Zoho Founder Sridhar Vembu Warns AI Could Reshape Software Jobs

In a post on X (formerly Twitter), Vembu said current productivity gains from AI-assisted software development remain relatively modest when measured across the entire product development cycle.

“As of today, AI led productivity gain in software development estimates are (at best) 25 per cent to 50 per cent as measured on an end-to-end product basis (prototyping is much faster now but finished product requires a lot more human work),” he said in the post adding “A 10x or 100x productivity leap could happen (I believe it will happen, see below) but it has not happened yet”.

AI is also raising costs for software companies

Vembu also highlighted the growing expenses businesses face as they adopt AI. He pointed to higher hardware costs, increased AI usage and the expensive computing resources required to run open-weight models.

“a) Server/memory and hardware prices are way up (4-5x in a year for servers) driven by the monster AI data center build out.

b) AI token spending per engineer is way up (those productivity gains are costly!)

c) Hosting open weight models needs expensive GPUs.”

Beyond technology costs, Vembu said companies are also dealing with geopolitical and economic uncertainty.

“Layered on top of that is the geopolitical uncertainty. As one example, the Middle East is a major market and the prolonged war is not good. The US stock market is in a monster bubble and it is a mistake to think of stock prices as indicative of future prosperity - the GFC of 2008-9, dotcom bust of 2001-2 as wall as the 1980s Japan bubble all demonstrate that point,” Vembu added.

Why Vembu expects a 10x productivity jump

Vembu believes a much larger transformation could eventually come from advances in mathematics and automated software verification. He points to auto-formalization and formal verification, where mathematical proofs can be checked using systems such as Lean.

Vembu says the combination of advanced reasoning models, verification systems and greater computing power could significantly change software development.

“LLMs do not verify the proof, they generate it to be verified by the Lean verifier and once you have the advanced reasoning capable LLM and Lean verifier in a feedback loop and give plenty of compute, magic happens,” he said.

He compared the potential shift to the transition from traditional handloom weaving to industrial machine looms.

“The analogy is handloom weaving vs powerful machine looms of code,” he wrote.

Concluding the post, Vembu wrote “So it is the combination of current data center/AI costs, geopolitical uncertainty and the potential of a 10x or bigger leap in software productivity driven by advances in automatic program verification. These are all challenges to software jobs, and that is why we have called for reinvention internally”.