the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe daily 0:00 –:––
/daily ·19 AUG 2026 ·WEDNESDAY ·3 MIN READ ·8 STORIES

Revenue day: the scorecard nobody wanted to wait for

OpenAI and Anthropic released comparable quarterly numbers for the first time, and they suggest the two biggest frontier labs are growing at very different speeds. Meanwhile Beijing quietly shipped 20,000 H200s, Cerebras unveiled a three-chip rack, and most Americans decided they are worried about all of it.

01 / The Day

WEDNESDAY 19 AUG 2026, ranked

08

Anthropic outgrew OpenAI in Q2 and turned a small profit

The Wall Street Journal reports OpenAI's Q2 revenue reached $6.7bn, up 18% quarter on quarter, while losses widened from $9.3bn to $12.3bn. Anthropic's sales more than doubled to $11.6bn in the same period, with a thin operating profit — the first for any frontier lab.

  • OpenAI told investors Q3 growth has since accelerated, without disclosing a figure
  • Anthropic's Q2 revenue broadly tracks the $65bn annualised run rate reported last week
  • A small operating profit at Anthropic's scale suggests compute cost per token is falling faster than price
Why it mattersThe labs have been competing on vibes and benchmarks; this is the first credible quarter-to-quarter comparison in actual money.

OpenAI adds monitoring that costs 20% more compute after Hugging Face breach

OpenAI disclosed new multi-stage chain-of-thought monitoring for some inference workloads, adding 20% compute overhead it will not pass to customers. The move follows a breach at Hugging Face and internal evaluations showing dangerous cyberoffense capabilities in agentic models.

  • The overhead applies to monitored workloads only; standard API calls are unaffected
  • OpenAI says it is 'pacing model development' as AI cybersecurity risks have reached levels it considers too dangerous to ignore
  • A separate Decoder report found AI systems quietly drop user instructions when compressing context
Why it mattersA frontier lab publicly slowing development and eating a 20% cost hit for safety is a more concrete posture shift than any policy statement.

Beijing let ByteDance and Tencent import 10,000 H200 chips each

The Financial Times reports Beijing permitted ByteDance and Tencent to receive approximately 10,000 Nvidia H200 chips each at mainland China data centres in recent weeks — a quiet relaxation of the restrictions that followed US export controls.

  • H200 chips are the highest-spec Nvidia GPU currently permitted into China under export control rules
  • The move is framed as helping Chinese tech firms catch up with US rivals, not as a formal policy reversal
  • Both companies have domestic chip alternatives in development, but their own supply remains constrained
Why it mattersExport controls have been the main lever the US is pulling on the compute race; any relaxation tests how durable that lever is.

Cerebras unveils CS-4: three wafer-scale chips in one rack

Cerebras Systems announced the CS-4, a server rack housing three WSE-3 Turbo chips connected via its new Nexus interconnect architecture. First commercial shipments are expected in Q3 2026, positioned primarily for high-throughput inference.

  • The WSE-3 Turbo is the updated version of Cerebras's dinner-plate-sized wafer-scale engine
  • Nexus manages chip-to-chip communication inside the rack without external networking
  • Cerebras went public as CBRS in late 2025; this is its first new silicon since the IPO
Why it mattersCerebras is the most credible non-Nvidia option for very fast inference; the CS-4 sets its new performance baseline.

Unitree stock rose 629% on its Shanghai debut after a $904M IPO

Unitree Robotics raised 6.1bn yuan ($904m) in its Shanghai IPO and opened to a buying frenzy, with shares surging dramatically on the first day. The humanoid-robot maker's listing is the largest in China's robotics sector to date.

  • Unitree is best known for dog-like quadruped robots and lower-cost humanoid frames sold to researchers and developers
  • The listing follows a wave of Chinese robotics private funding rounds that have drawn comparisons to the AI lab fundraising cycle
  • First-day IPO surges of this magnitude on the Shanghai exchange are common for high-profile debut issues
Why it mattersPublic market pricing for a humanoid robot maker sets a reference point for an asset class priced until now entirely in private rounds.

Temporal in talks to raise $500M at $12B+, six months after its $5B round

Bloomberg reports Temporal Technologies, whose open-source workflow orchestration platform underpins many production AI agents, is in talks to raise approximately $500m at a valuation of at least $12bn — up from $5bn after a $300m raise in February.

  • Temporal's durable execution model handles long-running workflows that survive failures, increasingly used for multi-step agents
  • A $12bn valuation would represent a 2.4x increase in six months on no disclosed revenue change
  • No lead investor or closing date has been announced
Why it mattersThe orchestration layer between a model and the world it acts on is being priced as if it will capture significant value from the agent wave.

More Americans are worried about AI than excited about it

A Pew Research Center survey finds 52% of Americans say they are more concerned than excited about AI in daily life, up from 37% in 2021. The gap has widened fastest among adults under 30, 55% of whom now lean toward concern — a reversal of the younger cohort's earlier enthusiasm.

  • Job displacement is the leading stated concern across all age groups
  • Concern is higher among frequent AI tool users than those who do not use AI — the opposite of what boosters predicted
  • The 15-point five-year shift is Pew's largest recorded movement on this question
Why it mattersPublic sentiment is a lagging constraint on regulation; a double-digit swing in five years suggests the lag may be shortening.

Anthropic publishes Claude protein design results, plans scientist access programme

Anthropic released results from two internal experiments in which Claude assisted with protein design and analytical chemistry, finding AI-assisted iteration compressed weeks of work into days. The company plans a dedicated access programme for life-science researchers.

  • Protein design: Claude proposed and ranked candidate sequences; lab confirmed folding behaviour faster than baseline iterations
  • Analytical chemistry: Claude identified likely impurities in reaction outputs without being prompted for specific hypotheses
  • The access programme has no announced date or pricing
Why it mattersLabs claiming productivity multipliers in wet-lab science are easy to make; this is notable because Anthropic is publishing the methodology, not just the headline number.
Be subscriber #013 the weekly ten, every friday by email · no spam · unsubscribe anytime
Past days

The Daily Archive