Power constraints now the binding limit on AI scaling

Inference economics reshape AI competitive landscape: OpenAI cuts costs 50%, Etched emerges with $1B orders, Nvidia dominance erodes

OpenAI achieved a 50% reduction in inference costs through architectural breakthroughs, cutting ChatGPT's GPU requirements from tens of thousands to a couple hundred and potentially saving $10–100M monthly. Simultaneously, Etched exited stealth with over $1B in orders for inference-optimized chips, signaling major capital reallocation toward the inference segment, which now represents roughly two-thirds of AI compute spend and 80–90% of lifetime model costs. Anthropic already achieves 80% gross margins on Opus 4.8 inference calls and is on track to be the first profitable AI lab; OpenAI targets 52% gross margins by December, up from 39% in Q1. Analysts project Nvidia's inference market share falling from 90%+ today to 20–30% by 2028 as Etched, Cerebras, Groq, and internal efforts from OpenAI, Amazon, and Google gain traction. Etched claims its eight-chip servers match 160 H100s. The narrowing gap between frontier and open-source models (~13 weeks) combined with inference-driven performance improvements (5–10x more thinking time than previous generations) means inference, not model weights, is becoming the primary competitive moat. Token demand is expected to grow 24x by 2030 despite falling prices, driven by agentic workloads.

Cursor and Warp pivot to 'software factory' platforms, scaling enterprise deployment teams and multi-agent orchestration

Cursor is expanding its Forward Deployed Engineering team tenfold through year-end, positioning the company to implement agents across entire software development lifecycles for enterprise customers in financial services, telecom, and semiconductors. Rather than point solutions, Cursor's FDE team is helping enterprises deploy long-running agents across design, code generation, testing, deployment, and maintenance stages, with feedback loops feeding back into planning. Warp CEO Zach Lloyd announced a parallel shift from a modern CLI and terminal-with-agents product toward automated development loops that continuously triage, implement, review, verify, and ship code. Warp's new Oz orchestration platform now supports multi-harness agent deployment (Claude Code, Codex, and Warp Agent), automatic multi-agent orchestration for long-horizon tasks, cross-harness Agent Memory for self-improvement, and expanded self-hosting options. Both companies are integrating with existing developer workflows (Jira, Linear, Slack, GitHub) rather than replacing them, positioning software factories as the next standard infrastructure layer alongside CI/CD. This represents a significant product strategy shift from individual developer tooling toward enterprise software factory positioning, competing directly with implementation-heavy AI infrastructure plays.

Genesis Molecular AI's PEARL model achieves sub-1Å accuracy on protein-ligand binding, crossing practical drug discovery threshold

Genesis Molecular AI's PEARL foundation model correctly modeled ligand-induced protein conformational changes without fine-tuning on target or homologous proteins, outperforming public models on 802 never-before-seen co-complexes of EV-A71. The model achieves 3.6× improvement on proprietary real-world drug targets at the more rigorous RMSD < 1 Å threshold, surpassing AlphaFold 3 and other open-source baselines on public benchmarks (14.5% and 14.2% improvements on Runs N' Poses and PoseBusters, respectively). This demonstrates that AI-driven structure prediction has crossed a practical accuracy threshold necessary for real-world drug discovery applications, moving beyond academic benchmarks that have been misleadingly lenient at 2Å RMSD.

Anthropic restores Fable 5 access after US review, establishes first shared jailbreak severity framework with Big Tech

Anthropic's advanced reasoning model Fable 5 is back online following government assessment, with the company drafting formal vulnerability standards alongside Amazon, Microsoft, and Google. Testing showed less capable models (Opus 4.8, GPT-5.5, Kimi K2.7) could exploit the same vulnerabilities reported by Amazon researchers, establishing that the jailbreak did not expose unique Mythos-level cyber capabilities. Anthropic deployed an improved safety classifier that blocks the reported technique in over 99% of cases. The incident catalyzed the first concrete governance framework for high-capability model releases, with the industry developing a shared standard for judging jailbreak severity to help AI developers triage findings, launch highly capable models with greater safety, and communicate risk consistently to government and industry partners.

Introspection launches infrastructure for self-improving agent loops in production

Roland Gavrilescu's new company, spun from xAI's agent infrastructure work, is building managed infrastructure to deploy "autoresearch" systems where outer-loop agents maintain and improve primary systems using feedback signals and evals. This addresses the gap between frontier lab agent technology and production deployment in vertical SaaS, with focus on cost control, security, and vendor independence.

Codex usage inside OpenAI surges 137x; product development bottleneck shifts from execution to curation

Codex usage inside OpenAI has grown 137x among individual users since August 2025, with research teams seeing 56x median use increases. The product development bottleneck has inverted from "how do we build this" to "which of 90 prototypes is actually good," shifting the critical skill from engineering execution to judgment and selection.

AI Is Running Out of Power

Limitless Podcast published a video on AI infrastructure power constraints and the emerging energy bottleneck in AI deployment. Main takeaways:

• US data center power demand projected to double from 80 GW (2026) to 150 GW (2028), creating a 5-year lead time for grid power access that is now the binding constraint on AI scaling
• New GPU generations (Vera Rubin) consume 100+ kilowatts per server rack, a 10x increase from prior generation, exacerbating infrastructure mismatch
• Bloom Energy (12.7% of Leopold Ashenbrenner's portfolio) reduces power provisioning lead time from 5 years to 90 days via modular on-site gas turbines; Q1 2026 revenue hit $750M (+130% YoY) with Oracle powering 2.4 GW data center primarily via Bloom
• Despite power infrastructure representing only 10% of AI data center budgets, it has become the primary bottleneck preventing deployment of already-purchased GPUs
• Major projects like OpenAI's Stargate have faced repeated delays specifically due to power constraints, not compute availability

Autoresearch: The feedback loop behind self-improving agents

Latent Space published an interview with Roland Gavrilescu, co-founder of Introspection, on autoresearch infrastructure and self-improving agent loops. Main takeaways:

• Introspection is building production infrastructure for "outer loops" where agents maintain and improve primary systems using feedback signals, evals, and human input—moving beyond agent harnesses to loop-as-product architecture.
• Agent "recipes" are proposed as portable containers encoding human expertise (evals, judges, signal processing, failures) that agents iterate on, analogous to data recipes in model post-training, enabling knowledge transfer and reproducibility across deployments.
• The framework separates inner loops (primary user-facing work) from outer loops (system maintenance), with humans remaining core to early-stage factories through "ask a human" tools and preference learning rather than immediate full autonomy.
• Introspection targets vertical SaaS and non-coding domains where enterprises want agent deployment without vendor lock-in, using open-source Pi framework (positioned as "Linux of agent harnesses") and Git-based audit logs for reproducibility.

Together AI Raises $800M at $8.3B Valuation; Open-Source Model Adoption Accelerates as Enterprises Reject Closed-Model Costs

New York Times Business/Dealbook in New York Times published a report titled "Together AI Raises $800M at $8.3B Valuation; Open-Source Model Adoption Accelerates as Enterprises Reject Closed-Model Costs". Main takeaways:

• Together AI processed 10,000x more tokens per month year-over-year; annualized revenue run rate reached $1.2B last quarter, driven by enterprise shift to open-source inference
• Chinese open-source models now cost one-fiftieth per-token versus US closed models (per JPMorgan); OpenRouter data shows open-source processing rose from 34% to 65% of volume January-June 2026
• Enterprises explicitly cite cost pressure: token-maxxing pushback and margin compression from closed-model costs driving adoption of open-source alternatives hosted on platforms like Together AI
• Competitor Baseten raised $1.5B at $13B valuation; Cerebras IPO'd in May at ~$50B market cap; Together AI executives signal public market trajectory
• Microsoft considering hosting DeepSeek (Chinese open-source model) on Copilot; enterprise customers report 5-7x cost reduction moving workloads from OpenAI/Anthropic to open-source via Together AI

Measuring developer productivity? A response to McKinsey

The Pragmatic Engineer published a collaborative essay by Gergely Orosz and Kent Beck responding to McKinsey's developer productivity measurement framework. Main takeaways:

• McKinsey's measurement approach is naive and ignores high-performing team dynamics; Kent Beck's experience at Facebook shows how survey-based metrics degrade when rolled into performance reviews and compensation decisions
• The essay introduces an effort-output-outcome-impact mental model for understanding software engineering value creation, distinguishing between what engineers do, what they produce, how customers respond, and what value flows back
• Different stakeholders ask different productivity questions (identifying underperformers, allocating headcount, managing performance, personal growth), each requiring distinct measurement approaches rather than a one-size-fits-all framework
• Measurement incentives create perverse behavior: when metrics are tied to compensation and promotion, engineers optimize for the metric rather than customer value, degrading organizational outcomes

Redeploying Fable 5

Anthropic published a release on the redeployment of Claude Fable 5 and Mythos 5 following export control lifting, including new safety frameworks and industry collaboration standards. Main takeaways:

• Export controls on Fable 5 and Mythos 5 lifted as of June 30, 2026; Fable 5 redeployed globally starting July 1 with 50% usage inclusion through July 7, then credit-based pricing
• Amazon researchers discovered jailbreak technique allowing Fable 5 to identify software vulnerabilities; testing confirmed less capable models (Claude Opus 4.8, GPT-5.5, Kimi K2.7) could replicate same behavior, indicating no unique Mythos-level capabilities exposed
• Anthropic deployed improved safety classifier blocking reported bypass technique in 99%+ of cases; new classifier increases false positives on benign coding/debugging tasks as intentional safety margin tradeoff
• Anthropic, Amazon, Microsoft, Google, and Glasswing partners developing consensus industry framework for assessing AI jailbreak severity across four criteria: capability gain, breadth of capability gain, ease of weaponization, and time-to-exploitation
• Deepened government collaboration including pre-release testing, information sharing, and research collaboration with US Department of Commerce's CAISI