@adlrocha - The 18-months Deadline

On surviving the Great AI Capital Shift, and our last chance to make useful work

For six months I’ve been trying to hedge my (and my family’s) access to intelligence. Open models, local inference, a box on my desk that runs a decent model without asking anyone’s permission, you know, the kind of things you’ve seen me write about in previous posts. Last week, I finally did the math on what realistic local local inference would require, and while it helped me make a better sense of the current state of affairs, it made me realise that there’s a limit on what local inference can achieve without the right budget.

Limiting myself to the 10k I wanted to make available for hardware purchases, with the current rising prices, and the increasing size of the models, wouldn’t end up in anything more than a baseline insurance for the worst case scenario. And my current hardware may already be covering for that.

If I wanted to secure my access to intelligence at a bigger scale maybe I was framing the problem in the wrong way.

---

Two powers, one weapon

To understand why everyone is feeling like we are living a “now-or-never” moment, it helps to stop reading this as a technology story and start reading it as a power one. Intelligence, and the underlying hardware required for it has become the new strategic resource, the way oil was in the last century, and (as you all know) two states have worked that out.

Let’s first have a look at what is happening in the US. In June the government issued an export directive suspending all foreign-national access to Fable 5 and Mythos 5, any foreign national, inside or outside the country, including Anthropic’s own foreign-national employees. Then this week, Anthropic was allowed to redeploy Fable 5 once the mechanics were worked out (and after making the model dumber through the attached guardrails). This chain of events were quite disturbing for me as I was following them: so you are telling me that a state can reach into a private lab and decide, at the level of individual passports, who gets the frontier and who doesn’t. Whether the lab wanted to is almost beside the point, the capability to gate access at the border now exists and has been used once, which means it can be used again.

That’s the first move in any resource war: capture the point of production, then decide who’s allowed to drink from it. But to make matter worse, market dynamics are also shifting who gets access to resources. Everyone is trying to get their hands into some hardware, which have forced Apple to raise base prices around 20% overnight, MacBook Pro from $1,699 to $1,999, Mac Studio from $1,999 to $2,499, and Blackwell cards jumped $400 to $900 in a single day at Microcenter. It is a fight between data centre operators, companies, individuals, and consumer electronics manufacturers for the limited resource required for intelligence (memory at this moment). Owning the compute is owning the means of producing intelligence, and we are all bidding for the same scarce memory at once.

So the states wall off the frontier from the top, while everyone else scrambles for the substrate from underneath, and the two dynamics feed each other. This is why people are rushing themselves (and urging others) to buy compute immediately, because it feels like the window of opportunity is closing, and access will become increasingly unreachable for individuals.

What makes me wonder, if we look at the situation with the lenses of a “race to get yourself access to intelligence”, is China actually the bad guy in this game? There are clearly two great powers doing what great powers always do when a new source of leverage appears, wielding it, walling it, and priming their publics to fear the other side’s version of it. The export bans, the fabs, the distillation cold war… they are both using the kind of nasty tricks once can think of but in the 21st century. None of them are free of sin, but the models coming from China at least are open (otherwise no one in the West would know about or be using them. This is the thesis from the Z.ai founder about why everyone knows their models but not Bytedance’s one).

And where does that leave a European? Nowhere, honestly. We have no frontier lab except Mistral, no leverage in the fight, and no seat at the table where these decisions get made. I feel like we are not even participating in this war. I’m a bystander in a country that either buys closed US models that can be export-controlled away from me by directive, or finds another way to hold intelligence that no directive can reach. If at least I was based in SF or NYC, where I could meet the right people, and be exposed to the right ideas, but this is not the case.

Which is the whole reason open source has to win, and why it is what I am optimising for (acknowledging that I have an obvious bias towards open source).

---

Why open source has to win

When we pay OpenAI or Anthropic subscriptions every month (or every year, as loyal readers may already know that I use yearly subscriptions as a hedge), we get in exchange access on their terms. The June directive reminded me that those terms include “we can cut you off if your passport is wrong”, “we can decide for what you are allowed to use intelligence”, and “how we use your data”. An open weight doesn’t work like that. I download it once, it sits on my disk, and nobody can take it back. When GLM-5.2 or Kimi drops, I pull the weights and I own that copy forever, ban or no ban (as long as I have the needed hardware). That’s the whole difference, and for someone in Europe it’s the only card I get to play.

The reason this matters now and didn’t a year ago is that the open models got good. GLM-5.2 is close to Opus, MIT-licensed, and I can (hypothetically, with enough hardware) run it myself. Coinbase, itself, now routes its load to Kimi as their defaults for most engineering work. The closed labs are still ahead on the really hard, open-ended stuff. But most of my work isn’t that, and for the rest the open models already do the job.

The labs keeping intelligence open and un-bannable right now are the Chinese ones. I have no fondness (or a formed opinion, really) for the politics behind them. But if the choice is weights I can keep versus a subscription someone can revoke, I’ll take the weights every time, and I’d rather be honest about where they come from than pretend I have a better option.

Unfortunately, to run these models you still need access to a lot of hardware.

---

Everyone AI-pilled is warning it

If you spend any time around people who have internalised where this is going, you hear one message on a loop: the window is closing, and you have to act now. This thread from 0xSero shares his view on why it is NOW or NEVER: buy hardware, you have one year, AGI fits on 24GB and superintelligence on 500GB, go.

But it is not just hardware, Kenneth Rogoff, wrote that nations failing to secure a place in the AI supply chain face mass job displacement without the tax base to contain the fallout. You’ve probably heard the meme about the “permanent underclass”, i.e. the inability to move in the social ladder once AI diffuses into the economy. There’s also Microsoft framing that got quoted everywhere and put most white-collar tasks as automatable within 12 to 18 months. Or what Emad Mostaque, the ex-Stability founder, calls it the Thousand-Day Window in his book The Last Economy (that I’ve already introduced in other posts of this newsletter).

One year, eighteen months, a thousand days… the point is that it looks like something will be changing soon if AI “makes it” and we need to get ready for it. At least that seems to be the consensus in the bubble I am living in lately.

It’s true that not everyone reads this situation as a crisis. There’s a calmer camp of people that are not as afraid or do not care much. Financial Samurai makes the clearest version: the safety net rises with productivity, an underclass is survivable, and if you’re really worried, just be a capital owner instead of a worker. It’s a reasonable piece, and I don’t think it’s wrong, but he talks from a comfortable position. It’s written by someone already FIRE’d with three-quarters of a million dollars in AI stocks, describing the view from a place he’s already reached. His prescription and mine turn out to be the same, own capital not labour. Where we part company is whether that’s a relaxed observation or a deadline.

I think it’s a deadline, and I’ll come back to why. But notice what both camps agree on, the panicked one and the calm one: the thing that matters shifts from labour to capital. That used to be a fringe take. It’s now the mainstream economist read, the doomer and the FIRE blogger and a Microsoft slide all landing in the same place. Labour in itself may end up becoming obsolete, and access to capital may make the difference moving forward (as access to capital may also become hard in the coming years).

---

I may be optimising the wrong variable

Here’s what last week’s post did to me. The entire hardware quest assumed access to intelligence was the bottleneck, and this is correct, but the amount of money required to make something that ensure a baseline level of intelligence for what I want to achieve were out of my reach. My open-tier usage cost me about $30 over two months, and, as I just argued, the open weights are closing on the frontier fast enough that access stops being the thing I need to worry about. I’d spent six months optimising the one variable that was already solving itself.

My realisation was that scarce things turned out to be capital, and an understanding of the system that AI is about to rearrange. Who cares if the price of hardware goes up if you have a way of finding the means of securing it for yourself? The deep pockets of the big labs will be able to get themselves the hardware that they need (and individuals are having a hard time to get).

I want to be clear, “local AI was not a mistake.” It wasn’t. Sovereign, self-hosted intelligence still matters, a lot, and I’ll keep running the box I own, and trying to upgrade my local rig. The honest reason I’m shifting my focus a bit is the need for capital, not conviction. I don’t have the money to move the needle on local infrastructure, and the people who do are already moving it faster than I would ever could. Exo Labs is pooling consumer hardware into clusters. And to be able to stand on their shoulders, and hoard myself all the hardware that I can get to provide open intelligence, I need to stop begging for donation, and find myself the capital that I need to get the hardware that hedges my access to intelligence.

---

Why capital is becoming leverage

Open weights hedge access. They do nothing about the thing Rogoff is actually warning about, which is ending up on the wrong side of the capital-versus-labour line when the music stops.

If intelligence becomes abundant, the return on selling your labour collapses, because the thing you were selling is now cheap and everywhere. What holds its value is ownership, of the compute, of the models, of the companies that run them, of the assets that appreciate as all of this reshapes the economy. We like it or not, we need to play by this new rules, but to do so, we first need to know which are these. There’s no point trying to change the whole game. Financial Samurai is right that the exit is owning capital. He’s just already through the door, and I’m arguing you have to sprint for it while it’s still open.

So my new focus to achieve AI independence is on amassing capital. Not as a life goal, but because capital is the only lever I see that an individual can have to navigate this fight out of individuals’ control. You can’t out-legislate a superpower. You can’t out-spend a lab. But you can build enough of a position that the transition happens with you rather than to you. We like it or not, money gives you optionality, and intelligence agency (I am on a rampage with this aphorism-like sentences the past few posts. Writing a post every week is paying off).

And I am aware that there is a ceiling on what an individual can do. I am not going to bend the global outcome of the AI transition. Nobody reading this is (and if you are, please hit me up, I would love to chat and learn more about you).

The scale of capital it would take to influence how this lands for humanity sits with states and a handful of firms, full stop. What the level of capital you can realistically reach does change is the radius of who you can protect. A little, and you shield your own family from the worst of the shock. A lot, and maybe your reach extends to your closest community, the people around you who won’t get a seat either. More than that, global impact is a fantasy for almost everyone, and I’d rather aim at something real. If I get to the community radius, I’ll consider it a life well spent. That’s my honest target. Protect my people, starting with my kids, because that’s the part of the board you can actually move.

---

What I’m actually going to do with this window

If we agree that there is an opportunity window of at most 1000 days where we can actually make useful work before everything changes, what should I be doing? AI alignment? Local AI? Those would definitely be useful, but being just me, would that achieve my goal of decromatising access to safe intelligence? I don’t think so. I realised that the best way to get access to the capital that has a chance to achieve that goal, I need to understand the machine that distributes it: the economy.

How I am thinking about this right now. First, I need to genuinely understand how the financial system works today, because I am getting increasingly convinced that it drives most of the real mechanics of the world, and it’s where AI’s impact shows up first, in wages, in asset prices, in who has leverage, and traditional economic theory and classical wisdom no longer describes how the machine works today (since the subprime crisis, traditional economics no longer explain the crazy market dynamics of the past few decades).

Second, model it as a complex system I can watch in something close to real time, so I can see the shocks as they happen and, ideally, position my family inside it rather than get positioned by it. Third, simulate what AI does to that system as it permeates everyday life.

That third thread would have been a fantasy two years ago. It might still be. But I have an idea of where to start, and auto-research loops and agents driving the right data feeds have got good enough that a personal-scale attempt is at least a real project now, not a daydream (I wrote about the lab that runs while you sleep a while back, and this is where that goes next). I’ll get to how I’d actually build this separately, that’s a post of its own, along with the alignment and formal-verification angle, why these same 18 months matter for embedding guardrails before the next generation of models diffuses. My backlog is filling up fast with ideas.

---

18-months left

I really don’t know if this 18 months window is real, and if the impact of AI in society is going to be as fast and critical as advertised. I don’t know if a guy in Spain modelling the economy at night can move anything at all, or whether this is the same fantasy every quant has chased and lost. Quite possibly I’ll abandon it in a month.

But the attempt teaches me the system, stress-tests these models at their actual limit, and, if it works even a little, funds the independence I was trying to buy the slow way. And I want to be able to tell my kids I tried. Doing nothing because the deadline might be fake is also a bet, and I think it’s the worse one.

If you half-believe the window is real, what would you actually do with it?

Until next week!