Anthropic's "Forced IPO": It achieves a revenue of 16 billion, incurs a loss of 80 billion, and is left with only 20 billion in cash.

At the AI gaming table in Silicon Valley, the most terrifying outcome is often not the failure of technical routes, but the depletion of cash flow.

Recently, a set of financial data about Anthropic has sent shockwaves through the venture capital circle: its revenue in the first half of 2026 was approximately 163 billion US dollars, with a net loss of around 807 billion US dollars, and cash and short-term investments as of the end of 2025 stood at roughly 202.8 billion US dollars. Harsh external judgments soon followed: the push for an IPO at this moment is not a coronation at a peak moment, but a last-ditch survival move; in the sluggish capital market environment, an IPO of this scale carries extremely high risks. Some even draw a final scenario deduction that Anthropic and OpenAI may only be cannon fodder in the transition phase, and the only players that will remain at the table in the end are Gemini (Google) with a super ecosystem and DeepSeek that achieves extreme cost reduction.

This line of argument has sensational elements, but it has torn open the most hidden wound of the large model industry: computing power means power, and cash flow is the lifeline. However, before embracing the "crisis narrative", we must first clarify the data calibers - otherwise the sense of crisis will be distorted.

Disassembling the financial black hole: How much of the 807 billion net loss is real cash outflow?

First of all, we need to examine the staggering "807 billion USD net loss" through business logic.

Of this 807 billion, approximately 720 billion USD comes from the fair value change of financing instruments such as convertible notes, which is a non-cash accounting provision and does not represent actual cash outflow. After excluding this item, the company still recorded a book loss of about 87 billion USD in the first half of the year, which includes non-cash expenses such as equity incentives; the real cash consumption should be checked against the cash flow statement, rather than only looking at the net loss on the income statement. Therefore, the headline figure of "losing 800 billion in half a year" is extremely impactful, but it is not synonymous with "burning 800 billion USD in cash in half a year".

The real pressure is hidden in two other figures: revenue of about 163 billion, and cash and short-term investments of 202.8 billion - the latter is a point-in-time data as of December 31, 2025, not the balance for the first half of 2026.

Large models are a "heavy industry" that consumes an extreme amount of computing power. The disclosed scale of rigid computing power commitments by Anthropic is staggering: cloud computing contracts total about 4500 billion USD (about 1110 billion USD from Google Cloud, about 1100 billion USD from AWS, and about 310 billion USD from Microsoft), the maximum lease of NVIDIA chips is about 860 billion USD, and the related Broadcom chip obligations exceed 1600 billion USD, adding up to more than 6000 billion USD; if pending agreements are included, the potential scale could reach 7500 billion USD. In the first half of 2026, its computing power and infrastructure expenditure was close to 90 billion USD, with an annualized figure of about 180 billion USD. For this single expenditure item, the 202.8 billion USD in cash can only cover a little more than one year of operation.

But this is a static caliber. Anthropic has disclosed to investors in the third quarter of 2026 that its "adjusted free cash flow has turned positive", although the definition of "adjusted" is still opaque and its sustainability is in doubt. Ignoring this marginal change and directly comparing the cash balance at the end of 2025 with the cash burn rate in the first half of 2026 will lead to a conclusion that is more pessimistic than the facts.

The urgency of the IPO is supported by public information: Anthropic confidentially submitted its S-1 filing in June 2026, targeting a listing in November; on the prediction market Kalshi, the probability of it officially announcing an IPO within the year is about 75%. A more accurate statement is that the narrowing of the primary market financing window and the need for early investors and employees to exit channels have made the IPO change from "optional" to "priority". "Forced" is a judgment, not a statement that can be directly verified by facts.

Going public as "running naked"? Why is the IPO at this moment extremely risky

Against the current macro environment of the US stock market, it is indeed very difficult to go to the public market to raise money with a "blood-loss" financial report.

In the past two years, as long as a company had large models, top scientists and impressive benchmark scores, Wall Street was willing to pay for the grand narrative of AGI. But the tide has turned. In the recent correction of tech stocks, the high leverage debt problem exposed by computing power leasing companies such as CoreWeave is an important trigger: its debt is about 520 billion USD, the 5-year CDS spread once exceeded 1000 basis points, implying a default probability of over 60%, and the cumulative negative free cash flow from 2026 to 2028 is expected to reach 710 billion USD. Investors have begun to re-examine the return cycle of AI infrastructure.

Microsoft lowered its capital expenditure forecast for the 2026 calendar year from about 1900 billion USD to about 1750 billion USD, partly due to accounting caliber adjustments; Meta cut about 8000 jobs in the first half of the year, but the absolute value of capital expenditure is still rising. Both the official and Goldman Sachs defined this as "optimizing the portfolio" rather than a retreat; Apple's AI-related capital expenditure accounts for about 2.5% of its revenue, far lower than cloud vendors. In July 2026, its market value once hit 5 trillion USD, and the market has indeed re-priced its asset-light AI strategy as a "risk-aversion" attribute.

Under this investor sentiment, when Anthropic goes to the public market with a financial report showing "a net loss of 807 billion in the first half of the year", it will face severe downward pressure on valuation, and the risk of issuance failure cannot be ruled out. When retail and institutional investors are no longer willing to pay a high premium for the "black box API", the IPO is essentially a large-scale stress test that tests the bottom line of the market.

Final scenario deduction: Why are Gemini and DeepSeek favored?

The final scenario deduced by the outside world is: Anthropic and OpenAI may fall into a fierce battle or even fall behind, and the market will eventually be dominated by Gemini (under Google) and DeepSeek (under High-Flyer Quant). Its business logic is: the large model industry is an infinite game that tests endurance, and the surviving players can be divided into two categories - the "conglomerates" with super cash cows, and the "geeks" with extreme cost reduction capabilities.

Google has the world's most profitable search advertising business, self-developed TPUs, huge cloud infrastructure, as well as the Android and search ecosystems. Its relative advantage is the combination of "self-owned chips + self-owned cloud + self-owned cash cow", and its marginal computing power cost is lower than that of independent model companies that purchase NVIDIA GPUs externally. However, a popular statement needs to be corrected: Google is not "free from the need to purchase computing power". The R&D, manufacturing and deployment of TPUs also require huge capital expenditure; and Anthropic itself is also a large-scale user of TPUs, with a commitment to purchase more than 1 million units, some of which are purchased directly and some of which are rented through GCP. To be more precise, Google's advantage is not "no computing power cost", but that its computing power cost structure and cash flow resilience are far better than independent model companies. As long as it does not make strategic mistakes, it can indeed rely on cash flow to outlast its competitors.

DeepSeek has taken an extreme engineering cost reduction route. Its R1 model claims a training cost of about 5.6 million USD using 2000 H800 chips, which compares to the cost structure of comparable Western models that ranges from 80 million to 100 million USD and uses 16000 H100 chips; the MoE architecture reduces computing costs by about 30%; the inference framework DeepInfer was fully open-sourced in October 2026. It does not rely on high-priced API subscriptions to support a large sales team, but relies on the reputation of "good to use and low cost" to penetrate the developer ecosystem. The existence of DeepSeek proves that under the background of limited computing power, extreme engineering efficiency can significantly close the hardware generation gap. Of course, this does not mean that it has locked in the final victory, but only represents a survival strategy that is more difficult to be crushed by cash flow pressure.

The dilemma of the intermediate state: The structural pressure on OpenAI and Anthropic

In contrast, Anthropic and OpenAI are in an awkward intermediate state: they do not have the abundant self-owned funds and underlying hardware like Google, but they have to maintain huge labor costs and computing power expenses.

OpenAI recorded a net loss of about 385 billion USD in 2025, paying 172 billion USD in related fees to Microsoft, and is highly dependent on Microsoft's capital injection; the urgency of its IPO push has been reflected in reports, but describing it as a "cashing out crisis for internal personnel" lacks direct evidence and is only an inference. Anthropic is stuck in the middle: it has no infrastructure, no extreme lightweight cost control, and is burdened with heavy cloud bills and R&D teams.

In order to maintain high valuations, they must continuously launch more powerful models; more powerful models require more massive computing power, which in turn leads to even larger losses. This is a real structural pressure, but calling it a "unstoppable death spiral" is a rhetorical statement rather than a proven financial trajectory. The "adjusted profitability and positive free cash flow" disclosed by Anthropic in Q3 at least shows that under specific calibers, scale effects may be generating marginal improvements, although its sustainability and the definition of "adjusted" are still in question. Once capital stops injecting, or technological iteration stagnates, huge sunk costs may indeed crush the company - this is a risk, not a foregone conclusion.

Conclusion: The fall from the AI altar and the return of business rationality

Anthropic's financial dilemma is the epitome of the "high leverage, heavy asset" model of the entire large model industry.

In the past two years, as long as a company had top scientists and models with high benchmark scores, it could get sky-high financing. But now, the patience of the capital market is running out. From "choking for dreams" to "worrying about bills", large model companies are going through a cruel coming-of-age ceremony. When huge computing power bills eat up revenue, and public market investors no longer unconditionally pay for the grand narrative of AGI, only those players that have self-owned cash cows (such as Google) or extreme engineering cost reduction capabilities (such as DeepSeek) are more likely to survive into the next era.

Anthropic's IPO may become the first loud bang of the bursting of the large model bubble, or it may just be a node for the industry to return to rational business logic - this judgment itself still needs time to verify. On this battlefield full of scorched earth from computing power consumption, surviving is more important than anything else.

This article is from the WeChat official account "Tech Business", written by Tech Business, and authorized for release by 36Kr.