[KongchangAI]
· 2 min read· 1,469 words

Anthropic's Sky-High IPO: $80.7 Billion Loss in Half a Year—Why Is It Worth $2 Trillion?

Anthropic's Sky-High IPO: $80.7 Billion Loss in Half a Year—Why Is It Worth $2 Trillion?

Anthropic's $80.7B half-year loss is actually an accounting treatment; operations are healthy, but the $800B compute bill and short-lived model advantage make it a huge risk bet.

Anthropic's $80.7 billion half-year loss includes $72.4 billion from accounting for early investor conversion vouchers—not a cash loss. Actual operating data is strong: half-year revenue of $16.3 billion, 14 times year-over-year; Q2 net profit of $970 million; $131.7 billion in cash reserves. Three key numbers reveal its business model: 2.08 (every $1 of compute cost earns $2.08, versus only 0.4 last year), proving it can now self-fund; 80% (over 80% of revenue from models released within half a year), showing an extremely short lead window and the need for continuous high-intensity iteration; and $800 billion (locked-in compute bills through 2045), meaning rigid costs against floating revenue—a huge risk. The IPO is essentially a bet: if AI demand keeps growing, Anthropic will become a money printer; otherwise it will carry long-term bills.

Introduction: The Truth Behind the Loss Numbers

A company lost $80.7 billion in half a year—equivalent to wiping out the entire market value of Nvidia in six months—and yet it's about to go public with an opening valuation of $2 trillion. This sounds like a cash-burning scam, but the truth is exactly the opposite: it hasn't actually lost a single real dollar. What's truly terrifying is the speed at which it makes money. The protagonist here is Anthropic.

To understand this sky-high IPO, you can't just look at the loss headline. You need to break down three key numbers: 2.08, 80%, and $800 billion. They correspond to the company's profitability, revenue structure, and risk exposure, respectively.

131.7 billion in cash on the books

The Accounting Truth Behind the $80.7 Billion Loss

Of the $80.7 billion loss, $72.4 billion comes from a single accounting treatment: early investors didn't buy shares with their money, but rather a conversion voucher that agreed to convert into shares at the low price of that year upon IPO. Because the company's valuation more than doubled in half a year, this voucher was recorded as a massive loss on the books. But this is just a numbers game—not a single cent is missing from the company's pocket.

The actual operating data is much healthier: half-year revenue of $16.3 billion, 14 times the same period last year; net profit of $970 million in the second quarter alone, achieving profitability for the first time; $131.7 billion in cash on the books, equivalent to $127 million coming in per day. The money it makes in one day is equivalent to 3% of its entire 2023 revenue. So don't be fooled by the "$80.7 billion loss" headline. The real question is: can this business keep going this crazy?

The accounting concept involved here is convertible preferred stock or similar convertible instruments. What early investors purchased at a lower valuation was not common stock, but a financial instrument that could convert into common stock at an agreed price upon a specific future event (such as an IPO). Under accounting standards (such as ASU 2017-11 under US GAAP or IFRS), when a company's valuation rises, the fair value of these instruments increases, and the incremental portion must be recognized as a loss in the income statement (or a loss on fair value changes of liabilities), even though the company has no actual cash outflow. This treatment is common in the financial statements of high-growth tech companies before going public, aiming to reflect the economic cost of potential equity dilution. Therefore, $72.4 billion of the $80.7 billion loss is a non-cash, non-operating book adjustment that does not affect the company's cash flow or actual operating health. When analyzing such companies, investors should focus on indicators such as operating cash flow, revenue growth, and gross margin, rather than being misled by accounting losses.

Number One: 2.08—Every $1 Spent on Compute Earns Back $2.08

To judge whether Anthropic's business model holds up, the first key number is 2.08. It means that for every dollar spent on compute, it earns back $2.08. At this time last year, this number was still 0.4—to sell a dollar's worth of goods, you had to front $2.5 on compute alone, and the more you sold, the faster you died. In one year, this ratio has increased more than fivefold.

In the second quarter of this year, revenue grew 1.4% while compute costs grew only 3%, indicating that the money from selling models not only covers compute costs but also earns more than double. How did they do it? On one hand, letting AI improve AI itself makes the same tasks increasingly compute-efficient; on the other hand, using chips from four suppliers—Amazon, Google, Nvidia, and AMD—whichever is cheaper. As soon as the new model dropped in price, usage exploded.

Not only covers compute costs

As long as this number stays above 1, it's a moat; the day it falls back below 1, it's a bottomless pit of cash burning.

"Compute" here refers to the computational resources needed to train and run AI models, typically provided by GPUs (such as Nvidia H100) or dedicated AI chips (such as Google TPU). Compute costs include chip procurement or leasing, data center energy consumption, operations and maintenance, and depreciation. Anthropic's 2.08 ratio is essentially the ratio of revenue to compute costs, similar to a variant of the "gross margin" concept in traditional manufacturing, but more focused on core variable costs. A ratio greater than 1 means that each unit of compute investment generates excess revenue, covering other fixed costs and yielding profit; less than 1 means that the more compute invested, the greater the loss. Anthropic improves this ratio by reducing per-task compute consumption through model self-optimization (AI improving AI) and by lowering unit compute prices through multi-chip supplier price comparison (Amazon Trainium, Google TPU, Nvidia GPU, AMD MI300). This metric is key to assessing the sustainability of AI companies' business models, because compute is usually the largest variable cost item for AI enterprises.

Number Two: 80%—Revenue Relies on Models Released Within Half a Year, and the Lead Window Is Extremely Short

The second key number is 80%. More than 80% of Anthropic's revenue in the most recent three months comes from new models released within the past half year. This means the shelf life of a model's lead is only half a year. Enterprise customers have no sentiment—whichever model is stronger and cheaper, they switch to it the next day. No old product can sit back and collect money.

It is itself the best example: more than 90% of the code in the company's codebase is written by its own AI. So its real moat is not any single model, but the machine that can continually produce stronger models. As long as the machine keeps running, it won't die; the moment the machine stops, 80% of revenue could vanish overnight.

It is itself the best example

Number Three: $800 Billion—Locked-In Compute Bills and Floating Revenue

The third key number is $800 billion. This is the compute bill Anthropic has signed, and it's "take-it-or-pay-it"—whether you use it or not, how much you use, not a cent less; if you don't use enough, you have to make up the difference, with contracts extending as far as 2045.

How desperate is it for compute? In February this year, Musk was still publicly calling its AI evil, but by May it had leased Musk's entire data center in Memphis. Once delivered in December, it would have to pay $3.1 billion a month. And all the cash it has on hand—$131.7 billion—sounds like an astronomical figure, but compared with the $800 billion bill, it's just a drop in the bucket. Costs are locked in, but revenue floats—if demand falls, revenue falls with it, while the bill doesn't shrink by a cent.

Entire data center

"The $800 billion compute bill" refers to the long-term compute procurement commitment (take-or-pay contract) that Anthropic signed with cloud service providers or chip suppliers. Such contracts require the buyer to pay a minimum amount within the agreed period, regardless of whether the resources are actually used, and unused portions still require make-up payments. This arrangement is common in large-scale AI infrastructure investments. To guarantee returns on massive capital expenditures (such as building data centers and purchasing chips), suppliers require customers to make long-term commitments. For Anthropic, this locks in compute costs for many years to come, but revenue fluctuates with market demand, creating a risk structure of "rigid costs, floating revenue." Similar models have appeared in the aviation industry (aircraft leasing commitments) and the energy industry (long-term supply contracts). If AI demand growth falls short of expectations, Anthropic will face enormous financial pressure and may even trigger a liquidity crisis. Therefore, the scale of this bill ($800 billion) far exceeds its cash reserves ($131.7 billion), highlighting the high-risk bet nature of the endeavor.

Putting the Three Numbers Together: A Self-Feeding Moat and an $800 Billion Shackle

2.08 shows that it can really make money now; 80% shows that it must run desperately and cannot stop for a moment; $800 billion shows that it has staked everything. This is a moat that is currently feeding itself, but an $800 billion shackle hangs on its edge.

It is betting on just one thing: demand for AI will only grow. If the bet pays off, it will be a money printer no one can catch; if it loses, it will be a bill that can't be paid off for lifetimes. And the answer isn't today—it's in the three numbers of each coming quarter. For readers following the AI industry, Anthropic's IPO is not just a capital feast, but an excellent sample for observing whether the AI business model can continue to work.

#Anthropic IPO#Anthropic valuation#Anthropic loss#AI business model#compute costs#convertible preferred stock#take-or-pay contract
Share:

Related articles