Yenta’s Tuchas and the Convenient Openness of AI

A coalition of 35 technology companies including NVIDIA, Microsoft, Meta, and OpenAI signed a July 24 letter advocating for open-weight AI models, arguing they strengthen safety, accelerate innovation, and support national sovereignty. The author, drawing on a grandmother's Yiddish proverb about the ease of advocating openness when it affects others, notes that while open models deliver real benefits like competition and lower costs, the same companies tightly control their own crown jewels such as CUDA, cloud orchestration, and frontier models.

TL;DR — Key Takeaways Everyone supports openness—when it commoditizes someone else’s business. NVIDIA wants abundant models, Microsoft wants interchangeable models on Azure, and application vendors want cheaper AI inputs. Open models deliver real benefits: More competition, lower costs, greater sovereignty and less dependence on a handful of frontier providers. The real test comes when openness reaches the crown jewels: CUDA, cloud orchestration, frontier models, proprietary workflows and customer data remain tightly controlled. AI routers will make models increasingly substitutable: Enterprises will reserve premium models for difficult tasks and route routine work to cheaper, specialized alternatives. Scarcity will not remain confined to the model layer: Demands for portability and competition will eventually spread to clouds, applications, accelerators and infrastructure. The AI industry says the future should be open. My grandmother understood the catch: It is easy to advocate openness when someone else’s business is being opened. My grandmother was born in Brooklyn in 1913. She attended City College at a time when going to college was hardly the norm for women, especially women from immigrant families. She was a wise woman in the formal sense of the word, but she also possessed another kind of intelligence. She had the practical common sense that came from being raised by Eastern European parents who arrived as part of the great wave of Ashkenazi immigration to the United States. When I was a child, she would say things that sounded silly to me. They were delivered in a mixture of English, Yiddish and whatever linguistic improvisation suited her purpose at the moment. As I have grown older, I have found myself returning to those expressions and realizing that many contained more durable wisdom than much of what passes for serious analysis today. One of her favorites was: “On yenta’s tuchas es good to schmisen.” I will not attempt to standardize the spelling, correct the Yiddish or make it sound more scholarly than it was. This is how I remember her saying it. Roughly translated, it means: It is easy to talk when it is someone else’s ass on the line. I thought about my grandmother when Jensen Huang and a group of the biggest companies in technology published a letter declaring that the future of artificial intelligence must include open-weight models. The July 24 letter, titled “ Open Weights and American AI Leadership https://www.microsoft.com/en-us/corporate-responsibility/topics/open-weight/ ,” began with 25 signatories and quickly expanded to 35. The current list includes NVIDIA, Microsoft, Meta, Mistral, IBM, Hugging Face, the Linux Foundation, Palantir, Cisco, Cohere, GitHub, ServiceNow, Palo Alto Networks, Perplexity and, after initially sitting it out, OpenAI. Jensen used his first post on X to promote the letter. He argued that open models strengthen safety and cybersecurity, accelerate innovation and support national sovereignty. The world, he said, needs both frontier closed models and frontier open models. He is right. Open models are an essential part of a healthy AI ecosystem. They create competition, lower costs, expand access and give enterprises and governments an alternative to depending on a handful of American frontier-model companies. They allow organizations to run AI on their own infrastructure, control their data, customize models for their own needs and avoid sending every token through someone else’s cloud. The letter makes an especially important economic argument. Organizations should be able to match “the right model to the right job at the right cost,” reserving expensive frontier-scale intelligence for the relatively small number of problems that truly require it while running cheaper, specialized models everywhere else. That is not merely an argument for openness. It is an argument for routing, substitution and multi-model AI architecture. It is also an almost perfect description of how the Indispensability Trap works. But before we get to the trap, we need to talk about yenta’s tuchas. The sudden industry consensus around openness looks considerably less ideological when you examine where each member of the alliance makes its money. NVIDIA wants the model layer to become open, abundant and competitive. The more models developers can download, customize and deploy, the more applications they will build. The more applications they build, the more inference they will run. The more inference they run, the more accelerated computing they will consume. That is excellent for NVIDIA. NVIDIA is not volunteering to commoditize accelerated computing, CUDA, NVLink or the tightly integrated hardware and software systems that constitute its own moat. It wants openness one layer above the layer where it captures its greatest value. Microsoft wants customers to have access to every important model through Azure. OpenAI, Anthropic, Meta, Mistral, Cohere, DeepSeek or whatever comes next—Microsoft is happy to host them all. The model can become more interchangeable because Microsoft expects Azure, Foundry, Copilot, enterprise identity, orchestration and its relationship with corporate customers to remain scarce and valuable. Application companies want models to become plentiful and inexpensive because models are an input cost. ServiceNow, Palantir, DoorDash and the other application-layer companies do not want to pay a frontier-model toll on every interaction forever. They want to capture the customer relationship, the workflow and the accumulated knowledge while forcing model providers to compete for the privilege of powering them. OpenAI can sign the letter, release an open-weight model and still protect the frontier models containing its most valuable capabilities. It gets political cover and strategic optionality without opening the crown jewels. Everyone wants openness, provided the opening happens in someone else’s business. They are all happy to open yenta’s tuchas. Gennaro Cuofano captured much of this in his Business Engineer analysis, “ The Open Weight Alliance https://businessengineer.ai/p/the-open-weight-alliance .” Policy positions in technology are rarely disconnected from business-model logic. Companies tend to advocate openness where openness commoditizes a supplier, competitor or adjacent layer. They tend to discover the virtues of control, safety, quality and intellectual property when the conversation reaches the layer where they possess scarcity and pricing power. The members of this alliance do not necessarily share a philosophy. They share a calculation. Each believes that making models more abundant will move value toward the layer it controls. That does not make the letter fraudulent. Self-interest and public benefit are not mutually exclusive. The internet did not require every company contributing to open source to possess pure motives. Open technology can produce enormous public value even when the companies supporting it expect to benefit. Mistral is a good example. I suspect Mistral’s support for openness is sincere, but it is also the smartest strategy available to the company. Mistral could not simply outspend OpenAI, Anthropic, Google or Meta at the absolute frontier. Openness gave it developer adoption, distribution, political relevance and a strategic identity it could not have purchased through brute force. Mistral looked at the hand it was dealt and played it well. Openness was not charity. It was competitive leverage. China’s emergence as a champion of open-weight AI deserves the same unsentimental analysis. Chinese companies have released increasingly capable open models at prices American frontier providers have struggled to match. Those models create real value for developers and exert badly needed competitive pressure on the market. But does anyone seriously believe China would be quite so enthusiastic about open models if its companies had established the early lead in proprietary frontier AI? China supports open-weight AI because it weakens American model-company dominance, accelerates domestic development and reduces dependence on technologies China does not control. It enables Chinese companies to distribute models globally even while U.S. export restrictions limit their access to the most advanced chips. Again, that does not make Chinese open models bad. It means China is playing the hand it was dealt, just as Mistral is. Still, the contradiction is hard to ignore. Open AI models, apparently. An open internet, not so much. The real test of the Open Weight Alliance is not who signs the letter. Signing costs nothing. Some members may make a token gesture—and yes, the pun is intended. The test comes when openness reaches something they actually consider indispensable. Will NVIDIA support genuine portability away from CUDA with the same enthusiasm it brings to open models? Will Microsoft open the orchestration and enterprise-control layers that make Azure sticky? Will OpenAI release a current frontier model rather than opening technology safely behind its commercial frontier? Will application vendors make their proprietary workflows, customer data and accumulated enterprise knowledge portable? Will China tolerate models and outputs it cannot censor, shape or control? NVIDIA deserves more credit than most because it has put at least some skin in the game. A recent CNCF post by NVIDIA’s Erin Boyd https://www.cncf.io/blog/2026/07/23/the-future-of-ai-is-community-driven-and-open/ detailed the company’s upstream Kubernetes work, its contribution of the KAI Scheduler to the CNCF Sandbox and a $4 million commitment over three years to let CNCF projects perform testing on real GPUs instead of emulators. That is more meaningful than another signature on another letter. GPU cycles have economic value. Contributing technology to community governance carries some loss of control. It does not open NVIDIA’s central moat, but it gives the company’s rhetoric substance. The letter itself may prove even more consequential than its authors intend. It does not merely defend open models. It endorses the operating principles that can turn intelligence into a competitive utility. It tells enterprises not to become locked into a single provider. It encourages them to control their own data and retain the knowledge they accumulate. It supports competition across models, clouds, chips, applications and services. It recommends reserving frontier models for frontier problems and sending ordinary work to cheaper, specialized alternatives. Reuters reported https://www.reuters.com/world/asia-pacific/nvidia-microsoft-other-tech-giants-back-open-source-ai-models-2026-07-24/ that controlling model costs and resisting restrictions imposed by closed providers were already central concerns behind the coalition. That is the Indispensability Trap springing into action. AI will become more deeply embedded in every business, government and institution. Token consumption will explode. More applications will be built, more agents deployed and more compute consumed. Intelligence will become indispensable. But that does not mean any particular provider of intelligence will remain indispensable. As enterprises adopt routers that can select the best model for each task, models become substitutable suppliers. Frontier capabilities may continue commanding premium prices, but only for workloads that actually require them. The enormous middle of the market will flow toward models that are good enough, cheaper, controllable and portable. Demand can rise dramatically while unit prices and margins fall. Railroads carried more freight after railroad capacity became abundant. Power consumption increased as electricity became a utility. Internet traffic exploded after fiber became plentiful. Greater usage did not protect every provider’s pricing power. Each member of the Open Weight Alliance believes it can commoditize the model layer while preserving scarcity in its own layer. But scarcity does not stay where companies put it. The same customers being taught to demand choice, sovereignty and freedom from lock-in at the model layer will eventually demand it everywhere else. They will want portability across clouds. They will want open orchestration. They will want ownership of their agent memory and institutional knowledge. They will want applications that do not hold their workflows hostage. Eventually, they will demand alternatives to NVIDIA as well. Hyperscaler-designed accelerators, AMD GPUs, specialized inference chips, sovereign silicon and open interconnects are all attempts to apply the same abundance and substitution logic to the infrastructure beneath the model. Jensen Huang may still be making the smartest bet in the alliance. Open models expand the total AI market, and NVIDIA is exceptionally well positioned to sell the machinery on which that market runs. Supporting openness can be both good for the industry and very good for NVIDIA. But no one gets to declare that the layer above should be open while assuming the layer below will remain scarce forever. My grandmother would have understood this immediately. She did not need an AI stack diagram, a theory of tokenomics or a coalition map. She knew that talk came cheaply when the consequences belonged to someone else. The Open Weight Alliance may accomplish something genuinely valuable. I hope it does. But we should judge its members by more than the signatures they place on a letter or the token gestures they make afterward. It is easy to proclaim that the future is open when someone else’s tuchas is exposed. Let us see what happens when it is yours.