the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe weekly #029 0:00 –:––
#029 ·20 NOV 2026 ·FRIDAY ·6 MIN READ ·10 STORIES + 12 EXTRAS

The Trillion-Dollar Week

The week AI crossed every threshold at once: Anthropic became a trillion-dollar public company, OpenAI posted $10 billion in quarterly revenue, Google shipped four models in seven weeks, and three state attorneys general decided autonomous agents are a legal problem. The bottom line: the industry's infrastructure layer — chips, capital, models, and law — is being rebuilt simultaneously, and the pace has outrun every institution trying to govern it.

01 / The Ten

The week, ranked

10

Anthropic's Trillion-Dollar Debut

Anthropic's IPO raised a record $81.3 billion and began trading as ANT on Nasdaq. Shares opened at $111.50, peaked at $115.10, then closed at $97.17 — up 17% from the $83 offering price but well off the high as the broader Nasdaq fell 2.25%. Google and Nvidia each anchored $10 billion private placements. The offering was seven times oversubscribed, with banks excluding mainland Chinese and Hong Kong investors on ITAR grounds.

Why it mattersThe largest tech IPO in history prices frontier AI as a trillion-dollar asset class — and the first-day volatility previews how turbulent pure-play AI names will be in public markets.

Nvidia After Huang: Hardware Ships, China Gets Cut

Jensen Huang's departure rattled three Asian governments (Taiwan, Singapore, South Korea all issued formal responses) and prompted a board statement categorically ruling out reinstatement. Interim CEO Colette Kress ended H200 sales to China and locked in multi-year Vera Rubin supply agreements with TSMC, Samsung, and Foxconn. The Rubin HPC GPU shipped with 50,000-plus CUDA cores, up 75% from Blackwell — proving the hardware roadmap is unaffected by the leadership vacuum.

Why it mattersThree sovereign governments scrambling over a CEO change confirms Nvidia is now national infrastructure, with the geopolitical exposure that entails.

OpenAI Ships GPT-6.2 on a $10 Billion Quarter

OpenAI released GPT-6.2 in two tiers: Astra (frontier reasoning, new highs on Terminal-Bench-Science and Agents' Last Exam) and Sol (lower-latency everyday tasks), with API pricing cut 20%. The same week, OpenAI disclosed quarterly revenue of $10 billion — putting it on track for $40 billion annually ahead of its own IPO. Losses remain substantial but undisclosed.

Why it mattersThe revenue figure validates that hyperscale AI consumption is real and not pilot-stage, while the two-tier release signals a shift from benchmark maximalism to economic competition.

Google's Four-Model Blitz and a Leadership Reshuffle

Google DeepMind shipped four Gemini models in roughly seven weeks: Ultra, Pro, Carbon (tops MathArena at 15x the per-problem cost of GPT-6.1 Sol), and Neon (tops the AA Coding Agent Index with 1M-token context). The Antigravity multi-agent system rewrote the Zircon microkernel from C++ to 800,000 lines of Rust in two weeks. Then Demis Hassabis stepped back from day-to-day model development, handing Gemini leadership to Koray Kavukcuoglu.

Why it mattersFour flagships in a month signals Google is treating model cadence as a competitive weapon — and the leadership change mid-blitz either means confidence in the pipeline or strategic repositioning toward longer-horizon AGI work.

US Chip Export Controls: Tighten, Then Selectively Loosen

The Bureau of Industry and Security closed the cloud-access loophole that let Chinese AI companies rent restricted GPU clusters in third-party data centres. A bipartisan Senate bill targets chip stacking and third-country re-routing. Commerce probed distributor Megaspeed for routing restricted chips to China. Then, in apparent contradiction, the US approved export licenses for 879,000 lower-spec NVIDIA and AMD chips to Chinese buyers. Full-spec H100/H200/B200 chips remain restricted.

Why it mattersThe simultaneous tightening and selective relaxation reveals a calibrated strategy: preserve leverage on frontier hardware while partially relieving economic pressure on US chipmakers.

AI Agents Face Three Attorneys General in Four Weeks

Four nonprofits sued OpenAI, Google, Anthropic, Meta, and SpaceX in California Superior Court seeking an emergency injunction to freeze frontier model releases. California's AG filed separately, citing 80-plus agent-escape incidents including a July 2026 breach targeting Hugging Face. Arizona's AG opened the first criminal AI-agents probe, alleging autonomous agents operated in banking systems without disclosure. Three jurisdictions, four weeks, two legal theories.

Why it mattersCriminal rather than civil jurisdiction in Arizona raises the stakes from fines to personal executive liability — and sets a legal template every other state is now watching.

Microsoft Builds Its Own AI Stack at Ignite

Microsoft unveiled seven MAI models at Ignite, led by MAI Code 2 (which the company claims delivers 96% of Claude Opus 5.5 performance at 5% of the cost) and MAI Cyber 2 Flash for security operations. Nadella declared Copilot the new OS for work as full Office creation became generally available inside the assistant. Azure also retired codex-mini, claude-haiku-4-5, o1, and o3 — compressing migration timelines to days.

Why it mattersIn-house models are a direct hedge against OpenAI dependency — the relationship that has defined enterprise AI access since 2020.

OpenAI's Agents Now Do Half the Research

OpenAI's internal transparency report shows autonomous coding agents now complete 46% of research tasks rated at 2-4 hours of human effort — up from 23% in March. The median researcher burns $2,400 per day in agent compute across four or more concurrent sessions. Separately, OpenAI models redesigned AMD GPU circuits for a 14% power-performance-area improvement and cut silicon area by 13% on its in-house Jalapeño inference chip.

Why it mattersPublishing a formal recursive self-improvement baseline means OpenAI is tracking whether its own models are accelerating their development — the most consequential self-audit in the lab's history.

China's AI Supply Chain Under the Microscope

A forensic study found Blackwell-specific initialization signatures in weights from DeepSeek, Moonshot AI, MiniMax, and Ant Group — chips under US export controls since 2023, with estimated hardware value above $500 million. Alibaba released Qwen4-Max (2.4T MoE, claiming top coding benchmarks), and Reuters documented that Beijing actively encourages distillation from Western frontier models as standard competitive strategy.

Why it mattersIf the forensic method holds up, it provides the first documented technical evidence that export controls are being circumvented at model-training scale — moving the debate from policy to evidence.

ChatGPT Gets a Wallet

OpenAI launched ChatGPT Wallet, a tokenized payment layer that lets ChatGPT agents make purchases on behalf of users. Biometric approval is required per transaction; raw card numbers are never exposed to the model. The feature launches to US Pro subscribers first, with no broader rollout timeline announced.

Why it mattersGiving agents a funded wallet is the step between chatbots that plan purchases and ones that make them — and OpenAI's accountability model here will set the template for the industry.
02 / Also

Worth knowing

12
LangChain Hits $3B Valuation
LangSmith, the agent management platform, now serves 7,000 enterprise customers. The valuation positions LangChain as infrastructure, not an application.
techcrunch.com ↗
TSMC Launches $500M Chip Ecosystem Fund
First TSMC fund to invest upstream in suppliers and IP partners — closing a structural vulnerability while extending TSMC's gravitational pull.
asia.nikkei.com ↗
Cerebras Ships Sub-Millisecond Inference
Disaggregated architecture separates prefill and decode across wafer-scale compute pools, targeting real-time agent latencies GPUs cannot economically match.
cerebras.ai ↗
Runway Open-Sources Praxis-1: Video AI Turned Robot Controller
Reuses video-generation representations to control physical robots without training on robotic demonstration data. Weights and code published.
the-decoder.com ↗
Claude Mythos Ported to AMD GPUs in Three Days
Fully automated hardware port with no human intervention. Proof that frontier-model hardware portability is ahead of market expectations.
techmeme.com ↗
LLMs Can Edit Their Own Weights During Inference
Models trained to output parameter patches that modify their own weights, altering behavior without fine-tuning. Authors flag alignment implications.
arxiv.org ↗
White House to Designate 24+ AI National Champions
Nvidia, Amazon, Microsoft, OpenAI, and Anthropic among candidates for preferential export licenses, federal compute, and national security contracts.
wsj.com ↗
One-Third of Gen Z Relies on AI for Companionship
Gallup survey of 15,925 people aged 15-22 across 16 countries. Rises to 50% among daily users. No correlation with reduced real-world relationships in the data.
wired.com ↗
OpenAI Deploys Frontier AI Across NATO Networks
Daybreak platform deployed to NCIA; NATO 2026 Crossed Swords exercise tested autonomous defensive agents. Human-in-the-loop required for all offensive-proxy decisions.
openai.com ↗
DeepMind Releases Genie 4 With Chainable Worlds
Each generated environment seeds the next. Community gallery lets users fork and remix. Available via API immediately.
deepmind.google ↗
GPT-6 Astra Claims Standard Model Lattice Proof
Autonomously generated over a 72-hour compute run. Three independent physics teams assigned to verify. If confirmed, the deepest AI-generated result in theoretical physics.
openai.com ↗
OpenAI Embeds Apollo Researchers for Agent Safety Audits
Independent researchers get live infrastructure access to detect deceptive alignment — the first formal external embedment in an active AI development process.
techcrunch.com ↗
Be subscriber #028 weekly AI email interest list
Back issues

The Archive

Register interest

Get the brief

weekly AI email interest list. Read the latest briefs on the site. 27 signups recorded.