# Invest Like the Best notes: a model only lives three to six months

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/iltb-2026-11-04-martin-casado-on-frontier-model-economics-ai-s-slo/

> In episode 498 of Invest Like the Best, published 4 November 2026, Patrick O'Shaughnessy talks with infrastructure investor Martin Casado about frontier model economics: he says the frontier has taken 80%+ of the market on a dollar-weighted basis for three years running, while a single model's lifespan is three to six months, so slipping three months on the next release can cost you the whole market. He also says the real shift is that capital can finally be pointed at engineering. Educational notes, not investment advice, no tickers, no price targets, and my reading may be wrong.

Published: 2026-11-04
Locale: en
Tags: venture capital, AI infrastructure, podcast notes, moats, M&A


![A farm chicken shed converted into a machine room, the camera low at the open doorway looking in, two rows of server racks with indicator lights receding down the long shed, a second doorway at the far end opening onto dusk fields and a single utility pole](/covers/iltb-2026-11-04-martin-casado-on-frontier-model-economics-ai-s-slo-cover.png)

> Study as though you will never catch up, and still fear losing what you have.
>
> —— *The Analects*, "Tai Bo" (Spring and Autumn period; my own translation)

In episode 498 of *Invest Like the Best*, published 4 November 2026, Patrick O'Shaughnessy brings on Martin Casado, founder of Nicira and now an infrastructure investor. On the show he says the frontier models have taken 80% or more of the market on a dollar-weighted basis for three years running, while a single model's lifespan is three to six months — slip three months on the next release and the whole market can be gone. He also says the big shift this cycle is that capital can finally be pointed at engineering: twenty people and a billion dollars, and they can now productively spend it. There's a condition attached, and he states it plainly: the market is still growing. What happens if use cases saturate in pockets, he says, nobody knows.

## Back to the 1970s: that term is being used wrong

Casado has been investing for a decade and done hundreds of deals, and he says he keeps noticing that people in those rooms have no shared definition for the words they use. His favourite one to pick on is network effects. The host's definition — a product that gets better as it adds a marginal user — is what a third of people say, and Casado calls it close to a tautology. Any new user will find a bug, give feedback, tell other people. None of that says anything about whether you can hold a market.

The original meaning comes from an economist at Bell Labs in the 1970s: on a telephone network, each user who joins adds connections that grow with the square of the number of users, so value grows faster than headcount, and you get natural monopolies. The root of the confusion, Casado says, is that we conflate the mechanism of a moat with the shape of a moat. There are many mechanisms — brand, scale, a two-sided marketplace, a network effect. There are only three shapes: superlinear, linear and sublinear with scale. Two companies each raise $5 million; if the leader's moat is superlinear, the gap widens, and if it's sublinear the challenger eventually catches up, because the leader has to spend ever more money to cover the same ground.

![Three curves rise from the same starting point: one bends further upward as it moves right, one stays straight, and one flattens out early, labeled superlinear, linear, and sublinear.](/figures/moat-three-shapes-en.svg)

Then he turns it around: how many genuinely superlinear, telephone-network companies have there ever been? His own answer is a handful. The telcos count — that's why they got broken up. LinkedIn counts. GitHub counts. All of them are interconnections between people. Uber he puts under scale effects and a two-sided marketplace: drivers drive for everybody and passengers ride with anybody, so neither side is locked in.

![Two panels each show two lines from the same starting point: on the left the lines spread further apart, on the right they stay close and meet.](/figures/five-million-each-two-outcomes-en.svg)

## The last three years: the frontier takes 80% of the dollars

Having built the vocabulary, Casado then flattens the whole subject: most of the time defensibility isn't worth discussing. When a market is growing, companies don't want to fight head-on, because there's so much new white space that they'd rather go take it. When a market slows, it consolidates regardless of moats. He remembers early conversations where investors told him they'd never invest in a model company because value wouldn't accrue there. Three years on, the frontier models take 80% or more of the market by dollars. Being on the frontier, he says, comes with tremendous pricing power even if you're only an epsilon better than the next guy, because demand is effectively infinite.

For penetration he gives the order of 0.1%. Enterprises say 3%; on a budget basis it's 1% or less.

![Two bars of equal length sit one above the other: the top bar is almost entirely filled, while the bottom bar has only a thin sliver filled at its left edge.](/figures/frontier-share-vs-penetration-en.svg)

The line of his that made me laugh: a junior investor told him some company's quality of revenue was bad, and he replied, "Homie, we didn't even know how to monetize the internet." Business models lag disruptive technologies. If you truly believe you're buying a disruptive technology, you should believe it creates its own TAM, and you should accept that nobody yet knows the pricing model — it took the industry years of hand-wringing to get from perpetual licences to recurring revenue. His warning is that screening on quality of revenue pushes you toward the non-disruptive end, because the genuinely disruptive thing hasn't grown a business model yet.

## Now: a model only lives three to six months

Then he asks a question I hadn't thought about. Suppose a frontier lab stumbles and doesn't ship for six months — what's the lifespan of the previous model? The host says: very short. Casado says that from the model-usage curves he sees in apps companies and pitch decks, it's three to six months. Software has always assumed a long tail, with people still running the old version for years. Models don't have that tail. Slip three months and the market may not be there.

![Two usage decay curves: the legacy software version trails off into a long tail extending to the right, while the model curve drops to zero between months three and six.](/figures/model-life-no-long-tail-en.svg)

The second step is the counterintuitive one. People say coding is already 95% automated, so you don't need a smarter model and can switch to a cheaper old one. Casado pulls the two claims apart: the first one — I don't need more tokens — may well hold, and the second doesn't follow from it. If you and I run competing startups, are you going to pick the dumber model to fight me with? In a competitive equilibrium you wouldn't; you'd pay ten times more for the smartest thing available. Can you ship faster, can you make it better, can you win in a market — all of that keeps pushing you toward more intelligence.

So his conclusion is that pricing holds while growth slows, because token consumption stops climbing. To hold revenue up you either raise prices or broaden beyond code into general knowledge work.

## Now: capital can finally be pointed at engineering

Here he picks up the other 1970s term, the mythical man-month. Fred Brooks's point was that software projects you try to speed up by adding people get slower. In Casado's own vocabulary, engineering output is very sublinear in headcount.

![An input-versus-output chart where the add-people curve flattens quickly while the add-money curve is lifted along its whole length, opening a vertical gap between them.](/figures/money-moves-engineering-en.svg)

Ten years ago, he says, if five founders had shown up asking for a billion dollars and he'd given it to them, what would have happened? The host: it sits there. Casado: or they detonate it — hire a pile of engineers, and then you're really stuck. Today he watches twenty people take a billion dollars and use it. Whether they make $500 million or $2 billion from it he has no idea, but the fact that the money can be productively absorbed is new, and his firm are capital allocators, which is why he calls this an industrial-revolution-level shift.

The other side of all that money is debt. He says he isn't a debt guy, and what he sees on the ground is available debt running out. If you're the Norwegian pension fund, do you lend to a random neocloud or buy T-bills? The result is that investment-grade companies get debt and the rest have a harder time, and the smaller players whose business models depend on debt will need their customers to prepay.

He's blunt about the real limitation: hardware supply, plus power, chips and materials. Under that constraint, 80% of the volume goes to the scale providers and 20% goes to the hard-to-reach corners — the person who happens to have a relationship with a government that has power available. He mentions a GPU marketplace where a good number of the GPUs sit in converted chicken sheds, with people on farms working out how to power them. The large operators aren't going to build a chicken-coop GPU business, and the market is big enough that 20% supports a wide, fragmented crowd. Meanwhile there are at least 500 pieces of proposed legislation aimed at stopping data centre construction in the United States alone, Texas included; if you can't break ground in Texas, he says, then it's international or it's space.

![A narrow bottleneck splits downward into two streams: one solid large block on the left and a scatter of small squares on the right.](/figures/hardware-bottleneck-eighty-twenty-en.svg)

## The deals of the past two months

Casado has been part of three acquisitions recently: Cursor, World Labs and OpenRouter. On Cursor: the company was close to finishing a raise of a few billion dollars when Mythos came out, and working backwards he estimates Mythos cost something like $3 billion to make — he says himself he doesn't know the real number. If Cursor was going to build models, it needed to raise more and get access to that much compute, and there aren't many counterparties with that much compute, which is what kicks off strategic conversations. The combination that closed it: the buyer had the compute, Cursor had by far the most users and the most data, plus a team that had already built models.

He offers three rules of thumb. First, these deals share deep personal alignment up front — both sides hold the same view of the future and think about it the same way. Second, take the buyer's point of view: the buyer is purchasing an option on a large market. His own exit to VMware was $1.26 billion, a billion-dollar option on a $40 billion market, so the seller's job is to complete the inductive proof for them — the base case is that we sell into this budget today, the inductive step is that attached to your product we can sell much more, and therefore you reach the big market. That conversation, he says, only founders can have.

![A very large box with a very small box nested inside it, where the small box is the acquisition price and the large box is the market the buyer wants.](/figures/acquisition-as-option-en.svg)

The third rule is the one he presses hardest: don't play games. A lot of investors' instinct is to leak it to the press and bring in other bidders. Casado says you're dealing with the most sophisticated people in the industry, and thinking you'll beat them at poker is kidding yourself. OpenRouter leaked constantly, he says; World Labs didn't leak, because the only people who knew were him, Fei-Fei and a few others. He's seen deals of that size fall apart because a board member got into a founder's head about bringing in competitors and leaking to the press.

## Two things to watch next

The first is saturation. Casado says three years of data already answer the question of what happens while the market grows: the frontier wins and does a great job. What nobody can answer is the other half — what if use cases saturate so you don't need more intelligence, or the market slows for some other reason? His read is that there's a very real world where the value comes crashing back down into open source, and because models lack the long tail software always had, that would be a shift we've never seen.

The second is imagination. He says the bottleneck has moved from the technology to our own imagination, because thirty years of bad software has beaten us down — hard to use, hated by older people, broken most of the time. His own project is getting Grok Bot to embody itself in his house: he wants the Star Trek version where you walk in and the house greets you, so he's been feeding it pictures and descriptions and letting it decide what to buy. The day before the recording it bought a few hundred dollars of gear, obscure open-source projects he says he doesn't even recognise. He also mentions World Labs' Atlas model, which generates 3D scenes with dynamics, and says we're this close to stepping into a movie — so why is there no immersive play you can walk into yet.

## When someone says a company "has no moat"

I've had a holding described that way, and my reaction was to panic first and then go looking for an article to decide it for me. After this episode I'd rather break the sentence into three questions I can check myself.

One: does its advantage grow or shrink with scale? That's Casado's shape, and it's more useful than arguing over whether something qualifies as a network effect. Two: is its market growing or shrinking this year? His reasoning is that while a market grows, companies go take white space instead of fighting, so a moat only gets tested once growth stops. Three: is the attacker a new technology or an old rival? His example is databases — more software written means more infrastructure needed, and the long tail of features, the performance tweaks and the actual operating of these things are all hard, so he doesn't think companies like that get disrupted by AI, while competition from a peer is a pre-AI phenomenon. That's his view; I'm not adding my own. The three questions share one virtue: each answer is a fact you can go look up, which "does it have a moat" never is.

## Which standard am I actually using

The other reader situation I know better: the same stock, one person calls it cheap and another calls it expensive, and I myself look at valuation one day and trend the next, and somehow I can always talk myself into it.

Casado enumerates four investing mindsets. Financial: the proxy for a business's health is market size and the financial metrics. Productive: you're buying a productive asset, and as it creates value it will create market, so what you diligence is the technical approach, the adjacent markets, how fast it evolves, whether this is the right team. Frontier: it's people, you believe the founders saw something you didn't, and you go. Speculative: you don't care about the underlying asset at all, you're timing a market. The problem, he says, is that with all four archetypes in hand you can justify any deal and reject any deal, and if you don't know what you believe, you'll end up rationalising whatever bias you brought.

![A horizontal axis runs from reasons you can check to reasons you cannot, with Financial, Production and Frontier placed along it while Hype sits alone below the axis.](/figures/four-investor-mindsets-en.svg)

My version of this: when a holding is down I slide toward "technology creates value," and when it's up I slide back to the financial frame. What I do now is write down which of those reasons I used to buy each position, and then only check it against evidence from that frame. Bought on financial reasons, the bad news is the numbers deteriorating. Bought on technical reasons, the bad news to watch for is the technical path being routed around — a weak gross margin this quarter doesn't count.

## Sources worth reading

- *Invest Like the Best*, episode 498, 4 November 2026 — the source for these notes
- Martin Casado's piece on the four investing mindsets
- Fred Brooks, *The Mythical Man-Month*, 1970s
- The Bell Labs observation on how connections grow on telephone and rail networks, later known as Metcalfe's Law
- World Labs' release notes for Atlas, its model for generating 3D scenes with dynamics
- Arthur Rock's writing on early venture capital, the comparison Casado draws in the episode
- Anthropic's S-1, filed the week the episode was recorded
- Roelof Botha's TechCrunch Disrupt talk, the one Casado argues against on the show

## One thing to take with you

Work out which standard you're actually using. With four of them in hand, you'll buy on the logic of one and sell on the bad news of another, feel rational both times, and the loss in between came from switching standards rather than from the company.

Something I've tried: take the most recent decision you made — a job change, a move, whether to stay in a relationship, buying something expensive — and write down the one reason you used to convince yourself. Then take the decision before that and write down its reason. Put the two side by side. If they contradict each other, the standard you're really using is whichever answer you wanted that day.