Wenfeng Liang shorts Anthropic

"Kill!" The sky-high valuations in AI rest on two kinds of maxxing: Tokenmaxxing and Benchmaxxing.

Tokenmaxxing means everyone mindlessly uses top-tier large models — and it has to be the latest Claude model. The market narrative creates an atmosphere where any company not giving employees unlimited API access is a dinosaur falling behind the times.

Yet Tokenmaxxing is fundamentally unsustainable. Even the world's largest companies can't afford API calls to Claude at scale.

This trend only had its moment for two or three months in the first half of this year. What we've seen since: major internet companies in both China and the US imposing quotas and repeatedly slashing them.

Once that two-month mania passed, even the people who kept saying "believe, love, be awed, be paralyzed" could figure out that demanding every company on earth — big and small — take out loans to fund Anthropic was impossible. Even if Dario ran the Catholic Church, this couldn't happen.

Medieval churches only collected tithes. Dario, packaging himself as an AGI cherub, fantasized about taxing through 2030 and beyond.

Without the religious fervor of Tokenmaxxing, Anthropic's unsustainable growth rate naturally gave way. Missing ARR targets was inevitable.

The other pillar of the AI bubble is Benchmaxxing.

The irrational exuberance around large models is built on benchmark scores. Every new model release, the labs somehow find a way to break the leaderboard, creating visual effects that shock and paralyze you like watching an atomic bomb explode. Leaderboard scores are like Pinduoduo's "chop a buck" — every time you see near-perfect scores, the labs and the ranking merchants cook up new methods for the next round of dramatic improvement, pushing ever closer to 100%.

But the mainstream benchmarks AI practitioners follow are severely disconnected from the masses, severely disconnected from productive society. This is classic common sense: once large models surpass human programmers in coding ability, the marginal returns from score-chasing diminish.

This is why we advocate for more real-world task benchmarks to evaluate models, replacing a single bloated leaderboard.

Large models are products. Products must satisfy user needs. Their capabilities should be demonstrated through real tasks and environments. Talking about how many points Claude leads GPT by is meaningless scoreboard theatrics.

Even if Opus 5.5 improved several-fold over Opus 4.8's benchmark scores, Anthropic's revenue growth won't multiply several-fold. American tech giants, model labs, and NVIDIA have long been chained together in an ironclad alliance — circular contracts, circular loans, accounting pushed out to 2045.

Anthropic alone has procurement and lease commitments of roughly $895.4 billion, with about $447.5 billion due before end of 2030, and the bulk of the remainder scheduled between 2031 and 2045.

A recent calculation struck me deeply.

OpenCode is a very popular coding agent. Its co-founder Dax tested and found that a single GB300 rack can process 6.82 trillion DeepSeek v4.1 Flash tokens per day. That's close to OpenRouter's entire daily DeepSeek usage.

Extrapolating linearly: if the whole world used only DeepSeek v4.1 flash, just 500 GB300 racks would satisfy inference demand — total value around $3 billion.

Of course this is unrealistic. Not everyone would use only DeepSeek v4.1 flash. People at model labs must use Claude. A minority of complex tasks do need top-tier models.

But Claude is only needed for a tiny fraction of demand. The vast majority of tasks at internet companies can be fully handled by DeepSeek V4.1 flash.

This is a demand-tiering problem. Claude serves the tiny peak of the pyramid. The vast base must be served by faster, better, cheaper models — and DeepSeek v4.1 flash occupies that ecological niche perfectly.

For ordinary people, AGI has essentially already arrived.

The Doubao app satisfies my parents' daily needs completely. Until AI applications innovate dramatically enough for ordinary people to feel an agent leap forward, my parents obviously have no need for office agents, personal agents, or other more token-hungry applications.

The same holds overseas. The vast majority of people have their AI needs met by ChatGPT. Most don't even pay the $20 membership — they use the free version or share family accounts.

OpenAI's February numbers: ChatGPT exceeded 900 million weekly active users, with over 50 million consumer subscribers — a payment rate below 6%.

Of course I know demand is continuously defined. With better AI applications, like super agents becoming widespread, people's needs will gradually increase.

But this logic presupposes major innovation in AI applications, and mass adoption of those new applications — clearly not something achievable within a year.

Computers, electricity, and steam engines each took decades to centuries to proliferate. Mobile Internet took ten years. Even with AI dramatically accelerating, agent普及 isn't a one- or two-year proposition.

So the AI bubble is entirely an expectations management problem.

American capitalists have united to compress an AI revolution needing ten to thirty years of普及 into two to three years of valuation multiplying tens of thousands of times, immediately dividing gains that should be released slowly across decades.

From primary to secondary markets is a massive threshold. Primary markets mean a few large companies footing the bill. Secondary markets require the world's largest index funds, pension funds, and global retail investors to take the other side.

In crossing this chasm, bubble bursts are inevitable. Because while Wall Street is greedy, it fears AI founders who are even greedier.

This year US stocks already absorbed a $177 billion SpaceX IPO, and are preparing to absorb two more trillion-dollar offerings from Anthropic and OpenAI. Fearful Wall Street capitalists even rated Anthropic as junk-grade, institutionally restricting some large funds from participating in its debt financing.

Wall Street's fear, translated into financial terminology: Anthropic's compute procurement and leasing spending is too massive, mismatched with revenue growth expectations. Once growth underperforms, these bills become cash flow pressure.

Before frenzied AI bros, even financial animals look more human.

I fully believe AI is a technological revolution as important as electricity. But technological revolutions unfold across decades, requiring models and applications to advance together — especially applications pulling models forward.

We're not saying don't do tech bubbles — do them slowly, gradually, rhythmically. Don't swarm in and stampede out while terrorizing the masses.

The current situation: coding agents led by Claude Code drove a leap in model coding capabilities, achieving localized Tokenmaxxing on the programming track.

But coding agents and their derivative office, personal, and other细分场景 haven't yet produced a next-generation agent awesome enough for broader populations.

Super agent hasn't arrived, let alone普及. Model labs' revenue growth has no engine. In the short term, compute spending becomes a castle in the air.

This article's conclusion is simple:

One: AGI for ordinary people is already here. The vast majority don't need top-tier models.

Two: The pendulum swings back to AI applications. Super agents are needed to drive the next round of model advancement.

Three: Wenfeng Liang is shorting Anthropic like Pinduoduo shorted Alibaba — pure revenge arc.

⬇️

Subscribe to our Substack: funeralai.substack.com