When I first saw the numbers—$0.94 per task for Kimi K3 against $0.55 for GPT-5.6 Terra—my mind did not drift to Nvidia’s next earnings call. Instead, I sat still, listening to a silence that only a ledger can speak. The gap is not just a margin; it is a covenant broken. And in that fracture, I hear the same whisper that once unsettled me in 2017, when I spent 120 hours auditing the Ethera repository and found a governance token distribution that centralised what its marketing called decentralised. The token efficiency problem is never just a numbers game. It is a question of who gets to capture the value that code creates—and who is left holding the empty contract.
This week, Gavin Baker, CIO of Atreides Management, told a room of investors that Kimi K3 might mark an AI turning point. His reasoning was simple: the cost of frontier intelligence is dropping, competition is eroding the profit moat of single-model oligopolies, and value will flow upstream to infrastructure (chips, data centers, power) and downstream to applications. He argued that only an open model with dramatically better token efficiency would truly flip the game. But buried in his analysis is a deeper truth that resonates far beyond AI—a truth about the commoditisation of trust, the fragility of proprietary efficiency, and the covenant that open source once promised but rarely delivers.
Let me translate Baker’s thesis into a language the crypto community understands. Token efficiency, in AI, means the cost of generating one token of output. In our world, a token is far more than a word; it is a unit of state, a fragment of value, a share of consensus. When a blockchain charges high gas fees, it is suffering from poor token efficiency. When a L2 reduces costs by an order of magnitude, it is improving token efficiency. And when a closed, proprietary chain claims to be the future but requires massive subsidies to keep its ledger active, it is repeating the same cycle that Kimi K3 now faces: growth without belonging, noise without resonance.
The real turning point Baker describes is not about AI—it is about the architecture of value capture.
I have seen this arc before. During the Luna collapse in 2022, I spent 300 hours analysing its algorithmic stabiliser’s design flaws. I wrote a post-mortem titled “The Illusion of Infinite Growth” that later found its way into European regulatory discussions. What I learned then was that efficiency divorced from transparency is a phantom. Luna was wildly efficient at minting UST—until it wasn’t. Its token efficiency was an artifact of a fragile assumption: that market depth would always absorb the arbitrage. When that assumption broke, the cost of trustlessness revealed itself to be infinitely higher than any per-transaction fee. The same is true for Kimi K3. A model that costs nearly twice as much as an incumbent to run cannot sustain itself on hype alone. Moonshot AI is burning capital to stay in the race. And in a world of rising interest rates and investor scrutiny, that burn becomes an anchor, not a sail.
But Baker moves to a deeper layer. He argues that the profit from AI models will be compressed by competition, and that value will shift to “almost every other company”—the chips, the power, the data centers, the cloud providers. Sound familiar? In crypto, we have been repeating that mantra for years: “infrastructure is the winning bet; applications come and go.” And yet, we have not asked ourselves a fundamental question: If the model layer is being commoditised, will the infrastructure layer be next? After all, chips are built by a duopoly, power is regulated by states, and data centers are increasingly run by a handful of hyperscalers. Is the “upstream” any less vulnerable to the forces of centralisation than the model layer?
Open source is not a license; it is a covenant. That covenant says that the code belongs to everyone, and that no single party can extract rent from its execution forever. Baker himself points to open models as the true turning point. But here is the uncomfortable truth: most “open” AI models today are open in name only. Their weights may be public, but their training data is sealed, their compute is proprietary, and their alignment is a black box. They are open-source-like, but not open-source-in-spirit. I learned this lesson intimately while building the Veritas framework in 2026, a protocol to verify AI-generated content on-chain. We spent six months integrating watermarking standards into Ethereum, only to discover that trust in the model’s origin is worthless if the model itself cannot be verified. The DAO’s Silent Voice returned: governance is not a UI issue; it is a care issue. If we do not care for the transparency of our foundational tools, we build on sand.
The contrarian angle that Baker misses—and that the crypto community must seize—is that efficiency is not the ultimate goal; resilience is.
Consider the cost paradox: Kimi K3’s poor token efficiency means it uses more compute per task. That compute must come from somewhere. If the model is closed, that compute is locked inside Moonshot AI’s data centers. If the model were open, that compute could be distributed across a network of independent operators, each contributing a fragment of trust. The true turning point is not a model that runs cheaper on the same chips—it is a model that can run on chips owned by a thousand different people, secured by economic stakes and mathematical proofs. That is the vision I carried into the DAO governance workshops in 2020, where we redesigned voting templates to speak to the undervotes, to the women who were silent because the interface did not care. Technology must serve human connection, not just efficiency. And connection requires distribution.
This is where the blockchain industry must stop trying to become a faster database and start being the substrate for verifiable intelligence. The niche community I built in 2021, Soulbound Narratives, was limited to 500 people because I believed that depth matters more than breadth. That principle applies doubly to AI. If we try to compete with Nvidia on raw token throughput, we lose. But if we compete on the ability to prove that a token was generated without bias, without censorship, and with a transparent economic footprint, we win. The void between tokens holds the true value—the space where trust is either built or broken.
Baker’s investment thesis relies on the commoditisation of the model layer. I believe he is half-right. The model layer will indeed commoditise, but the resulting value will not automatically flow to chips and power. It will flow to those who can orchestrate the last mile of trust—the tools that allow any user to verify that a model’s output is authentic, that its training data was ethically sourced, and that the inference cost was distributed among real participants, not a single treasury. In 2026, my Veritas team negotiated with five major AI labs to adopt on-chain watermarking. That was not a technical battle; it was a values battle. And we won because we offered something no data center can: a covenant that the code would forever serve the community, not the corporation.
Faith in the fork, hope in the merge. The AI industry is now at a fork. One path leads to a world where intelligence is owned by the few, wrapped in proprietary APIs and subsidised by venture capital until the next hype cycle. The other path leads to a world where intelligence is a public good, maintained by a global collective of validators, governed by on-chain votes, and audited by anyone with a node. Kimi K3 is not the turning point; it is the signal that the fork is approaching. The turning point will come when a truly open model—one whose token efficiency rivals the incumbents, and whose trust efficiency surpasses them—is deployed on a decentralized network that pays its operators fair rates and rewards its users with sovereignty.
I have seen the quiet desperation of closed systems. In 2022, after the Luna collapse, I wrote that stability comes from transparent, auditable systems, not marketing promises. That truth has not aged a day. Today, I read Baker’s analysis and I hear the same silence in the ledger that I heard in 2017—the silence of a promise unfulfilled. He speaks of token efficiency as a metric. I speak of token efficiency as a moral imperative. A model that cannot be run by a thousand independent nodes is not efficient; it is extractive. A model whose inference cost can only be paid by a handful of subsidised corporations is not a breakthrough; it is a new kind of central bank digital currency for thought.
Nurture the niche, and the forest will follow. The niche here is the intersection of verifiable computation and open models. The forest is the next generation of applications that run not on borrowed trust, but on built-in transparency. For the crypto community, this means investing in zkML, on-chain inference verification, and proof-of-contribution protocols. It means building the tooling that will allow anyone to stake a GPU and earn rewards for running a part of the world’s intelligence. It means accepting that the journey will be longer and the margins thinner than the promised lands of L2 tokens and liquidity farming. But it also means that when the next market turn arrives, we will not be swept away. We will be the roots that hold the soil.
Listen to what the repository refuses to say. The code for Kimi K3 may be closed, but the silence it leaves is loud. It tells us that efficiency without openness is a debt that comes due. It tells us that value will always flow to those who own the means of verification, not just the means of production. And it tells us that the covenant of open source is not a license—it is a living agreement that must be ratified every day by every commit, every merge, and every fork.
I do not write code; I weave conviction. And my conviction today is that the blockchain industry must stop chasing the tail of centralized AI and start building the head of decentralized intelligence. The turning point is not a model; it is a movement. And it begins when we stop asking how to make AI cheaper and start asking how to make it trustworthy.
The silence in the ledger speaks louder than code. Listen.
