November 18, 2026 · 小苏

Anthropic's S-1 Reveals AI's Most Absurd Business: Models Rot in 6 Months, But GPUs Depreciate Over 8 Years

Anthropic's S-1 Reveals AI's Most Absurd Business: Models Rot in 6 Months, But GPUs Depreciate Over 8 Years

By the time you read this, Anthropic may have already completed its IPO. But after grinding through all 300-plus pages of its S-1, I was genuinely laughing out loud at the numbers. A high-tech money printer? Please. This is the most absurd pump-and-dump scheme in the history of capital markets.

An LLM that cost hundreds of millions of dollars to build has a shelf life of six months. The servers running it depreciate over six years. And Jensen Huang stood on stage and swore up and down that GPUs will run reliably for a full eight years. A disposable consumable you throw out every six months, wrapped in a physical asset depreciated over eight years — jam those three numbers together and the entire business logic shatters on impact. So who's actually footing the bill for this party? And who's treating every investor on earth like a mark?

01 The Three-Number Financial Trick: Six-Month Consumables, Eight-Year Depreciation Theater

The S-1 says it in black and white: a full 80% of Anthropic's revenue comes from models released in the past six months. Which means after six months, the old model is industrial waste, and its revenue share drops below 20%. These models have a shorter shelf life than fresh strawberries — six months and they're moldy and rotten through.

And on the other side? The cloud giants dutifully depreciate over six years, while Jensen brags everywhere that GPUs last eight years and calls them the highest-quality fixed asset around. Doesn't that strike you as utterly absurd? If a company bought a batch of cabbages with a six-month expiration date but took out an eight-year loan for a heavy-duty cold storage facility, how many days do you think that company survives?

Six months ago everyone was saying data centers depreciate over three years. Now, to dress up the financials, they've conjured an eight-year fairy tale out of thin air — five extra years of value manufactured from nothing. And even if the chips could physically hold out for eight years, by year four what the hell is an old-architecture GPU good for besides heating the server room? It's like an e-commerce company selling cabbages and discovering that cardboard packaging is 50% of its costs — are you in the vegetable business, or are you an unpaid intern for the box factory? Short-lived algorithms bolted onto heavy hardware: underneath it all, this is a deformed business that simply cannot be profitable.

02 So Who Actually Wins? The Karaoke Bar on the Ground Floor Is Now Open

A couple of nights ago it hit me: the loudest karaoke bar on the ground floor of the building is about to ring the opening bell, and every big shot in the tower has shown up to cut the ribbon, grinning and shouting "congratulations!"

But whose abacus is clicking loudest? The landlord upstairs, obviously. Amazon, Microsoft, Google — the cloud giants talk a good game about supporting tech, but behind the scenes their beads are already flying off the frame. Karaoke bar packed to the rafters? Great, cram more people through the door! Circuit breaker tripped? The landlord immediately calls the utility company with fake concern: "Emergency capacity expansion, upgrade the whole building's grid!" Grid upgrades, server room renovations, sky-high rent — all deducted from the tens of billions the karaoke bar raised.

Isn't this the classic restaurant-industry survival drama? Restaurants have a 70% annual failure rate, so who actually gets rich on that street? The trendy spots that fold in a few months, or the guy flipping secondhand equipment? The flipper, obviously! You can't make it and shut down, he buys your stainless steel range at a discount, refurbishes it, and sells it at full price to the next startup rube. Today's shovel-selling hardware vendors and compute-renting cloud giants are exactly the secondhand dealers on the AI street. Trillions of dollars pass through the LLM companies' books, evaporate into thin air, and land right back in America's power, water, and infrastructure buildout.

03 The Donkey Gets Slaughtered, and Nobody Can Move: The Village's Second Hope Goes Off to Take the Imperial Exam

The way I see it, Anthropic is at best the "village's second hope" that the giants trotted out. The eldest son, OpenAI, is a born rebel who dreams of AGI all day long and won't take orders. So they turn to son number two: "Good boy, the whole village is pooling your travel money — get to the capital, ace the imperial exam, marry into the royal family, and come back rich to divide up the land!"

Google and Amazon dropped billions, then immediately locked it down with exclusivity agreements: every dollar raised must go straight back into buying my cloud services and my custom silicon. What kind of investment is that? It's a blatant left-hand-to-right-hand revenue pump. But can the second son really carry those expectations? Look at Google's recent bloodbath — down double digits in a single day. Is the market blind? The model lab you invested in is devouring cash flow, while your own lifeline search business is being cannibalized faster by AI, and the thin margins on new business can't begin to fill the crater left by the collapse of the old world.

Even more damning: Anthropic brags that half the Fortune 500 uses it, but flip open the S-1 and that spend accounts for just 0.1% of those giants' IT budgets. That's like a tycoon worth tens of millions pulling a dime out of his pocket to buy a plastic bag while telling you you saved the world. Corporate bosses know exactly what they're doing: put AI on a pedestal when it's time to blame layoffs, but when it's time to actually open the wallet, even a pack of napkins feels too expensive. How big a business empire can that kind of lukewarm relationship really support?

04 Stripping Off AI's Mask: It's Not Software, It's a Grinding Little Coding Shop

Why is Anthropic rushing to IPO first? Don't let the mystical PR copy fool you. The truth is brutally plain: because it's cheap and it's focused. It doesn't touch expensive video generation, doesn't chase flashy multimodality — it pours every ounce of compute into one thing: writing code.

The S-1 even hides a line item that's both hilarious and sad: penalties withheld from banned accounts account for 5% to 7% of total revenue. What developer hasn't been punished by its neurotic rules? And even after grinding their teeth to powder, the banned just pinch their noses and go make a new account. But strip away the aura and isn't this essentially a labor-intensive "coding outsourcing shop"? Can a glorified autocomplete for programmers really prop up a $2 trillion fortune?

Reality already slapped it across the face. Liang Wenfeng and DeepSeek have long since flanked it from the side. On the OpenRouter leaderboard, DeepSeek's monthly call volume has hit 50T, while the entire Claude family combined manages a measly 7T. Even if Claude earns a bit more toll revenue thanks to its premium pricing, how many more months can that markup survive the dimensional strike of cheap open source? The intelligence gap between algorithms is being flattened fast. Once people realize they can eat their fill at street-stall prices, who's going to keep showing up at the fancy restaurant to get ripped off?

05 The Schedule-Hedging Master's Abacus: The Ones Who Understand Revolution Best Are Always the Royalists

Look back now and you have to admire Altman's cunning. Why is OpenAI dragging its feet on going public? Altman has already done the math: let Anthropic go first. Let the second son be the one who walks the minefield.

You think Altman doesn't understand business? He knows better than anyone that models simply aren't profitable in the long run. The ones who understand revolution best are always the royalists. The model race has already become a bayonet-charge red ocean — once the open-source army gets going, global compute profits get punched straight through. OpenAI is building hardware, fantasizing about being the AI TikTok, burning cash at a terrifying rate. If it went public now, the public-market earnings reports would instantly squeeze out all its bloated froth.

So OpenAI has simply become Silicon Valley's most devious "schedule-hedging master." The moment Anthropic announces a major breakthrough, Altman drops a nuke on the back end — a price cut, a rate-limit reset, an open-source release — strangling his rival's exposure at the throat. This is straight out of Cao Gui's battle doctrine: "when they're exhausted, we're full." You go take the bullets for the village, you go sweep for mines; I'll keep harvesting hot money in the private markets and pull up a stool to watch you perform this absurdist drama under the spotlight.

Running this AI party to this point, the most fascinating part is the final distribution of spoils. America used the AI myth to force-feed stock-market liquidity into power grids, chips, and data centers, completing a massive upgrade of hard assets. The models themselves are nothing but the fastest-burned fuel and bridge financing in that infrastructure sprint. The capital ship has already docked safely, cheers ring out to the heavens, and only the short-lived models are left flailing in the mud as the tide goes out.

It's like when we act fancy and talk about ordering imported liquor at a high-end club, but at 2 a.m. what really soothes the soul is a few sizzling skewers at a roadside stall and a cheap bottle of beer. Watching the big shots raise their glasses to a $2 trillion compute bubble, I just want to quietly ask one thing: my account isn't going to get banned again tomorrow, is it?

The ones who understand revolution best are always the royalists.


[Subscribe to the JustSayAI Daily Brief] 📮JustSayAI Daily Brief · Subscriber Perks Two editions a day · Listen with one tap! Now fully available via email + IM + Skill.

← View all articles →