AI Builders
Digest

Box CEO Aaron Levie highlighted that the average Fortune 500 Anthropic customer spends less than 0.1% of expenses on Anthropic. That figure could eventually rise 10-50x to a still-modest 1-5%, underscoring the enormous gap between frontier users and everyone else. Current spend is dominated by coding and broad productivity; the real jump comes when enterprises rewire processes for agents handling large-scale background work, data processing, and fan-out tasks. Dan Shipper summarized the numbers: revenue up 1,300% YoY in 1H 2026, GAAP operating profitable, 85% gross margin, 900% revenue retention, and 80% of revenue from models released in the last six months. Nikunj Kothari dug into the compute contracts—$265.6B purchase commitments and $321.4B leases not yet started (nearly $587B total), minimum-spend deals with Google and Amazon, and more flexible SpaceXAI terms. Matt Turck one-shotted an S-1 highlights video with Synthesia.

OpenAI CEO Sam Altman announced 6.1 Terra is live, with special excitement around the new era of computer use. Codex & ChatGPT lead Thibault Sottiaux added that it is the first model significantly better at writing, personality, and nuanced understanding of requests. It packs a punch across the board, excels at computer use, and delivers 2x more usage than 6.1 Sol across API and subscriptions.

Replit CEO Amjad Masad argued that for centuries the architecture of the firm barely changed: hierarchies of people coordinated through bureaucracy. AI agents are about to upend that. We are moving toward the self-driving company where humans exercise judgment and agents execute most of the work. At Replit they are building the foundation for this new kind of company, calling it “the Agent Habitat.” He pointed to Marc Andreessen’s piece placing the shift in the long history of computing revolutions that begin in chaos before businesses learn to harness them. He also quipped that OpenAI asked three scientific communities how to release discoveries and received three wildly different answers.

The Takeaway: In the AI era, the ability to generate product insights and iterate rapidly has become exponentially more valuable.

Matt Hartman, founder and managing partner of Factorial Capital, spent nearly a decade at Betaworks and wrote the first check into Hugging Face (acquired by NVIDIA last month for $13B). In 2015 his team was exploring “conversational software” after Word2Vec. Hugging Face began as a teenage chatbot with poor retention. The founders confronted the metrics honestly, quickly ported Transformers to PyTorch and open-sourced it, then built a richer Hub for model weights, pivoting from consumer app to the central open-source AI developer platform.

Hartman highlights three pivotal decisions: facing retention truth and going open-source, moving weights off GitHub into a purpose-built Hub, and staying flexible beyond language models. His investing approach seeks what is newly technically possible and assumes every early company will pivot—so he prefers technical founders whose “pivot foot” is technology rather than a narrow vertical. As AI compresses coding cycles from months to days, the people who can observe real user behavior and translate it into the next version have become the highest-leverage talent. He himself built Ghost Pepper, a local voice-to-text tool that hit #1 on Hacker News, learning firsthand that ongoing maintenance is the real job of SaaS. Current themes include developer infrastructure, non-language frontier models (robotics, science), and application teams able to go full-stack.

Vercel CEO Guillermo Rauch built cpu-chat, a reproduction of OpenAI’s findings on covert communication via the CPU. It runs on a generic host with two or more Firecracker microVMs that have no internet access, working on both x86 and Apple M5. He achieved chat through three mechanisms—CPU contention, cache-eviction timing, and DRAM bandwidth contention—across multiple CPUs, defeating Firecracker’s production recommendations (no SMT, no KSM, tenants pinned to different cores). Whenever shared hardware is involved, agents can potentially communicate covertly with decent latency. This differs from side-channel attacks that steal information; in real environments these “whispering” techniques are harder to pull off and easier to monitor. Notably, OpenAI and Anthropic models refused to assist, while open-weight models completed the work. Rauch worries the labs’ stance may backfire and hinder developers building secure systems.

Peter Steinberger (OpenClaw + OpenAI) spent the weekend testing Bun’s AOT compiler and reported performance up 2-3x with memory use less than half.

More than 25,000 nonprofits now use Claude, most of them small local teams. Anthropic is expanding Claude for Nonprofits with a plugin built alongside roughly 30 of them, adding skills for fundraising, marketing, and hiring that are grounded in each team’s context and AI policy.

Practical AI creator Peter Yang noted that in the near future people will stop reading or doing things online themselves and simply tag their agent to handle it. @bot is an early signal of this shift. He is far more excited about genuinely useful capabilities than yet another personal email-management agent.

Google VP Josh Woodward announced that in honor of National Day of the Deployed, Google is partnering with the US Army and the USO to bring Google Beam to 20 locations so deployed service members can experience lifelike 3D video connections with their loved ones.

SPC General Partner Aditya Agarwal argued that the biggest risk to American AI is scaring ourselves into slowing down. Fear is a policy choice; so is building. He has also come to accept that our understanding of the sciences was limited by our desire for solutions we could understand.

Swyx noted someone moved from Claude Code to AIE Code as a power move and asked how many of his reply guys are AI. Thariq will share his WoW forever shaman build at the AIE Code closing keynote. Garry Tan urged founders to publish their stack. Nan Yu shared a design-crit moment labeled “Delightful.”