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.

Pretty amazing to see 99% of ppl lost and not getting what I was tryna say. Obviously thought we were all on the same page, but we are clearly not. So let me explain. The focus has been on GPU rental rates and @GavinSBaker ’s “no dark fiber” argument, which basically is that unlike telecom in 2000, the infrastructure is being used. GPUs are busy. Demand must therefore justify the buildout. If you look at my chart, around 72% of Anthropic’s compute/infrastructure spending goes to training, research and internal use. Only c.25% goes to serving paying customers. Thus today, around 1/4 is used to serve customers. This is really important, because training is essentially an investment in FUTURE demand, a bet that future demand will materialize. Let's say today there's around 25GW in use, that means today we're using around 6.25GW for inference. As per @SemiAnalysis_ there's another 50 GW coming up in 2027. A 75 GW installed base would be roughly 12x that. So basically, what this tells you is that potentially supply to serve customers could 12x almost instantly, as inference/training allocation is an allocation decision that can be switched instantly. In the telecom bust, around 97% of dark fiber was dark, so that is 33x (from 3% to 100%). Not as extreme today, but it is pretty big still and something cuts pretty directly towards Gavin's argument of dark fiber, which up until recently I found really solid. The take here is that supply to serve customers can increase much faster than ppl thought about due to this, and we do not need to wait to 2028 or beyond for new GPUs to arrive and be energized. If funding tightens (happening a bit) or the economics of training the next model deteriorate, that allocation can change pretty quickly. Suddenly, hardware that was already “fully utilised” becomes additional inference supply. Two implications follow. First, inference supply can expand sharply without an equivalent increase in installed hardware. Second, rental rates tell you how tight the market is under TODAY’S allocation. They tell you much less about durable scarcity when a large pool of compute could be redirected instantly. I always found rental rates pretty useless as an indicator of future economics or pricing power durability and this dramatically worsens that argument. If you can 12x supply instantly pretty clear that prices will drop, a lot. Now the question is if sufficient demand is created by such drop in prices (you need to look at QxP). This basically brings us back to the Q x P formula, which is more difficult to solve. Essentially the underlying question is pretty clear (not that the answer is): can downstream customer spending ultimately cover operating costs and provide an adequate return on ALL consolidated upstream capex? Every GPU can be busy while the system as a whole is earning an inadequate return and if market conditions tighten and you run out of money, well, you're gone, despite how useful that model or whatever you were doing with those GPUs was. This is why I think hyperscalers are much more favoured as you at least have a massive printing press online today with really good balance sheets that can help finance all this while others need to see quicker results. My concern here is that while I believe AI is utterly useful (and trust me I spend quite a lot of money as % of my income) there's lot of ppl that max out at $200/month and don't want to pay more. As @benthompson and many other's have pointed out, consumers do NOT pay for being more productive, they pay to be entertained. So if budgets are constrained and have a top, the question that follows is, where are we in the penetration cruve? how many users are spending $0-50 and can go to $200/month? It is pretty obvious there is no clear answer on this and due to the capex figures being thrown around for 2026 and 2027 the market is starting to get nervous. $META cutting it's workforce by 15% YoY and $MSFT trying to cut 60k employees yesterday (27% of workforce) is an example even companies that never thought about "budgets" or "capital limits" are starting to see them. I think in this environment, the players that have the biggest optionality in monetization and have massive cash gushing machines are better positioned than the rest, and will survive. For me these are $GOOG and $META followed by $AMZN and $MSFT (it has enterprise). $SPCX has certainly stepped up its game with Grok Bot and I believe it has "entered the chat". Probably at par with amazon and just behind $META . The rest are pretty much levered compute vendors that I personally wouldn't touch. In the meantime, as all star CEOs are keen on spending, and given they pretty much control their companies, memory players seem really attractive at today's valuations. They certainly will capture most of that spending among the whole value chain.

UPDATE. Ferrovial +10%. Al parecer los de Ferrovial no estan tan locos. La pasta que tienes que adelantar es mucho menos de la que se pensaba la gente, lo apalancas un poco mas y ademas segun El Economista estan buscando socios (fondos de infra y de pensiones) para meterlos antes de empezar a construir, sobre todo en Atlanta, que es donde tienen mas exposicion. Lo mismo que ya hicieron en la I-77 con GCM Grosvenor y John Laing. Cosa que parecia bastante evidente cuando nuestro querido Del Pino se juega toda su fortuna y ademas tiene mas datos que nadie sobre los queridos managed lanes en USA con un track record en rentabilizarlos bastante bueno. Mas likely eso que haya hecho 2 fat fingers seguidos (aqui no se hacen fat fingers) o que se les haya ido la cabeza a todos, no? Que es un poco lo que decia todo el mundo. En la I-24 de los famosos $24,8bn, durante la obra (2027-2033) solo se pagan $7,7bn, que es practicamente lo mismo que ofrecian en total ACS/Acciona ($7,25bn) y Sacyr ($7,4bn). El resto lo van pagando de 2036 a 2085 con los propios peajes. En valor presente todo junto son $21bn (no cambia tanto). La diferencia de 3x, como mas o menos intuia, radica en lo que pueden cobrar a los conductores durante los 50 años. Aqui la unica liada (a toro pasado todos somos manolete) es que podias haber apretado mas en la subasta dado que el resto iba tan perdido. Por contrapartida, como es a sobre cerrado y no tienes otra oportunidad, si crees con tu numero ya tienes una buena TIR es mejor ir con eso que bajarlo por tratar de apurar y quedarte sin el negocio. Las empresas no estan para especular y hacer una TIR de 600%, con una del 15-25% vas que chutas. Mañana call. Lo que parece esta bastante clarinete es que cuando un tio con el 21% de la compañia y los datos de 5 managed lanes en la mano te hace un bid que nadie entendemos, lo mas probable es que los que no tenemos ni idea somos los que estamos en el otro lado. Ahora la gracia va estar en si se podia jugar asi en el pliego (reperfilar) o no.

Anthropic will likely have around $100bn/yr in compute costs in 2027, so revenue better ramp up, because costs surely are.

Pretty easy to guess $META capex and equity raise, or at least have a ball-park estimate. Sell-side consensus though is sleep at the well. FY26 guide at $135–145bn implies $49–59bn in Q4 or $200–240bn run-rate. Given how capex is scaling up quarter on quarter, it wouldn't be crazy to estimate 2027 could be in the $250–300bn range. My guess is it's more towards the $300bn figure. Consensus is at $205bn, LMAO. OCF this year is around $130–140bn (cons $133bn) and 2027 consensus number is at v$168bn. That put's, the gap at $80–130bn, which as said by mgm, it's too big to plug with debt alone. Equity raise could thus be around those figures, which matches $GOOG and $SPCX equity raises.