TLDRocket
20 July 2026
The story of AI in 2026 is increasingly one of infrastructure and pragmatism displacing hype. While OpenAI and Anthropic still dominate headlines, the real action has shifted to the mundane work of serving models efficiently, cutting token costs ruthlessly, and confronting the gap between capability and actual utility. Netflix built its own LLM serving stack atop vLLM and NVIDIA Triton rather than rely on external APIs, achieving constrained decoding at scale through careful optimization. Developers are learning that raw model intelligence matters less than context quality—Augment Code's semantic retrieval approach beats Anthropic's simpler grep-based system by 33 percent on token efficiency—and that slashing token spend by eight-fold requires moving coordination logic into deterministic code rather than letting multi-agent workflows sprawl. The tension underlying all this: as AI becomes ubiquitous, the bottleneck shifts from generation to human judgment. Code review capacity now chokes pipelines faster than coding speed, and engineers report simultaneous productivity gains and motivational erosion as automation swallows the satisfying parts of their work. Meanwhile, Chinese competitors have closed the gap with stunning speed. Kimi K3 delivers Claude-quality outputs at a third the API cost and without US regulatory constraints, while Moonshot and Alibaba's new models rank above most American systems in internal benchmarks. The smartphone itself is becoming agentic—ZTE's NaviX Ultra integrates ByteDance's Doubao agent natively. Yet adoption remains puzzling: intelligence alone hasn't driven explosive growth, suggesting that solving real problems requires more than ideation, and that organizational AI projects often fail spectacularly, buried under performative metrics and employee dissimulation. The long view favors obscurity: like PostgreSQL in databases, the LLMs that win may be the ones nobody argues about—open-weight models running locally on 24GB GPUs, embedded systems powering a trillion devices, the invisible infrastructure that actually ships products.
Read the full briefing →