AI Companies Face Monetization Trap as Free Consumer Models Drain Cash

@nikesharora In your opinion, what will be a true cash cow for them in the next 5-10 years? Consumer ads or Enterprise with FDEs? Or both?

More than half of FAANG makes high $ through ads

If open source is good and free - how many will bite the bullet and use it for use cases where they aren't worried about upgradability or total security?

@nikesharora Yup, hold or restrict free AI to consumers, divert all that tokens for enterprises, make AI work for enterprises, earn some money. From that money, give away some to consumers. Not vice-versa currently.

The pricing point is right, but it solves the wrong half. Cheap tokens unlock experimentation. They don't unlock deployment.

A CIO isn't restricting AI because tokens cost too much. They're restricting it because a confident wrong answer in a real workflow is a liability someone has to sign for. That's a trust problem, not a pricing one.

Which means the paradox might not be one. Refusing to ship a system that can't show its work is the rational call. The Phase 2 unlock is verifiable output: a source on every claim, and "I don't know" as a real answer instead of a fluent guess. Cheap tokens get you the pilot. Only trust gets you to production.

High quality FDEs + Domain experts + Agent/workflow/UI builder platform = Enterprise/Model Labs' new best friend India is uniquely positioned to produce a world class company here imo considering FDE talent/cost advantage and tribal business domain expertise via IT services heritage

@nikesharora Interesting take.

But if token pricing trends toward zero over time, is the real moat actually FDEs + enterprise context + workflow ownership?

In that world, model labs start looking less like software companies and more like global consulting firms with AI attached.

@nikesharora I believe all 3 are playing out 1. Enterprise deals, assuming the use of model orchestrations for token burn. 2. Deployment companies showing ‘how’ to use AI and training on context 3. Harness around use case/edge cases.

Inertia arnd ‘accepting’ AI-native means 2 folks not 50!

@nikesharora Enterprise workloads will be 10 to 50x heavier than customer traffic, and not sure platforms are ready to scale. Plus frontier AI models are expected to absorb the blame whenever something goes wrong, whether the fault lies with a human or an agent. GenAI hype is over 😅

@nikesharora +1 I had a long conversation with an AI consultant yesterday who said increasing # of their clients (mid market) are shifting from open token budgets internally to fixed-price outsourced projects to deal with pricing shock and budget management.

Well said, and #1 is already playing out with enterprises we are working with. Especially outside of US, given where the token prices are combined with the inability to make any major modifications without incurring huge costs, we see enterprises increasingly trying out open source for experimentation. If it works out well, moving them away from open source will be challenging.

@madhuspuranik @nikesharora The last line “More than half of FAANG makes high $ through ads”

@nikesharora Besides pricing, it's also about vendor lock-in and data retention. So perhaps frontier labs aren't best suited to deploying FDEs for enterprise adoption and that role goes to the traditional deployment companies

How should labs change their strategy for that world?

@nikesharora Do you think open source models in a continual learning harness loop on production traces will push OAI/Anthropic out of 90%+ enterprise use cases?

In that case, the move would be to push into use cases where absolute frontier wins - cybersecurity, bio etc..

@nikesharora Why is he tagging harry stebbings😂😂😂

@sohilgupta @nikesharora Very well said. The only business model that has definitely found its PMF as of now is the FDE model. I am currently working at the edge of deployment for OpenAI and I am seeing that first hand with enterprises rushing to get rid of SaaS and build their own software in-house.

@nikesharora Token usage in tooling like CC or Codex is a black box and drift are real. So for continuous heavy use it will need overall tokenomics to be cheaper. Else part of the productivity gain moves to token optimization problem, as cost AI increases it bringing back traditional options.

@nikesharora Most companies use openrouter to save on LLM costs today.

Maybe PANW starts offering model hosting with embedded security over private network similar to tailscale?

Isolate the network, build private vpc, make it secure by default.

@nikesharora Speaking for myself, we are still far from the point to say open source is good.

I use perplexity (ie router) and I always go for Claude or Chat GPT. Even Gemini rarely and every time I use Kimi or others i end up re-doing it on Claude or Chat GPT.

We are far from it.

@nikesharora This —> “Build skill libraries - enter FDEs. Yes,FDEs will train the enterprise Waymos of the world.”

@nikesharora