proposes Restoring Cholesterol Transport or Protecting Neurons
  • Yoshua Bengio: AI Misuse Could Be Catastrophic, Developer Choices Shape Risk
  • Sarah Guo: Beyond Red Teaming, Can Formal Verification & Secure Distribution Tip Defense?
  • Kai-Fu Lee: US & China Chart Different AI Paths, Human Connection Grows Vital
Unlock all 10 headlines + deep analysis →Free 7-day trial · cancel anytime
Browse all past briefings →