The $145 Billion Question

For the past three years, the technology industry has mostly rewarded one thing: spending more.

More GPUs. More data centers. More models. More capital.

This week, I noticed a subtle change in the conversation. Meta can now spend as much as $145 billion in a single year. Samsung can report record profits from the AI memory boom. A startup can raise $30 million to let autonomous agents spend money. Yet none of those facts, on their own, tells you who will build the best business.

The question is no longer whether AI demand exists. It clearly does. The question is who can turn that demand into durable margins before the bill arrives.

Here are three signals I think are worth your attention.

Meta narrowed its expected 2026 capital expenditures to between $130 billion and $145 billion. At the upper end, that is roughly equivalent to spending almost $400 million every day.

The easiest interpretation is that Mark Zuckerberg has become extraordinarily confident about AI. I think the more useful interpretation is that Meta’s existing advertising machine gives it something most AI companies do not have: a way to monetize intelligence immediately.

If better recommendations make Instagram slightly more engaging, Meta sells more ads. If generative tools help advertisers create more campaigns, it gains more customers and inventory. Meta does not need to invent an entirely new business model before AI produces revenue.

That makes its spending different from the capital deployed by a standalone model laboratory. The infrastructure may look similar, but the economic feedback loop is much shorter.

Still, $145 billion changes the standard of proof. What I will watch is not the quality of Meta’s next model, but whether profit grows fast enough to prevent AI infrastructure from becoming a permanent tax on the advertising business.

Samsung Electronics reported quarterly revenue of 171.5 trillion won and operating profit of 89.5 trillion won, both records. Its semiconductor division alone generated 127.5 trillion won in revenue as constrained capacity and rising prices strengthened the memory business.

This is what the AI boom looks like several layers below the chatbot.

Large models need high-bandwidth memory close enough to their processors to keep data moving. A powerful GPU waiting for data is an expensive idle machine, which is why memory has become one of the industry’s most valuable bottlenecks.

And yet Samsung shares fell after the results. That reaction is more interesting to me than the record itself.

Markets price what today’s profits will attract tomorrow. Samsung and SK Hynix are preparing enormous capacity investments, while Chinese competitors are trying to close the technology gap. High prices create exceptional margins, but also finance the supply that can eventually destroy them.

This is the classic semiconductor problem. Being indispensable during a shortage is not the same as having permanent pricing power.

So I would resist the temptation to read every supplier’s record earnings as evidence of an equally durable moat. The winners will be the companies that can stay technically ahead as capacity expands—not simply the companies enjoying the highest spot prices today.

Natural, a one-year-old startup, raised a $30 million Series A to build payment infrastructure for AI agents. The idea is straightforward: an agent may be able to find a supplier, negotiate a price and organize a delivery, but it still struggles to complete the transaction without handing control back to a human.

At first glance, this sounds like a narrow fintech feature. I think it points to a much larger constraint.

Money is permission. Human payment systems assume that a person initiates a transaction, approves it and can later dispute it. An autonomous agent complicates every step. Who is liable if it pays the wrong vendor? How much can it spend? Can it hold funds? What evidence proves that it acted within the authority it received?

The difficult product is not moving money, but the control layer around it: identity, limits, audit trails, fraud detection and disputes. Agent payments must define a new commercial actor that banks, merchants and regulators can trust.

If agents begin purchasing software, logistics or cloud capacity for companies, the most valuable part may not be the payment rail. It may be the system deciding when a machine can use it.

If you want to understand where part of this enormous infrastructure budget may flow next, I recommend reading “The $73 Billion Fiber Optics Thesis” from Macro Notes.

GPUs receive most of the attention, but those chips must exchange an extraordinary amount of data. The analysis explains why optical connectivity is moving deeper into AI architecture, who may retain pricing power and how a $73 billion market could develop around this bottleneck.

I think it is a useful companion to this edition because it asks the question that matters after every spending forecast: who actually gets paid?

AI is moving from a demonstration phase into an accounting phase.

The next winners will not necessarily be the companies making the most ambitious announcements. I would look for businesses that sit between unavoidable demand and a difficult constraint: Meta’s direct monetization loop, advanced memory that keeps processors working, or permission systems that let agents transact.

You do not need to predict the most powerful model to understand the economics. You need to identify the invoice that cannot be avoided.