the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe daily 0:00 –:––
/daily ·15 NOV 2026 ·SUNDAY ·3 MIN READ ·7 STORIES

Sunday Reckoning

Anthropic's CEO admits the company is a bet on an unknowable revenue curve, Nvidia's exit reverberates through three Asian governments, and Google's best maths model is impressively capable but prohibitively expensive to run.

01 / The Day

SUNDAY 15 NOV 2026, ranked

07

Anthropic CEO admits 'huge uncertainty' ahead of history's largest IPO

Days before a reportedly six-times-oversubscribed offering targeting a $1.4 trillion valuation, CEO Dario Amodei told prospective investors that buying compute is 'the deepest business dilemma' Anthropic faces — and admitted there is 'a huge uncertainty in what's going to happen next year' around whether revenue will keep pace with roughly $901B in committed long-term compute spend.

  • IPO six times oversubscribed; $1.4T valuation would be the largest in market history
  • Amodei: committed compute runs into the tens of billions in the near term alone
  • Banks blocking mainland China and HK investors on ITAR grounds
Why it mattersWhen a CEO admits before an IPO that revenue trajectory is genuinely unknown, the market is pricing the dream rather than the fundamentals.

Gemini 4 Carbon tops MathArena — at 15x the cost per problem

Google's Gemini 4 Carbon scored highest on the MathArena competition-maths leaderboard with 85.4% expected accuracy, displacing GPT-6.1 Sol and Claude Fable 5.2 — but it costs roughly $15.70 per problem versus $1.07 for Sol, a 15x gap driven by extended internal reasoning loops.

  • Carbon leads MathArena at 85.4% expected accuracy, ahead of GPT-6.1 Sol and Claude Fable 5.2
  • Per-problem cost: ~$15.70 (Carbon) vs ~$1.07 (GPT-6.1 Sol) on the MathArena eval
  • Carbon uses internal long-loop reasoning and ignores standard benchmark time limits
Why it mattersBenchmark supremacy is becoming an expensive crown — the cost gap tells you which models labs will actually reach for in production.

TSMC launches $500M fund to seed the chip ecosystem it needs

TSMC's board approved a $500 million strategic venture fund — announced at the Nikkei Forum in Tokyo by Deputy Co-COO Kevin Zhang — that will back emerging chip designers and, in a break from precedent, invest directly in upstream suppliers and IP partners.

  • Fund size: $500M (NT$16.45-16.49B); first TSMC fund to invest upstream in suppliers
  • Stated aim: cultivate the next major chip company from within TSMC's design ecosystem
  • Board resolution confirmed via TSMC Form 6-K SEC filing
Why it mattersThe world's most critical semiconductor manufacturer funding its own supply chain closes a structural vulnerability while extending TSMC's gravitational pull across the chip ecosystem.

Applied Digital asks Oracle and Meta for billions to cover its power bets

Bloomberg reports that Applied Digital has requested multi-billion-dollar corporate guarantees from anchor tenants including Oracle and Meta before finalising data centre lease agreements — revealing that hyperscale AI infrastructure operators are demanding risk-sharing from the very tech giants creating compute demand.

  • Guarantees requested since at least mid-2026 to secure long-term power and infrastructure costs
  • Oracle and Meta named as tenants asked to backstop Applied Digital's capital commitments
  • Reflects the widening mismatch between AI demand timelines and infrastructure capital cycles
Why it mattersWhen data centres need guarantees from the companies deploying AI, the infrastructure financing model has reached an inflection point.

Anthropic's Mythos can be retrained at fraction of original compute

Former Anthropic researcher Max Coxon claimed Anthropic can already re-run the full Mythos training process with a fraction of the compute originally needed — a claim that, if verified, would signal the per-dollar commitments in Anthropic's IPO prospectus systematically overstate future training costs.

  • Coxon: Mythos retraining possible at a fraction of original compute; exact ratio not disclosed
  • IPO context: Anthropic committed ~$901B long-term compute; cheaper retraining reduces risk
  • Claim is from a former researcher, not an official Anthropic statement
Why it mattersIf training costs have collapsed internally, the economics of frontier AI shift from one-time bets to repeatable, affordable cycles — a material change for the whole field.

Runway open-sources Praxis-1: video AI turned robot controller

Runway released Praxis-1, an open-source model that reuses learned video-AI representations to control physical robots — converting the temporal and spatial understanding built for video generation into a generalist robot policy without training on robotic demonstration data.

  • Open-source: weights and training code published to Runway's GitHub
  • Architecture reuses video-generation representations, not trained on robot demos
  • Generalisation to unseen tasks claimed; independent robotic-manipulation benchmarks pending
Why it mattersIf video models transfer to robotics without additional robot training, the billions spent on video AI may have already trained a robot brain.

Asia scrambles as Huang exit rattles chip supply chains

Governments of Singapore, Taiwan, and South Korea issued formal responses after Jensen Huang's departure from Nvidia: Taiwan's MOEA warned of significant supply-chain implications, Singapore's MTI summoned Nvidia AP executives, and South Korean Naver Data Cloud alerted customers to enterprise support uncertainty.

  • Taiwan MOEA: Huang exit has significant supply-chain implications for the island's chip ecosystem
  • Singapore MTI summoned Nvidia Asia-Pacific executives for supply-chain clarification
  • South Korea's Naver Data Cloud issued customer alert over Nvidia enterprise support uncertainty
Why it mattersThree governments formally responding to a CEO change at a chip company confirms Nvidia is now national infrastructure — with the geopolitical exposure that entails.
Be subscriber #028 weekly AI email interest list
Past days

The Daily Archive