Upneet Grover

Now: Building LH2 AI Labs - Data Training Infra for frontier AI models
Prev: Founded and Sold Getmyuni, IELTSMaterial

[Upneet Grover @grover_7up]

@vasuman Low margin businesses are mostly commoditized eg: logistics, services and AI will pass equal benefits to all and hence further create margin pressure again at equilibrium. Entire thesis is assuming only some will benefit from AI, which is not true.

[Upneet Grover @grover_7up]

@AnjneyMidha Lot of alpha in pretraining your own models if you're sitting on prop.y data over and above what's available on the open web. for eg: if I have a Tn tokens of code extra than the public internet, my model is going to be much better in reasoning and coding

[Upneet Grover @grover_7up]

@HarryStebbings @AnjneyMidha @JonathanRoss321 Model companies want to own the entire supply chain without having dependencies
Power + Compute + Model + Data > Full Stack

[Upneet Grover @grover_7up]

@tanayj Post training infra [ open model hosting, deployment, vertical specific SFT / RLHF, data ] companies will be huge.

[Upneet Grover @grover_7up]

This is a lazy generalisation about @mercor_ai ...My guess is only about half of the 2B is GMV. They have datasets that are produced once and then sold multiple times, same with raw data (eg: codebases), evals, RL env etc [all high margin stuff]. Maybe half of this 2B is high quality rev.

[Matt Mickiewicz @MattMickiewicz]

Unpopular Opinion: The CEO of a $10b+ unicorn should know the difference between marketplace GMV and ARR.

[Upneet Grover @grover_7up]

@LuckyInvest_ARK @laxdotsh Sir you are 47...so in 2006 you were 27..how were you bankrupt and worth 100 cr at the same time :)

[Upneet Grover @grover_7up]

@sriramk @satyanadella Karp was shilling Palantir and Satya was saying if Microsoft doesn't win, let no closed model win.

[Upneet Grover @grover_7up]

The inference layer is the new cloud layer and companies like @togethercompute and @OpenRouter are the AWS and Cloudflare of the model era

[Upneet Grover @grover_7up]

@Trace_Cohen It's the same company with physical AI layered on top

[Upneet Grover @grover_7up]

@quxiaoyin I'd add companies that help with the deployment, fine tuning and running evals for open source models - as a productised service.

[Upneet Grover @grover_7up]

@chamath I think the Hyperscalers / Closed sourced models will be seen doubling down on the execution and application layer eg: claude design, claude tag, claude code as selling tokens for intelligence seems harder by the day

[Upneet Grover @grover_7up]

@adityaag Which in turn are powered by SOTA US models' distillation. Here, I completed the thought.

[Upneet Grover @grover_7up]

@mkanapaka @roshanonline @Infosys

[Upneet Grover @grover_7up]

@mkanapaka @roshanonline @Infosys He means data is used to train stuff...you can't train data

[Upneet Grover @grover_7up]

@heysamir_ @pitdesi @OpenStore It works with patient capital, not with VC capital because then you're forced to buy suboptimal siht to grow fast. Best buys come by sitting on capital and saying no to everything except for homeruns.

[Upneet Grover @grover_7up]

@Suhail noone doing ref checks in the valley ? or does he have that figured out as well ?

[Upneet Grover @grover_7up]

@iNikhilsaini Never have - never will go near anything remotely close to "adventure sports" in India. Banana boat, rafting, paragliding - No, No, and No !!!

[Upneet Grover @grover_7up]

@rohitdotmittal Small exits often have lock-ins/earnouts, leaving founders with little or no $ upfront. Spending 2 years to make a few M may feel off, especially in once-in-a-lifetime times like now, as they may feel time is too critical to wait. Not defending.

[Upneet Grover @grover_7up]

@saransh_sidhu That is rookie level garbage. Every street in Bangalore beats this.