The talent bottleneck framing is the most underrated finding in this survey. India produces ML engineers but not enough people who can evaluate models in production, run cost-quality tradeoff analyses, or build proper evals. That's a different skill set entirely ..... closer to reliability engineering than research. The 74% API-first statistic makes sense given the cost profile you describe, but it also means most Indian AI companies are one API price change or terms-of-service shift away from a forced re-architecture. The 16% citing data sovereignty as a real blocker is the number I'd watch most closely ..... it's a structural demand signal for India-hosted inference, and Anthropic's Bengaluru move suggests the frontier labs are starting to take that seriously.
The shift toward simpler stacks also feels telling, especially when teams move away from tools that add layers without clear payoff. If most builders are staying API-first due to cost and speed, what would need to change for more Indian startups to seriously invest in building or owning their own models?
Yeah, and here’s what nobody talks about: a new “better/cheaper” LLM launches like every other day now.
If you’re hardcoding provider logic, you’re constantly refactoring just to test new models.
The smarter move is using something like Fastrouter.ai where you can swap between any LLM (GPT, open-source, SLMs, whatever) on the backend without changing a single line of production code.
Throw in guardrails, observability, and cost tracking, and you’re actually set up to adapt instead of constantly playing catch-up.
With software development getting ultra cheap and ultra fast, there will be generational AI-native technology services agencies who builds, implements, and responsibly runs highly customized software systems for small and medium businesses.
Super valuable data on what's actualy shipping vs what's being hyped. The Anthropic preference gap (48% OpenAI usage vs 42% Anthropic preference) basicaly confirms what I've been seeing, once teams hit the limits of GPT4 for code generation they switch and dont look back. What jumped out most was the 24% Copilot abandonment rate, thats a masive churn signal for something that came bundled with enterprise deals.
Yep, as we called out in the post as well - Copilot’s loss of market share was very surprising that speaks to the preference for really best of breed products. For us as investors, we also need to keep that in mind & avoid the tendency to let “What happens if this incumbent introduces this as a feature?” creep into our mental model.
Claude Code’s edge over copilot can become a generational advantage. Thanks this article was very helpful, also how do I get into this builder community btw?
Great Insights based on actual developer inputs- One area I believe Orchestration Layer or Customized Frameworks would still be relevant would be the Agentic Workflows - Not sure native frameworks could seamlessly support multiple model providers.
Real onground visibility provided here. The trivial GST invoicing is actually a real problem
Thank you so much, that was pretty much the goal to break down the things we see & hear on the ground.
Cursor solved our infra outage in 20 minutes, that could easily have taken more than a day if done manually
Yeah, it’s an amazing product - we’re power users of the same.
This is so good thank you.
Thanks for reading, glad that you found this one useful. We had a lot of fun doing this.
The talent bottleneck framing is the most underrated finding in this survey. India produces ML engineers but not enough people who can evaluate models in production, run cost-quality tradeoff analyses, or build proper evals. That's a different skill set entirely ..... closer to reliability engineering than research. The 74% API-first statistic makes sense given the cost profile you describe, but it also means most Indian AI companies are one API price change or terms-of-service shift away from a forced re-architecture. The 16% citing data sovereignty as a real blocker is the number I'd watch most closely ..... it's a structural demand signal for India-hosted inference, and Anthropic's Bengaluru move suggests the frontier labs are starting to take that seriously.
The shift toward simpler stacks also feels telling, especially when teams move away from tools that add layers without clear payoff. If most builders are staying API-first due to cost and speed, what would need to change for more Indian startups to seriously invest in building or owning their own models?
Natzis
Yeah, and here’s what nobody talks about: a new “better/cheaper” LLM launches like every other day now.
If you’re hardcoding provider logic, you’re constantly refactoring just to test new models.
The smarter move is using something like Fastrouter.ai where you can swap between any LLM (GPT, open-source, SLMs, whatever) on the backend without changing a single line of production code.
Throw in guardrails, observability, and cost tracking, and you’re actually set up to adapt instead of constantly playing catch-up.
Aligned on the direction, not sure if a lot of people we spoke to use something like Fastrouter.ai
Why do you think that might be the case?
You should check out bloomr.world - it's an upcoming platform for young founders
With software development getting ultra cheap and ultra fast, there will be generational AI-native technology services agencies who builds, implements, and responsibly runs highly customized software systems for small and medium businesses.
Super valuable data on what's actualy shipping vs what's being hyped. The Anthropic preference gap (48% OpenAI usage vs 42% Anthropic preference) basicaly confirms what I've been seeing, once teams hit the limits of GPT4 for code generation they switch and dont look back. What jumped out most was the 24% Copilot abandonment rate, thats a masive churn signal for something that came bundled with enterprise deals.
Yep, as we called out in the post as well - Copilot’s loss of market share was very surprising that speaks to the preference for really best of breed products. For us as investors, we also need to keep that in mind & avoid the tendency to let “What happens if this incumbent introduces this as a feature?” creep into our mental model.
Fun times.
Claude Code’s edge over copilot can become a generational advantage. Thanks this article was very helpful, also how do I get into this builder community btw?
There’s a form link on our website for you to fill post which the team shall get you added to the crew.
Great Insights based on actual developer inputs- One area I believe Orchestration Layer or Customized Frameworks would still be relevant would be the Agentic Workflows - Not sure native frameworks could seamlessly support multiple model providers.