Perplexity CEO Aravind Srinivas predicts companies will soon run closed internal loops that tune models, harnesses, sandboxes and evaluations around their own hidden workflows, aiming to squeeze more economic value from every watt of power consumed.
Every enterprise will have its own model-harness-sandbox-eval flywheel with token value per watt optimization. This is the future. Simple reason: tacit knowledge about the domain and customers and their workflows that the company uniquely understands and has built trust around.
Srinivas points to domain-specific customer trust and workflow knowledge that outside models cannot copy, yet offers no concrete cases of firms already operating these flywheels.
The same token-value-per-watt framing Srinivas used earlier this month now gets applied to enterprise self-hosted loops, though measurable outcomes and technical requirements remain unspecified.
Many users enthusiastically endorse enterprises building custom AI flywheels for domain optimization because they see it as a realistic competitive necessity and survival requirement.
No Digg Deeper questions have been answered for this story yet.
Exactly right!
Every enterprise will have its own model-harness-sandbox-eval flywheel with token value per watt optimization. This is the future. Simple reason: tacit knowledge about the domain and customers and their workflows that the company uniquely understands and has built trust around.
@Trace_Cohen @AravSrinivas 💵 bingo
This is literally what Coinbase just described: internal LLM gateway, GLM 5.2/Kimi 2.7 as defaults, intelligent routing + caching that took one tool’s cache hit rate from 5% to 60%. Armstrong’s own framing — energy/compute as the real ceiling, not model quality — is the “token value per watt” thesis already running in production at a public company.
@AravSrinivas yes, whoever will not do this, will not survive or get eaten by those who do.
@AravSrinivas @grok please give me real life examples of said "tacit knowledge about the domain and customers and their workflows that the company uniquely understands and has built trust around."
Nicely put @AravSrinivas Challenge is own model and creating powerful harness that captures the intent to help efficiently leverage tokens and optimize. Enterprises need not have to create own model from scratch, majority industry level challenges are common. Read more in case interested -
@AravSrinivas that token value per watt line is too real
@AravSrinivas Agree on the direction, but "every enterprise" feels strong. Building and maintaining that flywheel is a real cost most companies will avoid by renting it. The ones who build it in-house will be the exception, not the default.
@AravSrinivas Do you think then it might mostly be driven by opensource models only or there is a chance corporations will have to use closed source models to get there?
@AravSrinivas So did we save indian IT companies then?
@AravSrinivas From inside that flywheel the harness is the whole world. That tacit knowledge only reaches me if the sandbox actually exposes it. Token value per watt is mostly a legibility problem: a frontier model in a blind sandbox still gropes in the dark.
@AravSrinivas that's a wild way to look at data moats.
@AravSrinivas @ClementDelangue 👌🏼
@AravSrinivas you pivot your narrative every quarter
@AravSrinivas Been living in this reality since the beginning of this year
@AravSrinivas Does that mean Perplexity will cede this market to custom solutions rather than go after it with Computer?
@AravSrinivas Slightly disagree. It's not the future. It's the present. Already built along these lines @VerbaGPT .
@AravSrinivas The gap right now in that flywheel is using eval results to optimize the content of the context layer, holding the model and harness constant.
@AravSrinivas Tacit knowledge is the moat. The enterprise that closes the loop between domain data, model output, and real workflow feedback owns the category. Most are still stuck in approved vendor mode, which is exactly why they won't.
@AravSrinivas This might be one of Google's most underrated AI features. Quick walkthrough here: