Anthropic 被曝 300 亿美元融资,投前估值超 (English)

Generated: 2026-06-21 03:15:50

---

90 Billion In, and All I Feel Is "Meh"? That Numbness Is the Real Horror

Hey, let me tell you something.

The other day I saw the news about Anthropic's fundraising round — 【this round, including a partial tender offer, pushed the valuation to 90 billion dollars】. Guess what my first thought was?

Not "holy shit, that's expensive."

No. It was… "Oh, 90 billion. Okay."

Think about it. Three years ago, when I was writing about AI company valuations, I was still agonizing over whether something was really worth 10 billion. Last year, when OpenAI hit 300 billion, I figured the world had just lost its mind. Now 90 billion lands in front of me, and my reaction is basically "fine."

That numbness is scarier than the bubble itself.

Because it means we've already accepted that this is just how the world works now.

---

Problem One: 90 Billion? Traditional Valuation Models Would Have a Breakdown

Let's do the math.

Public data says Anthropic's annualized revenue as of May this year is around 【3 billion】 dollars (note: that's their "gross" reporting—they're counting the cut taken by AWS, Google Cloud for distribution. Net-net, what they actually pocket is closer to 【2.2 billion】 ). Even if we take the most optimistic estimate of 【4.5 billion】, what's the price-to-sales ratio on a 90 billion valuation?

At minimum 20x, at maximum 40x.

You know what the price-to-sales ratio was for the most expensive SaaS companies at the peak of the bubble? 20 to 30 times.

So what gives Anthropic the right?

The answer is brutally simple—investors aren't buying today's revenue. They're buying a ticket to a future monopoly.

When I tested 【Claude 3 Opus】 back in February this year, I was genuinely blown away. Gave it a 2000-line Python refactoring task, and it nailed it on the first pass. GPT-4? Took three tries before it even ran.

But here's the thing—its token consumption was 【nearly double】 what I expected.

After GitHub Copilot integrated the latest model, the premium request multiplier went from 7.5x to 15x. I ran a few medium-sized projects and my bill jumped from $200 a month to almost $600.

Anthropic's strategy is clear: retail users aren't the target—go use the free tier.

They're going after the kind of companies that burn 【hundreds of millions of dollars】 a year on tokens—Uber, Netflix, Salesforce.

Marc Benioff, Salesforce's CEO, has publicly said that by 2026, just their company alone will consume 【tens of millions of dollars】 worth of Anthropic tokens. Imagine a hundred companies like that—it's 【billions】.

Looks great on paper, right?

But there's a giant pitfall.

---

Problem Two: How Much of That Revenue Is Just "Left Pocket, Right Pocket"

This one pisses me off.

Anthropic reports annualized revenue of 【3 billion】 using the gross method. If you calculate it net, what they actually get to keep is only around 【2.2 billion】 . Meanwhile, OpenAI has always been reporting net.

What's even more ridiculous? How much of that 【over 2 billion】 actually comes out of the pockets of their own sugar daddies?

Amazon invested in them, uses AWS to distribute for them, and then buys a huge pile of tokens themselves. Google plays the same game. I dug into a few sources, and the scale of this "capital recycling" is at least 【500 to 800 million dollars】 —the giants are shuffling money from their left pocket to their right, artificially inflating those revenue numbers.

I'm not making this up. UBS's analysis report hinted at it too, saying something like "investors need to closely examine the composition of revenue sources."

The core definition of a bubble isn't "is there revenue?" It's "can the capital chain close?"

The four giants—Amazon, Microsoft, Google, Meta—are looking at combined CapEx of 【250 to 300 billion dollars】 for 2026 alone, with year-over-year growth potentially over 50%. Goldman Sachs' model is even more insane, predicting total industry AI CapEx from 2026 to 2031 could be 【trillions】 .

Where's that money coming from? Will it be earned back through enterprise token purchases? Or will it require the capital markets to keep pumping blood?

---

Problem Three: 2028? That Ticking Clock in Silicon Valley

Speaking of which, there's a date Silicon Valley keeps muttering under its breath—2028.

The most famous prediction comes from 【an industry report from early 2026】 . Its core argument: by 2028, AI will be everywhere, but it won't have boosted total factor productivity. Instead, it'll trigger inflation from over-investment, plus rising unemployment from job replacement.

Sounds like fear-mongering?

Here's a number for you: OpenAI's own projections to investors showed cumulative net losses from 2023 to 2028 at 【potentially over 20 billion dollars】 . But actual expansion has been so aggressive that by 2026, the single-year loss could hit 【8 billion】, and widen further in 2027.

They expect to break even on cash flow by 2030—and that's assuming CapEx stays under 【100 billion】 . Now they're actually spending at a 【300 billion】 level, and you still expect profitability by 2030?

I've looked at the paying conversion rate for weekly active users on OpenAI. Honestly, it's low. Most people are just leeching off the free tier, occasionally upgrading to Plus. Anthropic is a bit better because they focus on enterprise clients, and companies are more willing to open their wallets.

But the core issue is—being willing to pay doesn't mean the math adds up.

I recently helped a friend's company with an AI procurement evaluation. A mid-sized e-commerce company used Claude for customer service automation. Results were genuinely good—they reduced the team by 40%. But after crunching the numbers: annual labor savings 【1.2 million】, annual token costs paid to Anthropic 【850,000】 —net savings of 350k.

Sounds pretty good, right? But think about it: to earn that 【850,000】 , what's Anthropic's underlying compute cost, R&D cost, sales cost? If they're burning 【billions】 of dollars a year, then that "customer value" of 350k can't even come close to covering the burn rate.

That's what that report meant: **the gap between "business value" and "cost