In Defense of Tokenmaxxing

May 13, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
May 13, 2026

HEADLINES

Google's Big Swing Before I/O: Gemini Intelligence

Google has announced a new agentic suite for Android users called Gemini Intelligence — and the fact that they're dropping this the week before Google I/O should tell you something about how much they still have under wraps. The suite includes a major Gemini assistant upgrade capable of handling multi-step tasks across apps, a Personal Intelligence memory system, and a new Android widgets feature for building lightweight automated tasks from a description. Google also unveiled the Googlebook, a Chromebook redesigned for AI with a Gemini Intelligence stack baked in — including a new mouse-jiggle shortcut to summon the assistant and a DeepMind-demoed AI-enhanced cursor that can act on whatever you point at. The competitive subtext is hard to miss: Google named its suite "Gemini Intelligence" even as it powers the new Siri, and Android ecosystem lead Sameer Samat took some pointed shots at companies "still working on their first iteration" of an AI assistant.

Google and SpaceX in Talks for Orbital Data Centers

Google is the latest company taking a serious look at data centers in space. The Wall Street Journal reports Google is in talks with SpaceX — and other rocket companies — to launch its first prototypes into orbit by next year, as part of an initiative called Project Suncatcher. This isn't a fringe idea anymore: Anthropic flagged orbital compute as a way around land permitting in its recent SpaceX deal, Robinhood co-founder Baiju Bhatt just launched Space Cowboy Corp at a $2B fundraise, and even NVIDIA is posting job ads for orbital data center architects. The structural case is real — terrestrial data centers are running into hard limits on power, land, and politics — but no one has yet put more than a single GPU in space, so the engineering questions remain wide open.

The Labs Go FDE: Google Joins the Consulting Race

Just a day after OpenAI confirmed its consulting plans and a week after Anthropic announced its initiative, Google is jumping on the forward deployed engineering bandwagon — hiring hundreds of FDEs to be housed within Google Cloud's go-to-market team. Google Cloud CEO Thomas Kurian framed the move around surging enterprise demand for agent development help. Google is also reportedly in talks with Blackstone, KKR, and EQT to deploy its AI products across their portfolio companies, taking the fight directly to Anthropic and OpenAI on the private equity front. The AI race has a major new dimension: it's no longer just about model performance, it's about deployment.

Anthropic is attacking its next vertical with the expansion of Claude for Legal. The legal plugin for Cowork launched earlier this year, and Anthropic says legal professionals have now become the most engaged users among all knowledge workers — with lawyers second only to software engineers in usage intensity. This release follows the same playbook as Claude for Finance last week: new connectors for dozens of legal tools including DocuSign, Trellis, and Thomson Reuters CoCounsel, plus 12 pre-built agents for specific practice areas. One notable addition is a direct connection to Harvey, which means Cowork can now serve as an agentic harness for Harvey's legal knowledge base. The broader pattern worth watching is how differently Anthropic and OpenAI are approaching knowledge work — Anthropic with branded, vertical-specific packages of agents and connectors, OpenAI routing everything through the Codex superapp.

MAIN STORY

In Defense of Tokenmaxxing

The tokenmaxxing narrative has reached a crescendo this week — fueled by an FT report on Amazon employees gaming internal AI usage scores, and a viral Slack screenshot about AI spending — and it's being used to resurrect two arguments that looked fairly silly by the end of last year: that AI isn't really that good, and that the whole thing is a bubble. The logic goes: if the token consumption driving AI's meteoric revenue growth is mostly people gaming leaderboards, then the demand signal is fake and the whole edifice comes apart. This is lazy thinking, and it matters — because every company right now is trying to figure out how to drive AI adoption, and cynical takes that reject experimentation out of hand can be actively harmful.

SORRY FOR TOKENMAXXING

How We Got Here
From seats to tokens — and the enterprise adoption problem that followed.
The shift happening right now isn't just a business model change for the frontier labs — moving from selling seats to selling tokens — it's a reflection of a deeper work shift from assisted AI (AI helping you do the things you do) to agentic AI (your job is to set up the conditions for agents to do things). In the agentic paradigm, success isn't measured by how many people have subscriptions; it's about what they're doing with them. That's a major challenge for enterprises, because the capability overhang — the gap between what AI can do and what organizations actually get out of it — is getting nothing but bigger. And the key premise from which everything else follows: there is no way to figure out the best ways to use agents without experimentation. You have to hack and build and see what works.

The Tokenmaxxing Timeline
From Silicon Valley status game to this week's FT report on Amazon.
Kevin Roose at the New York Times first wrote about this back in March, documenting an OpenAI engineer who processed 210 billion tokens in a single week, a Claude Code user with a $150K monthly bill, and companies like Meta and Shopify factoring AI usage into performance reviews. The Information then reported on Meta's internal token leaderboard — covering all 85,000 employees, with top users earning titles like "Session Immortal" or "Token Legend" — which The Information described as Silicon Valley's newest form of conspicuous consumption. Business Insider reported Disney had a similar AI adoption dashboard. Visa gave internal awards for highest usage. This week, the FT added Amazon to the pile, reporting that employees were using an internal tool called MeshClaw to automate non-essential tasks just to inflate their token scores. That landed at the same moment a viral Slack screenshot — almost certainly a joke, but resonant enough to rack up two million views — showed someone being praised for a $600 Anthropic bill while getting dinged for a $23 Uber Eats order.

Why the Critique Is Preposterous
Goodhart's Law is real, but it's not the whole story.
Yes, Goodhart's Law applies: when a measure becomes a target, it ceases to be a good measure. That's obvious and worth acknowledging. But the conversation has gone somewhere much sloppier than that. The logical leap being made is that the gaming behavior reported by the FT is somehow broadly reflective of the majority of token consumption — that most of the demand is people using AI for silly, non-consequential purposes rather than real, valuable work. That leap requires stacking three logical fallacies on top of each other. First, selection bias: tokenmaxxing fraud is a story precisely because it's the deviation. People using AI to create genuine value isn't a story right now. Second, hasty generalization: treating a visible extreme as the norm. Third, category error: using gaming behavior as evidence about the quality of the technology, when the only thing it's evidence of is the incentive structure. Meanwhile, the actual facts about AI remain unchanged — models have done nothing but improve, revenue is growing at a pace without precedent in business history, and demand for tokens radically outstrips supply.

The Affirmative Case
Incentivizing experimentation is R&D translated to the unit level.
Here's what the critics miss. The single biggest documented barrier to enterprise AI adoption from 2023 to 2025 was employees not feeling like they had time to learn it on top of their existing jobs. Incentive structures around experimentation may therefore be a practical necessity. Beyond that, the shift to agentic AI is a fundamentally bigger disruption than the ChatGPT moment. Managing agents is a new knowledge work primitive — not just a new skill, a new primitive — and right now there are no experts, only people who have experimented more than you have. The critics' framework assumes that unless token consumption produces specific, discernible financial value immediately, it's waste. But that leaves zero room for the kind of experimentation that is, right now, the only path to figuring out how to remake your business for this era. To make it personal: a billion tokens used last month, almost none of which led to direct financial gain. Are you really going to argue those were wasted? As for the fakers — companies aren't stupid. Jim from accounting going from zero to a billion tokens in a month isn't going to get a trophy. He's going to get asked to show what he built. This is highly traceable activity. The two dreary, cynical views stacked on top of each other — that non-financially-immediate token use is worthless, and that companies are too inept to spot gaming — deserve to be called out as such. Yes, there are more sophisticated ways to incentivize experimentation; Salesforce's "agentic work units" metric is an example. But the bottom line: companies that incentivize token experimentation, even with some fraud and waste along the way, are going to be light years ahead of the ones that sit it out. Do not be afraid of burning tokens on valuable mistakes.