Tracing the ghost in the ledger, byte by byte.
On June 1, Anthropic quietly submitted its IPO filing. By August, the market was already pricing a potential $2 trillion valuation. The numbers are staggering: an annualized revenue run rate that surged from $14 billion in February to $47 billion in May. A private valuation of $965 billion, up from $380 billion in the same period. Forbes columnist Jim Osman called it a critical juncture. I call it a forensic opportunity.
I have spent the last decade dissecting hype cycles—not just in crypto, but in any market where narrative outpaces fundamentals. The Tezos ICO, the Curve liquidity wars, the Luna collapse, the FTX house of cards. Each time, the pattern is identical: a small set of metrics are inflated to create a self-reinforcing story, while the underlying structural weaknesses are buried in footnotes. Anthropic's IPO is no different. It is a litmus test for whether the AI boom can generate real, distributable profits—or whether it is simply the largest capital-consuming machine ever built.
Let me be clear: I am not shorting AI. I am quantifying the gap between promise and reality. The numbers do not lie. The chain never lies, only the observers do.
Context: The AI Arms Race and the Piper's Bill
Anthropic was founded in 2021 by former OpenAI employees. Its mission: build safe, interpretable AI. Its product: Claude, a large language model that competes with GPT-4 and Gemini. Its growth has been explosive. The company claims an annualized revenue run rate of $47 billion as of May 2025, up from $14 billion in February. That is a 235% increase in three months. If true, it would be the fastest-growing enterprise software company in history.
But revenue run rate is not revenue. It is a forward-looking extrapolation of the most recent month's revenue multiplied by 12. It assumes linear growth, no seasonality, and no churn. In crypto, we call that a vanity metric. It is the equivalent of a DeFi protocol quoting its total value locked (TVL) at a peak snapshot without accounting for impermanent loss or wash trading.
Anthropic has also raised massive capital. In May 2025, it raised $65 billion in a private round. It has committed to spending over $100 billion on Amazon Web Services over the next decade. It has agreements with Amazon for up to 5 GW of new computing power, with Google and Broadcom for another 5 GW of next-generation TPU compute, and it is also using SpaceX's GPU capacity. The company is essentially building its own hyperscale data center empire—on leased infrastructure.
This is the context. A company that is burning cash at an unprecedented rate to acquire compute, while its revenue is still a fraction of its capital commitments. The IPO is not a celebration; it is a necessity. The company needs public market capital to continue funding its negative cash flow.
Core: A Systematic Teardown of the Numbers
Let me start with the revenue run rate. The figure of $47 billion is sourced from anonymous reports and leaked internal documents. I have seen this playbook before. In 2022, I analyzed the Terra/Luna ecosystem and found that the Anchor Protocol's 19% APY was derived entirely from new depositors. The revenue was synthetic. The same principle applies here: if Anthropic's growth is driven by a single customer, or by aggressive discounting, or by a temporary surge in API usage, the run rate is meaningless.

To test this, we need to look at the actual revenue composition. Anthropic does not disclose customer concentration. But based on industry reports, the majority of its revenue comes from enterprise API subscriptions and a partnership with companies like Bridgewater and Salesforce. The question is: how sticky is that revenue? In my 2020 Curve investigation, I built a Python tracker to analyze CRV token emissions against liquidity retention. I found that 40% of the reward tokens were being farmed by flash loan exploiters, not real liquidity providers. The revenue was inflated by short-term incentives. Anthropic's revenue could be similarly inflated by promotional pricing and free credits.
Next, the valuation. A $965 billion private valuation implies a multiple of 20.5x on the $47 billion run rate. That is already high for a company that is not profitable. But the market is now discussing a $2 trillion IPO valuation, which would be 42.5x the same run rate. For comparison, Nvidia—a company with actual profits and a dominant hardware moat—trades at roughly 35x trailing earnings. Anthropic is a software company with no earnings, no intellectual property moat (its models are open-weight to some extent), and significant competitive pressure from OpenAI, Google, Meta, and open-source alternatives.
Impermanent loss is not luck; it is mathematics. The same is true for valuation. Let me run the math.
Assume Anthropic achieves a $2 trillion market cap at IPO. To justify that, it would need to generate roughly $80 billion in annual net income within five years (assuming a 25x P/E ratio). That implies a profit margin of 30% on $267 billion in revenue. But Anthropic's cost structure is brutal. The company has committed to spending over $100 billion on AWS alone over the next decade. That is $10 billion per year in cloud costs, not including the compute from Google, Broadcom, and SpaceX. Add in talent costs (top AI researchers command $10 million+ packages), data center operations, and marketing. The gross margin for AI model inference is thin—often below 50% after compute costs. And that is before amortization of the $65 billion raised in May.
Let me be more precise. In my 2021 retrospective on the Luna collapse, I mapped six months of transaction logs to prove that 92% of Anchor's yield was synthetic. Here, I will map the capital flows. Anthropic raised $65 billion in May. It is spending $10 billion+ per year on AWS. It is building 5 GW of new compute capacity with Amazon and another 5 GW with Google/Broadcom. The capital expenditure required to build that capacity is estimated at $50-100 billion upfront, depending on the efficiency of the chips. The company is essentially a pass-through for capital: it takes money from investors, spends it on compute, and hopes to eventually sell AI services at a margin.
But the margin is being squeezed. The chip manufacturers (Nvidia, Broadcom) and cloud providers (AWS, Google) are the real beneficiaries. They capture the majority of the value. Anthropic is the middleman—a thin layer of software that sits on top of expensive hardware. In the 2023 FTX forensics, I traced $8 billion in customer funds through 400 wallets and found a circular transaction pattern designed to hide insolvency. Anthropic's financials are not circular, but they are linear: money in, compute out, revenue in. The question is whether the revenue can ever exceed the compute cost plus the cost of capital.
Osman noted that much of Anthropic's future success may already be priced in. I would go further: the current valuation assumes that Anthropic will capture a disproportionate share of the AI market, while ignoring the fact that the market is a commodity business. Models are becoming cheaper to train and deploy. Open-source models are catching up. The differentiation is shrinking. Anthropic's safety-first branding is a marketing angle, not a technical moat. In my 2017 Tezos audit, I found three critical logic flaws in the delegation mechanism that the team had overlooked. The same oversight applies here: the market is overlooking the structural flaw in the business model—the lack of pricing power.
Flaws hide in the decimal places. The revenue run rate of $47 billion is reported as an annualized figure. But if we look at the monthly revenue implied by that number, it is about $3.92 billion. That is a massive number for a company that was essentially pre-revenue two years ago. But what is the marginal cost of that revenue? If the cost of compute is 60% of revenue, then Anthropic is spending $2.35 billion per month on compute alone. That leaves $1.57 billion for salaries, R&D, sales, and overhead. With a staff of perhaps 5,000 employees (AI companies are notoriously lean, but Anthropic has been hiring aggressively), the average cost per employee is $314,000 per month—impossible. So either the revenue is inflated, or the costs are deferred. Deferred costs are a liability. They will come due.
I have seen this movie before. In 2025, I analyzed the MiCA compliance reports of the top 20 stablecoin issuers. I found that 60% of them were using opaque reserve structures that violated transparency standards. The gap between declared and actual reserves was significant. For Anthropic, the gap is between declared revenue run rate and actual cash flow. The company is not yet profitable. It may never be profitable at scale. The IPO will force it to disclose its financial statements, and when it does, the market will see the truth.
Contrarian: What the Bulls Got Right
I am not here to dismiss the entire AI thesis. The bulls have a point. Anthropic's growth rate is real in the sense that its customer base is expanding. The model is improving. The enterprise demand for AI is not a fad—it is a structural shift. The partnerships with Amazon, Google, and Broadcom provide a level of infrastructure commitment that most startups cannot replicate. And the company's focus on safety and interpretability gives it a regulatory advantage as governments begin to legislate AI.
Moreover, the market for AI is enormous. If the technology becomes as ubiquitous as electricity or the internet, the revenue potential could be in the trillions. In that scenario, a $2 trillion valuation today is cheap. The bulls would argue that Amazon's $100 billion commitment is a vote of confidence, not a cost. Amazon is effectively subsidizing Anthropic's compute to lock in its AI business. The same logic applies to Google. The hyperscalers see Anthropic as a strategic partner, not a vendor.
Sifting through the noise to find the signal. The signal is that Anthropic has a real product and real demand. The noise is the hype around the valuation. The bulls are right that the technology is transformative. But they are wrong to assume that the company will capture the lion's share of the value. The history of technology is a history of margin compression. The hardware and infrastructure providers usually win. Nvidia, not OpenAI, is the most valuable AI company. AWS, not Anthropic, owns the cloud. The contrarian insight is that Anthropic's IPO may be a peak for the AI hype cycle, not a new beginning.
Takeaway: The Litmus Test
History is written in blocks, not headlines. The blocks here are cash flows and capital expenditures. Anthropic's IPO will be a test of whether the market can distinguish between growth and value creation. If the IPO prices above $2 trillion, it signals that the AI hype cycle has reached its climax. If it corrects, it shows discipline. I am not making a prediction. I am pointing to the data.
The company must continue to invest heavily while preparing for the IPO. That is a contradiction. The IPO is supposed to be an exit for early investors, not a funding round. But Anthropic needs the money. The $65 billion raised in May will be gone within 18 months at the current burn rate. The IPO is a lifeline.
In my 2023 FTX forensics, I learned that the chain never lies, only the observers do. The same applies to financial statements. When Anthropic files its S-1, the numbers will be there for everyone to see. The revenue run rate will be replaced by actual revenue. The capex commitments will be quantified. The profitability timeline will be stated. Until then, every valuation is a guess.
Every exit is an entry point for the truth. The IPO is an exit for early investors, but it is an entry point for public market scrutiny. The truth will come out. I will be watching the decimal places.
— Nathan Williams, On-Chain Detective.