The most consequential sentence in CITIC Securities' recent tech-sector adjustment report isn't about valuations, Treasury yields, or even the compute arms race. It's a single phrase buried in the competitive dynamics section: "anti-distillation" โ identified as the largest potential variable in AI equity pricing. Read that again. Not model capability. Not commercialization velocity. Not even the GPU supply curve. The report's authors flagged the act of preventing competitors from training on your model's outputs as the single most consequential unknown in the AI market.
That's not a market observation. That's a cryptographic statement wearing an equity analyst's suit.
I've spent the last two years excavating the intersection of zero-knowledge proofs and AI inference โ building circuits that verify model outputs without revealing proprietary weights. When I read the CITIC report's framing of anti-distillation, I didn't see a competitive strategy. I saw a verification problem that the market hasn't yet learned to price. Every bug is a story waiting to be decoded, and this particular bug is hiding in plain sight across the entire AI equity complex.
The Framework Shift: From Macro to Micro
The report's core contribution is reframing AI stock pricing from macro-driven to industry-fundamental-driven. The authors argue that Treasury yields are not the root cause of the tech correction โ that even if the rate environment improves, AI companies lacking commercial validation won't get valuation relief. This is a bold claim, and it's largely correct. But the framework they propose โ three verifiable pricing variables (commercialization pace, compute conversion efficiency, model gap evolution) plus the anti-distillation wildcard โ deserves deeper technical scrutiny than the sell-side typically provides.
Let me excavate each variable, because the report's framework is sound but its execution is shallow. And the gaps matter.
Variable One: Commercialization Pace โ The Unit Economics Problem
The report correctly identifies commercialization pace as the first pricing variable. OpenAI's annualized revenue crossing $4 billion while inference costs remain stubbornly high. Anthropic's revenue growing fast but gross margins under pressure. Microsoft's Copilot penetration controversy. Salesforce's Einstein GPT adoption rates below early optimism.
The report frames this as a timing mismatch: the technology investment curve is steeply rising while the revenue realization curve hasn't hit its exponential inflection. The capital market is repricing this mismatch. Fair enough. But the report misses the deeper structural issue: AI's unit economics have not been validated at any scale.
I've been tracking this through a different lens. In my 2020 DeFi composability cartography work, I mapped 150+ protocol interactions and discovered how liquidation cascades propagated across chains. The lesson that stuck: when unit economics are unproven, systemic risk hides in the interconnections. The same applies to AI. The report asks whether LTV/CAC is improving โ but the more important question is whether AI's "liquidity mining" phase (subsidized inference, below-cost API pricing, free tier expansions) is creating artificial adoption that will reverse when prices normalize.
The report's hidden implication is that the market's "patience window" is narrowing. If the next 2-3 quarters don't deliver beat-expectations commercialization data, the valuation regime could shift from PS multiples to PE logic, triggering systematic downward revisions. This is correct, but it's also incomplete. The report doesn't quantify what "beat expectations" means. It doesn't provide the specific metrics โ revenue growth inflection, gross margin improvement, customer retention rates โ that would serve as signal lights.
Based on my audit experience, I'd suggest three concrete metrics to watch. First, inference cost per million tokens trending down while API volume trends up โ this indicates real demand elasticity rather than subsidized adoption. Second, enterprise customer expansion revenue (existing customers spending more) as a percentage of total revenue โ this distinguishes "incremental customer acquisition" from "deep monetization of existing customers," the exact distinction the report gestures at but never operationalizes. Third, gross margin excluding compute subsidies โ this reveals whether the business model works without artificial pricing support.
The report also notes that AI pricing models remain cost-plus (per token, per seat) rather than value-based. This is a critical observation. Until AI companies establish pricing power directly tied to customer value creation, their commercialization quality remains unverified. The transition from cost-plus to value-based pricing is the single clearest signal that an AI company has crossed from "technology validation" to "scaled monetization." None of the major players have made this transition yet. That's not a bug in the market โ it's a feature of an industry still in its subsidized growth phase.
Variable Two: Compute Conversion โ The Fragile Moat
The report's transmission chain โ compute advantage โ market share โ model gap โ is the most technically sound part of the analysis. Compute advantage enables faster iteration, lower-cost service delivery, and more flexible customer response. All three convert to market share. Google DeepMind's Gemini series and Anthropic's Claude series both validate this: compute investment intensity correlates positively with model market performance.
But here's where the report's analysis gets lazy. It treats compute as a monolithic moat, when in fact the compute efficiency curve is bending in ways that undermine the moat's durability. Algorithmic innovation โ Mixture-of-Experts architectures, quantization techniques, speculative decoding, continuous batching โ is compressing the effective compute gap by 10-100x in inference scenarios. A model that requires 10x less compute for the same output quality doesn't need 10x less compute advantage to compete; it needs a fraction of that.
I've seen this pattern before. In the 2022 bear market, I spent months analyzing Celestia's Data Availability Sampling mechanism, identifying potential sybil attack vectors in the node distribution. The thesis I developed โ that security is secondary to availability in rollup ecosystems โ was controversial. But the underlying principle applies here: in resource-constrained environments, efficiency innovation can offset raw resource advantages. The question isn't whether compute advantage creates model gaps. It's whether algorithmic innovation can compress those gaps faster than compute scaling can expand them.
The report asks whether compute advantage will significantly widen future AI model gaps. The answer depends on two factors the report doesn't fully explore: the duration of the compute gap (how long before supply catches up) and whether non-compute factors (algorithmic innovation, data quality) can partially offset compute disadvantages. My read: the compute gap is real but not permanent. TSMC's CoWoS capacity expansion, HBM supply increases, and the emergence of alternative chip architectures will erode the moat within 12-18 months. The companies that win won't be those with the most compute โ they'll be those with the highest compute-to-value conversion efficiency.
The report's framing of compute as "the new oil" of the AI economy is evocative but misleading. Oil is a commodity โ interchangeable, price-driven. Compute is more like a strategic asset with diminishing marginal returns. The first 10,000 GPUs transform a company's capabilities. The next 10,000 add incrementally less. The marginal value of compute declines as the ecosystem matures, which means the moat built on raw compute advantage is inherently self-limiting.
Variable Three: Model Gap โ From Generational to Intra-Generational
The report correctly notes that model capability gaps have narrowed from "generational differences" (GPT-3 to GPT-4) to "intra-generational differences" (GPT-4 to GPT-4o). But inference cost gaps and long-context capability gaps are widening. This is the nuance the report captures well: even if model capabilities converge, cost and capability boundary differences can sustain competitive advantages.
But the report misses a critical dimension: the verification gap. As models become more capable, the ability to verify what a model is actually doing โ its reasoning, its data provenance, its output integrity โ becomes more valuable and more difficult. This is where my ZK research intersects with the AI equity story. Zero-knowledge proofs can verify that a model's output was produced by a specific model with specific weights, without revealing those weights. This is the technical foundation for anti-distillation โ and it's also the foundation for a new class of AI trust infrastructure.
Navigating the labyrinth where value flows unseen: the AI market is pricing model capability, commercialization, and compute. It is not yet pricing verifiability. But it will.
The report's competitive landscape analysis โ "one superpower, multiple strong players" โ is accurate but incomplete. OpenAI maintains leadership in model capability and ecosystem maturity, but Anthropic, Google, and Meta are closing faster than the market expects. The competition has shifted from "whose model is stronger" to "whose combination of model, compute, and ecosystem is superior." This multi-dimensional competition favors vertically integrated players (Google's full-stack advantage) and deeply partnered players (OpenAI-Microsoft, Anthropic-Amazon).
The Anti-Distillation Wildcard: A Cryptographic Problem in Disguise
The report's most provocative claim is that anti-distillation โ preventing competitors from training on your model's outputs โ is the largest potential variable in AI pricing. The logic: if leading model vendors successfully implement anti-distillation (output watermarking, API usage restrictions, legal enforcement), smaller AI companies lose the "standing on giants' shoulders" path to catch up. They'd be forced to train base models from scratch, dramatically raising industry entry barriers and accelerating market concentration.
The report treats this as a competitive strategy question. It's actually a cryptographic question. Anti-distillation requires proving output provenance โ demonstrating that a given output came from a specific model โ without breaking the model's utility. This is precisely what ZK proofs can do. And it's precisely what the market isn't pricing.
Let me be concrete. In my 2021 ZK-SNARK protocol sprint, I implemented three distinct proof generation algorithms from scratch and forked the Circom compiler to create a simplified tutorial that helped 5,000 developers deploy their first zero-knowledge circuit. The key insight from that work: ZK proofs are not just for privacy โ they're for provenance. You can prove that a computation was performed correctly without revealing the inputs. Applied to AI: you can prove that a model output was generated by a specific model with specific weights, without revealing the weights or the training data.
This has two implications the report doesn't explore. First, anti-distillation is technically feasible โ but it's not free. Output watermarking degrades output quality. API usage restrictions create friction. ZK-based provenance adds computational overhead. The question isn't whether anti-distillation can work; it's whether the cost of implementation exceeds the competitive benefit. Second, if anti-distillation becomes standard practice, it creates a new market: verification infrastructure. Companies that can prove model provenance, verify inference integrity, and audit training data will become essential service providers โ the "auditors" of the AI economy.
The report's hidden concern is that anti-distillation will freeze the model gap, slowing AI innovation diffusion. This is a legitimate concern, particularly for China's AI industry, which has relied on the open-source-plus-distillation path. But the report doesn't address the countervailing force: open-source models (Llama, Qwen, Mistral) are improving faster than the closed-source gap is widening. The open-source ecosystem is itself a form of anti-distillation resistance โ a distributed training network that no single vendor can control.
There's also a deeper question the report doesn't ask: does anti-distillation increase or decrease total compute demand? If it leads to redundant training (everyone training from scratch), compute demand explodes. If it accelerates industry consolidation, compute demand concentrates in fewer hands. The answer determines whether anti-distillation is inflationary or deflationary for the compute market โ and by extension, for the entire AI supply chain.
The Valuation Regime Shift: From Narrative Premium to Execution Proof
The report's deepest contribution is reframing the valuation question. AI stocks have entered an "expectation verification period" where valuations will depend more on verifiable industry progress than on macro liquidity. The investment strategy shifts from beta-driven sector allocation to alpha-driven stock selection.
This is where the report's analysis connects to my broader thesis about verification economics. The market is moving from "paying for imagination" to "paying for execution" โ and execution, in the AI context, requires proof. Not just revenue numbers, but verifiable evidence that the business model works, that the compute is being converted efficiently, that the model gap is sustainable.
I've seen this transition before. In crypto, the shift from narrative-driven valuation to fundamentals-driven valuation was brutal. Projects that couldn't demonstrate real usage, real revenue, or real security were repriced to zero. The survivors were those with verifiable metrics โ on-chain usage data, revenue streams, security audits. The same dynamic is now playing out in AI. The market is learning to ask: "Can you prove it?"
The report's "K-shaped divergence convergence" mention hints at a trading strategy: dollar weakness and reduced rate-hike expectations could trigger capital rotation from US AI leaders to other markets, including A-shares. But this rotation's sustainability depends on whether AI industry fundamentals support valuation convergence. My read: the rotation will happen, but it will be selective. Only A-share AI companies with real revenue and verifiable technology will benefit. The rest will be repriced downward.
The report's warning against "over-grand narratives" is essentially a warning about AI narrative bubble risk. The market's AI expectations include substantial "grand narrative" components โ AGI approaching, productivity revolution, autonomous agent economies. When these narratives fail to materialize as concrete business outcomes, valuation correction risk amplifies significantly. The report is right to flag this, but it doesn't quantify the narrative premium embedded in current valuations. That quantification is the next frontier for AI equity analysis.
The Blind Spots: What the Report Misses
The report's framework is sound but incomplete. Three blind spots stand out.
First, the report underestimates macro factors. The claim that Treasury yields aren't the root cause of the tech correction is partially correct โ but in a high-rate environment, even companies with strong fundamentals get crushed. The 2022 crypto bear market taught me this: when liquidity contracts, everything falls, regardless of fundamentals. The report's dismissal of macro factors is convenient for its narrative but dangerous for investors who ignore rate risk.
Second, the report doesn't discuss AI safety and ethics as pricing variables. This is a significant omission. Regulatory risk โ EU AI Act, China's model filing requirements, potential US executive actions โ could reshape the competitive landscape faster than any of the three pricing variables. A single major AI safety incident could trigger regulatory intervention that reprices the entire sector. The report's silence on this is telling.
Third, the report's anti-distillation analysis is shallow. It identifies the variable but doesn't analyze technical feasibility, implementation paths, or industry impact. It doesn't ask whether anti-distillation increases or decreases compute demand. It doesn't consider the open-source counterforce. It treats a cryptographic problem as a business strategy question.
There's also a fourth blind spot the report shares with most sell-side analysis: it doesn't address the China AI compute bottleneck directly. The report's discussion of "whether compute gaps will significantly widen future AI model gaps" is a euphemism for a more specific question: under export controls on high-end GPUs, can Chinese AI companies narrow the gap through algorithmic innovation, compute optimization, or domestic chip alternatives? The report's silence on this is politically understandable but analytically incomplete. The answer to this question will determine the global AI competitive landscape for the next decade.
The Convergence: ZK and AI Verification
Here's where I'll make my contrarian bet. The next major AI market battleground won't be model capability โ it will be verifiability. As AI systems become more integrated into economic infrastructure โ autonomous agents, automated decision-making, algorithmic trading โ the ability to verify what these systems are doing becomes as valuable as the systems themselves.
This is the AI-ZK convergence I've been building toward. Zero-knowledge proofs can verify AI model outputs without revealing proprietary data. They can prove that an inference was computed correctly. They can audit training data provenance. They can enable anti-distillation while preserving model utility. They are the trust infrastructure for the autonomous agent economy.
The market isn't pricing this yet. But it will. When AI agents start transacting autonomously โ managing portfolios, executing trades, negotiating contracts โ the demand for verifiable computation will explode. The companies that build this infrastructure will capture value disproportionate to their current market caps.
Composability is not just function; it is poetry. The composability of ZK proofs and AI inference is the most exciting architectural development I've seen in a decade of building. It's not just about privacy โ it's about trust. It's about creating a world where you don't have to trust an AI system; you can verify it.
The report's three pricing variables โ commercialization, compute conversion, model gap โ are all ultimately verification problems. Can you verify that commercialization is real and not subsidized? Can you verify that compute is being converted efficiently? Can you verify that the model gap is sustainable? The market is moving toward asking these questions, and the infrastructure to answer them is being built right now.
The Takeaway: Verification Becomes the Premium
The CITIC Securities report is a useful framework, but it's incomplete. It identifies the right variables โ commercialization, compute conversion, model gap โ and flags the right wildcard โ anti-distillation. But it misses the deeper structural shift: the AI market is moving from narrative premium to execution proof, and execution proof requires verification infrastructure.
The companies that win the next phase won't be those with the best models or the most compute. They'll be those that can prove what their models are doing โ to regulators, to enterprise customers, to the market. Verification becomes the premium. Trust becomes the moat.
I've spent 22 years observing this industry, from the ICO chaos of 2017 to the DeFi summer of 2020 to the ZK revolution of 2021 to the modular blockchain research of 2022. Every cycle follows the same pattern: narrative premium inflates, then collapses, then fundamentals reassert. The AI cycle is no different. The question isn't whether the narrative will collapse โ it's whether the fundamentals can survive the collapse.
Excavating truth from the code's buried layers: the truth here is that AI's valuation problem is fundamentally a verification problem. The market is learning to ask for proof. The infrastructure to provide that proof is being built right now, in the intersection of zero-knowledge cryptography and AI inference. The investors who understand this convergence will be positioned for the next phase. The ones who don't will be left holding narrative premium with no execution underneath.
The market is asking: "Can you prove it?" The answer will determine which AI companies survive โ and which verification infrastructure builders thrive.