From packet to token, and back to square one

That forgotten Net Neutrality...

I hope I'm not the only one who remembers the battle over net neutrality. That not-so-distant time when telecom operators discovered with horror that flat-rate internet was eating into their margins. And since something had to be done, they decided that, while investing in improving their networks, the solution was to try to control who used the data packets for what, and how much they paid for it.

The talking points were always the same. In our country, my admired Julio Linares, then Chairman of Telefónica in Spain, declared that the flat-rate model was "not sustainable" and warned of "a serious decoupling between traffic growth, network cost, the revenue it generates and the strain on investment". On the other side of the Atlantic, his counterpart at AT&T put it less diplomatically: "There are companies that want us to be a big, dumb pipe that just keeps growing. Nobody gets a free ride. Those who want to use this will pay." The European telecom industry association, ETNO, proposed that the internet be billed like gas or electricity: the more you use, the more you pay. And the chairman of the FCC tried to strike a chord with us (a fiber-optic one, no doubt) by defending the prioritization of paid traffic to benefit telemedicine and emergency services. Of course, nobody ever explained who would bell the cat. But those were minor, unimportant details... which may well have helped stir people up rather than calm them down.

The danger of the wrong argument

Honestly, I think they got the argument wrong. If instead of appealing to emotion they had appealed to the apparent rationality of the status quo in utilities, the story might have ended differently. Because with other companies we calmly accept that we pay "x" for having a certain capacity available, plus "y" for what we actually use... without anyone raising the absurdity of "paying just in case I use it". A concept similar to supermarkets weighing us on the way in and charging us according to how fat we are, on top of the food we then consume. And one that conveniently forgot that the telcos, after their Capex investments, had a very manageable Opex. Their costs rose very little and depended little on how intensively their networks were used. It was the magic of the packet versus that of the circuit.

But things went the way they went, and the story didn't end as the telcos wanted: they ended up as the only utility that operates under an absolute flat-rate regime. As flat as their income statements and share prices are today in most cases. The flat rate raised their revenue floor but almost wiped out their usage-based variables. A sad consequence of having focused on defending what they believed had value, instead of betting on the services others built on top of their capabilities.

Net neutrality became a regulatory battleground, the operators learned to be subtler, and in the end the infrastructure grew enough for the debate to lose its urgency. But the pattern is the same as today's: when the infrastructure can't keep up, whoever controls it tries to decide who goes first and charges for the privilege. Even if they don't succeed, because the regulator (or should we say the politicians?) never stands up to popular pressure once the idea that something is the future takes root in people's heads. And it's curious that it was the telcos themselves who created that conviction as a sales pitch: the Internet was the future. Today the hyperscalers seem to be up to the same thing: there is no future without AI... we'll see whether that makes them victims of their own demagoguery.

I'm telling you this from first-hand experience: I lived through that battle from the inside, as CEO of Terra Spain when we were, by far, the largest ISP in our market.

Well then: if we swap "packet" for "token" and "operator" for "hyperscaler", we are at exactly the same point. Swap "traffic" for "tokens", "network" for "datacenter" and "revenue" for "subscriptions" in Linares' sentence, and you have exactly the discourse of today's artificial intelligence providers.

Devour me (even more) again

OpenAI's retirement of GPT-4o in February 2026 was not due to a technical failure. It was a decision of financial architecture. OpenAI loses $1.35 for every dollar it brings in: the cost of serving billions of inference requests a day is unsustainable. What the technical documentation reveals is the implementation of dynamic inference arbitrage: when selecting a model, the customer is not accessing a static asset, but a gateway that routes the request to lighter variants depending on the datacenter's load and the real-time cost of the compute cycle. An "Uber-style" algorithm that, instead of raising the price, cuts the quality of the response. Researchers at Stanford and Berkeley documented performance drops from 97.6% to 2.4% on specific tasks between successive versions of GPT-4. Sam Altman himself admitted it bluntly with reference to GPT-5.2: "I think we just screwed that up."

In my Terra days, the telcos routed traffic too. But they did it to avoid the collapse of the international connectivity outlet. With AI, the hyperscalers route inference to different models to avoid the collapse of their gross margin in real time. They are renting us a Ferrari but throttling the fuel flow through software according to what it costs them at any given moment.

And they are not alone in this war to optimize the use of their capacity. Last week, Anthropic, the maker of Claude, signed a deal with SpaceX to access more than 220,000 GPUs and 300 megawatts of additional capacity. Dario Amodei and Elon Musk, who three months ago were publicly calling each other every name in the book except "sweetheart", have sat down at the same table because Anthropic didn't have enough infrastructure to meet the demand of its own paying customers... and because Musk's Grok barely captures between 3% and 5% of generative AI traffic, and had a datacenter oversized for its real demand. The result is a deal that would have been unthinkable without the brutal pressure to obtain (and be able to pay for) computing capacity.

It also sounds as if the hyperscalers are fully absorbed in infrastructure and, with a few honorable exceptions, are not looking toward where the real value might lie in the near future. With one terrible difference compared with the telcos: the telcos had a business with brutal Capex but low per-use Opex, whereas for the hyperscalers, operating the infrastructure is also tremendously expensive and has no economies of scale with usage. And this will remain the case, at least, as long as they insist on pushing users to use frontier models first and foremost... even if those models then turn out not to be quite so frontier after all.

From that dust came this mud.

Meanwhile, every provider's solution is the same as the telcos': usage tiers. OpenAI has introduced a $100 Pro plan, has imposed stricter limits on the Plus plan, and has begun redirecting heavy usage toward a consumption-based credit model. When the flat rate can't stretch any further, the open bar is over. Exactly as the European telcos' association proposed fifteen years ago: the more you use, the more you pay.

With an additional "attraction": how are we users now supposed to know how far I can trust the answer my model is giving me today? Worse still: will we be able to reliably use that agentic AI they tell us is the future? Will I dare to automate functions in the hands of an "employee" who is brilliant one day, but the next may turn out to be a "borderline" individual?

Exciting, isn't it? Don't you get the feeling that someone may be shooting themselves in the foot? Some companies forget far too quickly that what for them is "a very cool technology" is, for an executive, the foundation on which to reinvent their business. And when a foundation is unstable, its appeal is lost completely.

... though there's mud, and then there's mud...

But there is an additional problem the telcos didn't have to face, and it makes this situation potentially more destructive. Their infrastructure was general-purpose: an optical fiber works just as well for telemedicine as for Netflix. The datacenters being built today are optimized for a very specific kind of computing (training and running LLMs) with a very specific kind of hardware. If the paradigm changes, those datacenters will be hard to repurpose.

The telcos, at least, had reusable pipes. The hyperscalers are building pipes that only carry one kind of water. But it is the water with which we have to rethink our future. Let's hope they know some history, because if they don't, they are going to repeat the mistakes of the past.

All that time learning to say pinícula, and now it turns out it's called flín! (an old Spanish quip: you finally master the word for "movie", and they go and call it "film")

This article is part of a series that began with "Less Wood, It's War!", and continues with its predecessor, "Where's the pea?". To be continued...