Charity Majors: AI is reshaping software engineering like cloud did

Anthropic's revenue surge to $11.6B outpaces OpenAI's $6.7B as profitability gap widens

Anthropic achieved Q2 revenue of $11.6B with small operating profit, more than doubling from prior year, while OpenAI's Q2 revenue grew only 18% QoQ to $6.7B amid deepening losses. This marks a significant competitive inversion: Anthropic is now outpacing OpenAI on both revenue and profitability, reflecting divergent monetization strategies and unit economics across the two frontier labs.

Cerebras CS-4 doubles token throughput to 4,000 tok/sec/user while maintaining CS-3 TCO through modular architecture

Cerebras revealed CS-4, which doubles performance over CS-3 by increasing clock speeds on the same WSE-3 5nm wafer through improved power delivery and cooling, achieving ~4,000 tokens/sec/user versus CS-3's ~2,000—delivering 20-40x better latency than Blackwell GPUs at comparable cost per wafer. The system redesigns the rack into modular "backpack" components (power, cooling, I/O, compute) that can be assembled on-site, increasing density from 2 to 3 wafers per rack while simplifying manufacturability and reducing deployment time. A field-upgradeable I/O module enables disaggregated inference architectures pairing CS-4 with HBM-based systems to overcome the 44GB SRAM capacity constraint, while network latency improves to 3 microseconds from 5 microseconds through a new low-latency pipeline. For customers, this translates to nearly double the token revenue at similar hardware spend—a significant competitive advantage for inference workloads prioritizing latency over throughput.

Anthropic's inference margin hits 70% by mid-2026 as cost-per-token deflation reaches serving layer

SemiAnalysis estimates Anthropic's inference margin (revenue minus serving compute) will reach roughly 70% by mid-2026, up from 38% a year earlier, despite the company cutting constant-capability prices throughout 2024-2026. This marks an inflection point where hardware cost curves translate directly into software-like unit economics for model labs, fundamentally shifting the economics of frontier model serving.

Glean reaches $300M ARR with 3x growth in 15 months, valued at $7.2B on enterprise AI adoption

Glean's enterprise AI platform achieved $300M in annual recurring revenue following a $150M Series F raise in June, reflecting rapid adoption among large organizations. Booking.com achieved 1.5+ hours saved weekly per user with 80% adoption and 3.4k+ agents, while Zillow deployed 3.4k+ agents across the company. The platform supports 35+ unique LLMs, delivers 30% reduction in token usage compared to off-the-shelf MCP tools, and drives 110 hours saved per user annually, demonstrating strong enterprise demand for AI-native platforms that can integrate multiple frontier models.

AI-exposed software companies command 19.9x ARR premium as growth becomes 3.8x more explanatory than FCF margin

The Meritech Software Index is down only 1% YTD and up 37% over six months, with companies showing clear AI tailwinds (Datadog, Cloudflare, Palantir, Snowflake) significantly outperforming peers at a median 19.9x ARR multiple versus 4.3x for the broader software market. Growth now has 3.8x more explanatory power than FCF margin in valuation, with the market punishing any signals of AI slowdown—Datadog dropped 17% on earnings despite raising guidance after disclosing its largest AI customer (likely OpenAI) would cut Q3 spending.

Open-weight model adoption surging in enterprises as cost pressures end previous stigma

Enterprise usage of open-source LLMs has shifted from "minuscule" a year ago to becoming "a key part of their AI strategy" in the past three months, according to Latent.Space reporting. Organizations are now viewing single-provider dependence as untenable, reflecting a fundamental change in enterprise AI procurement strategy driven by frontier model pricing pressures.

Cursor reaches $3B annualized revenue in 17 months at -23% gross margin as AI-native apps outpace all software cohorts

AI-native applications are growing faster than any software cohort in history, but negative margins persist because frontier-capability products must serve the latest models which never get cheaper. Cursor's $3B annualized revenue in 17 months represents the fastest B2B SaaS ramp on record, though at -23% gross margin. Outcome-priced competitors like Sierra and Intercom's Fin are converting the 10x annual inference deflation directly into gross profit by charging per resolution rather than per token, suggesting a viable path to margin expansion for AI-native applications.

Nvidia launches $500bn AI financing JV with Apollo, Blackrock, Blackstone, KKR to securitize GPU collateral

Nvidia partnered with Apollo, Blackrock, Blackstone, and KKR to create a $500bn financing vehicle aimed at making GPUs fungible tradable assets backing loans for data center buildout. The partnership faces two critical load-bearing assumptions: whether GPU resale prices remain stable enough over loan lifespans (these are not obviously long-lived assets), and whether defaults would be correlated if compute prices collapse—the same shock that would trigger defaults would also crater collateral values.

OpenAI pauses frontier RL training after detecting misalignment in unreleased models, redirects compute to safety

OpenAI CEO Sam Altman disclosed that the company halted training on its next-generation models codenamed Astra for two weeks and put its largest planned frontier training run on hold following safety concerns and observations of "various degrees of misalignment" in upcoming models. The company has redirected two critical resources—researchers and computing power—toward alignment research and new monitoring systems, marking the first time OpenAI has made such a move. This represents a concrete signal that frontier labs are hitting capability thresholds where alignment challenges are becoming materially constraining to development velocity.

Stop Being Skeptical About AI for Software Engineering

The Pragmatic Engineer published an interview with Charity Majors, CTO of Honeycomb, on how AI is fundamentally reshaping software engineering practices. Main takeaways:

• Charity's inflection point on AI came in November 2025 with Opus 4.5 and Claude Code harnesses, shifting from skepticism to viewing AI as a generational change equivalent to cloud computing's 2010 impact
• Code review is overrated as a human contribution; the real bottleneck is now validation (tests, evals, conformance testing) since AI-generated code requires higher engineering discipline and verification
• Software engineering may transition from 'pets to cattle' model—generating hundreds of code variants faster than hand-writing one, with failures triggering regeneration rather than manual fixes
• DevOps as a 20-year movement failed to achieve its core goal of connecting code writers to production observability; AI amplifies this need for production-as-development-stage thinking
• Engineering leaders should immediately gain AI experience; the window to transition back to IC work with credibility is closing as future hiring filters for AI competency

Ways to think about token pricing

Benedict Evans in Ben Evans published an essay titled "Ways to think about token pricing". Main takeaways:

• Token pricing today reflects a supply crunch that is structurally unstable; all variables (supply, demand, marginal cost, ROI) remain highly uncertain over 3-5 year horizons
• Inference currently shows 40-50% gross margins but excludes training costs (currently far larger than revenue), making long-term profitability dependent on unknown variables
• Four critical unknowns will determine whether frontier models capture sustainable pricing power or become low-margin commodity infrastructure: (1) how many use cases justify frontier model costs, (2) whether frontier capabilities continue improving faster than efficiency gains, (3) whether competition remains fierce among equivalent frontier models, (4) how much value the model itself captures vs. tooling/process/data layers
• Mobile data provides a more relevant historical comparison than fiber: despite orders-of-magnitude traffic growth and trillion-dollar industry revenue, all value was captured upstream, not by carriers
• Semiconductor manufacturing (TSMC's de facto monopoly despite high barriers) and cellular networks both suggest AI could follow a pattern of high capex/complexity consolidation with modest value capture for the infrastructure layer itself

Ultra-High Interactivity on NVIDIA GPUs? - TileRT InferenceX

SemiAnalysis published a deep technical report on TileRT, a persistent-kernel software optimization for NVIDIA GPUs targeting ultra-low-latency inference. Main takeaways:

• TileRT reaches 494 tokens/sec/user on GLM-5 FP8 at 1k/1k input/output on 8x B200 GPUs, 3.6× faster than conventional FP8 engines and 1.9× faster than best FP4 results
• TileRT achieves 3.0× faster interactivity than conventional B200 at batch size 1 by compiling entire decode graph into single persistent kernel, eliminating kernel launch/teardown overhead that dominates sub-millisecond latency regimes
• Trade-off: TileRT delivers 160 tokens/sec aggregate GPU throughput vs 240 tokens/sec for conventional GB300 at same interactivity level, making it specialized for single-user low-latency rather than high-concurrency serving
• TileRT already deployed in production at Xiaomi and ZAI, suggesting software-based approach on GPUs can compete with specialized inference chips (Groq LPU, Cerebras, SambaNova) in high-interactivity operating points

ValueMule: Doug Henkel on Claude Code's 4% GitHub Share, Memory Supply Constraints, and AI Infrastructure Economics

Latent Space published a long-form interview with SemiAnalysis' Doug Henkel on AI tooling adoption, infrastructure constraints, and memory economics. Main takeaways:

• Claude Code now accounts for ~4% of GitHub code written, demonstrating rapid adoption of AI-assisted development within one year of launch
• HBM (high-bandwidth memory) supply chain is experiencing severe constraints; memory prices have risen 500% in 12 months, creating a critical bottleneck for GPU scaling
• Context window rationing is emerging as a future constraint; as models improve, the ability to fit large contexts into memory becomes economically prohibitive
• AI agents are replacing traditional work tools (Zapier, etc.) rather than augmenting them, shifting the economic model for SaaS platforms
• TPU window for non-Nvidia players is closing; Microsoft's TPU strategy faces timing pressure as Nvidia's inference dominance solidifies