Pushing the limits of Microsoft Excel.
Not Financial Advice
Joined January 2023
- Tweets5.9K
- Following466
- Followers3.6K
- Likes2.6K
Pinned Tweet
Here is a brief NotebookLM video summary of Aggregator Wars.
The short version: AI agents go straight for the one asset that made the internet giants rich, the direct relationship with the customer. That is why nobody will take their foot off the accelerator. Data wins the customer, recurring cash flow keeps you in the fight, and profits migrate to hardware providers nobody can copy and to whoever owns the customer. Google and Meta are best placed, consensus has Meta upside down, and the AI labs and compute sellers face a funding squeeze into 2027 as capital markets run dry.
If you want a deep dive on what an aggregator is, how agentic AI changes legacy aggregation theory, who is best positioned to be the next aggregator, what each company's profit engine means in this capex cycle, what it means for AI labs and compute sellers, and where value migrates to, I recommend reading the full note. It is as condensed as possible, but packed with information.
I'd invite anyone to push back on any angle of this framework.
Ahora que en Madrid anochece a las 18.00 de la tarde y todos nos queremos tirar por la ventana uno entiende 2 cosas: la primera es porque cuando sale el sol en el norte de Europa o Londres la gente no pega palo al agua. La segunda es, por si nadie se había dado cuenta ya, la remada que es vivir en esos sitios.
Here is the take. In 2027 according to @SemiAnalysis_ +50GW of compute are coming online, which roughly doubles total compute. If only 25% of compute is used to serve customers today, that means you can 4x compute instantly (if you want to) and could 2x again in 2027 (i.e compute used for inference could 8x).
Anthropic allocates 72% of compute to training. Crazy. This confirms what @insane_analyst and others have been saying about 75% going to training. Lot of compute not being used to serve customers but to improve models, LOL.
One of the craziest facts on the S-1 filing form Anthropic is that among the $800bn in compute commitments, more than HALF (400-425bn) is $GOOG related, either via Google cloud or TPUs. $NVDA feeling the pressure here.
@GavinSBaker is basically stating the obvious here. As models commoditise and compete, their purchasing power falls, and value shifts to their suppliers. Chips also face heavy competition too as they specialize, so across the supply chain the capture rate ends up highest at the memory makers (scarce supply) and at the owners of the end customer (the aggregators).
If you want more detail I would recommend you read this piece⬇️
xitter.cf/The_Analyst_Lab/status…
Anthropic allocates 72% of compute to training. Crazy. This confirms what @insane_analyst and others have been saying about 75% going to training. Lot of compute not being used to serve customers but to improve models, LOL.
Crazy amount of detail in Anthropic ($ANT) filing on compute commitments.
Basically they have AT LEAST (minimum) the following commitments on compute (regardless of usage) and all are non-cancelable:
Google $114.5bn
Amazon $110bn
Microsoft $44.9bn
SpaceX is up to (not *at least*) $120.5bn and it is cancelable by either party with 90 day's previous notice.
One interesting bit is that they expect to spend ANOTHER $125-150bn (although pricing is not secured) that is not in the contractual commitments table above on TPUs alone. Crazy bullish for $GOOG
LBO_Guy retweeted
The most mind blowing stat we've seen so far is this:
"As of June 30, 2026, we had raised an aggregate of $141.1 billion of capital and had cash and cash equivalents of $112.4 billion and short-term investments of $19.3 billion."
That means to date, they have burned 9.7b, they are growing at a rapid pace, and are profitable on an operating income basis as of the June quarter! This is likely the most efficient scaling of a business in the history of capitalism.
xitter.cf/AnthropicAI/status/211…
Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.
Read more: anthropic.com/news/s1-sec
Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.
LBO_Guy retweeted
TSMC is finally set to resume SRAM scaling with its A14 node, following a stall that began after N5, according to Korean media reports.
SRAM bitcell sizes have remained at 0.021 µm² for roughly eight years since the introduction of the N5 (5 nm) process in 2020. With A14, however, the size of a high-density 6T SRAM (static random-access memory) bitcell is expected to shrink to below 0.017 µm².
Korean media noted that this would put TSMC ahead of Intel’s 18A process, for which Intel reported a high-density bitcell size of 0.021 µm² and a maximum SRAM macro density of 38.1 megabits/mm² at last year’s ISSCC.
SRAM is the ultrafast cache memory used inside CPUs, GPUs, and AI accelerators. Smaller bitcells—the smallest individual memory cells—allow more memory to fit into the same area, directly affecting performance and power efficiency.
etnews.com/20261026000222
Por muchos fallos que tenga la red actual, el mapa que hace el bueno de Maxim no tiene ni pies ni cabeza. Se pasa por el forro la orografia (la linea cantabrica y la costa entre Malaga y Almeria son de coña y no se me ocurre un mundo en el que eso sea viable, quizas en Minecraft) y encima te deja Sevilla a mas de 1.300 km de Madrid 🤣
Otra cosa que, por lo que veo, casi nadie tiene en cuenta es que el AVE, como medio de transporte, compite con otros como el avion!! Que implica esto? Pues que por norma general, en trayectos de menos de 3h gana el AVE casi siempre y por encima de 3h suele ganar el avion. Se ve bastante clarinete. En Madrid-Sevilla, por ejemplo, el tren se lleva cerca del 90% de los pasajeros entre avion y tren. Bastante claro tambien que nadie va a ir de Barcelona a Almeria o de Bilbao a A Coruña en tren por la costa.
El problema de la red actual no es que sea radial. De hecho tiene todo el sentido una vez uno se da cuenta de que Madrid es la mayor area urbana del pais, esta en el centro, en llano y a 2-3 horas tienes Barcelona, Valencia, Sevilla, Malaga y Zaragoza. Pasa que aqui mucha gente no piensa en que estas cosas tienen que salir a cuenta. Incluso en estos trayectos, que son los que mas pasajeros tienen de España, cuesta justificar la inversion!! A grosso modo, a una linea de 500km le hacen falta unos 10 millones de pasajeros al año para que salga a cuenta. En 2025, con un trafico que esta booming a lo bestia, sabeis cuantos cumplen esto? Solo una. Madrid-Barcelona con 14,4 millones. Madrid-Valencia (6,2m), Madrid-Sevilla (5,9m) y Madrid-Malaga (5,8m) se quedan bastante lejos de los 10 millones. Ya no te digo el resto.
El problema de verdad es que el AVE se alargo a sitios que no tienen sentido (por poca demanda), donde lo que se recupera de cada euro invertido es practicamente nada (unos 10 centimos por euro invertido). Casi que prendiendole fuego a un fajo de billetes recuperas algo mas entre las cenizas. El norte de la red (Galicia, Castilla y Leon y Asturias), por ejemplo, es un autentico disparate. Tienes mas de 1.000km para 4,6m de pasajeros. Tambien tienes atrocidades como la de Toledo-Cuenca-Albacete, que llevaba 9 pasajeros al dia y lo cancelaron a los 6 meses u otras como Barcelona-Valencia, la segunda y la tercera ciudad del pais a tiro de piedra (350km), sin AVE.
Como no todos somos ingenieros de caminos y esto suena complicadete vamos a resumirlo un poco. En el fondo es bastante facil. El AVE hay que ponerlo donde salen las cuentas, o al menos donde no sea el equivalente a prenderle fuego a los billetes. Eso mas o menos se traduce en trayectos con mucha gente, donde el terreno es llano y el viaje dura entre 1 y 3 horas. En que se traduce eso? Pues Madrid con Barcelona (via Zaragoza), Valencia, Sevilla, Malaga, Valladolid y Bilbao, el Mediterraneo (Barcelona-Valencia-Alicante-Murcia) y el valle del Ebro (Bilbao-Pamplona-Zaragoza), mas o menos como el mapa de abajo. Para el resto, lo que tienes que poner es una linea convencional a 200-250 km/h que funcione (que hoy es mucho pedir), que es mucho mas barata, como sociedad salen las cuentas y ademas sirve para mercancias.
LBO_Guy retweeted
Las verdes en UK han parado la construcción de un Lidl (un Lidl!) usando el kafkiano planning process que paralizar el país (y el resto de Europa), y hay gente hablando de reindustrializar… ja ja ja
No pun intended but the Economist is a pile of 💩
Truly unbelievable how many *smart* and well educated ppl still read this. Is just beyond comprenhension. Astounding!
JPMorgan CEO Jamie Dimon mocks The Economist during a tense exchange after being told he’s sounding like Trump: “I’m not Trumpian. I’m factual.” 🇺🇸
Interviewer: “You’re sounding Trumpy.”
Jamie Dimon: “I’m not Trumpian. I’m factual. I don’t agree with this binary shit that goes on all the time.
I think The Economist is the absolute best, most admirable thing in the world, and you should continue that.”
Dimon delivers the last line with apparent sarcasm, poking fun at the publication while pushing back against the idea that agreeing with Trump on certain issues makes him a Trump supporter.
LBO_Guy retweeted
Dylan Patel @dylan522p of SemiAnalysis considers power as a manageable constraint compared to memory, substrate, and optics:
"Power is definitely a big constraint, but one thing that our research sort of shows that I think is counterintuitive to much of the market is that next year there are data center constraints but a lot less than people realize because there are nearly 50GW data centers being built globally. 7-8GW in China, 35GW in the US, and a bunch everywhere else - Australia, India, Malaysia."
"All over the world data centers are being built at a tremendous pace."
"While power is a huge bottleneck, people can find creative ways around that, whether it be turbines on site, reciprocating engines on site, or fuel cells on site. All of these things are helping alleviate the power bottleneck."
"As far as what people most people think, oh data centers and power are the biggest constraint? I think they are constraints but I think they are a lot less than substrates, memory, lasers, and other things in the optics supply chain."
Via CASPA on YouTube
$META $GEV $MSFT $AMZN $GOOGL $NBIS $CRWV
LBO_Guy retweeted
CEO of Amazon, Andy Jassy, on the cost of inference:
"The cost of inference has come down quite a bit in the last few years, but it is still an order of magnitude too expensive, in my opinion, for it to be used as pervasively as possible. If you are asking enterprises to rewrite their applications that use CPU today with AI accelerators, if the difference in those costs are very significant. The amount of incremental benefit they have to get from their AI accelerators on there is a much higher bar to want to rewrite those apps."
"If you can take the cost of inference down, that bar is much lower and all applications are going to be better and smarter and more capable with AI accelerators and inference. That is a really important part for customers to being able to use AI pervasively."
Via @a16z
$AMZN
LBO_Guy retweeted
$AMZN CEO Andy Jassy says AWS capacity is largely sold out through 2027, with a significant portion of 2028 already committed,
"There's only so much data center and power today. There's only so much memory, and I don't think that shortage ends in 2027. I think that, you know, one of the reasons, coming back to the capital investment, why I think we could invest a lot more and it wouldn't be a bad use of capital is that, you know, 2026 is obviously at this point sold out. 2027 is largely sold out. We always leave a certain amount for on-demand and our inference service, Bedrock or EC2, but largely sold out, all the big chunks. A good chunk of 2028 is sold out. Like, there's so much demand."
LBO_Guy retweeted
Listening to the earnings call knowing that it will determine how attractive my future wife will be
LBO_Guy retweeted
Citi: Muse secured over 10 million users in about 6 weeks following its launch in the US and Canada, with approximately 25% of users using it daily—more than double the engagement rate seen at the launch of Threads.
$META
Ferrovial...
FERROVIAL: U MAD?
En dos meses Ferrovial ha ganado las dos mayores managed lanes de USA ofreciendo una pasta que triplica la de sus rivales (vuelve a leer esto). Los rivales btw no es que fueran unos paletos (ACS, Acciona y Sacyr).
Entre obra y canon, las dos concesiones suman unos $71bn de compromisos, de los cuales a Ferrovial le toca el 70% (65% en Tennessee, >75% en Atlanta), frente a un market cap de $35bn. Para los newbies, es una burrada. Ya se que no se pone de golpe, que parte es con deuda de proyecto y parte del canon se paga con los propios peajes durante 50 años, pero aun asi repito: es una burrada. De todos modos, lo que mas me sorprende no es el tamaño, que tambien, sino la diferencia con el resto.
1/🧵
