The landscape

Eighty minutes, five sections. Sections one and four are durable and should survive several years. Sections two and three contain the perishable material, confined to slides that carry their check date. Each section ends with four discussion questions; use them or skip them depending on the room.

Start with the object. Part IV showed that a language model is a very large function whose behaviour lives in billions of numbers, the weights, set once during training and frozen afterwards. Those numbers are a file. Everything in this deck follows from one question: do you get the file? If the answer is no, the model is closed. The vendor keeps the file in its own data centre and sells you its output by the token through an API. That is GPT, Claude, Gemini and Grok as products, and it is a perfectly good way to buy intelligence; most of this room uses it daily. If the answer is yes, the model is open weights. The file is published, usually on Hugging Face, and you can download it, run it on your own hardware or a cloud you control, inspect it, shrink it, adapt it with your own data and keep a copy for as long as you like. The third term is the one people confuse with the second. Open source, in the strict sense the Open Source Initiative defined in October 2024, requires the weights and the training code and enough information about the training data for a skilled person to rebuild the system, all under terms that let anyone use, study, modify and share it. Almost no serious model meets that bar, and the ones marketed as “open source” usually mean open weights. The distinction matters because the things you do not get, the data and the training recipe, are exactly the things that would tell you what the model is likely to be bad at. Next slide.

This table is the deck’s first durable tool. Across the top are three ways a model can reach you. Down the side are the six things you might want. A closed API gives you none of the artefacts and all of the convenience; your rights are whatever the terms of service say today, and the vendor can retire the model, change the price, or change the behaviour whenever it likes. Open weights, in the typical case, gives you the file and the code needed to run it, almost never the training code, and almost never the data; what you get instead is a technical report with totals and a model card with a paragraph. Your right to use it for any purpose depends entirely on the licence attached, which is the subject of the next slide. The right-hand column is the Open Source Initiative’s definition, version 1.0, published 28 October 2024: parameters, complete training and inference code, and a full description of the training data including provenance, selection and processing, with the public portions listed, all under terms that grant the four freedoms to use, study, modify and share. The OSI has said Meta’s Llama licence fails that definition, because it discriminates between fields of use and users, and has accused Meta of openwashing; Meta rejects the definition. The Free Software Foundation classed the Llama 3.1 licence as non-free in January 2025. For a business, the last row is the one that matters most. Once you hold the file, nobody can switch it off. That single fact is the source of every advantage on slide seven.

An open-weights model is only as open as the text file sitting next to it. There are four families worth knowing. Apache 2.0 and MIT are the permissive standards from the software world: use it for anything, commercially, forever, keep the copyright notice, and in Apache’s case accept a patent clause. Alibaba’s Qwen3.8-27B, DeepSeek’s models, Z.ai’s GLM-5.2, OpenAI’s gpt-oss, Mistral’s Small 4 and Large 3 and Meta’s Muse Glimmer all ship this way, and Hugging Face’s August 2026 report counts 59% Apache and 22% MIT among the 178 Chinese releases above twenty billion parameters. The second family is the custom community licence, which looks permissive until a threshold. Meta’s Llama licence has long required a separate agreement from anyone with more than 700 million users; Meta’s Muse Spark Open licence of October 2026 requires a separate deal once you earn more than twenty million dollars in a year from offering AI models as a service, and a “Built with Muse Spark Open” badge; Moonshot’s Kimi K3 licence has a twenty-million-dollar model-as-a-service gate and an attribution requirement at a hundred million monthly users; Alibaba’s largest Qwen3.8 model carries a fifty-million-dollar gate for model-service businesses, with internal use exempt. All of these come with an acceptable-use policy. The third family is research-only or non-commercial, common for speech and image models. The fourth is proprietary: no file at all. The honest summary is that the gates are set high enough that most businesses will never trip one, and that you must still write down which licence you are under, because the day someone asks is the day you have no time to find out. The licence terms here are as reported in October 2026 and change with each release; check the LICENSE file in the repository you actually download.

Part I listed the reasons in a sentence each; here they are with their costs attached. The first reason is price. A free model that is good enough for a job puts a ceiling on what any vendor can charge for that job, and every closed price list on slide eighteen was written with that ceiling in view. You benefit from the ceiling even if you never run an open model, which is why the cheapest use of open weights is as a number in a negotiation. The second is sovereignty, the subject of Part II: when the file is yours, you can run it in a VPC, on-premises or on a machine with no network at all, and “our data never leaves” becomes a fact rather than a promise. The third is forkability. You can fine-tune the file for your format and tone (Part I’s rule: fine-tuning is for form, not facts), shrink it to fit cheaper hardware, or build on someone else’s fork; Hugging Face counted 151,448 derivatives of Qwen alone in August 2026. The fourth is independence. A version you hold cannot be retired, repriced or silently changed, and the model you tested is the model you run. The costs are real and the bottom half of the slide is the honest part. Operating a model means GPUs, drivers, quantisation choices and somebody awake at 3am; the fix is to rent hosted open models first and self-host only when sovereignty or volume demands it. Security moves onto your plate: the model is software you run, so sandbox it, patch its runtime and control where it can send data. And somebody has to read the licence and write down what it says. None of this is heavy. All of it is yours.

Part I put this in one line and it has held up well enough to repeat. There are two races, with different scoreboards. The capability race asks who has the single most capable model this quarter. It is measured by benchmarks and by the things you could not do last month, it is mostly run by closed-weight US labs, and it is loud and expensive. The lead is real: when the US government’s Center for AI Standards and Innovation tested DeepSeek’s V3.1 and R1 against GPT-5, gpt-oss and Claude Opus 4 in September 2025, the best US model was ahead on nearly every one of nineteen benchmarks, by twenty to eighty percent on software engineering and cyber tasks, and by very little on science and knowledge. The lead is also short, measured in months, and it decides what is possible rather than what you will buy. The distribution race asks who supplies the tokens the world actually runs. It is measured by tokens routed through platforms like OpenRouter, files downloaded from Hugging Face, derivative models built, and products shipped. It is mostly run on open weights, and as of 2026 those weights are mostly Chinese: CNBC reported on 26 September 2026 that models from Chinese labs took between 57 and 67 percent of OpenRouter tokens in the week of 14 September, depending on region, up from 6 to 13 percent in February. This race decides what is cheap. Most business jobs, the summarising, extracting, classifying and drafting that Parts III and VII describe, are in the second race. The occasional hard reasoning job is in the first. Knowing which is which is most of model selection.

Three numbers, each with its date and its definition, because this is the material that will be wrong soonest. The first is the headline. CNBC reported on 26 September 2026, from OpenRouter data, that models from Chinese labs accounted for between 57 and 67 percent of tokens in the week of 14 September, the range spanning companies in the United States, Europe and OpenRouter’s “Global South” grouping of 82 countries, where the share was 67 percent. In February the same measure was 6 to 13 percent. Vercel’s AI gateway reported 55 percent in August, from 11 percent in January. Part I quoted a June figure of roughly 61 percent; that was a share of the top ten models only, a different denominator, so the two numbers should not be compared directly. The second number is about building rather than running. Hugging Face’s August 2026 report counts 151,448 derivative models built on Qwen, 4.7 times the Llama family, and Alibaba says Qwen has passed three billion downloads across platforms, against Hugging Face’s figures of 418 million for Google and 227 million for Meta. The third number is the sober one. The OpenRouter and a16z study published in January 2026, covering a hundred trillion tokens from November 2024 to November 2025, found Chinese open models averaging 13 percent of weekly tokens over that year, from a start of 1.2 percent, and open weights overall at about a third of usage by late 2025. The jump from 13 to 60 happened in 2026. The caveats on the slide are not decoration: OpenRouter is one router with a developer-heavy customer base, token share is not revenue share, and Anthropic was still the largest single US provider on it at 14.8 percent in July. The appendix walks through why the various published percentages disagree.

Part I’s map, redrawn in October 2026. The point is the shape, not the names. On the American side the frontier is closed and the money is in chips and distribution. OpenAI sells the GPT-6 family through its API and released gpt-oss under Apache 2.0 in August 2025, its first open weights since 2019. Anthropic sells Claude Opus, Sonnet and Haiku 5.5, all closed, and was the largest single US provider on OpenRouter in July 2026 at 14.8 percent. Google has Gemini, the open Gemma models, its own TPUs and the distribution of Search and Android. xAI, now inside Musk’s SpaceXSI, shipped Grok 4.7 in September; its Grok-1 was Apache 2.0 but later releases are source-available with commercial restrictions. Meta is the interesting one: Muse Spark, closed, replaced Llama in April 2026, then Meta released Muse Glimmer, a 30-billion-parameter model under Apache 2.0, in August, and Muse Spark Open, 816 billion parameters under a custom licence, on 15 October. Nvidia sells the pickaxes, has pledged about 26 billion dollars over five years to open-weight models, ships the Nemotron 3 family and agreed in September to buy Hugging Face for 12.93 billion dollars. Microsoft, Amazon and Apple matter more as hosts and channels than as model labs; Azure and Bedrock serve the Chinese open models from US data centres. On the Chinese side the pattern is open weights at low prices: DeepSeek’s V4 line under MIT, Alibaba’s Qwen with more than 460 open models and the largest derivative count on Hugging Face, Moonshot’s 2.8-trillion-parameter Kimi K3 with weights released in July under a custom licence, Z.ai’s GLM-5.2 under MIT in June, MiniMax, Tencent’s Hunyuan and ByteDance’s Seed. Everyone else includes Mistral in Paris, which raised three billion euros in September 2026 on a sovereignty pitch and ships Apache 2.0 models, plus labs in Canada, the UK, Israel and the Gulf that “US versus China” erases.

Part I said the US was responding rather than setting the table. Three months on from Part I, the response has a shape, and it is money rather than law. OpenAI released gpt-oss in August 2025: two open-weight reasoning models, about 117 and 21 billion parameters, under Apache 2.0, the company’s first open weights since GPT-2 in 2019, delayed several times for safety evaluation; the large one runs on a single 80-gigabyte GPU and the small one in 16 gigabytes. Nvidia, which profits whichever lab wins, has pledged roughly 26 billion dollars over five years to open-weight models according to a 2025 filing reported by Wired, shipped Nemotron 3 Super in March 2026 and the 561-billion-parameter Nemotron 3 Ultra, which Hugging Face records as the largest American open release of the year. Meta’s path has been the strangest. It replaced Llama with the closed Muse Spark in April 2026, after CNBC reported in December 2025 that its next model, codenamed Avocado, might ship proprietary; then it released Muse Glimmer, a 30-billion-parameter multimodal model under Apache 2.0, on 10 August, and Muse Spark Open, an 816-billion-parameter mixture-of-experts model with a one-million-token window under its own Muse Model License, on 15 October. In September Nvidia agreed to buy Hugging Face, the platform where all of this is published, for 12.93 billion dollars, with Jensen Huang promising it would stay open to every framework, cloud and model. And the market built its own answer to the geopolitics: OpenRouter’s US In-Region Routing, generally available in September, serves Chinese open models only from providers operating in US data centres. The takeaway is the honest one from Hugging Face’s data: in almost every month of 2026 the largest open model from a Chinese lab was larger than any American release. The response is real. The gap is not closed.

Every API prices a token four ways, and only the first is on the billboard. Input tokens are everything you send: the instructions, the documents, the conversation so far, re-sent in full on every call as Part IV explained. Cached input is the same prefix sent again within a few minutes; vendors keep it warm and charge roughly a tenth of the input price to read it back, which matters enormously for agents that call the model forty times with the same twenty-page system prompt in front. Output tokens are what comes back, priced about five times input across the major vendors, a ratio that has been stable for years; reasoning models bill their hidden thinking as output, which is why a “thinking” answer can cost ten times a plain one. Batch or flex pricing is for work that is not urgent: submit a file, get the answers back within hours, pay half; OpenAI, Anthropic and Google all offer a fifty percent discount for it in October 2026. The three cards turn this into a buying rule. First, price the job, not the token: a very cheap model that needs three retries and a human check can cost more than a mid-priced model that gets it right, and the only way to know is to run your own test set from Part VII and multiply. Second, for any job your test set can score, the right model is the cheapest one that passes; most business jobs pass well below the frontier, and the frontier is for the ones that do not. Third, remember ratios rather than numbers, because the numbers on the next slide will be stale by the time anyone reads this chapter.

The single most important economic fact in this series is that the price of a fixed amount of capability keeps falling, fast. Epoch AI measured it properly in March 2025: take a benchmark score, say GPT-4’s level on the GPQA Diamond science test, and track the cheapest model that reaches it over time. The price to hit a fixed score fell between nine and nine hundred times per year depending on the benchmark and the level, with the GPT-4-on-GPQA milestone falling about forty times per year. Epoch’s own caveat is on the slide: the steepest declines were the most recent, and they may not persist; the series stops in early 2025, and nobody has published an equally careful update that we could find. The second number is the open-weights discount in 2026. OpenRouter’s Justin Summerville told CNBC in July that Chinese open models were sixty to ninety percent cheaper than the leading Anthropic and OpenAI models; at that time DeepSeek’s Flash model cost fourteen cents per million input tokens against five dollars for GPT-5.5. The third is a worked example. Lindy, an AI-agent company, wrote up in June 2026 how it moved most of its agent traffic from Claude Sonnet to DeepSeek V4 Flash on a third-party host: replay real tasks offline against several candidates, discover that the same model scored differently on different hosts, re-tune prompts until offline scores matched, roll out gradually, then watch retention for weeks because users go quiet rather than complain. It reports about ninety percent lower inference cost on the migrated routes, and that an earlier candidate, Kimi K2.5, passed offline evals and then failed with real users. Self-reported, but the method is exactly Part VII’s and the warning is worth the price of the slide.

This is the one slide in the deck with prices on it, and it carries its date for that reason. All figures are list prices per million tokens on the standard tier as published by the vendors on 26 October 2026. At the top, OpenAI’s GPT-6 Astra at ten dollars in and fifty out, and Anthropic’s Claude Opus 5.5 at four and twenty. In the middle, Claude Sonnet 5.5 and OpenAI’s GPT-6.1 Sol are both two and ten; Google’s Gemini 4 Argon is also two and ten, with the vendor’s own note that it doubles on 1 December 2026. At the bottom, OpenAI’s GPT-6 Luna and Claude Haiku 5.5 are both ten cents in and fifty out, and DeepSeek’s own API serves V4.1 Flash at thirty cents in and a dollar twenty out, halved off-peak, with cached input at less than a cent. DeepSeek’s page did not render for us; those figures are from a price tracker that cites it, dated 25 September 2026, so check them first. Cached input is about a tenth of input everywhere, batch is half price at all three US vendors, and hosted open models on OpenRouter vary by host; GLM-5.2 was around a dollar forty and four forty in June. The thing to notice is the spread: a hundred times from top to bottom for jobs that, on your test set, may come out the same. That spread is the entire business case for having an eval. Treat every number here as a snapshot. The ratios on slide sixteen are what to remember.

One slide, deliberately non-partisan, on the only constraint in this industry that is made of atoms. Training a frontier model takes tens of thousands of accelerators; Nvidia designs most of them, TSMC fabricates them, and ASML makes the only lithography machines that can. That chain is the choke point, and the United States has used it: export controls on advanced chips to China since October 2022, tightened in October 2023 and December 2024. The 2025–2026 story is messier than “banned”. Nvidia’s H20, built to comply, was put under licence in April 2025, costing the company a 4.5-billion-dollar charge; licences were granted in August 2025 on the condition that fifteen percent of China revenue went to the US Treasury; Beijing then told its companies not to buy it. The more capable H200 was permitted in December 2025 with a twenty-five percent share; Chinese customs blocked shipments in January; by August the Financial Times reported ByteDance and Tencent had each received about ten thousand, and Nvidia wrote down inventory and removed China data-centre revenue from its outlook. Brookings’ Mark MacCarthy summed it up in June: the US is out of the AI chip market in China. China’s answer is Huawei: the Ascend 950 and an Atlas 950 SuperPoD of 8,192 chips at eight exaflops, shown in March 2026; a statement in April that the Ascend supernode fully supports DeepSeek V4; a rule since November 2025 that state-funded data centres use domestic chips; and a reported plan at DeepSeek for a 160,000-chip Huawei cluster. Smuggling continues and is prosecuted. For a buyer the lesson is simple. Chips constrain training, the capability race; they do not constrain serving a model whose weights are already public, the distribution race. Export controls will not raise your token price. They protect the frontier lead, so keep frontier work with frontier vendors and bulk work wherever it is cheapest.

Here is the durable part of the deck: seven gates, in a fixed order, and you stop at the first no. The order is the method. The first gate is quality on your own evaluation, the hundred real cases with known answers that Part VII told you to build, because a leaderboard measures someone else’s job and a model that fails yours is worth nothing at any price. The second is legal: the licence gate from slide six, the acceptable-use policy, and the data terms of whoever hosts it, especially whether they train on your inputs. The third is deployment fit: can it run where the data is allowed to be, whether that is a public API, a virtual private cloud, an in-region host, your own rack or an air-gapped machine. Your regulator sees the deployment, not the model, which is why this comes before price. Only then do you ask what the job costs, as tokens per run times price, including retries and the human check; the cheapest adequate model wins. The fifth gate is shape: the context length you actually need, tool calling that actually works, structured output, speed and languages. The sixth is who runs it at three in the morning, which for a self-hosted model is a named human and for an API is a status page and a rate limit. The seventh is what happens if the vendor vanishes or reprices: version pinning, a second model that also passes the eval, the weights file if it is open, and the abstraction layer from slide twenty-eight. No model name appears anywhere in the method, which is the point. Run it on three candidates each quarter, one of them open, and the quarter’s model is whichever passes all seven at the lowest cost per job. When two tie, prefer the one with the closest runner-up. Being switchable is a feature.

Part I left this question open on purpose. Here is the answer, said like an adult. A model is a file of numbers and the code to run it. A file does not phone home; it has no network code to do so. Data goes wherever the software running the model sends it, and whoever deploys it chooses that software. So “is this model spying on me” has three answers depending on which thing you mean. DeepSeek’s consumer app and its own API send your data to DeepSeek’s servers in China, which is the concern the Department of Homeland Security cites. DeepSeek’s weights running on a server in Virginia, or on your laptop, send nothing anywhere. Most government bans, from the US Navy, NASA and the House in January 2025 to the Department of Homeland Security’s October 2026 order removing DeepSeek products and services from federal civilian systems, are written about the first two, though the DHS wording is broad enough that government suppliers should assume it reaches hosted versions as well. The real risks are quieter and each has a mitigation. Behaviour: the US government’s CAISI evaluation in September 2025 found DeepSeek models echoed about four times as many Chinese-state narratives as US models and that R1 was hijacked by malicious instructions roughly twelve times as often as GPT-5 and Claude Opus 4; the answer is to include your own red-team cases in your eval, keep the model away from political content, and remember from Part III that security lives in the harness, not in the model’s manners. Continuity: keep the file, pin the version, have a second model that passes. Your regulator or customer: if you sell to government, check the list for both model and host; otherwise document the deployment fit. And the licence, same as any model. It is an adoption question with a checklist attached, and the checklist is mostly about where it runs.

Part I ended with a sentence that has turned out to be the organising principle of this whole deck: the names will be wrong in six weeks and the questions will not. This slide itemises it. On the left are the things that belong in your architecture document because they will still be true when the document is next read. A model is a file, and holding the file is what open buys you. Open weights are not open source and the licence is the product. There are two races, one for capability and one for distribution, and most business jobs live in the second. Output tokens cost about five times input, cached input about a tenth, batch half, and open models a fraction of frontier. The price of a fixed amount of capability falls every year, so plan on cheaper. Chips constrain training, not serving. Where a model runs matters more to your lawyer than where it was made. And the seven gates, in order. On the right are the things this deck confined to dated slides. Which model is best this quarter. Any price on the board, including the one Google has already announced will double on 1 December. The OpenRouter share, which is a weekly figure from a single router with a developer-heavy user base. Which lab has the largest open model, which this month is Moonshot at 2.8 trillion parameters. Which agency has banned which product. Which Nvidia part may be shipped to whom and what percentage the Treasury takes. Whether Meta is open this quarter; it has changed its mind twice in a year. Every name on the map. When you present this deck in a year, replace the right-hand column and leave the left one alone.

Three stories, one per section of this deck. DeepSeek is the distribution race in one afternoon. A lab in Hangzhou released its R1 reasoning model as open weights in January 2025, with a technical report pricing the final V3 training run at about 5.6 million dollars, a figure that excludes research, staff and failed runs. Its app went to the top of Apple’s US store. On Monday 27 January Nvidia’s shares fell seventeen percent and the company lost 589 billion dollars of market value, the largest one-day loss in US stock-market history according to Bloomberg; Broadcom, ASML and the Japanese chip-equipment makers fell with it. The capability gap to the US frontier was months. The price gap was the story. Meta is the lesson about vendors. Llama defined open weights in 2023 and 2024, under a community licence the Open Source Initiative says is not open source. Llama 4 in April 2025 underwhelmed and Behemoth never shipped; CNBC reported in December 2025 that the next model, Avocado, might be proprietary; Muse Spark, closed, replaced Llama in April 2026; and by October Meta had released Muse Glimmer under Apache 2.0 and the 816-billion-parameter Muse Spark Open under its own licence. Meanwhile Qwen accumulated 4.7 times Llama’s derivatives. A vendor’s openness is a strategy that can change twice in a year; what you hold is the file. Lindy is the method working. In June 2026 the agent company moved most of its traffic from Claude Sonnet to DeepSeek V4 Flash on a third-party host, by replaying real tasks offline against several candidates, re-tuning prompts, rolling out in stages and watching retention for weeks. It reports roughly ninety percent lower inference cost, self-reported, and the warning that an earlier candidate passed every offline eval and failed with real users within days.

Every story in this deck about switching cheaply had the same thing in the middle: a layer between the application and the vendor. Here it is in one picture. Your application knows the job: the prompts, the tools, the documents. It should not know which model is on duty today. Between it and the model sits a layer with one interface, which in practice usually means the OpenAI-style chat API that nearly every vendor and open-model host now speaks; one configuration file naming the model for each job; a log of every call; and the cost per job, computed from the log. In front of the configuration sits the gate from Part VII: a candidate model gets into the file only after it passes the test set at a lower cost per job, and only after shadowing the incumbent for a while as Part III’s rollout path requires. Behind the layer is the model of the quarter, whatever it is: a frontier API, a mid-tier one, a hosted open model in a region you chose, or a file on your own hardware. They are interchangeable because the layer is. You can switch this on by using a router or gateway, such as OpenRouter, Vercel’s AI Gateway or the gateways inside Bedrock, Azure and Vertex, where region pinning is a setting; or assemble it from a library such as LiteLLM or the Vercel AI SDK, or write a hundred lines yourself. The rule is the thing, not the product: no prompt calls a vendor directly. What the layer is not is a quality guarantee. Lindy found that the same nominal model scored differently on different hosts, probably because some served quantised copies, so the eval runs per model per host. Part III listed this under plumbing: pick once, reuse everywhere. It is also, quietly, your negotiating position with every vendor you have.

Four actions, each a direct consequence of a section of this deck, each doable in a week. Write the checklist: the seven gates from slide twenty-two as a single page, in that order, filled in honestly for the model you use today. Most teams discover they skipped gate one or gate two when they chose it. Run your eval on three models: the hundred real cases from Part VII against the incumbent, a cheaper tier from the same vendor, and one open-weight model served from a host in your region, recording quality and cost per job for each. The result is usually that a cheaper model passes for most jobs and not for a few, which is exactly the information you need. Put the layer in: a router, a gateway, or a hundred lines of your own code, with the rule that no prompt calls a vendor directly, the version pinned, the cost logged, and the weights file kept where you have one. Then diarise the re-check: once a quarter, re-run the three-model eval, re-read the price board, re-read the licence of anything you run, and if you sell to government re-check the ban list; thirty minutes, then update one slide. None of this needs a budget. All of it needs the thing most organisations still do not have, which is a test set that says what good means for their own jobs. That test set is the asset. The models are rentals.

For the reader who noticed that this deck, Part I, CNBC, Bloomberg and OpenRouter’s own blog all quote different percentages for the same phenomenon. They are measuring different things. The OpenRouter and a16z study of January 2026 reports Chinese open models at 1.2 percent of weekly tokens in late 2024, rising to nearly 30 percent in some weeks and averaging about 13 percent across November 2024 to November 2025. The “roughly 61 percent” that Part I quoted in August was a June 2026 snapshot of the top ten models only, a much smaller denominator. CNBC’s July report gave 46.4 percent of all routed tokens to Chinese-origin models, against 35.7 percent US-origin, with DeepSeek the largest single vendor at 17.6 percent, Qwen at 13.9 and Anthropic at 14.8. Bloomberg’s July figure of 58 percent, briefly 63, was the share of tokens that US companies specifically ran through Chinese models. OpenRouter’s own September blog said open-weight models of any origin were about 60 percent of tokens from US requests, with Chinese labs most of that. And CNBC’s 26 September report gave 57 to 67 percent by region for the week of 14 September. Every one of these is defensible; none of them is “the market share”. OpenRouter routed roughly twenty trillion tokens a week in April 2026 and its customers skew to developers and agent workloads, which favour cheap models. And token share is not revenue share: a frontier token costs ten to a hundred times an open one, so by dollars the picture is very different. When you quote any of these, quote the definition with it.

The fine print behind slide six, for the person who has to put it in a compliance document. Meta’s Llama Community licence requires anyone with more than 700 million monthly active users on the release date to obtain a separate licence, imposes an acceptable-use policy, and requires a “Built with Llama” notice; the Open Source Initiative says it is not open source, and the Free Software Foundation classes it as non-free. Meta’s Muse Model License for Muse Spark Open, released 15 October 2026, triggers a notification and separate agreement once a licensee earns more than twenty million dollars in any twelve months from offering AI models as a product or service, requires a “Built with Muse Spark Open” notice, prohibits military and weapons uses among others, and is not intended for anyone under eighteen. Moonshot’s Kimi K3 licence, which accompanied the weights released on 26 July 2026, requires a separate agreement once model-as-a-service revenue exceeds twenty million dollars a year and requires the name “Kimi K3” to be displayed once a product passes a hundred million monthly users or twenty million dollars a month; Kimi K2 had been under a modified MIT licence with the attribution rule but no service-revenue gate. Alibaba’s Qwen3.8-Max licence on its 2.4-trillion-parameter model requires a separate licence once model-service or AI-work-assistant revenue exceeds fifty million dollars in twelve months, exempts internal use, and sits alongside the plain Apache 2.0 licence on Qwen3.8-27B. DeepSeek, Z.ai’s GLM-5.2, OpenAI’s gpt-oss and Mistral’s Small 4 and Large 3 are MIT or Apache 2.0 with no gates. Everything here is as reported by the cited sources in October 2026; the LICENSE file in the repository you actually download is the only authority, and a new version can arrive under a new licence while the version you hold keeps the old one.