Jason A. Hoffman, PhD
The previous post, On Binary Narratives, used the coverage of Moonshot AI’s Kimi K3 release to show how a story about many interacting variables gets compressed into one variable with two values. That post was about the method. This one applies it to the second binary inside the same story: open versus closed.
The sharpest form of the question is one I have been asked directly: is a strong open-weights model actually good for Anthropic and OpenAI? Asked as a binary, the question has no answer. Enumerated, it has a useful one: seven mechanisms, some helping the closed labs and some hurting them, with a net that depends on two conditions that are measurable and mostly unmeasured.
What Open Means Here
First, two precisions that the coverage skips. “Open source” is doing a lot of work in these stories, and most of these releases are open weights, which is narrower. Open weights means you can download the trained model, run it on infrastructure you choose, inspect it, and modify it, subject to the license. It does not mean you get the training data, the training code, the infrastructure, or the ability to reproduce the model. Moonshot’s chief executive, Zhilin Yang, described the offer accurately in a talk at GTC 2026: you can deploy anywhere, on your own servers or any cloud, and you can access every bit of the weights instead of using a black box.
The second precision is temporal. As of July 20, K3 is available through Moonshot’s products and API, but the weights have been promised for release by July 27 and the full technical report is still forthcoming. K3 is currently a hosted model with an announced open-weights release. The distinction matters because neither the downloadable artifact nor claims that depend on self-hosting it can yet be tested.
Downloadable will not mean cheaply or casually runnable. Moonshot recommends deploying K3 on supernodes with 64 or more accelerators. That is not a spare-server installation; it is a datacenter-scale system with high-bandwidth networking, memory, power, cooling, software, and operators. Open weights make Moonshot optional as the server. They do not make the serving infrastructure optional.
The offer is real and valuable: inspectability and the option to choose the operator. It is not cheap execution, automatic workload portability, or the capability to make the next model.
Why Give the Model Away
A company that spends heavily to train a model and then publishes the weights is not confused about money. Openness is a strategy, and it is rational under specific conditions: when you are behind on distribution and need adoption more than margin; when you are recruiting researchers who want their work public; when you want to set the standard that the tooling ecosystem builds around; when you want to deny rents to a competitor whose paid tier you can match; or when a sponsor with objectives beyond model-layer profit is paying for it. Moonshot’s May round was backed by China Mobile, the state telecom company, and the investment arm of Meituan. Those are not investors underwriting API margins.
What openness is not, so far, is a profitable business for its practitioners. The Times reported that the difficulty of converting global adoption into revenue is fracturing some of China’s top labs, and that Alibaba is steering customers toward paid models. Open weights need an adjacent business or a strategic sponsor to pay for them. That fact belongs in any analysis of who the strategy serves. So does the state’s stated posture: the release coincided with an address by Xi Jinping casting China as the champion of the open approach, telling the Shanghai conference that AI development “should not be a solo performance by a single country.” A sponsor with those objectives measures the return on open weights in worldwide adoption, developer dependence, and standards position, not in model-layer margin. Judging the strategy a failure because it loses money at the model layer is measuring the wrong business.
The Mechanisms
It removes the middle of the market. Anything a free model does at parity cannot be sold at a premium, so open weights destroy the profit pool at every tier they match. The companies that die are not the frontier labs; they are the second-tier closed labs whose whole business sat at those tiers, and they are also the challengers who would have used mid-tier margins to fund a climb toward the frontier. The ladder from below loses its rungs. What remains as a paid market is the capability the open ecosystem cannot yet match, which is where the frontier labs live. Open weights narrow the field to the frontier and free, and they clear out everyone in between, including, so far, the open publishers’ own income statements.
It subsidizes the compute-rich. The efficiency research comes out in public. Moonshot published the first demonstration that the Muon optimizer scales to LLM training, with roughly a two-times gain in token efficiency over AdamW. A technique like that is worth more to whoever has more compute and more data to apply it to. The Times reported that Moonshot raised $2 billion in May while Anthropic raised $65 billion in the same month. A two-times multiplier applied to the larger training budget widens the absolute gap even as it narrows the relative one. Publishing the technique transfers the innovation; the export controls keep the constraint that forced it. The condition on this mechanism is that the published technique has to be additive to whatever the closed labs already run internally, which outsiders cannot know. The same subsidy also runs the other way: Anthropic and OpenAI have accused Chinese companies of training on harvested outputs from their systems, which, if true, means the closed labs’ capability partially funds the catch-up. The Times also quoted Graham Webster of Stanford saying the recent advances cannot be explained by distillation alone. None of this makes the openness altruism. The open strategy buys Moonshot distribution, researchers, and a worldwide developer base, which is what its sponsors are paying for. The subsidy to the compute-rich is a side effect of a strategy aimed at something else.
It compresses the monetizable window. This is the mechanism that actually threatens the closed labs, and it is not “your models are less valuable.” Webster told the Times that estimates of China’s distance behind the United States cluster around six months. Read that as a business quantity rather than a race quantity: six months is the approximate premium-pricing window for each new capability tier before a free equivalent arrives. The training investment for each generation has to be recovered inside that window, and training investments are growing. If the window shortens faster than the market grows, the frontier business becomes a treadmill: enormous recurring capital expense for a shrinking period of exclusivity. Whether the window is shortening is measurable, generation by generation.
It anchors every negotiation. Even where an open model is never deployed, its existence reprices the closed alternative. The Times reported that developers adopted Z.ai’s GLM-5.2 largely because it was cheaper, arriving just as American businesses were looking to cut AI spending. An enterprise buyer who can credibly say “the free model is close enough” pays less for the closed one. Price pressure at the matched tiers arrives through the option, not only through the substitution.
The word credibly carries the mechanism. Nate B. Jones’s model-replacement test identifies the operating constraint: the alternative matters only if the work can move without being rebuilt. Moonshot warns that K3 expects the agent harness to preserve its complete thinking history and that generation can become unstable if a session is switched to K3 from another model; it recommends a verified harness such as Kimi Code. Moving an agentic workload is therefore not just changing an API endpoint. The history format, prompts, tools, evaluations, and surrounding workflow have to move too. Open weights make the model portable. A buyer has a real negotiating option only to the extent that the rest of the system is portable as well.
None of this is hypothetical at my own desk. Within a week of the release, K3 had taken a seat in my writing workflow, including on drafts of this post, because reviewing prose in a terminal is about the most portable workload there is: plain text in, plain text out, no harness to rebuild. It displaced a free tool, not a paid one, and the workloads wired into agent harnesses stayed where they were. One desk is not a market, but that is the mechanism: good enough to switch to, portable enough to switch.
It expands the total market. A free capable model makes AI economical for uses that could not carry the cost before. Some of those uses stay on the free tier forever. Some grow into workloads that need reliability, integration, longer horizons, or the newest capability tier, and those graduate to paid. The shared effect is larger than the graduation effect: the tooling, the agent frameworks, the serving stacks, and the developer habits built on open models carry across providers even when individual tuned workloads do not, and the closed labs plug into that matured ecosystem rather than building it alone.
It moves the politics. A near-frontier model promised for open release by a Chinese lab strengthens the American buildout argument and gives the closed labs their most effective line in Washington: containment of released weights is impossible, so policy attention shifts toward compute, deployment, and capability thresholds rather than restraint of domestic research. At the same time, enterprises and governments with provenance, security, and liability concerns default to the closed suppliers, because “we know exactly who trained it, on what, and who is accountable” is a product feature that open weights structurally cannot offer. For government and critical-infrastructure buyers there is a harder version of the same point: the question of whose jurisdiction a training lab answers to does not go away because the weights are free and inspectable. Treating that question as paranoia mistakes procurement discipline for politics.
It feeds the complements. Open models are unambiguously good for whoever sells what models run on. Every open deployment is demand for accelerators, memory, networking, power, and cloud capacity, with no model-layer margin attached. It is worth noticing that the slide Moonshot’s chief executive presented to argue that open models are reaching the frontier was, by his own description, taken from Jensen Huang’s CES talk. Nvidia’s enthusiasm for open models is the oldest strategy in the industry: commoditize your complement. Everyone in this stack is running some version of that play. The model labs want compute cheap and abundant. The chip and cloud vendors want models cheap and abundant. Open weights are a move in that game, and asking who benefits from openness mostly reduces to asking whose complement just got cheaper.
What Decides the Net
For the closed frontier labs, the mechanisms above net out on two conditions.
The first is the durability of the frontier gap. Every helpful mechanism assumes there is a paid tier the open ecosystem has not reached. If the gap holds at some interval, the open tier keeps clearing out the middle, expanding the market, and maturing the ecosystem underneath a premium the frontier labs still own. If the gap goes to zero and stays there, there is no premium tier, and no amount of ecosystem growth compensates the companies whose product was the premium tier.
The second is revenue mix. “Anthropic and OpenAI” is itself a small binary error, because the two are exposed differently. A consumer subscription business with a brand, a product surface, and daily habits is not directly attacked by downloadable weights; an open model does not cancel anyone’s chat subscription. Raw API volume at tiers an open model matches is attacked immediately when the workload can move. Enterprise contracts sold on reliability, integration, security posture, and accountability sit in between. Each company’s exposure is the fraction of its revenue at tiers an open model both matches and can credibly replace, and that fraction is different for each and changes every quarter.
Two Analogies, One Question
The history of open-versus-closed offers two endings, and they turn on exactly the first condition.
In one, the gap persists. Android was the free operating system that was supposed to destroy Apple. It destroyed Nokia, BlackBerry, and Windows Phone, expanded the smartphone market by billions of users, and left Apple holding the premium tier of a vastly larger market. Open killed the middle and enriched the top.
In the other, the gap closes. Linux reached parity with commercial Unix, and Sun, whose premium was the thing Linux equaled, did not survive it as an independent company. The value did not disappear; it moved to the layers around the commodity, which is why Microsoft, having lost the server operating system battle, ended up monetizing Linux quite happily through Azure.
Which ending applies to frontier models is not a matter of opinion. It is the measurable question from the previous post, made specific: track the length of the premium window each generation, the substitution rates at each tier, the price curves, the premium the newest tier commands over the one below it, and where the serving volume physically lands. The binary “open beats closed” has no answer. The enumeration has one arriving on a schedule, roughly one data point per model generation.
So: is the open-weights approach helpful to Anthropic and OpenAI? It removes their cheapest competitors, matures their ecosystem, subsidizes their training efficiency, and strengthens their political position, while compressing the window in which each of their models earns a premium. Both directions are real, and neither makes the strategy behind the openness friendly. The openness is aimed at adoption and standards, a more patient target than the frontier labs’ revenue line. The companies that should be unambiguously worried are the ones in the middle, and the companies that should be unambiguously pleased are the ones selling the compute. For the frontier labs, the answer is a race between a shortening window and a growing market, and that race has a scoreboard, which is more than can be said for the one in the headlines.
Leave a Reply