When I first read the news that Anthropic is targeting a $965 billion valuation for its 2026 IPO, I felt a familiar pang of unease. It was the same feeling I had in 2017 when I declined advisory roles for ICOs that promised the moon but delivered vaporware—only this time, the vaporware is an AI model trained on a mountain of venture capital and locked into a single cloud provider. The headline screamed “largest IPO in history,” but beneath the surface, I saw a pattern I’ve been tracking for years: the quiet centralization of power under the guise of innovation. As someone who has spent the last decade auditing smart contracts, building DeFi protocols, and watching the rise of crypto-native governance, I know that the most dangerous risks are the ones that look like progress.
Let me be clear: Anthropic’s technology is impressive. Their Constitutional AI framework is a genuine advance in safety alignment, and Claude’s performance on benchmarks like SWE-bench and GPQA puts it in the top tier. But the narrative around this IPO has nothing to do with technical merit. It’s a story about capital, control, and the illusion of decentralization. The $965 billion figure is not a reflection of value—it’s a reflection of how much the market is willing to pay for a centralized oracle that promises to be trustworthy. And I’ve seen this movie before. In 2017, we called it “trust the code.” In 2025, we call it “trust the model.” Both are a bet on a single point of failure.
Context: The Architecture of Centralization
Anthropic is a private company with a mission: build safe AI. It has raised over $18 billion from investors including Amazon ($8 billion) and Google ($2 billion+). Its primary revenue comes from the Claude API, Claude Pro, Claude Team, and the recently launched Claude Code—a developer tool that has become a darling of the coding community. The company’s core differentiator is its focus on safety and interpretability, embodied in constitutional AI and its “mechanistic interpretability” research. But here’s the rub: all of this runs on AWS. Amazon is not just an investor; it’s the primary compute provider, the cloud infrastructure, and the strategic partner. The IPO prospectus will likely reveal that Anthropic’s ability to scale is entirely dependent on Amazon’s willingness to provide GPU clusters at a favorable price.
This is not inherently bad. Every startup needs infrastructure. But when a single entity controls both the capital and the compute, the “decentralization” narrative of AI becomes a farce. The valuation of $965 billion implies that Anthropic will generate $32–$97 billion in revenue by 2026 (using a P/S ratio of 10–30x). That’s a 5–10x increase from its estimated $7 billion run rate in 2025. To achieve that, Anthropic must maintain its position as a top-tier model developer, fend off OpenAI and Google, and avoid any disruption to its AWS relationship. The margin for error is zero. And as someone who has watched the Terra-Luna collapse destroy billions in trust, I know that zero-margin bets are the most dangerous.
Core: The Hidden Costs of the AI Oracle
Let me drill into the technical details that the IPO hype obscures. Anthropic’s architecture is a modified Transformer, similar to GPT-4, but with a crucial difference: its safety alignment is baked into the training process via constitutional AI. This is a significant achievement—it reduces the need for post-hoc filtering and makes the model more resistant to jailbreaks. But it also introduces a dependency on the quality of the constitution itself. Who writes the constitution? Who audits it? In the crypto world, we would call this a governance attack vector. If the constitution is flawed, the entire model’s behavior is flawed. And the constitution is controlled by a single company.
From a crypto perspective, this is analogous to a centralized oracle. Chainlink, for all its flaws, distributes its data sources across multiple nodes. Anthropic’s “oracle” is a single model, trained on a single dataset, hosted on a single cloud provider. The truth is immutable, unlike the price action. But in this case, the truth is whatever the model says it is. If Amazon decides to throttle compute, Anthropic’s entire operation slows down. If the SEC decides that the model’s outputs are financial advice, the company is liable. The $965 billion valuation is a bet that none of these contingencies will materialize. Based on my experience auditing 14 critical vulnerabilities in the Tezos mainnet during the 2017 ICO boom, I can tell you that the most overlooked risks are often the ones that seem most stable.
I want to highlight a specific number: the compute cost. Anthropic’s training runs consume hundreds of gigawatt-hours of energy. The next-gen model, likely to be released before the IPO, will require an estimated 1e26–1e27 FLOPs. That’s a massive commitment to NVIDIA GPUs and AWS infrastructure. The company has signed multi-billion-dollar contracts with Amazon to secure that compute. But what happens if the AI market cools down? If the revenue growth slows, the fixed costs of compute become a death spiral. I’ve seen this in DeFi: protocols that over-leverage on liquidity mining rewards often collapse when the incentives dry up. Anthropic’s compute obligations are a form of financial leverage. The IPO is a way to refinance that debt with public market capital.
Another hidden cost is the talent war. Anthropic’s safety team is one of the deepest in the industry, but after the IPO, key employees will have lock-up periods. Once those expire, the temptation to cash out and join a competitor (or start a new venture) is enormous. The very people who built the constitutional AI framework could leave, taking the institutional knowledge with them. In crypto, we call this “rug pull.” In traditional tech, it’s just “executive turnover.” But the effect is the same: the value of the company depends on the continued presence of key individuals. The $965 billion valuation assumes that the team stays intact. That’s a fragile assumption.
Contrarian: The Pragmatic Test of Decentralization
Now, let me play the contrarian. The crypto community loves to hate centralized AI. We paint it as the enemy of sovereignty, a monopolistic force that will concentrate power in the hands of a few. But is a decentralized alternative really better? Let’s look at the data. The so-called “AI on blockchain” projects—like Bittensor, Render, or Akash—have a combined market cap that is a fraction of Anthropic’s valuation. Their models are not as capable. Their inference is slower. Their governance is messy. The reality is that training a frontier model requires hundreds of millions of dollars in compute, and no decentralized network currently provides that. The volatility is noise; utility is signal. And the signal is that centralized AI, for all its flaws, produces better products today.
But here’s the catch: that advantage is temporary. The same way that centralized exchanges dominated crypto in 2017 but have since been partially displaced by DeFi, centralized AI will eventually face competition from decentralized alternatives. The key insight is that the value of AI comes from data and compute, both of which are becoming more commoditized. As GPU costs drop and open-source models improve, the moat of a company like Anthropic narrows. The $965 billion valuation is a bet that the moat is wide and deep. I’m not convinced. I’ve written about this before: long-term vision > short-term pumps. The IPO is a short-term pump for early investors, but the long-term vision of decentralized AI is still in its infancy.
Another contrarian point: Anthropic’s relationship with Amazon is a double-edged sword. On one hand, it provides a built-in distribution channel via AWS Bedrock. On the other hand, it creates a conflict of interest. Amazon is both a competitor (with its own AI services) and a supplier. If Amazon decides to promote its own models over Claude, Anthropic’s growth could stall. The IPO prospectus will need to disclose the terms of the AWS contract, including any volume discounts or exclusivity clauses. Based on my experience with the 2022 bear market, when I retreated to a cabin in Virginia to write “The Soul of Sovereignty,” I learned that the most important question to ask is: who controls the exit? In this case, Amazon controls the compute exit. That’s a risk that the market is currently ignoring.
Takeaway: A Vision for a Different Future
I’m not saying Anthropic is a bad company. On the contrary, I respect their commitment to safety and interpretability. But the $965 billion IPO is a symptom of a larger problem: the financialization of a technology that was supposed to be about empowerment. The same forces that centralized the internet—advertising, monopoly, and venture capital—are now centralizing AI. The crypto community has a responsibility to build alternatives. We need decentralized compute networks that are cheap enough to train frontier models. We need on-chain governance for AI alignment. We need open-source models that cannot be captured by a single corporation.
This is not a pipe dream. I’ve seen the power of decentralized communities in DeFi and DAOs. The same spirit can be applied to AI. The question is whether we will act before the centralized players become too entrenched. As I wrote in my 2024 op-ed “Institutionalization vs. Ideology,” the ETF approval was a double-edged sword: it brought legitimacy but also centralization. The Anthropic IPO is the same. It will bring capital and attention, but it will also solidify a model of AI that is owned by a few. The choice is ours. Resilience is the only alpha. We must build the infrastructure for a truly decentralized AI, or we will wake up one day to find that the oracle has become the master.
Truth is immutable, unlike the price action. The price of Anthropic’s IPO will rise and fall, but the truth of what we build—whether it is centralized or decentralized—will shape the future of intelligence. I choose to build for the latter.
