In the middle of a bull market that has turned every AI startup into a unicorn, a recent analysis by SemiAnalysis dropped a bombshell: Anthropic, the company behind Claude, is reportedly generating $65 billion in annualized recurring revenue. But as I dug into the numbers, something felt off.
I’ve been in this space long enough to know that headline ARR numbers often mask the real economic engine. In 2017, I spent nights auditing the Gnosis Safe multisig contract, finding 12 critical flaws that would have let a few signers drain the entire treasury. That experience taught me to look past the glossy marketing and into the actual code—or in this case, the actual revenue structure.
The Context: Cloud Platforms as the New Multisig
Anthropic’s distribution model is heavily reliant on three cloud giants: AWS Bedrock, Microsoft Foundry, and Google Cloud Platform. According to SemiAnalysis, over 40% of Anthropic’s ARR flows through these channels. On the surface, this seems like a savvy move—piggybacking on existing enterprise relationships to achieve rapid scale. But the economics are far more sinister.
When a customer accesses Claude through Bedrock, Anthropic pays AWS a commission (likely 15–30%) plus compute costs for the GPU instances used during inference. The same applies to Azure and GCP. The result: every dollar of ARR from cloud channels yields significantly less gross profit than a direct API sale. This is the “cloud tax”—a silent drain on profitability that ARR numbers alone cannot reveal.
The Core: How Profit Dilution Works
Let me break this down with numbers. If Anthropic’s direct sale gross margin is 80% (typical for a pure API service), then a channel sale might see margins drop to 40–50% after commissions and compute overhead. For a company with $65B ARR, even a 10% margin difference represents $6.5B in lost profit potential. But the real kicker is that $65B figure itself.
Publicly available data suggests Anthropic’s actual 2024 revenue is in the range of $1–2 billion. The $65B figure is almost certainly a misinterpretation—perhaps a long-term target or a unit error. Yet the channel dependency remains. Even at $2B ARR, 40% or $800M coming from cloud channels means $300–400M in lost profit compared to direct sales. Over time, this compounds.
I’ve seen this pattern before. In DeFi Summer 2020, I watched protocols like Compound boast massive TVL, but the actual yield was consumed by gas fees and impermanent loss. The same principle applies here: the metric that glitters (ARR) is not the metric that matters (sustainable profit).
The Contrarian: Why Cloud Channels Are a Trap
Some will argue that cloud channels are necessary for reaching enterprise customers who already have AWS contracts. But this is a dangerous oversimplification. The cloud providers are not neutral distributors; they are competitors. Google runs Gemini, AWS has its own AI services, and Microsoft is heavily invested in OpenAI. By relying on these platforms, Anthropic is essentially handing over its customer relationship and pricing power to potential rivals.
Moreover, the cloud tax creates a perverse incentive. As Anthropic pushes more revenue through channels to hit growth targets, its margins erode. The company may end up trading long-term profitability for short-term ARR bloat—a classic “growth at all costs” trap that has felled many startups.
In my own NFT project, “On-Chain Diaries,” I deliberately avoided large platforms and minted only 50 artifacts on a custom contract. The result was a smaller community but a genuine one, with royalties that actually flowed to artists. The same lesson applies to AI: direct relationships with users are more valuable than distributed volume through middlemen.
The Takeaway: Follow the Fear, Not the Chart
The cloud tax on AI is a solvable problem, but only if founders and investors recognize it. The solution is not to abandon channels entirely, but to build a mixed model where direct sales grow faster than channel sales. This requires investment in enterprise sales teams, self-hosted deployment options, and perhaps even decentralized infrastructure for AI inference.
If you can’t see the fees, you can’t trust the revenue. The next time you see a $65B ARR headline, ask yourself: how much of that is actually making it to the bottom line?
Follow the fear, not the chart. If you can.
