Theo examines why AI costs aren't falling as expected and what that means for the economics of products built on LLMs.

Hrishi Olickel argues that cheap intelligence lets you build ontologies on demand instead of designing schemas up front, with concrete data-pipeline implications.

A security talk on how a co-tenant on a shared GPU can leak information about your model, which matters if you serve on shared hardware.

Cerebras and Gimlet Labs discuss how inference infrastructure can keep up with rapidly growing token demand. Relevant to anyone serving models at scale.

Fireworks AI lays out a train-to-deploy loop for specialized models, making the case for owning your model instead of renting frontier APIs.

An AI Engineer talk on taking GPU utilization from 15% to 90% by fixing the data pipeline instead of touching the model.

Shows how a relational context engine can cut token burn compared with plain RAG, a direct cost and retrieval win for agent builders.

A builder talk on running serverless AI agents with zero cold starts on Akamai Functions, with concrete serving trade-offs.

H Company explains how it optimizes VLM serving for computer-use agents with NVIDIA Dynamo, with real serving and latency details.

Dynatrace makes the case for observability built for agents, showing how to trace and monitor agent behavior in production.

AWS CEO Matt Garman with a16z on what the cloud has to become for AI agents: compute, identity, sandboxes and the economics of inference at scale.

NVIDIA engineers walk through what's new in CUDA 13.4, useful for anyone tuning GPU workloads for inference or training.

NVIDIA's live session on smart routing for hybrid local and cloud AI on DGX Spark, with hands-on serving details.

Hugging Face explains HF Buckets and Xet-backed uploads that only push what changed, cutting the cost of moving model weights and datasets around.

A short, focused look at TokenOps for cutting AI agent token spend, relevant to anyone watching inference bills.

The MongoDB .local NYC 2026 keynote introduces MongoDB 9.0, Atlas Infinite and the Atlas Agent Engine for running agents against your data.

OpenAI DevDay session on cutting AI cost per completed task with model and reasoning-effort choice, prompt caching, programmatic tool calling and the Batch API.

NVIDIA shows a standalone VLM kiosk deployed on Jetson, built with Codex, so you see edge deployment and AI-assisted coding together.

Sequoia sits down with Google's AI infrastructure chief Amin Vahdat on the physics and economics of frontier-scale compute, power and the next decade of datacenters.

Modal's Charles Frye explains what an inference engine actually does, from batching and KV cache to scheduling, in a talk every builder paying per token should watch.

Nearly two-hour Braintrust workshop that takes an OpenAI Agents SDK support agent from zero tracing to custom scorers, online grading and topic clustering over 968 real…

Akamai's Sheilah Kirui benchmarks speculative decoding with vLLM on a Blackwell GPU and shows when a draft model pays off (1.6x on structured output) and when it doesn't.

IBM's nine-minute explainer on distributed inference: data, pipeline, tensor and expert parallelism, KV cache and prefill/decode bottlenecks.

PlanetScale's Ben Dicken on keeping databases up when agents multiply traffic overnight: isolation, redundancy, sharding and back pressure.

Modal CEO Erik Bernhardsson's Runtime 2026 keynote: the new serverless primitives for running agents, sandboxes and GPU workloads at scale.

PlanetScale argues that modern Postgres and MySQL on fast NVMe can replace Redis for most caching and queueing workloads, with benchmarks.

Datadog's Marina Petzel on LLM observability: catching the silent failures that still return 200 OK and wiring tracing and evals into production monitoring.

Nebius's Sujee Maniyam explains what actually makes open models fast in production, from serving-stack choices to batching and caching trade-offs, with real numbers.

Crusoe engineers show how they keep large training runs alive through GPU failures with automatic detection, checkpointing and node replacement.

Docker's take on running autonomous agents with full permissions inside disposable microVM sandboxes, so YOLO mode stops being a security gamble.

Fractile founder Walter Goodwin tells Sarah Guo how a full-stack chip startup bets against NVIDIA, Broadcom and AMD on next-generation inference hardware.

SemiAnalysis walks Latent Space through ClusterMAX 3.0, its ranking of GPU clouds on reliability, networking, pricing and support, so you know who to actually rent…

YC Paper Club works through the research on running frontier workloads without GPUs and what a post-GPU inference stack could look like.

Breaks down why inference pricing looks nothing like compute cost, from batching and KV-cache economics to who actually makes money serving models.

Langfuse's own full walkthrough of agent tracing, evals, experiments and prompt management — the open-source observability stack many Lounge builders self-host.

Braintrust's intro to evals: datasets, scorers, experiments and how to stop shipping prompt changes on vibes. The single most useful infra habit for anyone with an LLM…

A conference-talk version of evals 101: what to measure, how to build a golden set from production traces, and how teams actually run evals in CI.

Side-by-side look at three tracing/observability stacks for LLM apps, with the trade-offs between hosted and open-source options.

Tour of the Vercel AI SDK primitives: streamText, tool calling, generative UI and provider switching between Claude, OpenAI and Gemini.