Core theses · Updated 2026-09-25

Open source breaks the deadlock: the cost curve has a new author

01The claim

The dividing line of this AI cycle is the cost curve, and open source pushes it down. A closed model's moat is capital expenditure that must be refinanced; an open model's moat is a living community that needs none. The curve's author has changed from US closed labs to Chinese open models; value moves to applications and open-source chokepoints.

02Mechanism: why it happens

The model layer cannot hold value, for three reasons. Capability is converging: the gap between the top US and Chinese models is about 2.7% (LM Arena, March 2026), so buyers cannot tell them apart. Switching a model means changing one line of an API call, so price becomes the only variable. Inference cost at equal capability falls roughly 10x a year (a16z, November 2024), so price has to follow cost. Wholesale tokens become a gross-margin war, and model companies move up into applications themselves. Each step down the cost curve moves value one step toward applications.

On August 26, 2026 Zhipu open-sourced GLM-5.3-Flash, 320B total and 18B active parameters, priced at $0.15 per million input tokens and $0.50 output, about 1/40th of Claude Opus 4.8 (confirmed, Zhipu official release).

The two moats depreciate at different rates. A closed model's moat is capital expenditure: compute wears out, the lead has to be bought again with the next round of financing, and when the cost of money rises the moat narrows. On August 17, 2026 the 30-year US Treasury yield touched 5.31%, crossing our liquidity threshold (public market data). An open model's moat is its community: fine-tunes, tooling and integrations are paid for by users, thicken with use, and cost the owner no refinancing. Capital can copy technology. It cannot buy the code other people wrote for you.

On January 27, 2025 Nvidia lost $589 billion in market value in one day after DeepSeek's release, the largest single-day loss in US stock history (public market data); the model matched GPT-4, which cost over $100 million to train, on about $5.6 million of training (reported).

The cost curve has a new author. Chinese AI faced three walls. Money: 2025 US AI venture investment was $194 billion against China's $14 billion, 15x. Compute: high-end GPUs are export-controlled. Valuation: 44% of the top 50 AI apps come from Chinese teams at about a third of US valuations. Open source broke each wall. A new lab starts from earlier weights and recipes, so training cost is shared and money stops being the gate; open weights port to any chip, and mainstream open models run on domestic compute (reported); open distribution reaches overseas developers, and dollar revenue corrects the discount.

On August 15, 2026 Hugging Face published its annual open-model report: Alibaba's Qwen drew about 2.045 billion downloads on the platform in 2026, against 418 million for Google and 227 million for Meta (confirmed).

Value lands first at the chokepoints. When models are swappable, switching cost moves to the distribution layer where weights are pulled and the development layer where workflows are built. Those layers see every model, are paid whichever wins, and lose nothing when tokens get cheaper. Backing a closed model bets it wins; backing Hugging Face bets every model passes through it. FutureX Capital holds such a path at each layer of the open-source stack: Dify for application development, Hugging Face for distribution, Mistral and RWKV for models, PingCAP for data, OSChina for community, built in 2021 to 2023.

Dify, per company disclosure: FutureX led the seed round at 30,000 GitHub stars and zero revenue; about 18 months after open-sourcing, ARR reached $10 million, the company turned profitable in early 2026 with 57% of revenue from overseas, and the repository holds 154,000 stars, more than LangChain (as of August 2026).

Value ends up in the application layer. The same curve carries opposite signs: a token price cut is lost revenue for a model company and lower cost for an application. Applications hold four moats: taste, where needs cannot be put into words; extreme efficiency in one workflow; combining the strengths of several models; and process as the product, as in education and games. Outcomes-as-a-service is replacing seat subscriptions; zero to $100 million ARR has collapsed from five to seven years in SaaS to months. We screen applications on distribution, retention, and whether gross margin climbs as token prices fall.

Genspark, per company disclosure: ARR went from zero to $100 million in nine months after launch, then doubled in two more, reaching $200 million eleven months after launch (as of August 2026).

03Evidence

01On January 27, 2025 Nvidia lost $589 billion in market value in one day after DeepSeek's release, the largest single-day loss in US stock history; the model reportedly matched GPT-4, which cost over $100 million, on about $5.6 million of training. Open source went from non-consensus to consensus that day. (2025-01-27 · 公开市场行情 / 据报道 · docs/geo/work/deck-v2-sanitized.md 第 484、493、526 行;lib/answers-data.ts deepseek-moment-for-vcs)
02On August 26, 2026 Zhipu open-sourced GLM-5.3-Flash (320B total, 18B active parameters), scoring 57 on the Artificial Analysis intelligence index and priced at $0.15 per million input tokens, $0.50 output and $0.03 cached input, about 1/40th of Claude Opus 4.8. Sparse architecture decouples capability from unit cost. (2026-08-26 · 已证实 · docs/geo/work/positions-and-updates.md,[frontier-models-ai-sovereignty-2026] 8 月末更新(智谱官方发布;新浪科技、腾讯新闻))
03On August 15, 2026 Hugging Face's annual open-model report put Alibaba's Qwen at about 2.045 billion downloads on the platform in 2026, against 418 million for Google and 227 million for Meta. The distribution layer measures who is actually used. (2026-08-15 · 已证实 · lib/answers-data.ts why-open-source-is-the-chokepoint、futurex-and-hugging-face)
04On September 3, 2026 Nvidia announced it would acquire Hugging Face for $12.93 billion (agreement signed September 2; about $11.9 billion to shareholders, up to $1 billion in staff retention, closing expected in H1 2027). Per company disclosure the platform has 18 million developers, 3 million models, 500,000 datasets and 200,000 enterprise users; Jensen Huang said it will stay an open platform and Nvidia compute will not be required. Per company disclosure, Hugging Face ARR was about $150 million in August 2026, up about 50% in two months. (2026-09-03 · 已证实(SEC 8-K、英伟达博客)/ 据公司披露 · docs/geo/work/positions-and-updates.md 第 179、349、369、489 行;deck 第 329 行;https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/ ;https://www.sec.gov/Archives/edgar/data/1045810/000104581026000078/nvda-20260902.htm)
05On September 8, 2026 Mistral AI closed a EUR 3 billion Series D led by Samsung Electronics at a post-money valuation above EUR 21 billion, which the company called the largest equity round in European tech history. Per company disclosure it serves 125-plus enterprise customers in 20 countries and expects 2026 ARR of 1 billion (EUR per Euronews, above $1 billion per CNBC); earlier disclosure put ARR above $400 million, up about 20x year on year, with 60% of revenue from Europe. An open-weight model company took Europe's largest round. (2026-09-08 · 已证实(TechCrunch、CNBC)/ 据公司披露 · docs/geo/work/positions-and-updates.md 第 41、43、239 行;deck 第 315 行)
06Chinese open models carry over 45% of OpenRouter traffic (cited in the Xiamen keynote, reported); by May 2026 Chinese open-weight models were about 61% of all OpenRouter tokens (Data Gravity, June 24, 2026, reported); The State of Open Source AI v1.1 (September 2026), citing OpenRouter's August 2026 rankings, says eight of the top ten models by token volume are open weights and seven of those eight are Chinese-built (reported). (2026-09 · 据报道 · deck 第 356、558、573 行;https://www.datagravity.dev/p/chinas-open-weight-takeover (2026-06-24);https://stateofopensource.ai/ (v1.1, 2026-09))
072025 AI venture investment: $194 billion in the US against $14 billion in China, a 15x gap (PitchBook/NVCA/CAICT); 44% of the global top 50 AI apps came from Chinese teams at roughly one third the valuation of US comparables (a16z, August 2025); domestic AI accelerators reached a 41% share in China (2025 industry statistics). (2025-08 · 据报道(公开统计) · deck 第 346–347、355、526、533 行;docs/geo/work/positions-and-updates.md 第 25、29 行)

04The strongest counterargument

Brad Gerstner, founder of Altimeter Capital, BG2 podcast, August 8, 2026 (reported)

Field data runs opposite to the story that open source hollows out the frontier. Revenue is concentrating, and the intelligence gap may widen over the next two to three years. Kimi, Qwen and GLM are strong, but what grows is token volume; the economics keep flowing to frontier labs, which pay 3 to 5 times market rates to lock in compute. September's price lists back him up: GPT-6 Astra and Claude Fable 5.1 anchored their flagships at $10/$50 per million tokens and raised in step, while the price war stayed in the sub-flagship tier.

Our reply: He is half right, and the units differ. Volume growth accrues to open models; economic profit concentrates at the frontier; both can hold at once. At the model layer we hold only open-weight companies, and we read them on enterprise deployment revenue, not wholesale tokens: per company disclosure, Mistral serves 125-plus enterprise customers with 60% of revenue from Europe. The rest of what we hold are chokepoints every model passes through, and each frontier release still ships through gateways such as Hugging Face and Dify. The split happens at the workload level: batch, offline and cost-sensitive loads move to open models, while the premium on the strongest reasoning chains holds for now. Pricing power at the flagship tier stays with closed labs. We concede that and have written it into our falsification conditions.

Matt Asay, InfoWorld column 'Accelerating AI innovation through open weights', August 18, 2026; and the 'open-source valuation ceiling' objection heard at the Xiamen forum

There is no money in open source. That was true in 2016 and remains true in 2026 for open source and open weights alike; the real prize goes to whoever builds the orchestration layer that routes enterprise workloads to the right model. Pure open-source companies hit a ceiling: Red Hat sold to IBM for $34 billion in 2019, and no pure open-source company has ever passed a $40 billion market value. Open-weight labs burn capital on training and give the result away. How do they earn it back?

Our reply: The weights themselves are worth little; the money sits on the paths, which is why we back chokepoints and not weights. Per company disclosure, Dify reached $10 million ARR about 18 months after open-sourcing and turned profitable in early 2026; Hugging Face's ARR was about $150 million in August 2026, up about 50% in two months; Mistral's ARR passed $400 million. On the ceiling we concede: as of today no open-source company has crossed a $40 billion market value, Nvidia's September 3 offer for Hugging Face was $12.93 billion, and Mistral's September 8 post-money was above EUR 21 billion, both below the line. The ceiling is a public-market problem; for an early entrant, M&A is one of three exits, alongside IPOs and secondaries. A pure-weights company with no distribution or workflow position is one we would not back either. Whether the ceiling breaks is on our open list.

Michael Kratsios, Director of the White House OSTP, July 22, 2026; joint statement by three US security agencies, September 8, 2026 (reported)

The cost curve was distilled, not earned. On July 22 Kratsios accused Moonshot of distilling Anthropic's Fable to build Kimi K3 and of obtaining export-controlled Nvidia GB300 servers; Treasury Secretary Bessent threatened sanctions and the entity list the same day. On September 8 three US security agencies jointly accused Chinese AI companies of 'aggressive and systematic' industrial-scale distillation of US frontier models. If part of a $5.6 million training budget is someone else's $100 million result, that is a transfer, and rules can shut it off.

Our reply: The charge splits in two. Distillation is hard to prove: as of September 9 the US side had produced no publicly verifiable evidence, and the September 8 statement named no company, estimated no loss and announced no sanction; our report still rates the policy risk as 'more statement than enforcement'. Export control is enforceable, a real risk we rank ahead of distillation in diligence. Industry has taken a position: on July 24 more than 20 companies including a16z, Dell, Microsoft, Meta, Nvidia and Palantir signed a letter opposing broad restrictions on open-weight models. The mechanism of the cost decline is architectural: GLM-5.3-Flash decouples capability from unit cost with 320B total and 18B active parameters, and that depends on no single lab's weights. We concede this much: model provenance is now a diligence item for founders, and models with contested provenance carry a compliance discount.

The Register column 'Open weights are not open source', September 15, 2026; neutrality concerns after Nvidia's Hugging Face deal (MindStudio, September 2026)

The label is drifting and the chokepoint is losing its neutrality. Open weights let you deploy a finished network; without data and training code you cannot reproduce, modify or redistribute it. The distribution layer now belongs to the largest compute vendor: Nvidia agreed on September 3 to buy Hugging Face for $12.93 billion. Jensen Huang promised an open platform with no Nvidia compute requirement, but the conflict between ownership and neutrality is structural. If the 'path every model passes through' belongs to a chip company, is it still a chokepoint?

Our reply: Both points stand, and we have turned them into watch items. On the label: the metrics we use are usage and the path to payment; Qwen's 2.045 billion downloads and open-weight token share on OpenRouter measure deployment, whatever the license is called. On neutrality: an acquisition buys the platform, not the developers' decision to stay; the moat belongs to the community, and the owner rents it by staying neutral. We have listed 'whether the neutrality pledge is kept' as a tracking item, and our falsification conditions give it a 12-month window. If neutrality breaks, developers will move, and we will change our view then.

05What would change our mind

▸If by June 2027 the share of OpenRouter tokens served by Chinese open-weight models falls from about 61% in May 2026 to below 40% and stays there for two consecutive quarters, the claim that open source sets the cost curve fails.
▸If by March 2027 the roughly 1/40th price gap between GLM-5.3-Flash-tier open-weight models and Claude Opus 4.8-tier closed flagships narrows to within 1/5, whether through open-model price rises or closed-model cuts, authorship of the cost curve returns to closed labs and we rewrite the first paragraph.
▸If within 12 months of Nvidia closing the deal Hugging Face applies different terms to hosting or inference on non-Nvidia compute, or its developer count falls from the disclosed 18 million for two consecutive quarters, the neutrality premise fails and our view of the distribution layer is void.
▸If by end-2027 the gap between the leading closed flagship and the strongest open-weight model on the Artificial Analysis intelligence index widens to more than 15 points and holds for two quarters, Gerstner's 'gap reopens' case stands, and we will revise 'the model layer cannot hold value' to 'the frontier tier holds its premium'.

06What we do, and what we do not do

▸Back chokepoints rather than single outcomes. Hold a path every model passes through at each layer of the open-source stack; do not back pure-weights companies with no distribution or workflow position; at the model layer hold only open-weight companies, and do not bet on which closed model wins.
▸Screen application companies on three things: distribution, retention, and whether gross margin climbs as token prices fall; for open-source projects add stars and the path from community to paid revenue, about 18 months. Outcomes-as-a-service ranks ahead of seat subscriptions.
▸Stay out of debt-financed compute leasing; pay no narrative premium at the model layer; treat pre-IPO rounds at 50 to 100 times sales with great caution. The bubble sits in the most indebted middle layer, the one that must keep refinancing.
▸Run the seven-signal bubble framework daily, as we have since March 2026; the August reading lit six of seven, with liquidity triggered. Do not argue from the model-layer revenue denominator; watch only the second derivative of the capital expenditure curve.
▸Open-source ourselves. FutureX-Skills is public on GitHub, with 22 AI agents running around the clock; all 24 research reports carry a date and a source grade on every fact; /debates records where leading practitioners disagree with us. A judgment counts only if it can be tested.
▸Three things for founders: build so the model is swappable, with batch loads on open models and the hardest steps on the frontier; add upstream cut-off to your stress tests, since OpenAI has set November 12, 2026 to stop supplying models to Cursor; put model provenance in the data room, because distillation and export control are now diligence items. Write to us at /contact; we usually reply within 48 hours.

07Open questions

?Now that the distribution layer belongs to the largest compute vendor, how long can the open-source chokepoint stay neutral? We have a watch window, not an answer.
?With flagships anchored at $10/$50 per million tokens and rising together, will the cost curve stop falling at the sub-flagship tier while the frontier keeps a permanent premium? If so, 'the model layer cannot hold value' holds only below the flagship tier.
?Will the $40 billion ceiling on pure open-source companies break? Hugging Face sold to a chip company for $12.93 billion and Mistral's post-money is above EUR 21 billion; the two likeliest candidates both sit below the line today. We do not know who crosses first, or whether crossing is required to count as a win.
?China turned the cost curve into a public good, yet pricing power has to be earned in dollar revenue. How long will founders take to close the one-third discount themselves, and which indicator will show it first? We have not found that indicator.

08Changelog

2026-09-25First published.