CoreWeave's stock has nearly halved in 12 days, falling from around $60 on Monday to roughly $54 on Friday, even after Nvidia chief Jensen Huang said demand for AI cloud firms is very strong. On Thursday, Meta dropped nearly 9%, the Nasdaq fell 2.6%, and fellow neoclouds Nebius and IREN fell 7.2% and 8.3%. Investors worry rising rates will make new debt more expensive and question whether CoreWeave can handle its existing debt. Meta, a $35 billion CoreWeave customer, plans a new share class in January to raise AI capital, but it is building its own data centers and chips and may not return. SpaceX, which reported surging demand for selling spare AI computing power, is now a rival with far deeper pockets and far less debt.
Why it matters
The AI boom is now being priced as a lending story, not just a technology story. Companies that borrow heavily to buy chips look fragile when rates rise, while cash-rich giants that own their own computing look safe. If you depend on a neocloud, ask how it pays for its next expansion, because its funding cost is now your risk too.
OpenAI described Habitat, the storage system behind its products, which serves over 1 billion users across 40 regions and holds more than 500 petabytes of data. Its largest Azure Cosmos DB account alone holds tens of petabytes and handles more than 20 million requests per second. After cycling from shared containers to separate containers and back, OpenAI now runs mostly shared containers with hot-key throttling and request tiers that protect interactive traffic from bulk jobs. A caching layer using change feeds clears stale copies within seconds, a Rust rewrite helped cut typical read time from about 120 milliseconds to about 10 milliseconds, and more than 90% of requests are reads. OpenAI is now building its own exabyte-scale storage system, based on RocksDB Cloud with Raft replication, to cut flash storage needs for bulky data.
Why it matters
At this scale, a managed database stops being managed. OpenAI had to co-engineer the system with Azure, page each other's teams directly, and test at production size. The lesson for any fast-growing company is simple: your storage design is a product decision, because a slow or fragile data layer will cap everything you build on top of it, and the fix is usually shared ownership with your provider, not a bigger machine.
In a 34-minute a16z Show episode on October 25, Box chief Aaron Levie argued that AI agents give a ten-person startup leverage that was not economically possible before. Founders are shifting from doing the work themselves to directing agents, which puts a premium on strategy, taste, and knowing exactly what to build. He said nearly every industry and job type could support specialized agents, potentially creating hundreds of new billion-dollar companies. He also argued that business software may capture much of the next wave of model gains, and that cheaper AI labor could create demand that does not exist today. One thing he said AI will not change: distribution. Building a good product still does not guarantee anyone finds it.
Why it matters
Small teams are a real threat now, but only if they pick a narrow job and own how customers find them. A ten-person team can match a big firm's output, but it cannot match its sales channels by accident. If you are starting now, spend as much time on how buyers discover you as on what your agents can do, because distribution is the moat that AI has not flattened.
In a 41-minute a16z Show episode on October 24, Ben Horowitz told Speedrun's Andrew Chen that AI changes how fast a company can reach users, but not the hardest parts of founding: making decisions when you cannot know if you are right, keeping momentum, and managing your own doubts. They discussed whether tiny teams can build billion-dollar companies, and whether the app itself matters less as agents work directly on data instead of through screens. They also covered what creates lasting advantage when everyone has similar models, what founders should look for in an investor, and why the best founders are defined by one exceptional strength or insight, not by having no weaknesses.
Why it matters
When tools are cheap, judgment becomes the scarce skill. Two founders with the same models will get very different results, because one will make the hard call under uncertainty and keep going when doubt hits. If you are hiring or backing founders, judge the one sharp insight and the nerve to act on it, not a polished all-round resume, and expect the app layer to keep thinning as agents go straight to the data.
