In the summer of 2017, I spent four months auditing ERC-20 token standards in Cape Town, watching project after project collapse not because their code was fundamentally broken, but because the narrative couldn't survive contact with reality. That experience taught me something I carry into every analysis: the gap between technological capability and commercial promise is where fortunes are made—and where they evaporate. Today, a similar dynamic is playing out in artificial intelligence, except the stakes are orders of magnitude higher, and the tools being deployed to manage that gap have taken a decidedly more aggressive turn.
The catalyst is something the industry has quietly begun calling "anti-distillation": a suite of technical and contractual measures designed to prevent competitors from using an AI model's outputs to train superior alternatives. If this sounds like building a moat around knowledge itself, that's because it is. And it may be the most consequential development in the AI arms race that most investors haven't yet priced in.
The Commercialization Reckoning
Let me be direct about what the data actually shows. OpenAI has crossed $4 billion in annualized revenue—a figure that would have seemed fantastical three years ago. Yet beneath that headline number lies a troubling reality: the company's inference costs remain stubbornly high, and the path to meaningful gross margins runs through territory no one has successfully mapped. Anthropic's revenue is growing at impressive clip, but every investor conversation I've had this quarter circles back to the same question that haunted DeFi in 2021: where's the actual profit?
This is the crux of what analysts at CITIC Securities have identified as AI stock pricing's primary variable—the能不能跟上市场预期 (whether commercialization pace can keep up with market expectations). The translation obscures the urgency: we're witnessing a fundamental reassessment of how to value companies whose technology is undeniably impressive but whose unit economics remain stubbornly unproven.
The market's patience is not infinite. Traders I've spoken with—who manage real capital, not theoretical portfolios—are已经开始担心 (starting to worry) that if the next two to three quarters don't deliver commercialization data that exceeds already-elevated expectations, the valuation framework for AI stocks could shift dramatically. Price-to-sales multiples have sustained the sector through uncertainty; a transition to price-to-earnings logic would trigger a systematic repricing that most current positioning isn't prepared for.
The Compute-to-Competitive-Advantage Pipeline
Here's where it gets technical—and where my experience auditing smart contracts becomes relevant. In blockchain, we learned that consensus mechanisms don't just secure transactions; they encode power relationships. The same principle applies to compute infrastructure in AI.
The transmission chain is elegant in its brutality: compute advantage enables faster model iteration, lower service costs, and more responsive customer adaptation. These three factors compound into market share. Google DeepMind's Gemini series and Anthropic's Claude lineup have validated this logic—投入强度与模型市场表现呈正相关 (investment intensity correlates positively with market performance). The companies with the most GPUs are winning the race to build the best models, which attracts more customers, which generates more training data, which trains better models.
But here's the wrinkle that the bullish narratives conveniently omit: model capability gaps are compressing within generations. The jump from GPT-3 to GPT-4 was seismic. The evolution from GPT-4 to GPT-4o was meaningful but marginal. Yet the gaps in inference costs and long-context capabilities continue widening. This suggests that even as raw model performance converges, the economic moats around cost efficiency and capability boundaries remain substantial enough to sustain competitive advantages for the companies that built them first.
The Anti-Distillation Variable
Now we arrive at what CITIC Securities correctly identifies as the biggest potential variable—and what I believe most Western coverage has critically underweighted.
Anti-distillation encompasses technical measures like output watermarking, API usage restrictions, and contractual prohibitions on using model outputs for training competitors. If implemented effectively, this creates what we in open-source advocacy call a "perpetual disadvantage trap" for smaller players. You cannot stand on the shoulders of giants when the giants have locked the ladder.
The implications for the industry structure are profound. Currently, mid-tier AI companies rely heavily on knowledge distillation—using outputs from larger models to fine-tune their own systems at fraction of the training cost. This has been the primary pathway for追赶 (catch-up) players. If anti-distillation becomes industry standard, that pathway closes. The choice becomes stark: build from scratch at prohibitive cost, or accept permanent second-tier status.
This dynamic affects blockchain and AI convergence projects disproportionately. Protocols building decentralized AI infrastructure depend on the ability to leverage open model weights and distillation techniques. An industry shift toward anti-distillation could effectively strangle the emerging field of on-chain AI agents before it matures.
The K-Divergence Paradox
Here's the contrarian angle that traditional analysis misses: the very mechanisms designed to entrench头部厂商 (leading vendors) may ultimately create the conditions for their disruption.
When competition narrows to a handful of players protected by compute moats and anti-distillation barriers, the innovation that made them dominant gradually suffocates. The history of technology is littered with incumbents who optimized for defense rather than exploration—and paid the price when paradigm shifts arrived from unexpected directions. Quantum computing, neuromorphic chips, and radically efficient training algorithms all represent potential discontinuities that could render current compute advantages obsolete.
Moreover, the regulatory response to anti-distillation remains entirely uncharted. European authorities have already signaled discomfort with practices that restrict AI knowledge diffusion. The EU AI Act's provisions on transparency could be weaponized against watermarking-based anti-distillation. China, facing its own compute constraints from semiconductor export controls, has every incentive to champion international norms around model accessibility. The geopolitical dimension of AI governance is entering territory that most technical analyses treat as beyond scope—but it isn't beyond consequence.
The Path Forward
Education remains the only true decentralized currency in this landscape. The investors who will navigate the next eighteen months successfully are those who understand the difference between technological capability and commercial viability—not as abstractions, but as measurable indicators they can track quarter by quarter.
The signals worth watching are concrete: revenue growth rates in AI company earnings reports, gross margin trends that indicate whether inference costs are truly scaling, customer retention metrics that reveal whether enterprise deployments are sticky or speculative. A company that demonstrates consistent improvement across all three dimensions deserves its premium. One that excels in capability but falters on economics deserves scrutiny that current market enthusiasm rarely provides.
The anti-distillation gambit may succeed in the short term. The companies with compute advantages may consolidate further, margins may improve, and the competitive landscape may ossify into the comfortable oligopoly that incumbents have always preferred. But technology has a way of finding cracks in even the most formidable walls. The question isn't whether disruption will come—it's whether the current leaders will be the ones to deliver it, or whether the next generation of builders is already plotting the obsolescence of today's giants from garages and labs around the world.
We build bridges, not just blocks, between what AI promises and what it delivers. That delivery is now the only metric that matters.