Ignacio de Gregorio

Most followed AI analyst on Medium (+200k). I also write on TheWhiteBox and TheImperative, explaining consumer and enterprise AI in first principles.

[Ignacio de Gregorio @thewhiteboxAI]

@antoniosarosi I don’t know the answer, but I would argue the contrary; Claude code mostly sees customer intent and the AI does all the coding, cursor is an IDE and still sees way more human code, which could be more valuable

[Ignacio de Gregorio @thewhiteboxAI]

@jdmce @stalkermustang Totally valid concern. OpenAI 2025 financials have leaked today and they had 42% gross margins… terrible

[Ignacio de Gregorio @thewhiteboxAI]

@airkatakana It’s the Europe tax

[Ignacio de Gregorio @thewhiteboxAI]

For sure. It's clear that Anthropic knew from the get-go they had to bet strong on APIs and it's going pretty well for them.

However, I do think prices will still need to go up. PitchBook claims Anthropic's gross margins are 44% based on internal documents. I can't attest that's true, but if it is, that's really bad.

[Ignacio de Gregorio @thewhiteboxAI]

I do agree for consumer subscriptions. But I don't think the average enterprise subscription makes money, but you could be correct and I can't prove otherwise.

I will say I could see overall gross margins from subscriptions being positive because of rate limits, quantizations, etc, but they aren't sustainable if they don't rise in price quite a bit (or if serving costs fall a lot, but I don't think they can if $/GW continues to rise as it's doing right now). To me the clear problem is not inference, it's capital costs

[Ignacio de Gregorio @thewhiteboxAI]

I was trying not to sound too extreme in my position, but I fully agree they aren't good. Pitchbook has Anthropic's gross margins at 44%. If that's accurate, you tell me.

But again, I think this entire conversation is misguided. If I'm valuing an AI Lab business, the only number I care about is free cash flow due to capital intensity

[Ignacio de Gregorio @thewhiteboxAI]

You're right, totally misread that data point and should not be considered in this discussion. But I still completely stand by what I said; subscriptions don't make money.

SemiAnalysis showed that APIs are simply a much better business, and usage-based pricing is the only thing that makes sense when a technology has non-negligible marginal costs (i.e., a single user can spend way more than the subscription fee), unless you rate-limit them, but that's just a usage-based pricing disguised as a subscription and at the very least a revenue opportunity loss relative to the API biz.

[Ignacio de Gregorio @thewhiteboxAI]

Buena suerte explicandole al contribuyente que no pueden invertir más en servicios públicos pero se van a gastar cientos de millones (asumo que miles de millones, que es lo que haría falta, no hay cojones) en una gigafactoría de IA.

Dicho sea de paso, ojalá se ejecutara esta inversión, porque es cierto que se debería tener soberanía tecnológica, pero no va a pasar.

[Ignacio de Gregorio @thewhiteboxAI]

But sir, you don't have to take Ed's numbers for granted, just look at what the Labs tell you.

Anthropic disclosed they were spending $13 per dev per day, averaging $150-$250/month. Most subscriptions are $20/month, so they are deep underwater, and even the max-20x subscriptions have, at best, a 25% gross margin. That is terrible, but even then, this is only the total picture; you have to add training costs (which Labs may or may not include in COGS), and ALL operating expenses, which include depreciation, salaries (including stock-based compensation, which is huge for Labs), and marketing and sales. Once you factor all of that in, the cost/token is super high, in the single-digit dollars per million tokens, which makes subscriptions unsustainable. I'm sorry, but a technology with high marginal acquisition costs like AI can't make money on subscriptions, ever.

The silver lining is that APIs are a much better business than most believe, but I genuinely don't care because with AI Labs, it's all about free cash flow. I don't care you're making $8 for every $10 you charge a customer if you're then spending $20.

[Ignacio de Gregorio @thewhiteboxAI]

@eliebakouch KV Cache magic

[Ignacio de Gregorio @thewhiteboxAI]

@htihle I suspect this has a lot to do with Anthropic's base model being much larger relative to GPT-5.5, which allows a higher degree of forward-pass computation

[Ignacio de Gregorio @thewhiteboxAI]

@Rmwhitem @MPlusDeportes Básicamente

[Ignacio de Gregorio @thewhiteboxAI]

@Sofascore @Pedri @SEFutbol This just goes to prove player scores are dogshit wrapped in catshit with a sprinkle of horseshit on top. Time to redo your algorithm brother

[Ignacio de Gregorio @thewhiteboxAI]

@Rob_Lowe_off @solofichajes123 @COPE Pero coño que lo ha puesto de mediapunta que es como si me pone a mí

[Ignacio de Gregorio @thewhiteboxAI]

@AlexCobas29 Pero sin duda eh, en 30 años no había visto algo peor, rozando alta traición

[Ignacio de Gregorio @thewhiteboxAI]

@JoseMan30451413 @MPlusDeportes Es imposible plantearlo peor. Un buen entrenador sabe cuando se equivoca. Este se niega porque se cree que ha inventado el fútbol, y así nos va

[Ignacio de Gregorio @thewhiteboxAI]

@PacoRinversion @MPlusDeportes Ha sido dantesco, hacía años que no me desquiciaba tanto un partido de futbol. Es la misma España de los últimos tres mundiales. Coño al minuto 30 ya tenía que haber metido cambios, pero es que es un terco de cojones y así le ha ido

[Ignacio de Gregorio @thewhiteboxAI]

@talento20Espano Es muy fácil; acabamos de presenciar el peor planteamiento táctico de la historia del fútbol. Es imposible plantear un partido peor, pero imposible y reto a cualquiera a que me demuestre lo contrario