The Last License War: The Clouds Never Paid — So What Is Moonshot Collecting?

Key takeaways

Pick up where the last deep dive left off. Our July 28 deep dive took apart the Kimi K3 license: the weights are free to download, but any company that "operates a model inference service" and books more than twenty million US dollars in group revenue must first sign a separate agreement with Moonshot — openness handles distribution, the license handles collection. That piece left one question unanswered: can this wall actually collect? Kevin Xu, the investor-circle writer who proposed the analogy, says this is not a new invention: open-source database companies fought exactly this war against the cloud giants in the 2010s, when the accusation of the day was "strip mining" — clouds packaging other people's freely open-sourced software into paid services while returning almost nothing to the makers. But his thread stopped before it reached how that war ended. This piece does three things: fills in the second half of that war company by company, pulls K3's first-month market prices vendor by vendor, and re-books the ledger of "who captures the value of Chinese open weights." Three conclusions:

>

- How the last war ended: the clouds never once paid a license fee for using open-source code. Five companies changed their licenses to set up tollgates. AWS's answer was to rewrite a compatible engine (the MongoDB case) or fork out a free version (the Elasticsearch and Redis cases); Google and Microsoft did come back and sign, but what they signed were managed-service partnerships, with revenue splits never disclosed. The original vendors recovered by going into the cloud business themselves — MongoDB's own cloud service Atlas climbed from 22 percent of revenue to 74 percent — not through license terms. The two companies that built the highest walls (Elastic, Redis) tore them down after 43 months and 13.5 months respectively, and their official reasons read like the same script: the rivals' replacements (OpenSearch, Valkey) had taken root, and the market no longer needed the original vendor's copy.

- K3's first-month market reading: third-party prices are pinned to the official rate. We pulled the K3 quotes of 12 third-party inference services; 8 match Moonshot's official price to the cent. In the previous K2 generation, third parties would still undercut by 12 to 33 percent; with K3, the discount space has gone to zero. Meanwhile, none of the three hyperscalers sells K3 tokens under its own name: AWS's and Google's catalogs carry only the previous generation, and Microsoft lets Fireworks resell while skimming a 10 percent markup. But this reading cannot settle its own cause: are third parties paying license fees and unable to cut, or does nobody want a price war in the first place (US hosting and data non-retention being the real selling points)? Even the person who raised the observation added the second explanation himself 93 minutes later.

- The "value shifts to America" ledger has to be kept in two layers. The license's own carve-out shaves "merely renting out GPUs" from its reach — so at the rack layer, the money stays with American clouds as before; the license's collection point opens only at the layer that sells model service by the token, and that layer is currently small in magnitude with zero rate disclosure. Stack the history and the readings together, and the provable function of K3's license wall is defending its own API's price floor and holding the whitelist switch over the channel — not collecting rent.

>

Why this is worth 20 minutes: this piece splits the denominator behind the daily brief item on K3's license tollgate — "value shifting to US neoclouds" used to be one blended account, and it is now booked separately as a rack layer (untouched by the license) and a token-service layer (where the collection point sits); we tested the inference "prices pinned means someone is paying license fees," and it did not hold up (its own proposer supplied the geography explanation, and we found the lone counterexample daring an 11 percent discount); and we wrote down five dated conditions for what would prove this wrong — the nearest one comes due next week.

The first half of the last war: five companies, one script

Start by finishing the thread Kevin Xu left hanging. In an earlier era, he wrote, there was constant conflict between open-source infrastructure software makers — MongoDB, Confluent — and the hyperscalers: "The hyperscalers, AWS in particular, were accused of 'strip mining' open source, provide it as a cloud service to sell to their customers without paying the original maker a dime" (Kevin Xu / X).

That conflict left five announcements, and we pulled every one of them. October 2018: MongoDB invented a license of its own, the SSPL, and stated its reason for opening fire plainly: "Once an open source project becomes interesting, it is too easy for cloud vendors who have not developed the software to capture all of the value while contributing little back to the community" (MongoDB). Two months later Confluent followed, but only moved the perimeter components it maintained itself behind a no-cloud-service license — Kafka proper stayed untouched (Confluent). January 2021: Elastic named AWS outright in its announcement: "we have lived this experience firsthand, from our trademarks being misused to outright attempts to splinter our community" (Elastic). August 2023: HashiCorp moved its entire product line to the BSL. March 2024: Redis fired the last shot, and said it plainest: "the majority of Redis' commercial sales are channeled through the largest cloud service providers" — and, under the new license, "cloud service providers hosting Redis offerings will no longer be permitted to use the source code of Redis free of charge" (Redis).

Five announcements, one script: the clouds didn't build it, took the value, and gave nothing back to the community. Set that script beside Moonshot's position — weights released for free, and the parties best positioned to profit from them being the services that wrap them into a paid API — and the shape of the problem is nearly identical. The first half of Xu's analogy holds. But all of the analogy's predictive power sits in the second half: how that war ended.

The second half: three roads, and none of them was a license fee

Dig out how each of the five cases ended, and the clouds' responses sort into three roads — none of which is "pay the license fee like you're told."

Road one: rewrite. AWS's answer to MongoDB was not to violate the SSPL but to spend more than two years rewriting a database of its own that is compatible with MongoDB's interface (DocumentDB, live January 2019) — a license reaches your code; it cannot reach "we don't use your code, we only imitate your interface" (AWS). The SSPL never blocked it for a single day.

Road two: fork. For Elasticsearch, AWS took the last open-source version before the license change and raised it as its own, named it OpenSearch, and turned the accusation back around: "Elastic's assertions that the SSPL is 'free and open' are misleading and wrong" (AWS). With Redis, the play had matured: the original core maintainers walked out together; AWS, Google, and Oracle endorsed simultaneously; and the foundation-governed Valkey was born — 150 contributors and nearly 50 participating companies inside a year, with AWS pricing the Valkey flavor of its own managed service 20 percent below the Redis flavor (node-based) and 33 percent below (serverless) (Linux Foundation; AWS). Each generation of fork hit harder than the last: from a single cloud raising its own, to foundation governance, multi-vendor feeding, and a price war thrown in.

Road three: sign — but not a license fee. This road does exist, and it deserves the close look, because it is the only real precedent for "the clouds came back and paid." In April 2019, Google Cloud signed seven open-source vendors at once (including MongoDB, Confluent, Elastic, and Redis Labs): their managed services were integrated as native services on Google's cloud, with unified billing and unified support (Google Cloud). Microsoft walked the same road: inside Redis's 2024 relicensing announcement sits a Microsoft executive endorsement, exchanged for exclusive features. Faced with the same license change, AWS and Google went off to feed a fork while Microsoft chose to sign — the cloud camp split. But look carefully at what was signed: commercial partnerships to co-operate managed services. Google's official announcement does not contain the words "revenue share" anywhere, and the split has never been disclosed; the oft-quoted line that open-source companies should be "fairly rewarded" comes from reporter coverage, not from any company document. That is a different thing from paying a license fee to use someone's code.

So what did the original vendors live on? Going into the business themselves. The year MongoDB changed its license, it booked $267 million in revenue, with its own cloud service Atlas at 22.5 percent; by the latest fiscal year, revenue was $2.46 billion with Atlas at roughly 74 percent (computed from the earnings supplement), and its market capitalization had gone from $1.6 billion at listing to $27.7 billion (MongoDB). The cleanest curve in this entire history is driven by "build a cloud service yourself and win at it" — it owes nothing to the SSPL. The SSPL didn't even stop DocumentDB. The cautionary tale runs the other way too: HashiCorp adopted the BSL in August 2023 and, roughly a year and a half later, sold to IBM for $6.4 billion (deal completed 2025-02-27) — against a first-day market cap of nearly $15 billion at its 2021 IPO, a shrinkage of more than half (TechCrunch); Elastic grew 42 percent the year it changed licenses and 17 percent five years later — no cliff, but no pricing power regained either.

Last come the teardowns. Elastic returned to an open-source license after 43 months; its founder's account: "Amazon is fully invested in their fork, the market confusion has been (mostly) resolved, and our partnership with AWS is stronger than ever" (Elastic). Redis held out only 13.5 months, and its CEO's confession was franker: "This achieved our goal—AWS and Google now maintain their own fork—but the change hurt our relationship with the Redis community" (Redis). RedMonk, the analyst firm that has tracked this territory for over a decade, closed the books with this verdict: such licenses "are not measurable in a statistically significant way. They remain extremely uncommon and are not trending" — and Valkey was the market's first real pushback (RedMonk).

Compress the second half into one sentence: the license wall never obtained "pay or else" leverage; what it obtained was the power to force the rival to pick a side. The money was earned back by first-party services. And the wall's measured half-life ran 13.5 to 43 months — with the teardown triggered, both times, by the rival's replacement taking root.

Which lesson K3 took: three switches in the fine print

Read K3's license with that ending in mind, and Moonshot looks like a company that studied the case law — all five predecessors opened first, watched the clouds eat their channel for years, and only then built the wall; Moonshot wrote the wall into the first version of its license. But the design is finer than "set a threshold." Three switches are worth reading word by word (license text verified by us against the raw file Moonshot published on Hugging Face).

The first switch is the definition of "Model as a Service": only a service that lets a third party exercise "meaningful control" over the model's inputs, parameters, or training data counts — and two categories are expressly excluded: end-user products that merely embed model capability inside specific features, and the mere relaying of requests to models hosted by others. Note who this definition shaves off: clouds that merely rent out GPUs are not in it. Rent a rack and stand K3 up yourself, and whether you cross the line is your business, not your landlord's; and if you are only serving yourself (no third-party access), a separate exemption waves you straight through. The second switch is how the threshold is mounted: the twenty million dollars counts aggregate group revenue, not the revenue of the model-service line — so any company of consequence that opens a model-service business clears it automatically. The real gate is not the amount; it is the yes/no question of whether you operate that business at all. The genuinely unwritten item is "certified inference partners." The exemption clause says use through Moonshot's official products or its "certified inference partners" is free of all the obligations above; but who a certified inference partner is, the text never defines. Who qualifies, who certifies, whether certification costs money — none of it is written. Which leaves Moonshot holding a fully discretionary whitelist switch over the channel.

One more thing the text settles by itself: the branding obligation — that very large products must prominently display Kimi K3 — is a zero-revenue clause. Xu's example makes it vivid: suppose the US carrier AT&T fine-tunes K3 into its customer-service system — plenty big, plainly commercial — but as long as it shows something along the lines of "powered by Kimi 3," in his words, "AT&T does not need to pay Moonshot a dime" (Kevin Xu / X). So the license's entire collection function hangs on the model-service clause — and that clause's rate appears nowhere in the text.

Month one: prices pinned to list — and what that reading cannot settle

The day the license went public, Jamin Ball of the investment firm Altimeter posted the first observation: the third-party inference providers Baseten and Fireworks were quoting K3 at exactly Moonshot's official API price — $3 per million input tokens, $15 per million output. "There was lots of speculation that as weights were released other would serve the model at much lower costs. Not yet!" (Jamin Ball / X)

We re-pulled a larger sample. Through the machine interface of the model-routing platform OpenRouter, we pulled all 12 K3 endpoints: 8 quote Moonshot's official price to the cent ($3 / $15 / $0.30 cached — identical all the way down the decimals); 3 are "fast" tiers priced 50 to 72 percent above; and exactly one small provider, called Morph, dares to quote 11 percent below. The control group is more persuasive: on the previous K2 generation, third parties (such as DeepInfra) quoted 12 to 33 percent below official — back then, they haggled; on K3, DeepInfra has not even listed the model, and the discount space has collapsed to zero across the board. Meanwhile, Moonshot's own list prices differ by only 1 percent between its China and international sites — no two-track pricing. The last deep dive speculated that the license threshold props a price floor under the official API; this month's machine readings are consistent with that speculation: the floor held through month one.

The three hyperscalers' posture is the same direction, read another way. AWS's per-token catalog (Bedrock) lists the previous-generation K2.5 — no K3; what AWS published instead, on July 30, was an official guide teaching you to stand K3 up yourself on its own compute; Google's equivalent catalog likewise carries only the previous generation, and all three paths it offers for K3 are self-hosting; Microsoft did put K3 on the shelf, but under the name "FW Kimi K3": the legal relationship is pushed onto the inference provider Fireworks, with Microsoft skimming a 10 percent channel markup (Microsoft). Read these three in two groups: AWS and Google are pure landlords — they sell compute, and they even teach you to self-host; Microsoft is a reselling counter — the sign is up, but the legal relationship belongs to Fireworks. What they share: not one of them is willing to carry the K3 license relationship under its own name. And the license's carve-out happens to exempt exactly the landlord.

Here we have to brake hard: this set of readings cannot settle its own cause. "Prices pinned" admits at least three mutually independent explanations. First: the third parties really are paying license fees, their margins are thinner, so they cannot cut — Xu's first explanation, along with his flat assertion that the day-0 cohort had all "already" signed revenue-sharing contracts (his unilateral claim, confirmed by zero providers; we carry it as a hypothesis). Second: nobody wants a price war, because the competitive axis is not price. The same Xu supplied it himself 93 minutes later: "Pinging Kimi's servers in China is a non-starter for many western enterprises. Pinging baseten's or Fireworks's infrastructure located in the US with better GPUs is totally ok. No need to compete on token cost" (Kevin Xu / X). Market behavior supports this reading: Fireworks sells a US-hosting-only compliance tier; Vercel, another third-party channel, headlines its K3 listing "ZDR and US-based providers" (ZDR: zero data retention); and Microsoft's entry is literally named "Data Zone" — three independent channels all selling location and data non-retention, not price. Third: it is day 11 of a new product; everyone anchors on list price and waits, and tiers form over 60 to 90 days — K2-era discounts did not appear on day one either. The three explanations lead to entirely different conclusions, and today's public evidence cannot separate them. Morph, the only provider daring to discount, is the counterexample most worth watching — if it comes to no harm, the wall has no bite; if it disappears or reprices upward, that is the wall moving.

This ledger takes two layers

Now back to the question this piece actually has to rule on. Two opposing judgments are on the table, and this piece rules between them. On one side is the commentator Teortaxes's "shift" thesis: Chinese open weights are not dumping — dumping means selling below price, and a party giving weights away is not selling at all; what actually happens is value moving house inside the American stack, from the US labs that sell tokens to the US clouds that sell machine time — "That's… not a big blow to the US" (Teortaxes / X). On the other side is Xu's "collection point" thesis: the license names precisely those American clouds; cross the threshold and they have to come back and sign with Moonshot — so the shift no longer terminates entirely inside American borders.

Layer the close reading of the text and the first month's readings on top, and this piece's ruling is: both are right, each about one layer, and both are narrower than as originally stated. The shift thesis was written on July 20 — seven days before the license went public; it holds for a world in which the license did not yet exist. But the license's carve-out happens to preserve its trunk: the rack layer (renting out GPUs) sits outside the license's reach, the hardware bills for self-hosted workloads still land with American clouds, and AWS's self-hosting guide stands squarely on that side. The collection-point thesis, meanwhile, has to be narrowed: the collection point opens only at the layer that sells service by the token — Fireworks, Together, Baseten, and services like Nebius's token factory have all listed K3; if money truly flows back, this is the layer it flows through. And how big is that layer today? We sketched a rough magnitude (taken from third-party aggregator readings; the two sets of numbers contradict each other and neither is verified, so treat this as direction, not figures): the entire third-party K3 token ecosystem's annualized run-rate may amount to no more than a few multiples of that twenty-million-dollar threshold itself; even at a double-digit percentage rate, the backflow would be far smaller than Moonshot's own API volume. History (the clouds never paid license fees), plus magnitude (the ecosystem is still small), plus rate (zero disclosure) — all three point the same way: they press the "rent collection" reading down and lift the "own the channel, hold the floor" reading up. The rate item is a total blank: Together's official partnership announcement contains no revenue-split language anywhere, the Goldman Sachs report describes the mechanism without giving a number, and Moonshot has not even filed its Hong Kong listing application — the statutory disclosure window is far away.

But three forces push in the opposite direction, and they must be booked together. First: the day after the license went public, Goldman Sachs published a report arguing that Chinese model developers may shift wholesale toward a "paid weights" model — a second institutional voice independent of Xu, with entirely different motives (SCMP). Second: Reuters and the Financial Times reported on July 21 that China's Ministry of Commerce is consulting on a tiered scheme to restrict the export of model weights — the key detail being that access through APIs and cloud services would remain available: collect the free download, keep the priced service — the same direction as the K3 terms, only scaled up from company terms to state control. If it lands, the shift thesis's premise that free weights are globally available simply disappears (still in consultation, nothing decided). Third: a self-defeating mechanism few have pointed out, from an analysis co-written by the Stanford China digital-policy scholar Graham Webster and Xu: "With revenue comes regulability" — freely downloaded weights may be treated as information flows, hard for the US government to reach; but the moment American service providers sign commercial contracts with a Chinese company, the relationship becomes a transaction, squarely inside the reach of US sanctions and information-and-communications supply-chain rules (Transpacifica). The more successfully the collection point opens, the higher the odds that policy severs it.

Finally, put the clock in. The last war's walls stood 13.5 to 43 months before coming down; a model generation's half-life runs about six months: Kimi's own K2 shipped July 2025, K2.5 in January 2026, K3 in July 2026 — that is the spacing between releases. This is a directional estimate drawn from a single product line, not a cross-vendor constant. K3 can hold its floor today on the strength of its terms — and of its capability position: the third-party evaluator Artificial Analysis scores it 57 on its composite index, first among open-weight models (verified in our last issue). The wall's actual lifespan depends on when the next lab ships a same-tier model under MIT or Apache. In the last war, that role was called Valkey — organized by a foundation, fed by multiple clouds. This time, no one needs to organize anything — a single release decision by any one lab is enough. That is the most asymmetric point in the entire analogy: weights can be downloaded and trained onward, but whoever continues the training cannot tap the original maker's next frontier upgrade, so the fork ceiling on the model side is lower than in software — while switching to another lab's same-tier open-weight model costs far less than raising a fork.

One line to read this tree by: once model vendors' per-token economics flip positive, whether the margin holds is contested along three paths; the last deep dive placed K3's high list price inside the "pricing power versus scarcity rent" ruling, and this piece grades the other guardrail — the license wall. History says this wall, on its own, never held pricing; what it holds is channel and brand. K3 is the first model vendor to build the wall on release day, and the wall's maximum lifespan is written on the release date of the next MIT-licensed peer.

Where we land

The provable function of an open-weight license wall is not collecting rent from clouds. In the previous war of this exact shape, the clouds never paid to use open-source code — AWS rewrote or forked, and what Google and Microsoft signed were managed-service partnerships; the original vendors recovered through first-party services, and the two who built the highest walls tore them down themselves after 13.5 and 43 months. K3's first month reads consistent with the same script: third-party prices pinned to the official rate (the floor held), none of the three hyperscalers carrying the K3 license relationship under its own name, and zero rate disclosure. The "Chinese open-weight value shifts to America" ledger is, from here on, kept in two layers: the rack layer stays in the American stack as before, while the token-service layer has opened a collection point — whose current magnitude looks less like rent collection and more like channel control. The axis with real predictive power, which the last deep dive placed at "where the license threshold sits," narrows one more step here: it is business form — the license's gate is the yes/no question of whether you operate a model service, not an amount; and the wall's lifespan ceiling is a model generation, which we estimate in half-years — not the 13-to-43 months measured in the license wars.

>

What would prove this wrong (observation windows are set by us unless noted; we return at expiry to mark the answer): One — any party publicly discloses its license rate with Moonshot and the amount is material; if Moonshot's prospectus single-lines licensing income in its revenue breakdown, the "channel, not rent" reading flips (no listing application has even been filed, so the odds of getting this soon are low). Two — AWS's currently blank K3 catalog page (the URL exists, the content is empty) turns into a formal listing with no public agreement, which would break this piece's landlord-versus-counter layering; we recheck within one to two weeks. Three — within 90 days (by 2026-11-05), third-party K3 quotes develop K2-style discount tiers (12 percent or more), which would weaken the "license floor" explanation and hand the win to "new-product anchoring"; conversely, if Morph's 11 percent discount disappears, that is the first behavioral evidence that the wall has bite. Four — the Commerce Ministry's weight-export scheme lands as formal state control, and this piece's company-level analysis has to be redone wholesale at the state level. Five — within 3 to 6 months, any lab releases a K3-tier model under MIT or Apache and the third-party ecosystem migrates: the model-side Valkey moment, the historical analogy paying out in full, and the wall's window closed. Two dated near-term data points besides: CoreWeave (August 11) and Nebius (August 12) report second-quarter earnings — the first read, at statutory-disclosure grade, on how much volume US clouds carry for Chinese open weights.

What this means for you

Inference providers and platform engineering. The scope check is three yes/no questions: does your service give customers meaningful control over the model's inputs and parameters (yes → in scope); do you only embed model capability inside product features (yes → exempt); do you only relay requests to models hosted by others (yes → exempt). Renting out raw compute does not cross the line; running inference endpoints for customers does. The same AWS self-hosting guide is a signal in the other direction for hosting providers: when a cloud writes the self-hosting path up as a tutorial, it is lowering customers' dependence on third-party hosting. Two more things to remember: the "certified inference partners" exemption lane is defined nowhere in the text, which leaves your compliance status to Moonshot's discretion; and the act of signing with a Chinese model vendor is itself what pulls you into the reach of US sanctions tooling — that is a question for your counsel before signing, not after.

Enterprise procurement. Self-hosting for internal use is expressly waved through by the exemption clause — fine-tuning K3 for internal systems or product features costs you nothing. The comparison worth running is not license cost but what the third-party hosting premium actually buys: this month's market says you are paying for US hosting and data non-retention, not for a lower token price.

Investors. In Moonshot's listing narrative, the quality of the "license monetization" half comes down to one indicator: whether the prospectus breaks licensing fees out as their own revenue line. If it does not, the license wall is a channel tool, not a revenue line. On the neocloud side, next week's two earnings reports provide the first statutory-disclosure-grade magnitude for "carrying Chinese open weights"; and MiniMax — the other Chinese open-weight model vendor, whose license was tightened and then loosened again (last deep dive) — remains the most honest gauge of monetization pressure.

Policy circles. The last deep dive argued that the label "open weights" can no longer separate commercial designs; this piece adds a layer: it cannot separate value flows either — for the same open-weight model, rack-layer money stays in the US while service-layer money may flow back to China, and any regulation drawn on "are the weights public" cuts neither layer. The thing to actually watch is the Commerce Ministry's collect-the-download, keep-the-API scheme: it and the K3 terms are the same direction at two different scales, and if it lands, what "open weights" means at the export end is rewritten wholesale.

US frontier labs. K3's license wall itself does almost nothing directly to US flagship pricing (the last deep dive's conclusion stands: the pressure on prices comes from Zhipu AI and DeepSeek, which are still sitting in the permissive low-price band). This piece's increment is the time structure: the license wall gives Moonshot a protected monetization window, but the window's length is set by the next MIT-licensed model of the same tier — and history says that when that moment arrives, the one who built the wall tears it down themselves.

Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.