Over the past seven days, a mid-cap DeFi protocol I have tracked since its launch quietly shed nearly forty percent of its liquidity providers. No exploit. No depeg. No governance scandal. Just a slow, silent drain as capital rotated toward whatever narrative was paying that week — which, in a bear market, is the only migration that ever really happens. Nobody wrote a post-mortem, because nothing technically died. The protocol is still there. It is just no longer trusted, and in this industry the two are the same thing.
That is the weather I was standing in when the AI headline arrived. A crypto-native outlet — not a research lab, not a policy journal — reported that Chinese AI laboratories were closing the distance with OpenAI and Anthropic, and that the narrowing was beginning to challenge American dominance of the frontier. No lab was named. No benchmark was cited. No training run was dated. Just a sentence, wrapped in the vocabulary of markets, released into a feed already primed to react.
I read it three times. By the third pass I understood the story was never really about AI. It was about trust — where it pools, who is permitted to verify it, and what happens when the machinery of verification belongs to a handful of firms in a handful of jurisdictions. That is a story I have spent the better part of a decade inside: first as a junior analyst auditing token distributions in Singapore, then as a builder of small, stubborn communities, and now as someone who has come to believe that the ledger is the only honest instrument we possess for measuring power.
In a bear market, survival matters more than narrative, and a story like this is not information — it is weather. And the weather has been moving in a direction that should concern anyone who cares about who owns the next decade of computation. Not because one bloc is winning, but because the frame itself — the idea that the only thing worth measuring is the distance between two monopolies — quietly forecloses the more important question. Who gets to own the intelligence, and on what terms?
Let me set the scene before I argue with it.
For two years now, the dominant narrative in technology has been a horse race. On one side, the American frontier laboratories: OpenAI, Anthropic, Google, Meta. On the other, a cluster of Chinese laboratories that, the story goes, are no longer a generation behind but are breathing on the leaders' necks. The proof offered is always the same — benchmark tables for mathematics and code, parameter counts, occasionally a training-cost figure that reads like a magic trick. The subtext is always geopolitical. Whichever side reaches the frontier first is said to inherit the century.
What this framing omits is the thing I actually audited for a living: the infrastructure of trust beneath the models. A language model is not a product. It is a claim — a claim about data, about provenance, about whose labor trained the weights and whose interests the outputs serve. And claims, unlike benchmarks, cannot be evaluated by looking at a score. They have to be evaluated by looking at the ledger: who contributed, who consented, who captured the value. On that ledger, both sides of the so-called race are, for the most part, equally opaque. The gap everyone is measuring is a gap between two closed systems.
I have written before about a project called OmniChain, back in the fever of 2017 — a decentralized identity platform whose tokenomics were engineered, in a single spreadsheet, to reward early insiders while its whitepaper preached egalitarianism. I spent months on a five-thousand-word exposé, and it was shared widely in the weeks before the project's rug pull. That experience taught me a permanent lesson: the ethics of a system live in its distribution mechanics, not its rhetoric. The same lens applies to AI. A model that benchmarks at the frontier but concentrates data ownership in three corporations is not ahead in any sense that matters to a human being trying to live under it. It is simply a more elegant enclosure.
So when I read that the gap is narrowing, my first instinct is not to celebrate the challenger or mourn the incumbent. It is to ask what unit the gap is measured in — and to notice that the unit is almost never sovereignty.
The single most important fact about the Chinese laboratories' recent progress is one the headlines routinely bury: it was achieved under a hard compute ceiling. Export controls denied the most advanced accelerators — the H100-class silicon that American labs treat as the baseline of existence. The response was not to build a bigger cluster. It was to build a more frugal one. Mixture-of-experts architectures that activate a fraction of parameters per token. Lower-precision training that trades a sliver of numerical fidelity for a large reduction in cost. Reinforcement-learning pipelines that substitute for the brute-force supervised fine-tuning Western labs leaned on. Every one of these is an efficiency innovation, not a paradigm break. And efficiency innovations are what you invent when you cannot buy your way to scale.
This is a pattern I know intimately, because it is the pattern of my own field. When a blockchain hits a throughput ceiling, the answer is never to use more chains. It is compression, batching, proving systems that make each unit of work carry more. The entire Layer2 roadmap is a study in constraint-driven engineering. And here is the uncomfortable parallel I keep returning to: the efficiency that looks like a triumph in one cycle becomes a trap in the next, because the constraint never lifts — it only moves.
Let me make that concrete, because I think it is the most under-discussed risk in the AI race and the most instructive one for anyone in crypto. When Ethereum shipped EIP-4844 and gave rollups a cheap, temporary data channel called blobs, the celebration was immediate and, in my view, premature. Blobspace was priced as though it were infinite. It is not. My working estimate, which I have held since Dencun went live, is that blob demand saturates within roughly two years, at which point rollups will be bidding for scarce data availability again and the gas fees that users were promised would fall forever will double back toward their old levels. The efficiency did not eliminate the constraint. It deferred it, and the deferral trained an entire ecosystem to build as though the constraint were gone. That is a dangerous kind of progress.
The AI laboratories are walking the same road. Frugality under a compute ceiling looks like a permanent advantage until you realize it is a temporary subsidy — one that vanishes the moment the ceiling moves, or the moment the efficiency curve flattens against its physical limit. Scaling laws are not infinitely elastic. There is a point where lower precision costs you more accuracy than it saves in dollars, where sparsity stops paying, where the data you have left to distill from is already contaminated by the outputs of the very models you are trying to beat. The laboratories that look like they are closing a gap may simply be the ones furthest along a curve that bends.
But let me not let the incumbents off the hook, because the deeper asymmetry is not in the weights. It is in the data, and specifically in who owns it.
Every frontier model is trained on a corpus assembled without meaningful consent and compensated through a legal fiction called fair use. The people whose writing, images, and code made the weights valuable received nothing, and they will receive nothing, and no benchmark table will ever show this because benchmarks measure capability, not justice. The Chinese laboratories are not worse on this axis. They are the same. Both are extracting from a commons they did not build, and both are racing to enclose it before the other does.
This is where I believe the ledger stops being a metaphor and becomes a mechanism. In 2026 I launched a pilot under the banner of The Algorithmic Soul — a series of essays that predicted, before it was fashionable, that without blockchain-based data ownership, AI would concentrate power in exactly the way we are now watching. To test the claim rather than merely assert it, I convened a hundred AI developers to contribute to a decentralized training dataset, with provenance recorded on-chain through smart contracts. Each contribution was attributable. Each license was explicit. Each contributor held a claim on the downstream value. The project drew fifty thousand dollars in grants from impact-focused funds — modest money, but the point was never the money. The point was to demonstrate that the unit of trust in AI is not the parameter count; it is the provenance record.
What that pilot proved is small and enormous at once. It proved that you can train on data whose lineage is auditable, and that when you do, the question of who benefits has an answer a smart contract can enforce. It also proved how far the industry is from caring. The same week I published the pilot's results, the market moved more on a benchmark chart than on the entire question of data sovereignty. The weather system again. This is why I find the gap-is-closing story so frustrating. It is a story about two arks racing across the same flood, and nobody is asking who built the ark or who is being left to drown.
Now let me talk about how this narrative transmits into markets, because I think the transmission is the actual news, and because it echoes something I have been saying since the spot Bitcoin ETFs were approved. When the ETF approval came, a great many people I respect celebrated it as validation. I did not, and I said so at the time: the approval did not bring Bitcoin to Wall Street; it brought Wall Street to Bitcoin, and the direction of that verb matters. What was once a peer-to-peer electronic cash system — a bearer instrument you could hold and spend without permission — became an asset class you rent from a custodian. The vision did not die because it failed. It died because it succeeded on terms that were never its own. The most dangerous outcome for a radical technology is to be adopted by the institutions it was built to route around.
I see the same transmission happening in AI. The China-is-closing-the-gap headline is not primarily a technical claim; it is a market-confidence claim, and it lands on a market that has been priced for a single outcome — that American frontier labs hold an unassailable moat, and that the compute beneath them is a perpetually scarce, perpetually appreciating asset. When a story arrives that muddies that moat, the reaction is not to re-underwrite the technology. It is to reprice the narrative. That is why the coverage lives on a crypto outlet and not in a systems journal. It is a trade, dressed as a trend.
And in a bear market, that matters enormously, because the trades that move first are the trades that hurt most. The protocols bleeding liquidity right now are not bleeding because their technology failed. They are bleeding because confidence is a shared resource, and confidence is being spent on a story that no one can verify. The same week the headline ran, I watched two lending markets and a bridge I have used for years tighten their risk parameters — not because anything broke, but because the mood shifted. Confidence is the only collateral that never appears on a balance sheet, and it is the first to be liquidated.
Here is the part of the analysis that the horse-race frame gets most wrong, and it is where I think crypto builders have the most to learn. The real competitive weapon of the Chinese laboratories is not a benchmark score. It is open weights. Models like Qwen and its peers are released under licenses permissive enough that a startup in San Francisco can — and many do — build on them without ever touching an American API. That is not a marginal event. It is the slow erosion of the pricing power that closed models have taken for granted. When a capable model is free to download, the marginal cost of intelligence collapses, and a business model built on charging per token for that intelligence has to justify itself on something other than scarcity. Some can. Most cannot.
I have watched this exact dynamic play out in DeFi, and it taught me to be suspicious of the way the industry manufactures its own crises. For years, the loudest voices in decentralized finance have described liquidity fragmentation as the central problem of the sector, and every solution to it has conveniently been a new product, a new token, a new layer — each one sold by the same venture capital that funded the last. The fragmentation was never the disease. It was the symptom the industry needed in order to keep selling cures. The same mechanism is now at work in AI: the gap is a product of the framing, and the framing exists to keep a market liquid in a story.
What open weights do is precisely what a functioning commons does — they refuse to let scarcity be manufactured. They make the moat visible, and once visible, a moat that is only a business model becomes very hard to defend. The incumbents' true defenses are not the weights at all. They are the developer ecosystem, the enterprise trust, the tooling integrations, the procurement relationships — the soft infrastructure of adoption, which is slower to erode than any benchmark. Capability can be copied in a quarter. Trust is built in a decade, and it is the only moat that does not depreciate.
Now the contrarian core, the part I have been circling.
There is a version of the gap-is-closing story that is genuinely hopeful — the one where constraint breeds ingenuity, where efficiency innovations propagate outward and lower the cost of intelligence for everyone. I want that version to be true. But I have audited enough systems to know that the hopeful version is usually the one that omits the load-bearing assumption. Here is the assumption: that the constraint stays constant, or eases. It will not necessarily. The entire recent progress of the challengers rests on a specific configuration of export controls, and export controls are the most volatile variable in the equation. A single policy revision — a tightening of the rules on the accelerators or the fabrication equipment beneath them — can reverse the trend faster than any research roadmap can respond. The laboratories that look ascendant today are ascendant on a foundation of borrowed time. This is not a prediction of collapse. It is a reminder that a lead built on someone else's supply chain is not a lead; it is a loan.
And there is a second fragility the headlines never mention: the domestic silicon substitution story. The attempt to replace restricted accelerators with indigenous alternatives is real, and it is the single most important variable in whether the efficiency thesis holds, and I cannot find a single piece of credible reporting on its actual maturity — its yields, its software ecosystem, its usable throughput. I have learned to treat the absence of data as data. When a story is told without its most load-bearing number, the number is probably unflattering to someone.
This is the pragmatism test I apply to every claim, and it is the test I would apply to this headline if it asked me to act on it: show me the provenance, show me the supply chain, show me who can verify the claim independently, and show me what happens to the claim when the constraint moves. A technology race that cannot survive a change in its own inputs is not a race. It is a weather pattern.
Let me now state the counter-intuitive claim cleanly, because I have been building toward it. The prevailing assumption is that the important distance in AI is the distance between the leaders and the challengers — that when that distance shrinks, the world becomes more balanced. I think the opposite is closer to true. When capability diffuses but ownership does not, the result is not balance. It is a more efficient enclosure. Two competing monopolies that benchmark against each other are still two monopolies. Their competition lowers the price of intelligence for a while, and then, once the race resolves or the capital runs out, the survivors consolidate the commons they were fighting over. The horse race is not the antidote to concentration. It is the most effective engine of it ever devised, because it persuades the public that concentration is a contest they are winning.
I hold this view because of what I watched in 2022, when I retreated to a cabin in Yilan and stopped writing about prices. The bear market of that year — first one empire, then everything — was not a story about bad technology. It was a story about what happens when an ecosystem is optimized for the confidence of the crowd rather than the resilience of the system. I spent three months journaling not about charts but about trust, and I emerged convinced that the only durable question in any technology is whether the people living under it have any power over it. That question is not answered by which lab is ahead. It is answered by who holds the keys — to the data, to the model, to the compute, to the ability to verify.
By that standard, both sides of the AI race are failing, and they are failing symmetrically. Neither offers provenance. Neither offers meaningful consent. Neither lets the people whose work trained the weights audit the result. The gap between them is real, and it is closing, and it does not matter nearly as much as the gap between both of them and us.
Let me not end on diagnosis alone, because diagnosis without a path is just another weather report. There is a practical program here, and it starts with refusing the frame.
For anyone building at the intersection of AI and crypto — and there are more of you than the market currently rewards — the work is to make provenance the default rather than the exception. That means training data whose lineage is recorded, licenses that are explicit, and value flows that reach contributors through mechanisms a contract can enforce. It means treating the model not as a product but as a governed commons. It means, when you deploy, choosing the stack that lets a user verify where an output came from, not merely one that produces a fluent output.
I know how abstract that sounds, so let me ground it in what I have actually done. Through The Alignment Circle, the community I founded in 2024 with fifteen thousand dollars of my own savings and a handful of believers, I have mentored fifty core members through the unglamorous work of DAO structuring — transparent treasuries, value-aligned decision-making, governance that survives the departure of its founder. Three of those mentees launched DAOs with community-first models that have outlasted several better-funded competitors. The lesson was not that our governance was clever. It was that governance built around stewardship rather than speculation is more durable in a bear market, because it does not depend on the crowd's confidence. We don't need more users; we need more stewards. The difference is the difference between a customer and a citizen, and only one of them is still there when the price falls.
This is also why the report I helped produce on a protocol I will call Harmony Bridge matters to me more than any benchmark table. My role was not to review the code — I am not that kind of auditor — but to assess whether the protocol's compliance mechanisms honored the sovereignty of its users. The conclusion we reached, and which the governance council adopted, was that true decentralization requires regulatory resilience, not evasion. Privacy-preserving identity is not a contradiction; it is the only version of compliance a decentralized system can honestly offer. That report changed a real protocol's identity design. It was not a headline. It was a fix. That is the unit of progress I trust.
So when the next headline tells you the gap is closing, I want you to ask a different question than the one it wants you to ask. Not who is winning, but who owns the record of what was learned, and who gets to verify it. The laboratories will keep racing, and the race will keep generating weather. But the ledger — the boring, unglamorous, auditable ledger — is the only instrument that lets a person outside the race check the score for themselves. We built not for the peak, but for the valley. And in the valley, where this bear market has left us, the only question worth carrying forward is the one the race was designed to make us forget: trust is the only protocol that cannot be coded, and it is the one we must build anyway.
