Today's News
Anthropic Q3 revenue of $16.8B annualizes to $76B, but 45% QoQ growth signals deceleration from 147% prior quarter
Azeem Azhar, Exponential View
Anthropic's Q3 results reveal a sharp slowdown in growth momentum despite strong unit economics. The company achieved $16.8B annualized revenue (implying ~$4.2B quarterly), representing 45% quarter-over-quarter growth—a dramatic deceleration from Q2's 147% QoQ expansion. The miss against leaked expectations of $65B annualized revenue has rattled AI trade sentiment, with Korea's KOSPI down 16.2% this month. However, underlying metrics remain robust: 85% gross margins and $1.70 revenue per compute dollar spent demonstrate pricing power and efficient capital deployment. Most significantly, Anthropic's 900% net revenue retention indicates existing customers are spending 9x more year-over-year, suggesting deepening product-market fit and willingness to scale usage even as the company cuts prices on lower-tier models like Haiku. This retention metric is critical for evaluating whether the company's $7bn training compute investment (Q1 2025–Q1 2026) will continue generating outsized returns, or whether the deceleration signals that early-adopter demand is normalizing.
ASML enters hybrid bonding with track-based die-to-wafer tool; HBM adoption slip to 2029+ weakens Besi's growth outlook
SemiAnalysis
ASML is building a die-to-wafer (D2W) hybrid bonding tool with a novel track-based architecture, signaling entry into a market currently dominated by Besi and threatening the latter's growth trajectory. Based on 27 bonding patent families filed since December 2022 and architecture design work with Prodrive and VDL ETG, ASML's tool uses independent die holders that ride past process stations—a departure from traditional pick-and-place designs. The track-based approach offers potential advantages in station utilization, alignment accuracy, and particle control by leveraging ASML's lithography expertise. ASML expects advanced logic to adopt first, where D2W hybrid bonding already runs in volume at TSMC and Intel; logic accounts for most of Besi's hybrid bonding business today.
Simultaneously, Besi faces headwinds from HBM adoption delays and Korean competition. The market previously expected HBM to become the major D2W volume driver from 2028, but SemiAnalysis now forecasts adoption will slip to 2029 or beyond. Korean competitors including Hanwha Semitech are closing the gap with tools claiming 0.1um placement accuracy at roughly half Besi's price, while the 4-hi/8-hi HBM de-spec removes the height problem hybrid bonding was designed to solve. HBM hybrid bonding represents up to 20% of Besi's 2028 consensus revenue, making this delay materially significant to the company's growth trajectory.
Meta AI Research via Techmeme
Meta released Muse Spark Open, an open-weight model with 816B total parameters and 1M-token context window, positioning it as the most capable US open-weight model available. The model runs efficiently on Mac Studios via partnership with Apple, with Meta executives highlighting implications for enterprise AI orchestration and increased token consumption across the stack. This represents a significant competitive move in open-weight models that could reshape how enterprises balance frontier vs. tailored models.
The Information via Techmeme
TSMC reported Q3 revenue of $46.36B, up 50.9% year-over-year, driven by sustained AI chip demand. Record quarterly profits and capacity expansion plans underscore TSMC's dominance in AI chip manufacturing. The company plans to add a new 3nm fab in Taiwan with production expected to begin in H1 2028, signaling confidence in continued AI infrastructure buildout. CEO comments on Elon Musk's Terafab project signal ongoing geopolitical complexity around advanced semiconductor production.
Applied Materials via Not Boring
Applied Materials opened a $5B EPIC (Enduring Process Innovation Center) in Sunnyvale, a 600,000 sq ft facility bringing TSMC, Intel, Samsung, SK Hynix, and Micron together alongside equipment and research partners. The center compresses chip development cycles by running materials research, equipment development, and pilot production in parallel rather than series. This directly addresses the slowest layer of the AI infrastructure stack, potentially accelerating Moore's Law dynamics through coordinated investment ahead of future bottlenecks.
Andrej Karpathy via Clouded Judgement by Jamin Ball
Andrej Karpathy's analysis reveals a sharp funnel in AI adoption: approximately 75% of the global population (~6B people) has barely come in contact with LLMs; ~20% (~1-2B) are casual users of free-tier products; ~0.2% (~20M) are professional users of frontier-grade LLMs in math and code; and ~0.00006% (~5,000) have access to frontier-grade systems internally. Only ~2% of US households pay for AI services. The gap is driven by impact scaling with problem size and ambition, jaggedness of capability (peaking in digital, verifiable, economically valuable domains), and access tiers. This widening gap between capability and diffusion directly impacts whether AI infrastructure capex will be justified by revenue growth, as the general public still perceives LLMs as derpy chatbots while a small sliver of professionals experience vertical capability curves.
Nick Cammarata / Goodfire via AINews
His essay "Neural Networks Are Already Clean" argues that agentic analysis resolves most cases previously thought to require model unfolding, potentially unlocking major progress on mechanistic interpretability and alignment. This directly challenges a foundational assumption in interpretability research and could reshape how researchers approach neural network analysis.