Open-Source AI

Confidence 0.90 · 19 sources · last confirmed 2026-09-21

The stance — and the class of open-weight models, datasets, and tooling — that treats AI as something organizations should own rather than rent, published transparently rather than gated behind closed APIs. The wiki uses “open-source AI” as an umbrella for four interlocking claims that recur across vendor-CEO sources: (1) the own-vs-rent economics of production AI, (2) open vs. closed as a safety question, (3) AI sovereignty (who shares the open source a country uses), and (4) concentration of power as the field’s top risk.

The motif is already latent in agent-harness — “the model is what you rent, the harness is what you own” — and crystallizes across two independent vendor-CEO interviews two days apart: Jensen Huang (NVIDIA) and Clem Delangue (Hugging Face).

The own-vs-rent thesis

Both vendor CEOs make the same core argument from different vantages:

• Delangue (Hugging Face): “companies need to own AI and own models instead of renting them and outsourcing them to someone else.” An AI/tech company shouldn’t outsource its core capability to “a black-box API you don’t control, don’t have visibility on, don’t have ownership” of. AI is “the next generation of software” / software 2.0, and software was always built in-house. The empirical flow he reports: enterprises start on frontier APIs to experiment and launch, then switch to open models when they hit production and scale because frontier-model cost “becomes too big.” His projection: frontier models for experimentation and high-value tasks; most production workloads on private or open-source models.
• Huang (NVIDIA): “you can’t possibly not continue to control it, improve it, make it better” — outsourcing your specialized intelligence “makes no sense,” whether for a person, company, or country. Huang’s framing is proprietary specialization on an open substrate: build domain-specific “super agents” on open weights (Nemotron) inside a harness you own. Same “start with the frontier, specialize once it’s good enough” sequencing.

The convergence is strong; the emphasis differs. Delangue foregrounds open source specifically and cost-at-scale as the switching driver; Huang foregrounds owned specialization and the harness as the ownership locus. Both cite NVIDIA’s Nemotron open-weight family as evidence open models are reaching frontier performance cheaply (Huang: Nemotron 3 Ultra 86% vs. Claude Opus 87% at ~10× lower cost).

Open vs. closed as a safety question

Delangue’s contested position: open source has historically been less dangerous than closed initiatives because it is transparent — capabilities are legible and defenders can patch them. API guardrails are “very shallow and quite ineffective” (“easy to jailbreak,” weights can be stolen), so keeping models “behind closed doors for just a few players” doesn’t make AI safe — it makes it more dangerous by creating asymmetry of power between actors who can access/steal/misuse weights and those who can’t defend themselves. His prescription: level the playing field, maximize transparency, and make the attacks illegal rather than the tools. This runs against the closed-labs framing that (per the interview’s own premise) halted the Anthropic Fable and OpenAI GPT 5.6 releases on cybersecurity grounds — see Debates, and responsible-ai.

AI sovereignty and China’s open-model lead

Per Hugging Face’s own Spring 2026 report (cited in the Delangue interview), Chinese models accounted for ~41% of downloads — the plurality, surpassing the US. Delangue’s causal claim: US AI leadership 2016–2023 came from open research (“the T in ChatGPT came out of Google sharing an open source transformer”), so open source “creates the conditions for AI leadership almost automatically” — and if China keeps leading in open source, it may “lead AI in general, next year or the year after.” He calls the “China is only good because of distillation attacks” rebuttal “reductive and simplistic” (distillation is a small factor everyone uses). The sovereignty worry: US scale-ups (Cruise, Airbnb) and all academia already build on Chinese open weights because “you can’t study an API — it’s a black box”; ideally more of the open source used in the US would be shared by US organizations (NVIDIA as “the king of American open source AI”; startups like Reflection).

Concentration of power as the top risk

Delangue: “the biggest risk in AI is concentration of power.” The AI companies becoming the most valuable are also becoming the most powerful (he cites an AI firm’s leverage vis-à-vis the US Department of War); a world where “a few companies completely dominate AI” is “basically similar to if there were just one or two companies able to do software” — “the real dangerous scary scenario.” Open source is the counter-force: it “enables innovation, competition, job creation; you don’t create monopolies.” The same logic drives his robotics argument — a home robot shouldn’t be “a black box controlled by a few,” “especially if these organizations’ CEO is not the most stable person in the world.”

Two non-vendor routes to the same conclusion (BBC AI Decoded + Bloomberg, August 2026)

Both of this page’s founding sources were open-model vendors, which capped its confidence at 0.72 for self-interest. Two August 2026 sources reach substantially the same conclusions without selling open weights — an enterprise-AI advisor arguing from data trust, and an academic economist arguing from national capacity.

Peter Grant, on the BBC’s AI Decoded panel, states the enterprise buyer’s version of the own-vs-rent argument, and reaches a stronger conclusion than this page previously held — a partial return to on-premise:

“Kimi just got released from China, which is an open model… you can take that model, you can install it yourself, you could put a firewall around it, you understand the weights, the algorithms and everything else. Where a closed model, you’ll be giving your data to that model. They can take it, they can use all that information. And that’s what scares enterprises the most… I think what you’re going to see is a hybrid. You’re going to see large organizations actually go back to on-prem. They will protect all their data at all costs and they decide where they send the query to for the agent — whether it goes outside the organization for a very basic query or internally for something to protect their own IP. Which ultimately comes back to trust.”

The routing decision he describes — per-query, by sensitivity, between an external frontier API and an internal open model — is a more granular version of the own-vs-rent flow this page holds from Delangue, and it is driven by IP protection rather than by cost at scale. Note the caveat: Grant sells enterprise AI advisory services, so he is not a disinterested party, but he has no stake in open weights specifically.

Carl Benedikt Frey supplies the page’s first genuinely independent vantage — an academic economist with no commercial position in the question — and adds a geopolitical trigger the page has not carried. His premise is that frontier-model access is becoming a lever of statecraft: “you saw that recently with the Trump administration imposing restrictions on foreign use of Anthropic’s latest model. We can expect to see similar things happening going forward, perhaps at greater scale.” The conclusion he draws for every country behind the frontier:

“You cannot really be dependent on the technology leader. You have to try to grow some domestic capacity, and with large language models the easiest way of doing that is through open source or open weights — and that’s how China has closed the gap. It really embraced an open-weight ecosystem, in large part because of export controls on chips, which essentially forced it to go there.”

Two things this adds. First, an independent causal account of China’s open-model position that matches Delangue’s observation without sharing his interest: not ideology but constraint — chip export controls pushed China toward an efficiency-and-openness strategy. Second, a forecast about the demand side of the open ecosystem: many countries “will be probably pivoting either towards Chinese or European technology if they feel that America is an unreliable trading partner in technology,” or attempting a domestic open-weight ecosystem, “although that is going to be a harder approach for most places.”

Four sources now, from four positions — an open-model platform CEO, a silicon vendor, an enterprise-AI advisor, and an academic economist. The last two have no stake in open weights, which is what lifts this page’s confidence from 0.72 to 0.80. What is still missing is empirical corroboration: no source on this page measures how much enterprise inference actually runs on owned or open models versus rented frontier APIs.

Open weights as an American competitiveness argument ( Washington Post Live, July 2026)

The page’s existing sources argue for open weights from cost (Huang), data trust (Grant) and national capacity (Frey). Andrew Ng adds the argument that has the most political traction and is the most contested: that open weights are a condition of American leadership, not a threat to it.

“To sustain competitive advantage in America, one of the most important things we have to do is support and sustain open models… the success of these businesses cannot be at the cost of shutting down everyone else’s access to open models. … I’ve been alarmed at the amount of lobbying that a handful of businesses have been doing saying that open models are dangerous. I think that’s false. … the lobbying against open models has hampered American AI development.”

He is careful to establish he is not hostile to the frontier labs first — “I think I’m the only person that both Sam and Dario have worked for… I really hope they do well and have fantastic IPOs.”

His mechanism for why China’s open strategy worked is knowledge diffusion, and it is the page’s clearest causal account: “when you release models freely for anyone to use, it helps the whole world. Yes, but it helps you even more than it helps the whole world.” The American counterpart is opaque by construction — “because a lot of American work has gone into closed proprietary models, it’s just much harder to know: if you have a question, who should I call up to understand how to do this modelling thing?” This complements Frey’s account (export controls forced China toward openness) rather than competing with it: Frey supplies the cause, Ng the compounding effect.

On distillation he is the page’s most direct dissenter from the standard American framing: “the concept that distillation is a major factor has been overstated, vastly overstated.” Two arguments. A symmetry argument: “AI labs all around the world took data off the open internet and used it to distil knowledge from the internet into their AI models. Is it fair for them to turn around and say, I’ve distilled the internet into my model; if anyone distils my model from here on out, that’s not fair?” And a timing objection to the claim that a recent Chinese release was distilled from a short-lived frontier model — “there just couldn’t have been that much data, and how could [it] have been trained in such a short time.” He also notes the traffic runs both ways: American labs read Chinese open research heavily, “you have to be dumb not to.” This corroborates Delangue’s near-identical dismissal from an independent vantage a fortnight earlier — but note both are open-source advocates, so the page holds this as a well-argued position of two interested parties rather than as a settled technical finding.

Two economic points. The market-share consequence of under-investing: “in places where they’re a little bit more price sensitive, Chinese models have really gained tremendous market share… for example in Africa, DeepSeek adoption is through the roof.” And the unit economics that drive it: “if an open model allows you to get intelligence at one fifth or one third of the cost… if your supply of intelligence costs three times more, that’s a very fundamental business disadvantage.” On the open-weights business model he is honest that it is unresolved — Red Hat and Linux as precedent, publicly traded Chinese open-source companies “doing just fine at least in the stock market,” but “the details of how to do this with open models, I think, are still being worked out.”

His firm-level prescription is the transferable one: preserve optionality. “I can’t forecast in a year or even six months what is going to be the top model. So one of the most important things to do is preserve optionality… I don’t let myself be locked into any one of them.”

Five sources now. Confidence moves 0.80 → 0.85: Ng is a third distinct argument (competitiveness) alongside cost, trust and sovereignty, and the distillation dissent is now double-sourced. It stays below the cap for the reason the page has always given — every source here is an interview, none is empirical, and Ng is the most interested party of the five.

”Not your weights, not your product” — open weights as the enabling condition for enterprise ownership ( Sequoia, August 2026)

A venture investor’s framing of open weights, which matters here because it is the corpus’s third own-vs-rent source and the first from someone who is neither a model platform nor a compute vendor.

Huang defines sovereign AI as “companies owning their own intelligence without external dependencies down to the weights”, and ports the slogan from crypto: “in the crypto days there was this meme for the DeFi degens — not your keys, not your crypto… I hereby present the AI version of this meme: not your weights, not your product.” She draws the boundary immediately, which is more careful than the slogan: “we are definitely not telling our companies to get off Opus or GPT… For coding agents, for desktop work, for frontier-level APIs, the closed model APIs are wonderful.”

The claim that is new relative to what this page already holds is about performance rather than cost or control:

“The beauty of owning your stack is that you can actually drive frontier-level performance now. And so this is somewhat new, and in large part this is thanks to the newest open-weight models, especially Kimi K3 and GLM-5.2, being extremely good. Because the weights are available, they’re actually much more malleable than working with the closed APIs. And so you start with a baseline that’s already close to frontier and then with a good enough technical roadmap — strong post-training, prompt harness engineering, online learning — you can actually reach better than frontier performance by owning your stack. And so this is new for 2026.”

This reframes the open-weights choice from a cost-and-control decision to a potential performance one, and does so specifically in-domain. It is the strongest version of the claim in the corpus — and it is entirely unevidenced: no benchmark, no company-plus-figure, no before/after, and the speaker twice flags that she is not technical and that the framework is opinionated.

Her four drivers, in stated order, map onto material this page already holds: cost (with the COGS inversion — the most advanced deployers switch first), speed (“a small distilled custom model can beat a large general one” in coding and security), performance (the new one, above), and controlling your own destiny — where she is notably generous to the labs: “Anthropic and OpenAI, I actually think to their credit, they’ve been really wonderful partners to a lot of the ecosystem, but companies are increasingly finding that they want their own set of independent legs to stand on.”

Her higher-floor / lower-ceiling characterisation is the most reusable thing in the talk and belongs on this page as a decision frame:

	Closed-API stack	Owned stack
Production layer	foundation model + out-of-box harness	open-weight base + heavy post-training; choose and configure an open harness, its logic, tools and context
Development layer	prompts, context, your own evals	eval monitoring and drift-watching, high-quality post-training data (expert trajectories / synthetic data / RL environments), online learning
Trade	”a higher floor, but a lower ceiling, because you don’t actually have the ability to take your own data, to take online production data and improve your own intelligence"	"like opening a Pandora’s box”

She also notes the policy dependency this page tracks: “just last week, Jensen led the charge in making sure that open-weight models remain available in the US, and it was awesome to see the near unanimous wave of support” — the same dependency Ng argues from the competitiveness side.

Source-quality caveat, load-bearing. This is investor content addressed to the speaker’s own portfolio at an event designed to change that portfolio’s behaviour, and the companies named as evidence are largely portfolio companies. The wiki now holds three own-vs-rent sources — a platform CEO (Delangue), a substrate vendor (Jensen Huang) and an investor (Huang) — whose shared blind spot is that all three sell something that becomes more valuable if enterprises stop renting, and none of them measures the outcome. Confidence on this page is raised to 0.87 on the strength of a third independent vantage converging, not on the strength of the evidence, which remains absent.

The three tiers, stated plainly (added 2026-09-03)

This page has carried open-weight models as an alias since it was created without ever laying the spectrum out. The GitHub Podcast, S02E02 supplies the cleanest short version in the corpus, from a developer-education vantage rather than a vendor-CEO one:

Tier	What you get	What you don’t
Closed	An API. “Everything is just exposed via API” — the default most people have used (OpenAI, Anthropic)	Weights, data, method — all of it
Open weight	”You can download the weights, you can use the model for free… hosting it on your machine"	"You don’t have access to the data set and the training method”
Open source	”Every part of the model that was used to train the model is available openly” — enough to build your own version	—

The framing underneath it is the transferable claim: “when we think of open source traditionally, we think of just software — the code. But now in this new AI phase there are more aspects to making a model work than just code, so having an open-source model would mean that every part of that is available to you.” Open weights are therefore “slightly different because it’s not as open, but still kind of has that same ethos.”

Two things this settles for the page. First, the wiki’s own-vs-rent sources are almost entirely arguing about the middle tier, not the top one — Delangue, Huang, Ng and Frey all say open source and mean open weight, and the distinction matters exactly where this page’s safety argument lives, because transparency claims that hold for a published dataset do not automatically hold for published weights alone. Second, it clarifies what agent-harness’s “the model is rented, the harness is owned” motif actually buys: open weights make the model run-anywhere, not inspectable, and specialisation on top of them is ownership of the harness rather than of the model.

The source adds no evidence and does not move this page’s confidence — it is a vendor-produced developer podcast with no measurement in it. It is cited for definitional precision only.

The experimentation argument for open weights (added 2026-09-04)

The page’s four claims are all about production: own-vs-rent economics, safety, sovereignty, concentration. Sokolenko at PyCon DE 2026 adds a fifth that is about learning, and it is the one a working engineer feels first. Reading the BFCL rankings on stage, he notes that the top three tool-calling models are Claude Opus 4.5, Gemini 3 and GLM 4.6 — and then:

“They are all proprietary. So you cannot use them for local experimentation. You cannot use them for experimentation in your environment.”

So the model he actually uses is the highest-ranked one he is allowed to run: xLAM-2 at 32B, 4-bit quantized, on a two-year-old laptop, at zero marginal cost. The whole talk’s thesis — the way through the agentic hype is to build one yourself — depends on open weights being available at a size that fits consumer hardware. Openness without smallness would not have delivered it; see small-language-models.

The NVIDIA SLM position paper reaches the democratisation claim from the compute side rather than the geopolitics side of Ng’s argument:

“When more individuals and organizations can participate in developing language models with the aim for deployment in agentic systems, the aggregate population of agents is more likely to represent a more diverse range of perspectives and societal needs.”

Worth noting who is making these arguments: an enterprise vendor’s research lab (Salesforce AI Research) open-sourcing both the xLAM weights and the synthetic training data, while the same company’s product line sells the agentic platform Sokolenko cites as the emblem of the hype. Open-weight releases are increasingly a competitive move by large vendors, not only a community one.

Sovereignty moved to its own page, and a layer-choice framing (added 2026-09-15)

WEF Radio Davos, September 2026 supplies a framing this page had been carrying implicitly and can now hand off. His claim is that autonomy is not achievable whole — “I don’t think any country has full-stack AI sovereignty anyway — we have semiconductor dependencies, manufacturing dependencies, cloud dependencies” — so the real decision is “what is the layer of AI that is most important for me to have strategic autonomy.” Open weights, on that reading, are not the sovereignty answer but one layer at which autonomy can be bought cheaply, which is a narrower and more defensible claim than the one Delangue’s national-leadership argument makes.

He also names a motive that has no open-weights component at all: cultural autonomy — “I don’t want a foreign technology teaching my second graders.” A country pursuing that is reaching for the post-training and application layer, not the weights.

Bookkeeping: the AI sovereignty alias moved off this page to ai-sovereignty on 2026-09-15, so wikilinks of the form [[AI sovereignty]] now resolve to the concept rather than here. The §AI sovereignty and China’s open-model lead section above is unchanged and stays — it is about open-model leadership, which is this page’s subject. The broader autonomy question, the state-versus-platform governance material, and the firm-level own-vs-rent decision now live on ai-sovereignty.

Sources consulted

• Hugging Face (TechCrunch Equity, Jul 2026) — the fullest single-source treatment: own-vs-rent flow, safety-through-transparency, China’s open-model lead, concentration-of-power, local AI / robotics.
• NVIDIA (LangChain, Jul 2026) — the own-vs-rent thesis from the substrate/silicon side; proprietary specialization on open weights; the Nemotron benchmark.
• BBC AI Decoded (Aug 2026) — the enterprise-buyer vantage: self-hosting an open model behind a firewall, per-query routing by sensitivity, and a predicted partial return to on-premise, driven by IP protection rather than cost.
• Washington Post Live (Jul 2026) — the American-competitiveness argument: open weights as the condition of US leadership; knowledge diffusion as the mechanism behind China’s gains; distillation “vastly overstated”; price-sensitive markets defaulting to Chinese open models; preserve optionality as the firm-level prescription.
• Bloomberg Trumponomics (Aug 2026) — the academic-economist vantage: open weights as the route to domestic AI capacity for countries behind the frontier; export controls as the cause of China’s open-weight position; frontier-model access as an instrument of statecraft.
• The GitHub Podcast S02E02 (Sep 2026) — the definitional vantage: the closed / open-weight / open-source three-tier spectrum stated plainly, and the observation that “open source” for a model means more than code. No evidence; cited for precision.
• Bloomberg Originals (Sep 2026) — OpenRouter usage share (Chinese models ahead of US from June 2026), a per-task price comparison, the Polsia switching case, the monetisation problem (“iPhone vs Android”), and GPT-5.6 Luna as the frontier price response.
• Reuters On Assignment (Sep 2026) — distillation as widespread and forced by compute scarcity, with its share of China’s progress undeterminable.
• Hugging Face Agentic Evaluations Workshop (Mar 2026) — the open-evals corollary: open weights as a precondition for studying, trusting, and improving AI (background support, not counted in source_count).

Debates and supersession

• Open-vs-closed as a safety question is unresolved in the corpus. Delangue argues transparency makes open source safer and concentration is the real danger; the closed-labs framing (which halted the Anthropic Fable and OpenAI GPT 5.6 releases “for cybersecurity concerns”) argues the opposite — that frontier capabilities need gating. The wiki holds both without resolving; see responsible-ai. No supersession.
• Vendor-advocacy discount, partially lifted (2026-08-12). The page’s two founding sources are vendor-CEO interviews (Hugging Face sells open-model infrastructure; NVIDIA sells the silicon and Nemotron weights beneath it), so their pro-open claims are self-interested. The August 2026 additions are not: Grant sells enterprise AI advisory rather than open models, and Frey is an academic with no commercial position. confidence moves 0.72 → 0.80 on that basis. It stays below the cap because no source on this page is empirical — all four are interviews. The AI Index open-weight-trend data and the primary Hugging Face Spring 2026 report remain the open ingest targets that would close the gap.
• Two different drivers of the same behaviour, not yet disentangled. Delangue and Huang argue firms move to open/owned models as cost bites at production scale; Grant argues they move for IP protection and data control; Frey argues states move for strategic autonomy under export-control and access-restriction risk. All three predict the same migration, so the page cannot currently distinguish them — and they imply different things about what would reverse it (cheaper frontier inference, better contractual data guarantees, and geopolitical détente respectively).
• The cost driver now has an observed reversal (2026-09-19). Of the three drivers above, cost is the only one for which the corpus now records the reversing event. Bloomberg’s Polsia case moved to Chinese open models when its frontier bill reached $1–1.5M a month. After OpenAI’s GPT-5.6 Luna cut prices by about 80%, the same founder began testing the frontier model again. One case does not measure how elastic the migration is, but it shows the migration can be reversed on price alone, which the data-control and autonomy drivers would not predict.
• “Workloads move to open/owned” ≠ “frontier labs decline.” Delangue explicitly hedges that OpenAI/Anthropic can remain “the most valuable companies” on frontier reasoning even if most workloads run on open/owned models. The claim is about where workloads run, not frontier-lab viability — keep the two distinct.

Related concepts

• agent-harness — the ownership locus of the own-vs-rent motif; open weights are the rented/owned model, the harness is the owned specialization.
• foundation-models — open-weight models are a subset of the foundation-model substrate; this concept is the openness dimension of that substrate.
• enterprise-ai-adoption — the frontier-experiment → owned-production flow is an adoption pattern.
• responsible-ai — the safety-through-transparency vs. closed-gating debate lives across both pages.
• generative-ai — the broader capability class these open models deliver.

A security asymmetry (added 2026-08-30)

Spracklen et al. measured package hallucination at ≥5.2% for commercial models against 21.7% for open-source models — a four-fold gap between model classes rather than a capability gradient. Worth holding alongside this page’s capability-convergence argument: on this particular safety-relevant behaviour, the classes had not converged as of the 2024–25 model vintage studied. See ai-generated-code-quality.

Ownership below the firm, and two limits the page has been missing (added 2026-09-09)

Three sources from the 9 September 2026 batch push this page in two directions at once: down an aggregation level, from the firm to the device and the individual, and sideways into two constraints the own-vs-rent literature here had not yet stated.

Down a level: the device (OpenJarvis)

Jon Saad-Falcon’s OpenJarvis segment runs this page’s argument at personal scale, and its economics are not the ones the firm-level sources use. The case against cloud-bound personal AI is four-part — cost (thousands of dollars a year aggregated), privacy (your most personal data leaves the machine), rented-not-owned intelligence, and energy — and the enabling claim is that locally runnable models are “only 6 to 12 months behind” the frontier, with the gap closing as consumer accelerators improve.

The load-bearing mechanism is new to this page: pay frontier prices once, at optimisation time, not per inference. A cloud model diagnoses the local stack, proposes changes and gates them; the resulting configuration then runs entirely locally. Optimised configurations beat out-of-the-box local deployment on cost, latency and quality, at a reported ~800× lower inference cost — and any strong optimiser works (Opus 5 and GPT-5.6 best, but Gemini, Kimi and GLM all functional), so the technique is not tied to one vendor. That is a different argument from “open weights are cheaper at volume”: it says the frontier’s advantage can be spent down into an owned artifact rather than rented continuously.

Down another level: the individual (Tan)

Tan’s keynote takes the identical structure to the level of one worker and reaches a conclusion this page has no analogue for. His equation keeps the frontier model rented — “model quality is rented but your brain is owned” — so the owned asset is not weights but accumulated context and executable procedure. That the two Own Your Intelligence events, Sequoia’s and YC’s, ran within a week of each other with the same slogan and different owned assets is the clearest sign the motif has become general.

He also answers an objection the cost-driven sources cannot reach: what happens when models improve. “The better the models get, the more the differentiator moves to context. When everyone’s engine is a 1000 horsepower, the race is won on the driver and the map… A better model makes your library worth more because a smarter reader extracts more from the same books.” Frontier progress becomes a free upgrade to the owned layer rather than a threat to it — which is a stronger defence of ownership than cost-at-scale, because it does not expire when inference gets cheap.

And it adds a labour dimension: skill files are externalised cognition, so who holds the repo becomes a question about careers, not just about margins. “Own your skills, because if you don’t, your job becomes a skill file.” See ai-deskilling for why this is a distinct risk from skill atrophy — the worker’s judgment stays intact and is captured anyway.

Two limits, from the regulated-profession vantage

The CFA Institute roundtable supplies the honest counterweight, and it matters because the speakers are practitioners with no position in the outcome.

Limit 1 — throughput, not capability, is where open deployment breaks. Tate did the comparison properly: proprietary model as baseline, open alternatives through an identical pipeline. “With a much smaller large language model that was open source, I was able to replicate more or less what the proprietary OpenAI model was able to do for the task.” The failure was scale — a ~36B model “still took a very, very long time” against a proprietary batch API absorbing 50,000 requests and returning within 24 hours. “It’s very difficult to replicate that locally unless you invest a lot in the infrastructure.” This page’s sources argue capability convergence; none of them had addressed throughput convergence, which is a different and largely unmet condition.

Limit 2 — the harness gap is a reason open substitution fails even when the model is fine. Pisaneschi: take an open model, “[throw] it into the open source harness, it is not going to quite be as good as Claude Code”, because the closed vendors have built real parallelisation and optimisation into theirs. His conclusion — “it’s a winner take all scenario” — and his framing of open source’s value as cost optimisation rather than capability is the older position, and it contradicts Huang’s claim three weeks earlier that owning the stack can now beat frontier performance in-domain.

The disagreement is genuine and this page should not settle it. The likely reconciliation is scope — Huang is talking about a narrow domain with proprietary data and a real post-training capability, Pisaneschi about general capability available to a firm without one — but neither addresses the other. Note also that Pisaneschi’s “three months” lag and Saad-Falcon’s “6 to 12 months” are probably not in conflict: the frontier of open weights is not the frontier of what runs on a laptop, and neither speaker draws that distinction.

The open-model plateau on a strategy task (added 2026-09-09)

A narrow but useful data point from [[2026-03-11-allen-mcdonald-how-well-can-ai-do-strategy-simulation-benchmark|Allen & McDonald (Strategy Science, 2026)]], who ran 13 open-source models (DeepSeek, Gemma, Qwen, Llama, GPT-OSS) alongside 21 proprietary ones through the same strategy simulation under identical conditions.

Open models scored substantially lower than proprietary systems, with performance less systematically correlated with release date, appearing to have “plateaued at a markedly lower level.” Notably, they did not show the frontier decline the newest proprietary models exhibited — they were simply flat and lower.

This is worth holding against the page’s capability-convergence argument without overreading it. The task is a specific one (multiperiod strategic resource allocation under uncertainty), the models are out-of-the-box with no post-training or harness work, and the paper’s own framing is that it tests “general out-of-the-box LLM capabilities” while “focusing attention through fine-tuning or prompting could yield very different results.” So it is evidence that convergence is uneven by task, not that the own-vs-rent case fails — and it is precisely the kind of domain-specific gap that Huang’s post-training argument claims to be able to close.

An equity-research argument for open weights: margin, not ideology (added 2026-09-16)

The open-weights case on this page is usually made on inspectability, safety, national positioning or cost-to-the-buyer. Goldman Sachs, August 2026 makes it from where the profit sits in the value chain, which is a vantage the page has not held.

His starting fact: “according to all of the survey work that we’ve done, all of the economic value is accrued to the semiconductor companies — and that’s great for now for the semiconductor companies. But that’s completely unsustainable unless the end customer… starts to make or save money.”

Open weights are how that changes: “open source really benefits the enterprise customer… it’s really good for the hyperscalers because it’s more likely then you’re going to be able to profitably fill up all this capacity that you’re adding. And then I think it’s more of a challenge to the semiconductor layer that’s benefited from the massive compute power that the frontier models require. If you can build models that don’t require as much compute power, then the customers can start to shift some of the economic value from the semiconductor companies… further up in the chain.”

The routing prediction, and the boundary he draws. “I don’t think enterprises will ever successfully implement AI only using frontier models. I think they’re going to have to use frontier models, because the high-consequence queries are going to demand the most powerful models, but then there will be open source and open weight models that can take care of a lot of the rest of the queries.” This is the same per-query own-versus-rent split Huang describes per capability; see ai-sovereignty.

He rejects the China-versus-US framing of open source outright — relevant to the §AI sovereignty and China’s open-model lead section above: “This has become a little bit of an issue of open source is China and frontier is the US. I don’t actually see that as being the long-term dynamic… Open source has been around forever in different flavors. Linux was an open source operating system. Red Hat on top of Linux was an open source software company. Meta’s model is an open-weight model. This isn’t a China versus US thing. I think ultimately you’re going to have US open source and open weight as well.” On distillation, which Delangue also dismisses: “it’s not to me about distillation or stealing code… smaller, faster, cheaper is what powered technology innovation since the beginning of time.”

Caveat. Goldman has a large business on every side of this trade, the survey work behind “all of the economic value” is never named or sized, and Covello has held a publicly sceptical position on AI economics for roughly two years — consistency that helps interpret the view and cautions against reading it as an independent update.

Usage share, a switching case, and the monetisation problem (added 2026-09-19)

Two same-day September 2026 pieces from wire and financial journalists, neither with a stake in the open-versus-closed question, add four things.

1. The first usage measure, not a download measure. Bloomberg Originals reports that on OpenRouter, global usage of Chinese models overtook US models in June 2026, and that Singapore, Germany and the US itself now route more traffic to Chinese than to US models. This complements Delangue’s ~41% download share. Downloads count interest, while routed tokens count work actually run, so this is closer to the empirical measure the page has been missing. It still is not that measure: OpenRouter’s users are developers who already choose models on price, so the platform over-represents the cost-sensitive traffic where open models are strongest. It is not a census of enterprise inference.

2. A switching case at the scale where cost binds. Bloomberg’s Polsia case is the own-vs-rent flow with numbers. An agent-automation startup started on Anthropic “because I need to give my customers the absolute best”. When usage went viral its bill grew from $10–20k to $1–1.5M a month, and moving to Chinese open models brought it to about $100k. The film’s line for the principle: “you don’t need God to write your emails.” That is Covello’s frontier-for-high-consequence split, stated as a slogan.

3. The monetisation problem, which the page had not recorded. Every source above argues open weights from the buyer’s side. Bloomberg names the seller’s problem: “Nobody has quite figured out how to make money.” Chinese providers are in “a race to the bottom for more than a year now,” and domestic consumers “are not willing to pay for AI services that much.” The analogy: “Anthropic and OpenAI have built the iPhone. The Chinese companies are more like Android. iPhone makes by far the most money, but Android has the larger market share.” This bears on how durable the open ecosystem is. An open-weight lead financed by state industrial policy and price war depends on those conditions lasting, which Frey’s export-control account of its origins does not address.

4. Distillation: confirmed as widespread, still unmeasured. Reuters’ China correspondents add a third position to the corpus’s distillation dispute, between the US government’s (distillation explains the catch-up) and Ng’s and Delangue’s (it is overstated). The practice is “pretty widespread” and structurally forced: China “can’t really access even a fraction of the most advanced compute,” so it “will keep relying on” distillation. But its contribution is “very hard to determine,” and there is a counter-example: “Meta spends millions of Anthropic tokens every month but their model has not really caught up.” That is the most defensible statement of the question the corpus holds: the practice is established and its effect is not.

These two sources do not change the page’s confidence (0.90). The usage figure is biased toward the population where open models win, and the rest is reporting rather than measurement. They do add the page’s first observed case of the cost driver reversing; see Debates.

A sitting enterprise-platform CEO on open weights, on premise (added 2026-09-21)

IBM on Bain’s Winning with AI, September 2026. Of three closing pieces of advice to CEOs, Krishna’s first is open-weight models, “especially run on premise, not just on the cloud” — to keep “really critical proprietary IP” in-house, because “people outside the US may worry a lot about where their queries and data is going”, and for cost. He calls the model market an “and world”: several LLMs for technical, political or diversity reasons, plus open weights, a view he has held for five years that has “come true in the last 9 to 10 months.” This matches Delangue’s own-vs-rent flow from the enterprise buyer’s side. Note that IBM sells the hybrid-cloud platform on which such models would run, and that co-host Andrew Ng makes the policy case for open weights elsewhere in the corpus.