From Open Source Software to Open Source Strategy

Bill Gurley, for 25 years a general partner at Benchmark, published this essay on the P3 Institute Substack in May 2026. It opens by quoting his own Above the Crowd post from July 1999, "The Rising Impact of Open Source", and spends most of its roughly ten thousand words arguing that open source has grown from a way of building software into a tool companies use against each other. He calls that use Open Source Strategy. AI arrives only in the last fifth. Nathan Lambert put it first on his open-source-ai-reading-list, as the walkthrough of how businesses have used open source and what that implies for AI.

Classic open source and what changed

Gurley's history runs through Stallman's GNU announcement in 1983, Torvalds' Linux kernel in 1991, and Eric Raymond's The Cathedral and the Bazaar in 1997. Raymond's "given enough eyeballs, all bugs are shallow", which he named Linus's Law, carries the whole essay: Gurley treats it as proof that a distributed open effort will always beat a closed one. Benchmark backed Red Hat, MySQL, SpringSource, Hortonworks, Elastic and Confluent, so the financial record is one he knows first-hand. IBM alone has spent more than $50 billion buying open source companies, with Red Hat at $34 billion in 2019, HashiCorp at $6.4 billion in 2025, and Confluent at $11 billion in March 2026. The "classic" playbook is the Red Hat one: start a project and, if it takes off, sell service, support and extra features around it.

Four developments set up the strategic version. Foundations such as the Linux Foundation and the CNCF act as neutral referees when competitors share a project. Enterprise IT went "open source first"; the Linux Foundation's 2025 survey puts open source at 55% of operating systems and 40% of AI/ML workloads. AWS showed what happens when a whole stack is commoditized: in 2000 most people would have picked IBM, HP or Intel to lead cloud computing, but the IP owners had no supplier power left to hold Amazon up. And China adopted open source as national policy, first in the 14th Five-Year Plan (2021-2025) and more explicitly in the 15th, approved in March 2026, alongside a Government Work Report calling for Chinese AI models to lead global open source. Gurley's reading is that open source is the obvious move for a technology follower accused of IP theft, because nobody can steal code that is already open.

Six plays

His definition is concrete: using open source on purpose to neutralize a stronger competitor, commoditize an expensive input, align an industry around a shared standard, or head off a regulatory crisis. Most such plays are defensive, a point he also takes from the 1999 post.

Android is the flagship. Apple's 2008 iPhone deal with AT&T frightened carriers and handset makers with the prospect of a Wintel-style monopoly in mobile, and Google had already gathered them into the Open Handset Alliance in November 2007. Android now runs about 73% of handsets, and Google took back control by tying Search, Maps, YouTube and the Play Store to certified builds. The alliance never had a neutral foundation, which is what made that recapture possible. The payoff Gurley stresses is defensive: Apple could have become a toll-taker between Google and its mobile search users. China got a side benefit, a full mobile OS with none of the tying.

Meta's Open Compute Project (2011) is the input-commoditization case: over 400 members, an estimated $132 billion of OCP-recognized spending in 2025, and cheaper hardware for a company guiding $115-135 billion of capex for 2026. Kubernetes, open-sourced in 2014 and donated to the CNCF, was Google's answer to AWS lock-in; 82% of organizations now run it in production, and AWS is a Platinum contributor because its customers demanded support. LF Networking (2018) is the case Gurley calls mixed. Cisco's gross margins are still around 65-68%, but Juniper sold itself to HPE, Nokia's mobile networks revenue fell 21% in 2024, and the telecom equipment market shrank 11% that year.

RISC-V began as a Berkeley research project (David Patterson tells Gurley the early slides joked about world domination) and now has about 25% of silicon by one analyst's estimate, more than 4,600 member organizations, a March 2025 Chinese government framework mandating it in critical infrastructure, and a line in ARM's own SEC filings naming it as a competitive risk. Overture Maps, formed in 2022 by AWS, Meta, Microsoft and TomTom under the Linux Foundation, is the newest: an open base map aimed at Google's map moat, with a stable place identifier (GERS) that Gurley expects to matter once AI agents need to ground answers about real places.

Across all six he sees the same shape: one incumbent with a structural lead, a coalition of also-rans, a neutral foundation, and a layer turned into a commodity nobody controls.

Autonomous vehicles

The first live case gets the most space. Waymo has absorbed more than $45 billion, was valued at $126 billion in February 2026 and operates in 10 US cities; Tesla pursues its own proprietary stack; GM shut Cruise after spending over $10 billion, and Argo AI is gone. Gurley lists seven reasons an open AV platform would be better, all resting on Linus's Law, and frames the choice for the fifty-plus automakers, ride-hailing and logistics firms as a prisoner's dilemma: chase Waymo alone at enormous cost, or back a shared standard so nobody becomes "the Microsoft of autonomy". Baidu launched Apollo in 2017 as "the Android of autonomous driving", but let its commercial Apollo Go eclipse it and never moved it to a foundation. He predicts China, with cheap EVs and at least six funded AV companies, is best placed to lead an open consortium. Waymo, in his framing, is Apple in 2008.

AI

For AI, Gurley means open weights and says so. Open weights, he argues, give three things: no lock-in, real academic participation, and room for small companies to build without paying platform tax to two vendors. Then he sets out four facts. China leads the open-weight frontier with DeepSeek, Qwen, Kimi, GLM and MiniMax; Demis Hassabis has said so, Cursor's Composer 2 is built on Kimi, and Airbnb's customer service agent runs on Qwen. OpenAI and Anthropic lead the absolute frontier and are closed, and OpenAI looks like Apple did in 2008. Hyperscaler commitment to open is mixed, and Meta has pulled back: Llama 4 disappointed in 2025, Llama 4 Behemoth was shelved, and Meta Superintelligence Labs shipped the closed Muse Spark in April 2026, so Zuckerberg's open-source-ai-is-the-path-forward "reads very differently today". And regulation could close the market. On 29 April 2026 two House committees wrote to Anysphere and Airbnb about their use of Chinese models, which Gurley sees as the start of a path toward banning Chinese open weights and leaving only closed US incumbents (us-scrutiny-of-chinese-model-use).

His search for a Western open frontier player finds Mistral closest, with Mistral Large 3 and ARR growing from $20 million to $400 million in a year, but a tier below the frontier at 675B parameters. Gemma is open only at small sizes, and nobody else counts. The consequence he fears is the internet era inverted: if the US restricts Chinese open weights and has no open frontier model of its own, the other six billion people pick the free, self-hostable Chinese stack, which becomes the global default by 2030. His advice to policymakers is to be careful whose national-security arguments they accept. Tom Bedor's arguments-against-open-source-ai, two months later, makes a similar commoditization case from the startup side.

What the essay leaves out

The argument rests on Linus's Law, and Gurley applies it to model weights without asking whether it carries over. Open weights are not source code that thousands of contributors improve together; almost nobody outside the releasing lab can retrain a frontier model. Lambert makes this point in what-comes-next-with-open-models: there is no fully open AI system yet, only open weights inside systems that are partly closed, and he sees Nvidia as the one company with a clean business reason to release. The claim that open source "will always" produce the better product also sits awkwardly with 2026 measurements that put open models a steady several months behind closed ones (open-models-in-perpetual-catch-up).

The safety case against open weights appears only as incumbent lobbying, with no engagement with misuse arguments or with work like a-safe-path-to-open-weights. And Lambert's open-and-closed-models-are-on-different-exponentials uses the same iPhone and Android analogy to argue that people who depend on coding agents will keep paying for the iPhone.