Truth is not given, it is verified.
Forbes columnist Jim Osman recently dissected Anthropic's potential IPO, and the numbers are staggering: annualized revenue run rate surged from $14 billion in February to $47 billion in May, private valuation now at $965 billion, and whispers of a $2 trillion IPO valuation. The company has committed over $100 billion to Amazon Web Services over the next decade, secured agreements for up to 10 GW of new computing power from Amazon, Google, Broadcom, and even SpaceX.
This is the narrative of the AI boom: exponential growth, massive infrastructure investment, and a belief that the model developer will capture the lion's share of value. But as someone who has spent years auditing smart contracts and building decentralized protocols, I see a different story. The same structural flaws that plague centralized finance are now being baked into the AI industry. The bull market euphoria around AI IPOs is masking a deeper technical risk: the value of artificial intelligence is being siphoned by its infrastructure providers, not its creators.
Context: The Infrastructure Trap
Anthropic's trajectory mirrors the classic crypto cycle of 2021: high revenue, but even higher capital expenditure. The company's compute commitments — $100 billion to AWS, 5GW from Amazon, 5GW from Google/Broadcom, plus SpaceX GPU capacity — are not optional. They are the price of admission to the frontier model race. Osman correctly notes that the key question is how much revenue converts to cash, how much must be reinvested, and whether pricing power can be retained.
From my experience auditing DeFi protocols, I've seen this pattern before. In crypto, the value often flows to the base layer (Ethereum, Bitcoin) or the infrastructure (L2s, data availability layers), while application tokens struggle to maintain value. The same dynamic is playing out in AI: the real profits accrue to chip manufacturers (Nvidia, Broadcom), cloud providers (AWS, Google Cloud), and data centers. Anthropic may be the face of the innovation, but it is essentially a tenant on rented land.
Core: The Decentralized Alternative
Let me be clear: I am not anti-AI. I am anti-centralized infrastructure that creates single points of failure and rent-seeking intermediaries. Blockchain's promise is modularity — separating execution, consensus, data availability, and settlement. Applied to AI, this means decoupling model training, inference, and data storage from monolithic cloud providers.
Projects like Render Network, Akash Network, and Golem are already building decentralized compute marketplaces. They allow anyone to contribute GPU power and earn tokens, creating a supply curve that is more resilient and cost-efficient than centralized hyperscalers. The key insight is that modularity is the architecture of freedom. In a decentralized AI stack, the model developer (like Anthropic) would not be locked into a single cloud provider; they could route workloads across a global network of nodes, reducing costs and increasing censorship resistance.
Consider the numbers: Anthropic's $47 billion revenue run rate sounds impressive, but what is the net margin after paying for compute? Based on public disclosures, AI companies often spend 60-80% of revenue on infrastructure. That means for every $1 earned, $0.60-$0.80 goes to AWS, Nvidia, or Google. In a decentralized compute network, that margin could be captured by token holders and node operators — a more equitable distribution of value.
Furthermore, the concentration of AI compute in a few companies creates systemic risk. What happens if AWS has an outage? Or if a regulatory crackdown targets cloud providers? A decentralized network with thousands of independent nodes is inherently more robust. This is not theoretical; it is the same logic that drove the shift from centralized exchanges to DeFi after the Mt. Gox collapse.
Contrarian: The Pragmatic Test
Now, the contrarian angle: Will decentralized AI actually compete with Anthropic? The immediate answer is no. The performance of frontier models requires massive, co-located clusters — thousands of GPUs in a single data center with ultra-low latency interconnects. Decentralized networks currently cannot match that. But that is a short-term limitation.
Over the next 3-5 years, as model architectures become more efficient (e.g., mixture-of-experts, quantization), and as decentralized compute networks improve latency and coordination, the gap will narrow. More importantly, the market may begin to price in the risk of centralization. We saw this in crypto: initially, none could compete with centralized exchanges, but after the FTX collapse, billions flowed into self-custody and DeFi. A similar event could catalyze demand for decentralized AI infrastructure.
Osman’s analysis hints at this: investors should focus on how future profits are distributed among model developers, chip manufacturers, cloud providers, data centers, and software companies. I argue that the current distribution is unsustainable. Anthropic’s $2 trillion valuation assumes it can maintain pricing power. But as open-source models (e.g., Llama, Mistral) improve, and as decentralized alternatives emerge, that pricing power will erode. Skepticism is the first step to sovereignty.
Takeaway: Build for the Long Tail
In the bear market, only code remains. The AI hype cycle is reaching its peak, and IPOs like Anthropic’s will likely be the top. But the lessons from crypto are clear: centralized infrastructure creates vulnerable points of failure. The true value of the AI revolution will not be captured by a single company, but by the open protocols that enable anyone to participate.
I challenge builders in the crypto space to focus on decentralized compute networks, privacy-preserving inference, and tokenized data markets. These are the primitives that will underpin the next generation of AI — not a $2 trillion IPO, but a network of sovereign nodes. The question is not whether AI will grow, but who controls the infrastructure it runs on.
As I always tell my students: We do not trust; we verify. The same applies to AI. Verify the compute, verify the model, verify the distribution. Anything less is just another centralized system dressed in hype.