Javi™

CTO @theagilemonkeys.
Building production AI agents & automation.
Writing about agentic engineering, software design, and what works in prod.

[Javi™ @javitm]

Hardware power per dollar keeps doubling, and open-source models keep closing the gap with frontier models. When those two curves cross, inference becomes virtually free. At that point, why would anyone keep paying per token?

https://t.co/Xl1d7q1Gjf

[Javi™ @javitm]

https://t.co/YgeM4ewVdz

[Javi™ @javitm]

The price tag is still high, ~$85K-$100K, but let Moore's law do its thing… a powerful SGI machine had even higher costs in the 90s, especially taking into account inflation.

[Javi™ @javitm]

This is the beginning of the end of the token-based AI economy. The future of AI is local, secure and performant with dedicated per-user hardware https://t.co/g59a6bIdkF

[Javi™ @javitm]

@Austen How does one even spend $5k/mo in tokens per employee? We all use subscriptions here, use AI all day long for everything we can imagine, and we’re all under usage limits 99% of the time…

[Javi™ @javitm]

@MatthewBerman Exactly, and junior devs are often the ones who adapt faster to new tools and workflows, so you may be surprised of what a smart junior dev can do with a Claude/Codex subscription!

[Javi™ @javitm]

@bitoshimoto @AskYoshik Exactly, and Moore’s law is pushing these hardware costs down really hard 😬

[Javi™ @javitm]

@iraw000 @AskYoshik No need to wait for China, the Nvidia Jetson Nano chip can already run Gemma 4, and Apple has been embeddings increasingly powerful AI cores in their M-series for a while. When everyone can run SOTA models for free, who’s going to pay for all these tokens?

[NVIDIA Robotics @NVIDIARobotics]

Exciting news for Jetson developers 🎉

Gemma 4 is now on Jetson. @GoogleGemma’s latest multimodal, multilingual models run across the full Jetson platform—from Orin Nano to Thor—bringing on-device AI to robotics, edge, and embedded systems.

Cut latency, manage costs, and keep sensitive data secure. Check out the tutorial and download the container to get started: https://t.co/xpq2tkLczl

[Javi™ @javitm]

@steipete We’ve recently started using @WorkOS and are quite happy so far.

[Javi™ @javitm]

@CDAnalista @marianorenteria Exacto

[Javi™ @javitm]

@marianorenteria Yo no sé qué estarán haciendo otros, pero nosotros tenemos varias posiciones abiertas para desarrolladores y FDEs. Para mí que solo hay más trabajo…

[Javi™ @javitm]

I guess the problem is not really the token price, but that nobody is really seeing a strong ROI yet. If you have an AI system that induces direct gains or savings, anyone would be happy to pay a good percentage of that amount in tokens. But it turns out that integration complexity and achieving acceptable eval scores are very hard. There’s a lot of work to be done to make this technology have the impact everyone is imagining!

[Javi™ @javitm]

@elonmusk @Starlink Any chance Spanish airlines are in? @AirEuropa @Iberia @BinterCanarias 🙏🙏🙏

[Javi™ @javitm]

@plainionist On common tasks, absolutely.

[Javi™ @javitm]

@NickSeagull I’m trying X posts, nothing fancy to say about them though 😅

[Javi™ @javitm]

@KaiXCreator I reach that amount sometimes, but it’s likely I’m only paying real attention to ~4 or 5 at any given moment. They’re becoming the new Chrome tabs for me 😅

[Javi™ @javitm]

Software engineering is not about writing code. Code is just the way we express software engineering.

[Javi™ @javitm]

@kapilansh_twt 1Password is our choice, although Vercel’s env pull is quite straightforward and low friction

[Javi™ @javitm]

Claro, es cierto que ese tipo de perfiles es muy difícil de encontrar; muchas veces preferirán emprender por su cuenta que trabajar para otro, pero vale la pena buscarlos…

Nosotros dimos una vuelta bastante grande haciendo bootstrapping, empezamos como consultoría pura, que es relativamente más fácil de poner en marcha (no te juegas todo a un proyecto), y al cabo de varios años pudimos crear el venture builder y empezar a invertir en producto también. No es un camino para impacientes 😅

[Javi™ @javitm]

It's not just that software engineering is strong; it's that real-world companies, the ones that operate outside the X bubble are still far from fully adopting AI, so the opportunity is massive.

Driving this adoption with safety, reliability, and integrating it into existing systems and workflows requires a massive amount of engineering work.