AI coding assistants are booming, yet the economics are tough: LLM inference is costly, competition is intense, and model providers are also competitors. The net effect is thin—sometimes negative—gross margins, forcing startups to rethink roadmaps, pricing, and infrastructure.
Why margins are so thin
Three forces drive the squeeze:
- High LLM costs: top models for coding are expensive, especially on complex, multi‑step tasks like refactoring, debugging, and repository‑scale analysis.
- Pressure to use the latest model: users expect the most advanced, coding‑tuned models—which often cost the most.
- Relentless competition: products like Cursor (Anysphere) and GitHub Copilot set a fast pace and compress margins.
Windsurf: fast growth, a failed deal, and a strategic exit
After exploring a financing round at around a $2.85B valuation, Windsurf considered a sale at roughly $3B. The OpenAI deal fell through; later, founders and key talent moved to Google in a transaction that delivered a sizable return to shareholders, while the remaining business was sold to Cognition. The takeaway: rapid growth is not enough if variable costs exceed unit revenues.
Cursor/Anysphere: rapid growth, dynamic pricing, and a homegrown model
Anysphere (Cursor) has scaled quickly (ARR in the hundreds of millions) and adjusted pricing to reflect the cost of running the latest models, especially for heavy users. The company is also pursuing an in‑house model to curb dependency on third‑party providers, amid R&D and hiring challenges.
The cost landscape keeps shifting: OpenAI introduced GPT‑5 with fees below premium alternatives like Claude Opus 4.1, and Cursor promptly made it available to users.
“Today’s inference cost is likely the highest it will ever be; over time it should come down.”
Erik Nordlander, General Partner, Google Ventures
Suppliers as competitors: the co‑opetition risk
Model makers are entering code‑assist directly (e.g., Claude Code; OpenAI solutions for code), while entrenched players like GitHub Copilot remain benchmarks. Building a proprietary model can improve unit economics but requires heavy investment and carries R&D risk. Operators report similar variable cost structures across products, with margins often neutral or negative.
Levers to improve unit economics
- Proprietary or hybrid models to reduce per‑token cost and tailor coding workloads.
- Multi‑model routing (cheap models for easy tasks, top‑tier only when needed).
- Caching, context trimming, and local tooling to cut LLM calls.
- Usage‑based pricing and fair limits for power users, with transparent communication.
- Inference optimization and capacity deals for GPUs and serving.
Implications for the ecosystem
If even leading “code gen” apps struggle to build healthy margins atop third‑party models, emerging AI verticals may face similar dynamics: volatile pricing, urgent differentiation, and a shift toward proprietary models or more integrated stacks.
Conclusions
AI coding assistants remain a hyper‑growth market, but sustainability hinges on hard architectural and business choices: controlling model costs, differentiating meaningfully, and aligning pricing with real usage. Those who execute will convert hype into durable margins.