High Model Costs Erode Margins at Fast-Growing AI Coding Startups

Start-ups that sell AI coding assistants are posting head-turning revenue gains but are struggling to turn those sales into profit. The Information reports that Replit’s annualised revenue reached $144 million in July, while TechCrunch cites sources saying rival Windsurf’s gross margins are ‘very negative’. Windsurf sought a $3 billion sale to OpenAI earlier this year after investors balked at a new funding round, underscoring concern that the high cost of licensing and running large language models can outstrip what customers are willing to pay. The economics are being strained by the price of the latest models from suppliers such as OpenAI and Anthropic, which coding platforms feel compelled to adopt to stay competitive. Analysts say margins across the fast-growing ‘code-generation’ sector are either neutral or negative, pushing some companies to consider building their own models despite the capital expense. Rising infrastructure costs add another layer of pressure. Cloud and internet giants—Microsoft, Alphabet, Amazon and Meta—are expected to pour about $200 billion into AI-related projects this year, much of it for new data centres. Consultancy estimates show U.S. data-centre electricity demand could jump 50% by 2027, and venture investor Karl Mehta projects the sector’s share of national power consumption could quadruple to 10% by 2030. Energy availability is now seen as a key bottleneck. Amazon chief executive Andy Jassy noted last week that 85%–90% of global IT spending is still on-premises but predicted that the mix will flip over the next decade as companies look to cloud-based AI tools—leaving operators racing to secure power and control costs even as the sector’s topline growth accelerates.