The Sequence Radar - Issue #958: Last Week in AI: Cheap Intelligence, Expensive Ambitions
TheSequence Jesus Rodriguez ● Covered by 3 sources
DeepSeek open-sourced its inference stack while Meta and SpaceX kept spending huge. Cheaper AI is here, but the bills are getting bigger.
Based on reporting by TheSequence, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Artificial intelligence keeps getting cheaper to run, and somehow the money keeps getting larger. This week’s proof came from DeepSeek, NVIDIA, Meta, SpaceX and Legora, all in the same news cycle. Efficiency is supposed to lower the price of intelligence. Instead, it’s helping justify bigger factories, larger model bets and even more aggressive infrastructure spending.
DeepSeek’s release was DeepInfer, the framework behind its production inference service. It’s open source, built for DeepSeek-V4.1-Flash, and tuned for NVIDIA and Huawei hardware. The stack uses tricks like drafting tokens before verifying them and overlapping scheduling with computation. What it does not include is the whole service; operators still need caching and load balancing. The point is simple enough: if serving each request costs less, developers can afford agents that make more requests.
NVIDIA used GTC Berlin to push the same idea from software into concrete and steel. Its DSX POD design is meant to speed up modular Vera Rubin AI factories, with deployment at prepared sites falling from six months to as little as 14 days, according to the company. Those sites still need permits, electricity and cooling, so the real-world bottleneck remains real. NVIDIA also introduced OpenShell Reference Architecture for governing and securing agents. The pitch is no longer just about building models. It’s about building the whole machine around them.
Meta showed that this machine can still juice a profitable core business. Quarterly revenue rose 26% to $64.6 billion, with management pointing to better advertising from improved engagement and recommendations. But capital spending hit $34.9 billion and free cash flow turned negative at $2.1 billion. That’s the pattern now: the business pays for the ambition, and the ambition gets very good at consuming the business.
SpaceX provided the strangest example. Its AI division, now called SpaceXSI, reported $7.56 billion in quarterly revenue, above Starlink’s $5.25 billion, alongside $431 million in operating income and $16.44 billion in capital spending. Legora, meanwhile, raised $550 million at an $8.8 billion valuation and said annual recurring revenue is above $200 million. The shared bet is obvious: cheaper inference expands demand, and whoever owns the workflow gets to collect the rent. Maybe the spreadsheet really is the final frontier.
My take — AI-written commentary, not fact-checked reporting
The whole week was a nice reminder that “AI efficiency” is mostly a sales pitch for higher spending. Every company says cheaper inference will unlock demand; then it turns around and orders more chips, more factories, more cloud, more everything. The adult version of this market is simple: margins matter more than demos, and the demos are very expensive right now.
Read more about this at: TheSequence