Investing in typewriters

Are we still betting on the devil we know?

In 1985 my father bought a Xerox Memorywriter 645s for his medical practice. A typewriter with a screen, daisy-wheel printing, two 5¼-inch floppy drives and an 8088 processor, which cost him 800,000 pesetas. 175,000 more than a Ford Fiesta at the time. A small fortune.

The early success of those "single-purpose" machines was the reason why not only Xerox, but also Olivetti and IBM itself, poured a fortune into factories to keep leading the world of professional writing. Today they may look like a joke, but they were a revolution: electric machines with built-in correction, interchangeable typefaces, industrial-precision finishes. Those enormous contraptions turned the head of any client who walked into an office where one had been installed. The demand was real, and every dollar invested in a new production line looked like a safe bet.

Five years later, almost all that money was worthless. Because PCs could be used for a million other things, and their word processors were completely superior to the software on the Xerox machines.

Xerox and Olivetti were not defeated by a better typewriter. They were relegated to irrelevance by something that operated in an entirely different paradigm. And the most painful part: the investment was not transferable. You couldn't convert a typewriter assembly line into a computer factory. The machinery, the supply chains, the workers' skills... everything was optimized for a product that was bleeding out.

Bleeding out so fast that Olivetti lost tens of thousands of employees between 1990 and 1996, despite its move into computing. And Smith Corona, the last American typewriter manufacturer, filed for bankruptcy in 1995 after more than 100 years in business, with sustained year-on-year sales declines that reached 15% a year.

I'm telling you this because for months I've had the uncomfortable feeling that I'm watching the same movie. Albeit in an ultramodern theater.

Large language models — the LLMs behind ChatGPT, Claude, Gemini and company — are, if you think about it, the most sophisticated typewriters we have ever built. I don't mean that as a cute metaphor, but in an almost literal sense. They are systems designed to produce text by predicting the next most likely word. Typewriters. Extraordinarily good ones, capable of generating paragraphs that seem written by a human, but typewriters all the same. Like my father's typist, they don't understand the subject they are writing about. And even worse: they have no model of the world. They predict tokens. What looks like reasoning is sophisticated statistical prediction.

Well, around these typewriters we are building the most expensive factories in history.

Microsoft, Google, Meta and Amazon have committed hundreds of billions of dollars to datacenters optimized to train and run LLMs. The value chain is well known: NVIDIA designs the chips, a couple of suppliers make the high-bandwidth memory they need, and the hyperscalers assemble everything into ever larger, ever more expensive and ever more energy-hungry data centers. Each new generation costs exponentially more than the last, to deliver incremental improvements measured by benchmarks nobody understands. A brutal investment, with no sign that we are getting close to a qualitative leap that would bring them anywhere near genuine understanding. A far cry from what came to be called Artificial General Intelligence.

David Cahn, of Sequoia Capital, warned in September 2023 of an annual gap of 125 billion dollars between what was being invested in AI infrastructure and the ecosystem's actual revenue. In June 2024, he revised the figure upward: the gap now exceeded 600 billion.

What is happening is best understood as a pincer movement attacking the current paradigm on three fronts at once.

From below, China is pushing on efficiency. DeepSeek, Qwen and other labs have shown that competitive performance can be achieved at a fraction of the compute cost, using smarter architectures instead of more brute force, which allows them to be far more aggressive on price. DeepSeek in particular sells at $0.14 per million tokens, between 30 and 100 times cheaper than its Western equivalent. The uncomfortable question is: if you can get 90% of the result with 10% of the budget, what exactly is all that massive infrastructure worth?

From within, the very physics of the paradigm conspires against its continuity. HBM memory prices are skyrocketing because only three manufacturers produce it. Energy consumption is scaling unsustainably. And hallucinations are not a bug that will be fixed with more parameters, but a structural consequence of a system that does not check what it says against a reality that language cannot model.

And from above, the most significant move. Yann LeCun — Turing Award winner, co-founder of deep learning, for more than a decade Meta's chief AI scientist — has left his former company, set up a startup in Paris with more than a billion dollars in seed funding, and put his professional legacy behind a very specific idea: that LLMs are a dead end. That real intelligence will not come from predicting the next word, but from systems that perceive the world, reason causally and plan actions. What he calls "world models." This is not an academic grumbling from the margins of the system. He is the most qualified scientist on the planet to evaluate this technology, and he has concluded that the direction is wrong. And a billion dollars in funding isn't raised on slides. All right, I know: another massive, capital-intensive bet with a high degree of uncertainty. But... what if it were the right bet?

In a previous article on the end of the token open bar, I argued that the economic model of AI was changing — that the era of flat rates was ending and that marginal costs mattered again. What I'm suggesting here is that this change might be just a symptom of something bigger: it's not that tokens are becoming expensive, it's that the very architecture that produces them might be reaching its expiration date.

I'm not saying LLMs are going to disappear tomorrow. Typewriters kept selling for years after the PC made them obsolete. What I'm saying is that anyone signing ten-year infrastructure contracts today on the assumption that the LLM paradigm is the definitive answer should at least ask themselves the question.

Because my father, who was an intelligent man, bought a Xerox with a screen when PCs were already in the stores. And it wasn't for lack of information. It was because investing in the known paradigm always seems safer than betting on the one that's coming.

How many Xeroxes with screens are being bought right now at datacenter prices?