- Do We Grow Software or Do We Design It?↗ 2 Oct 2026
Software development splits into two paradigms: organic 'vibe-coding' with iterative exploration, and deterministic 'design' via state machines and formal verification—each suited to different problems in the AI era.
- The Fastest Growing Market in Tech isn't AI↗ 30 Sept 2026
Tokenized real-world assets (RWA)—stocks and commodities trading on blockchains—are growing at 53% monthly, outpacing AI startups, with major financial institutions and blockchain infrastructure like Allium powering institutional adoption.
- Segmentation Drives Market Share Wins in AI↗ 29 Sept 2026
Anthropic and OpenAI are competing on business model innovation—segmentation and price discrimination—as much as technical capability. Anthropic's enterprise metering doubled revenue in a quarter; OpenAI's 80% price cut on its cheapest tier keeps them competitive, with both approaching $100B revenue run rates by year-end.
- How GPU Prices Can Double While AI Gets Cheaper↗ 28 Sept 2026
GPU costs have doubled while AI inference prices plummet—a seeming paradox resolved by rapidly improving model efficiency that's outpacing infrastructure inflation, though the balance remains precarious as capital becomes more expensive and demand explodes.
- Thinking in Systems, Shipping in Loops↗ 24 Sept 2026
AI has eliminated hand-coding as economically viable; software engineering now means designing systems and feedback loops that enable AI agents to write and validate code at scale. Donella Meadows's framework of resilience, self-organization, and hierarchy defines how these loops should be built.
- The Most Important Market in AI is the Middle↗ 23 Sept 2026
AI model pricing is collapsing fastest in the middle market, not at the frontier. Demand follows a normal distribution favoring capability-per-dollar over peak performance, forcing pricing dynamics that could commoditize the entire market as intelligence costs plummet.
- AI Comes for the If Statement↗ 21 Sept 2026
AI models optimized for specific decision-making primitives—like if-then classification—can achieve 80%+ accuracy at 76–209x lower cost than frontier models, suggesting a major margin opportunity for production-layer AI infrastructure.
- The Harness Margin Opportunity↗ 17 Sept 2026
Berkeley research shows AI harnesses—the systems controlling agent workflows—determine pricing power: identical models cost 71% less on different platforms with no quality loss. Startups building efficient harnesses achieve 75% gross margins vs. 38% for those calling expensive models directly, creating defensible software businesses through customer understanding, evals, and optimization factories.
- When Inbound Sells Itself↗ 15 Sept 2026
Vercel compressed its inbound sales development team from 10 to 1.25 people using AI agents, achieving 90% automation, 93% support case handling, and 32x ROI—demonstrating that the bottleneck is now workflow design, not model capability.
- What Does Pacing Mean?↗ 14 Sept 2026
Dario Amodei's call to pace AI development lacks a shared definition. Five constituencies—interpretability researchers, labor advocates, growth-focused economists, geopolitical strategists, and regulatory skeptics—each frame pacing differently, and proposed compute thresholds have proven unworkable as innovation accelerates past them.
- Is the 3x AI Productivity Gain just a Computer that Never Sleeps?↗ 8 Sept 2026
OpenAI's claimed 3x productivity gain from AI agents may simply reflect running machines 24/7 while humans work 8 hours—costing $2.5M/year per researcher in inference costs, with a 50% defect rate requiring constant human oversight and debugging.
- The Three Waves of AI Consumption↗ 7 Sept 2026
AI consumption arrives in three stacked waves—chat, single agents, and meta-harnesses orchestrating many agents in parallel—each orders of magnitude more token-intensive than the last. Agent token consumption already dwarfs human usage and is accelerating toward billions of tokens daily.
- Concrete, Silicon, & Leverage↗ 4 Sept 2026
The AI infrastructure buildout will require ~$4 trillion in debt over five years—equivalent to 34% of the US corporate bond market—but AI revenue must reach $1.2–1.5 trillion annually by 2030 to service it, raising macroeconomic questions about credit availability and feasibility.
- The Ads Model for Prompts Vertically Integrates AI↗ 3 Sept 2026
Meta's new tiered pricing for Muse Spark—offering a 92% discount for data-sharing versus privacy—formalizes an "ads model" for AI inference: subsidized compute in exchange for training data, vertically integrating the data supply chain just as search and social platforms did for digital advertising.
- AI Productivity Doesn't Mean What I Thought It Means↗ 2 Sept 2026
AI doesn't reduce effort in knowledge work—it raises the ceiling of output quality. The author's writing workflow shows constant editing time (136 median edits per piece) but improved quality floor, shifting human labor from structural scaffolding to rhetorical precision.
- AI Productivity Doesn't Mean What I Think It Means↗ 1 Sept 2026
AI productivity doesn't eliminate human effort—it automates mechanical tasks and redirects human energy toward higher-value work. The author's writing tool moved edits from structural fixes to rhetorical precision, mirroring how chess players after engines focused on deeper strategy, not less work overall.
- The Price of Entry to the Frontier↗ 31 Aug 2026
Frontier AI labs are consolidating access through exclusive partnerships, government tiers, and revenue-gated open weights—transforming AI from a utility commodity into a scarce, controlled resource. Nvidia's $5T push for open ecosystems offers the primary countervailing force.
- Revenue per Megawatt & The AI Model Factory↗ 27 Aug 2026
AI model companies are becoming compute factories: selling inference at high margins to fund training of progressively cheaper, more efficient models. Anthropic's path from −94% to 40–50% gross margins per megawatt shows the model—profit from serving customers funds the R&D flywheel.
- NVIDIA's $108b Quarter↗ 26 Aug 2026
NVIDIA's Q2 revenue hit $96B with Q3 guidance of $108B—annualizing to $432B and making it the world's sixth-largest company. But hyperscaler growth has slowed to 13% sequentially while non-hyperscaler revenue grew 25%, forcing NVIDIA to extend payment terms and build $581B in supply commitments to fund emerging AI labs and startups.
- NVIDIA's $108b Quarter↗ 26 Aug 2026
NVIDIA's Q2 FY27 hit $96B revenue with $108B guidance, but hyperscaler revenue growth slowed to 13% sequentially while other buyers (AI startups, enterprises) grew 25%, forcing NVIDIA to extend payment terms and build a $581B stack of commitments to sustain demand from weaker-balance-sheet customers.
- How Long Should an AI Agent Live?↗ 24 Aug 2026
AI agents should live for 24 hours—not perpetually—to avoid memory degradation, security vulnerabilities, and context rot; dispatch narrow specialists for individual tasks rather than holding multi-year access.
- How Long Should an AI Agent Live?↗ 24 Aug 2026
AI agents should reset daily rather than run perpetually—long sessions degrade memory, create security vulnerabilities, and persist stale instructions. Deploy a 24-hour coordinator that delegates specific tasks to ephemeral single-purpose helpers, consolidating durable learnings to disk at midnight.
- The AI Bullwhip↗ 23 Aug 2026
AI infrastructure exhibits classic bullwhip dynamics: demand shocks in GPUs cascade through memory, CPUs, and storage with multi-year lags, locking in higher costs at each stage and risking severe overcapacity when software revenues fail to justify $20B/GW data center buildouts.
- The AI Bullwhip↗ 23 Aug 2026
AI infrastructure bottlenecks cascade through a multi-year supply chain, creating a bullwhip effect: GPU scarcity starved server CPUs, memory constraints diverted wafer capacity, CPU demand shifted to storage, and now data center power and cooling infrastructure face three-year lead times and potential overcapacity when software revenues fail to keep pace.
- Mainframes became personal. So will your data center.↗ 21 Aug 2026
Local AI models now match frontier cloud models on 89% of everyday queries while consuming 80% less energy and 74% less cost. As intelligence-per-watt improves, edge deployment will displace cloud inference for routine tasks—mirroring how mainframes gave way to PCs.
- Mainframes became personal. So will your data center.↗ 21 Aug 2026
Local AI models now match frontier cloud models on 89% of everyday queries while consuming 80% less energy and 74% less cost, suggesting a mainframe-to-PC shift where data center workloads migrate to edge devices—but cloud retains advantages for complex reasoning and batch processing.
- Previously Unmanufacturable↗ 19 Aug 2026
AI agents translating natural-language intent into software's domain-specific grammars—CAD, Salesforce, Figma—democratize powerful tools by eliminating steep learning curves. But expertise doesn't vanish; it shifts from interface mastery to systems architecture and documentation depth.
- Previously Unmanufacturable↗ 19 Aug 2026
AI agents that translate English intent into application-specific grammar unlock previously inaccessible tools—enabling non-experts to operate complex software like CAD—while expertise still matters for deep system understanding and edge cases.
- Birds Don't Fly Like Planes. Neither Does AI.↗ 18 Aug 2026
Smaller open-source language models running locally can match cloud models' output quality by reasoning more intensively, using different inference strategies rather than just scale. A practical benchmark of venture-capital tasks shows Qwen3.8-27B achieving parity with DeepSeek V4 Cloud, trading latency for capability.
- Birds Don't Fly Like Planes. Neither Does AI. 18 Aug 2026