Over the past 72 hours, a narrative has rippled through the AI investment community: Kimi K3, a new model from Moonshot AI, marks a turning point. Gavin Baker, CIO of Atreides Management, argues that its arrival signals the beginning of the end for model-level profits, shifting value toward infrastructure and applications. But a single data point from Artificial Analysis reveals a different story: Kimi K3 costs $0.94 per task, compared to GPT-5.6 Terra at $0.55. That’s a 71% premium for a model whose performance—by Baker’s own admission—is merely “close” to the frontier.
This cost inefficiency is not just a pricing anomaly. It is a structural weakness that echoes a pattern I have observed repeatedly in blockchain Layer2 scaling. Dozens of rollups and validiums have launched, each claiming to be the ultimate Ethereum scaling solution. Yet they fragment liquidity into silos, forcing users and developers into a game of hopscotch that raises, not lowers, total system cost. The narrative that competition inevitably drives down costs is only half the truth; it can also drive up complexity and friction.
Tracing the hidden vulnerabilities in the code—whether in a transformer model or a smart contract—requires understanding where costs actually accumulate. In Kimi K3, the token efficiency shortfall means that every inference consumes more compute than a comparable GPT request. The root cause is opaque: is it an architecture that trades parameter count for precision? A suboptimal inference engine lacking FP8 quantization? Or simply a smaller team’s inability to match OpenAI’s engineering depth? Without answers, the model remains a prototype, not a product.
Baker’s thesis—that model profits will compress and value will flow to power, chips, data centers, and software—is intellectually appealing. It mirrors the “pick and shovel” logic of the gold rush. But it rests on an assumption that competition among models will be intense enough to erode margins, a status that Kimi K3 does not yet achieve. In blockchain, we saw a similar illusion during the 2021 DeFi summer: dozens of forks launched, promising lower fees and higher yields, but the majority bled liquidity faster than they could attract it. The survivors were those with actual efficiency gains, not just marketing narratives.
Let us examine the cost structure more carefully. Kimi K3’s per-task cost of $0.94 is derived from the model’s total compute requirement divided by an assumed batch size and average prompt length. In practice, a developer sending a long-context query (e.g., 128k tokens) could see costs spike significantly higher. Meanwhile, GPT-5.6 Terra benefits from years of optimization, including custom hardware co-design with NVIDIA and a proprietary inference engine that reduces latency and improves throughput. This is not a level playing field. It is a contest between a specialized entrant and a vertically integrated giant.
Redefining what ownership means in the digital age—in AI, as in blockchain—requires acknowledging that infrastructure vendors capture the majority of value when the application layer fragments. NVIDIA’s GPU sales and Microsoft Azure’s cloud revenue have soared while even successful model companies struggle to break even. Baker’s own portfolio likely reflects this: he invests in power, chips, and cloud, not in OpenAI or Anthropic. His commentary is thus a sophisticated form of position disclosure, not an objective technical judgment.
From the blockchain practitioner’s lens, Kimi K3 resembles a new Layer2 that claims to be faster and cheaper but, upon inspection, requires users to lock capital for a week to achieve those benefits. The user-centric cost analysis is damning: if your average transaction (or inference) costs 71% more than the incumbent, you are not a turning point—you are a curiosity. The turning point arrives only when a new system demonstrably reduces total economic cost for at least one significant use case without introducing unacceptable risk.
Quietly securing the layers beneath the hype means focusing on the protocol mechanics that determine true resilience. In Kimi K3, the vulnerability lies in its cost curve. If the model cannot improve its token efficiency by at least 50% in the next six months, it will remain a niche product for users who prioritize absolute capability over price—such as researchers or compliance teams who cannot use OpenAI due to data sovereignty concerns. Similarly, many Layer2s live or die by their ability to reduce gas costs for DeFi users without compromising security. Those that fail, fade.
Baker points to “open models” as the real inflection point. Here, the parallel to blockchain is sharp: open-source models like Llama 3 have a community-driven optimization cycle that can rapidly reduce costs. In blockchain, permissionless protocols allow anyone to fork and improve the code, creating a Darwinian environment that eventually yields efficient designs. Building trust through rigorous, unseen diligence—whether auditing a smart contract or benchmarking an AI model—is the only sustainable path to credibility in either industry.
The contrarian angle is uncomfortable but necessary: Baker’s “turning point” may be a self-serving prophecy. If too many investors believe model profits are doomed, they will stop funding model companies, reducing competition and allowing OpenAI and Anthropic to solidify their duopoly. This is the opposite of the intended effect. In blockchain, the narrative that “Layer2s will make Ethereum obsolete” led to a proliferation of underfunded projects that ultimately failed, strengthening Ethereum’s position as the settlement layer. The ecosystem is now more centralized on Ethereum’s core than it was before.
Blind spots abound. Baker mentions “product, toolchain, and internal models” as OpenAI/Anthropic’s moats, but he does not quantify how strong those moats are. ChatGPT has hundreds of millions of users; Claude Pro has a growing enterprise base. Switching costs are high. Kimi K3, even if free, would struggle to overcome network effects. In blockchain, we saw this with EOS—high performance but negligible user adoption because the existing Ethereum developer ecosystem was too entrenched.
Takeaway: Kimi K3 is a signal, not a turning point. It signals that the cost of training a frontier-level model is falling, but the cost of running one efficiently is not. Until token efficiency improves by an order of magnitude, the model adds noise, not value, to the AI ecosystem. The same vulnerability applies to countless Layer2s: they launch with fanfare but deliver marginal improvements at best. The infrastructure layer—power, chips, cloud—continues to accrue value regardless, because it is indifferent to which model or which chain wins. The next real turning point will come from an open model that matches GPT-5.6 Terra’s cost while matching its quality. Until then, the narrative remains a reflection of investor positioning, not technical reality.