How many R's in strawberry?
THE COMPUTE / BY TOKENANDO Issue V / June 4, 2026
Last week we promised to report back: we had a full house.
On a warm Thursday afternoon in Barcelona, 15 people gathered in a spacious meeting room, with the late-spring light pouring in through the windows. I asked a question that sounded like a primary school riddle: how many letter R’s are there in the word strawberry? Most people said three. For the past two years, a large language model, when asked the same thing, would have confidently answered “the word strawberry contains 2 R’s”. That was the starting premise of the session: an AI model breaks text into tokens, fragments that rarely line up with single letters, so counting the R’s in a word is not the sort of task it is built for. Follow the Token, our 60-minute map of the AI economy, took its cue from that one riddle.
My worry was that the subject would feel abstract or irrelevant and people would drift to their phones within the first 20 minutes. None of that happened, everyone stayed a full 90 minutes, well past the hour we had advertised. The attendees were a deliberately mixed group, from the education sector all the way to global consulting, through the healthcare and legal fields. Every one of them uses AI in some form; only two are actually building with it.
The aha-moments were abundant, and not always expected. The strawberry trick won some people over; for others it was the realisation that an agentic system, left to work on a task, consumes a lot more tokens than a chat conversation, and runs up a bill that scales with its own autonomy. One question came back a couple of times: which kind of companies, which real use cases, and what figures can you point to? The audience welcomed the theory but they wanted it tied to real applications and numbers they could relate to.
The session definitely raised more questions than it answered, which I have come to see as the right outcome: raise awareness even slightly and curiosity follows; once curiosity is there, the appetite for proper education and usable insight grows quickly.
That is the premise of what we are building at Tokenando: awareness first, then education, then insight. Everyone walked in for the first and left asking for the other two.
The meter is running
On June 1, GitHub moved every Copilot plan to usage-based billing. Instead of counting requests, it now meters the tokens each interaction consumes, input, output and cached, and charges them against a monthly allotment of credits priced at one cent each. Developers who had grown used to a flat $10 a month did not welcome the change quietly.
Meanwhile at Uber, the company had handed roughly 5,000 engineers an agentic coding tool late last year, and by April it had spent its entire 2026 AI coding budget. Staff even had a word for the behaviour: tokenmaxxing. Uber’s chief operating officer Andrew Macdonald said he cannot connect the spending to anything a passenger would notice, and the company has since capped each engineer at $1,500 a month per tool.
The same force is being felt by the firms whose product is the billable hour. Accenture has lost about 41 per cent of its value over the past year as AI compresses work that used to fill a multi-year engagement into a few months, and the market has stopped paying a premium for headcount alone. All of which is the lesson the Barcelona group arrived at: the more an AI workflow or an agentic system is left to run on its own, the faster the meter ticks.
Which use cases justify the spend? Our attendees asked it first; finance teams everywhere are asking it now.
Sold out
Earlier this morning in Taipei, at the chipmaker TSMC’s annual meeting, chief executive C.C. Wei told shareholders the company will not meet AI demand for several years, even as new production plants come online. Mr Wei said its most advanced capacity is booked through 2028, and the next plant it is building in Arizona is sold out before construction has finished. A fab (short for fabrication plant) takes two to four years to build, which means we will be facing a multi-year structural shortage.
The demand behind that warning surfaced in the week’s financial results. Dell guided to $167bn in revenue on the strength of a $51.3bn order backlog for AI servers, and its shares rose about 40 per cent; Lenovo’s market value doubled over the month, its best run since 1999, on AI revenue up 84 per cent. Storage and networking names that rarely reach the headlines, Kioxia and Marvell among them, were repriced within days as the market extended the bet across the whole AI supply chain. But such scarcity invites alternatives. Enter Cerebras chief executive Andrew Feldman, announcing the company now works with every major hardware maker except Nvidia, a pitch aimed at buyers trying to escape a single point of dependence.
Meanwhile, Broadcom, whose AI chip revenue rose 143 per cent, still lost more than $300bn of market value in a day when its guidance “only” met expectations, a sign of how much perfection the sector’s prices now assume.
For anyone planning capacity, the key question is whether the chips can be secured at all.
Local heroes
At Computex in Taipei this week, Qualcomm chief executive Cristiano Amon in his keynote estimated that in 2026, global token demand will reach 31.7 billion tokens… every 10 seconds, projecting a 40x increase to 1.27 trillion tokens… for the same 10 seconds, because AI agents are going to consume a lot of tokens. That trajectory puts total demand into what Mr Amon jokingly called “four gazillion” tokens by 2030, confirming tokens as “the currency for AI”. He reassured the audience however that distributing work intelligently between device and cloud can achieve the same task with 30 per cent fewer tokens and a quarter of the cost.
On Monday, Nvidia chief executive Jensen Huang had unveiled RTX Spark, a laptop chip with enough memory to run a 120-billion-parameter model on the device itself, with no connection required. Two days later Google DeepMind released Gemma 4 12B, a model that reads text, images, audio and video and runs on a laptop with 16GB of memory, under a licence that lets a company use it commercially at no cost. At the same show, Perplexity demonstrated a system that decides, task by task, what stays on the device and what travels to the cloud, keeping sensitive material local and sending only the heavy reasoning out.
Lines on the map
China’s new trade-secret rules, in force from June 1, now treat algorithms, source code and training data as protectable secrets, and reach beyond its borders to do so.
Washington went the other way, with President Donald Trump signing an executive order on June 2, prompted in part by the cyber capabilities of Anthropic’s unreleased Mythos model, that invites labs to “voluntarily” submit frontier models for a 30-day safety review and practically “orders” them to do nothing. OpenAI wanted it firmer: in a policy paper it pressed for the reviews to be mandatory and run by a civilian testing agency rather than the security services, while still resisting any rule that would force it to clear a model with the government before release.
Europe is not standing still. At the Choose France summit on May 30, SoftBank pledged up to €75bn to build 5 gigawatts of AI capacity in the country; Mistral, the Paris frontier lab, spent the week signing Airbus and BMW to sovereign-AI deals while beginning to court American customers.
A note to our readers
The Barcelona pilot session of Follow the Token was the first of several planned iterations. We will be in Paris the week of June 15, for a private session alongside VivaTech, with a Barcelona executive briefing planned later in the month (seats are limited, register through the link) and an online edition in the second half of July. We are also building a handful of longer formats: a half-day executive brief that gives all eight topics the time the 60-minute version can only point at, and sector-specific masterclasses for groups or teams that want the economics applied to their own industry.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours.
While the Follow the Token session was filled with friends and neighbours, it was actually the series Friends and Neighbours which led us to this week’s soundtrack with the discovery of the extremely talented actress and singer Lena Hall. Another TV series we have been relentlessly binge-watching, the very witty Rooster brought back Michael Stipe to our ears with I Played The Fool and unleashed a torrent of R.E.M. nostalgia. The closing quiz ran on Kvistly, a platform built by a friend who came along that afternoon, which I should flag as the friendly conflict it is; and the neighbour who won it went home with a fresh lavender plant.