AngelList Co-founder Naval Ravikant: Pure Software is Rapidly Becoming Non-Investable
AngelList co-founder Naval Ravikant wrote on February 28 that pure software is rapidly becoming non-investable.
He later referred to this statement as a weakened version, originally meaning that pure software is non-investable, period. There are two reasons: those who can write requirements can piece together applications on the same day. The advancement of coding agents is rapid, and scalable, architecture-compliant software could be written within a year.
The turning point he pointed out is Anthropic's Claude Opus 4.5 at the end of 2025. He said this model can maintain tasks, complete applications from scratch, and handle tricky problems, like a fast, nearly free, always-on junior programmer. The genie is out of the bottle.
His alternative list for venture capital includes hardware, network effects, and AI models. Model training is described as new software building, valid until automated research and automated training are operational. Individual creators are said to be in a renaissance, while pure software supported by venture capital has no market.
Y Combinator co-founder Paul Graham responded that he would not say that, emphasizing that the most important aspect of a startup is the founder; smart and determined individuals can still do pure software, and he would still invest. Cresta co-founder Tim Shi added on October 7 that some software still requires knowledge of coding subfields. Tim Shi worked early at OpenAI in 2016, and Cresta is backed by Sequoia, Greylock, and a16z, with a self-reported valuation of over $1 billion.
The sellers are application companies relying solely on code barriers, while buyers are turning to computing power, hardware, and models with data loops. The event is driven by the decreasing cost of coding agents, with funds shifting from seat-based software to model training and distribution networks. Beneficiaries are model companies and vertical software with proprietary workflows, while companies whose functions can be replicated in a single session are under pressure.
Source: Public Information
ABAB AI Insight
When Naval Ravikant was at AngelList, he earned from standardizing software platform fees and syndicate channels for angel rounds. His early bets were on software and network effects like Twitter and Uber, not model training. By 2026, this judgment negates the replicable code premium from his rise to fame, shifting investability to hardware and models.
The capital path has already diverted as per his list. He views application layer financing as easily replicable in a day, suggesting money should go into computing power, robotics, distribution with network effects, and training that can maintain barriers. Cresta is taking a different path: scripts, quality control, and industry processes in customer service scenarios, with investors including Sequoia, Greylock, and a16z. Tim Shi's response places the remaining premium on domain knowledge written into software, rather than on the interface.
The analogy is between enterprise software post-2000 and SaaS post-2014. The former still had implementation barriers on licensing fees, while the latter relied on seats and sales teams. Proximity to cloud computing has thinned server software margins, and the difference this time is that the code itself is being replicated. Y Combinator, represented by Paul Graham, continues to bet on founders, while Naval's asset class veto represents two pricing strategies within the same market.
This is a technological substitution. As agents drive the marginal cost of junior programmers to near zero, the entry barrier for pure functional software disappears, and pricing power shifts from those writing code to those owning workflow data, distribution, and computing power. Domain knowledge can still retain a premium, simply because it has not yet been included in training data.
ABAB News · Cognitive Laws
When code can be replicated, the premium shifts to data.
Tool software sells functionality, while vertical software sells processes.
After marginal costs reach zero, barriers remain only in distribution.