the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe weekly #014 0:00 –:––
#014 ·21 AUG 2026 ·FRIDAY ·4 MIN READ ·10 STORIES + 10 EXTRAS

The quarter Anthropic won

Anthropic turned a quarterly profit while OpenAI’s losses widened, giving investors what frontier AI has never produced: a lab in the black. Nvidia guaranteed $105 billion of data-centre value, Stripe acquired the model-routing layer for $7 billion, and Washington told 35 countries to choose sides. The bottom line: the financial centre of gravity in AI moved this week, and every infrastructure bet — from Ohio gigawatt campuses to $60 billion rocket-company acquisitions — is being repriced against the new baseline.

01 / The Ten

The week, ranked

10

Anthropic outgrew OpenAI in Q2 and turned its first profit

Anthropic Q2 revenue more than doubled to $11.6B with a thin operating profit — the first for any frontier lab. OpenAI reached $6.7B, up 18% QoQ, while losses widened to $12.3B.

Why it mattersFirst quarter a frontier lab turned a profit reframes the IPO conversation from growth-at-all-costs to growth-with-margins.

Nvidia guarantees up to $105B for OpenAI Ohio campus

Nvidia disclosed residual-value guarantees covering the first 4.25 GW of SB Energy PORTS-Pike campus, which OpenAI will lease for 20 years. $1.5B went directly into SB Energy. Total design: ~8 GW.

Why it mattersCompute risk migrating from labs to chip suppliers. A chip maker backstopping the real estate that houses its products is a new financial structure.

Stripe acquires model router OpenRouter for over $7B

Bloomberg: Stripe agreed to buy OpenRouter, the gateway routing API calls across 400+ models to 8M users, for over $7B — 5x the $1.3B valuation three months earlier.

Why it mattersThe neutral switching layer between apps and labs turns out to be a payments problem. Whoever owns routing sees which models get used.

SpaceX closes $60B Cursor acquisition

SpaceX completed its acquisition of AI coding assistant Cursor for $60B, one of the largest software acquisitions on record.

Why it mattersA rocket company owning the most widely used AI coding tool has no clean precedent. Whether integration materialises determines the deal logic.

Anthropic projects $190–200B in 2028 revenue ahead of IPO

Bankers pricing Anthropic IPO working from an internal forecast of $190–200B in 2028 revenue, ~4–5x current run rate. Last valued at $965B.

Why it mattersIPO pricing anchors expectations for every subsequent AI listing. The gap between current revenue and the 2028 target is what the market is being asked to fund.

US tells 35 partners: join Pax Silica or China framework, not both

Washington plans to notify 35 partner countries that joining China rival AI chip framework means exclusion from Pax Silica. Countries asked to make a formal choice.

Why it mattersGeopolitical bifurcation of AI supply chains moved from theoretical to documented policy.

Groq raises $350M at half its 2025 valuation

Groq took $350M at $3.5B, down from $6.9B in Sept 2025, pivoting from custom-silicon to Nvidia-based inference cloud. Nvidia hired its founder.

Why it mattersThe custom-silicon challenge to Nvidia keeps ending with the challenger taking Nvidia money and selling inference capacity instead of chips.

Anthropic launches Cowork: Claude that edits your files without code

Anthropic shipped Cowork, a Claude Desktop agent operating directly on local files with user-set permissions. First consumer-facing agentic product beyond the API.

Why it mattersCowork is the first Anthropic product that literally opens your folders, positioned against Copilot and Gemini for Files.

OpenAI adds 20% compute overhead for safety monitoring

OpenAI disclosed new multi-stage CoT monitoring adding 20% compute overhead it will absorb. Follows Hugging Face breach and internal evals showing dangerous cyberoffence capabilities.

Why it mattersA frontier lab publicly eating a 20% cost hit for safety quantifies what responsible deployment actually costs at production scale.

Beijing permits ByteDance and Tencent to import 10K H200 chips each

Beijing allowed ~10,000 Nvidia H200 chips each at mainland data centres — a quiet relaxation of restrictions following US export controls.

Why it mattersExport controls have been the main US lever on the compute race. Any relaxation tests how durable that lever is.
02 / Also

Worth knowing

10
Alibaba releases Qwen 3.8-27B open-weight model
Dense 27B multimodal model, 262K context, Apache 2.0. Frontier-competitive at its parameter count.
huggingface.co ↗
Amazon destroying rare books to train AI
404 Media tracked a book to an Amazon facility where spines are cut and pages scanned for Nova model training. Copies do not survive.
404media.co ↗
Cerebras unveils CS-4: three wafer-scale chips in one rack
First new silicon since 2025 IPO. Three WSE-3 Turbo chips via Nexus interconnect for high-throughput inference.
reuters.com ↗
Anthropic ships watermark detection API
Third parties can identify Claude-generated text via invisible watermark. Dual use — verifying authorship and catching users — defines adoption.
the-decoder.com ↗
Google demonstrates practical homomorphic encryption for AI
Working code for AI inference on encrypted data. Expensive but now viable for specific enterprise workloads.
blog.google ↗
DOJ probes a16z over board seats at rival AI firms
Justice Dept examining whether a16z partners violate Clayton Act by directing competing AI companies. No decision after nearly a year.
fortune.com ↗
Etched transformer-chip valuation doubles to $21B
Custom silicon for transformer inference doubled valuation in a month, with no disclosed revenue or production customers.
techcrunch.com ↗
GLM-5.3 tops open model rankings, weights delayed
Zhipu AI model leads Chatbot Arena, undercuts rivals on pricing, but public weights delayed — API-only for now.
the-decoder.com ↗
Callosum raises $100M seed for AI model routing
London-based Callosum closed $100M seed from Atomico and UK Sovereign AI Fund to route tasks to cheapest capable model.
bloomberg.com ↗
52% of Americans now more worried about AI than excited
Pew: concern rose from 37% in 2021. Widening fastest among under-30s. Higher among frequent AI users than non-users.
pewresearch.org ↗
Be subscriber #013 one email a week · no spam · unsubscribe anytime
Back issues

The Archive

Every Friday · 10:00

Get the brief

One email a week. The ten things in AI that mattered, and why. Choose your channel.