By Turning AI Into a U.S.–China Race, We Told the Wrong Story

Open weights, AI rents, and enterprise sovereignty: Who should own intelligence trained on humanity’s knowledge?

Table of Contents

1. Why the Man Selling the Chips Is Talking About Openness
2. Models Do Not Grow Out of Nothing
3. When Learning Is Allowed to Flow in Only One Direction
4. Why Those Who Take More Will Lose to Those Willing to Take Less
5. Chips, Memory, and the Next Tollbooth
6. Rented Intelligence Can Get Dumber Overnight
7. Safety Cannot Mean Trusting a Handful of Companies
8. AI Is Not a Mountain Where Only One Flag Can Fly
9. Charge for the Road, but Do Not Close It

Prologue: Why the Man Selling the Chips Is Talking About Openness

Early on the morning of July 24, 2026, Jensen Huang published his first post on X. When someone steps into a public square of hundreds of millions of people for the first time, the usual move is to introduce himself or say something safely forgettable. Huang did neither. He used that unusually valuable piece of attention to share an industry letter supporting open-weight AI. AI, he wrote, would transform every industry, power every company, and be built by every country. Open models would spread innovation, strengthen cybersecurity, and give institutions and nations greater sovereignty. The world needed frontier closed models and frontier open models at the same time. [1]

There was a wonderful contradiction in the gesture. Nvidia is one of the most successful toll collectors of the AI boom. Nearly everyone who wants to train or run a serious model passes through its gates. Yet the man running the busiest toll road was now arguing that the road ahead could not belong to only a few companies. Open models will, of course, sell more chips. Jensen Huang is not disinterested. But commercial self-interest does not erase public value. A restaurant may support a new road because it expects more customers; the village still gets a road. We do not need to inspect Huang’s soul before judging the idea. We only need to ask whether the door opened a little wider.

The letter, titled “Open Weights and American AI Leadership,” was signed by more than two dozen organizations, including Microsoft, Nvidia, Meta, Palantir, Hugging Face, IBM, Mozilla, the Linux Foundation, and Mistral. Its meaning can be explained without much technical language. Using a closed API is like calling a car from the same taxi company every day. You can tell the driver where to go, but you cannot open the engine, and you do not know whether tomorrow’s car will behave exactly like today’s. With open weights, you at least receive the engine. You can install it in your own vehicle, maintain it in your own garage, and drive it along your own route. Open weights are not the same as fully open source: the blueprints, materials, training data, and factory process may remain hidden. But you no longer need the manufacturer’s permission every time you turn the key. [2][3]

At almost the same moment, a long transcript of a conversation between DeepSeek founder Liang Wenfeng and investors began circulating on the Chinese internet. It is not an official DeepSeek transcript, and the company did not confirm it when reporters sought comment, so its figures should not be treated as audited disclosures. As a record of a founder’s thinking, however, it carries a candor that is rare in corporate speech. Liang said openness was not a launch-day pose but a requirement of the company’s mission. He said a company needed profit without making profit maximization its purpose. Then he offered one plain sentence: “Those who take more will be defeated by those willing to take less.” [4] Across the Pacific, Palantir CEO Alex Karp was making a related complaint in a much angrier register: enterprises were spending more and more on tokens, the basic units used to meter model consumption, without being able to explain what value those tokens had created. [5]

A man who sells chips, a man who trains models, and a man who sells AI systems to large enterprises had reached a similar conclusion from three different positions. That is not a coincidence. They had all seen the same wall rising. If the most important form of intelligence can only be rented through remote interfaces controlled by a few companies, and if those companies can change the price, quality, rules, and supply relationship whenever they wish, then the deeper AI enters the economy, the more fragile the economy becomes. Every argument in this essay eventually returns to one simple question: Will intelligence grown from humanity’s accumulated knowledge flow into society like water and electricity, or will it become a private road where every mile carries a toll and the gate may close without warning?

Chapter One: Models Do Not Grow Out of Nothing

When a book is finished, its author’s name appears on the cover. When code is committed, an engineer leaves a trail in the repository. When a bridge opens, we know who designed it and who built it. A large language model is different. Ask it how to comfort someone who has just lost a parent, and its answer may carry the language of novelists, the experience of therapists, and the kindness of countless strangers. Ask it to repair a program, and its ability comes from papers, manuals, open-source code, and two decades of programmers answering one another’s questions. Ask it to explain a war, an illness, or a law, and it draws on knowledge that generations of people wrote down, disputed, and corrected. No single person created the whole model, but every river inside it has an upstream somewhere in the human world.

Model companies have done extraordinarily difficult work. They buy expensive chips, invent training methods, clean data, hire researchers and annotators, and absorb the cost of one failed experiment after another. Without those efforts, human knowledge would remain scattered across libraries, websites, and databases. It would not spontaneously turn into a system capable of conversation and reasoning. That is why I do not accept the breezy claim that “the data was free, so the model must be free.” An author’s work is not air. An engineer’s labor is not air. Training a frontier model is not a button someone presses. But the opposite claim is equally unreasonable: that paying the training bill gives one company permanent ownership over everything the model absorbed, along with the right to lock general-purpose intelligence inside its servers forever.

Anthropic’s copyright case with a group of authors shows why slogans fail here. The judge separated two questions. Could a model read books and learn from them? And how had the company obtained those books? The court found that training on lawfully acquired books could qualify as transformative fair use. A model was learning patterns across many works, not simply photocopying one of them. But books taken from pirate shadow libraries did not become lawful merely because the company later used them to train AI. [11] A later $1.5 billion settlement covering roughly 465,000 books primarily addressed how the material had been obtained. It did not declare that machines were forbidden to learn. [12]

That distinction saves us from two bad answers. We do not have to tell creators that anything done to their work is acceptable in the name of progress. Nor do we have to pretend that learning itself is theft in order to protect copyright. Human civilization has always been a long process of reading, imitating, arguing, and surpassing. Novelists read earlier novelists. Scientists use earlier theorems. Programmers build every day on foundations laid by open-source software. AI should not receive unlimited privileges that people do not have, but it should not be banned from entering the same river of knowledge. The law should focus on how materials are acquired, when outputs infringe, and where real harm occurs. It should not turn the act of learning into a territory owned by a few firms.

This is where the harder question begins. If society allows model companies to learn from a vast body of human knowledge, should those companies leave society a path for learning in return? I believe they should. Every service does not need to be free, and every lab has the right to keep some secrets. But a healthy frontier ecosystem must include genuinely competitive open-weight models, or the flow of knowledge will end in a cul-de-sac. Model companies can still earn money from the easiest API, the fastest inference, the best enterprise support, the most reliable security, and capabilities that keep advancing. Universities, hospitals, startups, and ordinary businesses should also have the option to bring a model into their own environment without reporting to the same company every time they use it.

The Microsoft-hosted letter reaches back to the history of open-source software. Today’s internet, cloud services, and enterprise systems all stand on software that earlier builders chose to share. Openness did not eliminate the software business; it made a much larger software economy possible. [2] Open weights may follow the same path. They will not make model services disappear, just as Linux did not make technology companies disappear. They simply prevent the lowest layer of capability from belonging entirely to one firm. Liang put the point bluntly in the investor conversation: if a company seeks a reasonable profit, openness does not destroy the business because the original team can still provide a cheaper and more reliable service. Openness becomes a mortal threat only when the company wants a hundredfold return. [4]

Chapter Two: When Learning Is Allowed to Flow in Only One Direction

When other companies begin learning from model outputs, the tone often changes. In February 2026, Anthropic accused DeepSeek, Moonshot, and MiniMax of using fraudulent accounts, proxy networks, and methods that evaded geographic restrictions to collect Claude outputs at scale and use them to train competing models. Anthropic did not claim that all distillation was illegal. It acknowledged that teaching one model with another model’s outputs is a common industry technique. Its allegation was that these particular companies had used deceptive methods to bypass the service’s rules. [8] If those allegations are true, Anthropic has every right to terminate accounts, enforce its contracts, and seek legal remedies. Giving a book away for free at the end does not entitle someone to break into a house to obtain the manuscript.

Yet the episode still leaves a deep discomfort. When model companies learn from millions of books, public websites, and open-source repositories, they emphasize the distinction between learning and copying. They tell us the model created a new capability. When a later company learns from model outputs, the pioneer can suddenly begin speaking as if it created the entire body of value from nothing. The world is an ocean from which the company may learn; the company becomes a private reservoir from which no one else may draw. The two acts may differ legally because of contracts, account fraud, and the origin of the data. Morally, however, the first mover cannot pretend its debt to human knowledge has been paid in full.

The open-weight letter shared by Huang offers a more sensible boundary. Distillation, it says, is widely used for model improvement, evaluation, and validation. When someone unlawfully extracts value from a closed model, targeted contract and legal remedies are appropriate. But the abuse of a technique should not become an excuse to ban the technique itself. [2] It sounds moderate, but the principle is sharp: rules should punish deception, not preserve a lead. They should pursue demonstrated harm, not treat a later competitor’s progress as the harm.

One of the materials provided for this essay was clawd.rip, a sharply written collection of controversies involving Anthropic. [6] I understand the anger that drives it. When a company builds its identity around “safety” and “responsibility,” people respond strongly when it also faces copyright disputes, product regressions, pricing changes, and conflicts with its ecosystem. Anger can tell us where to look. It cannot replace evidence. For that reason, I have used the site only as an index and returned to Anthropic’s own postmortems, court reporting, and the accounts of parties directly involved. Criticizing a powerful company does not require turning it into a cartoon villain. Put the one-way rule on the table, and readers can see the problem for themselves.

What I oppose is the attempt by model companies to decide where learning must end. Suppose a later entrant lawfully studies model outputs, builds a smaller and less expensive system, and releases its weights for universities, businesses, and countries with fewer resources. The project may reduce the pioneer’s profit, but it also gives society more options. Public benefit does not legalize misconduct. Commercial loss does not automatically become a loss for humanity. Model companies deserve a period of advantage and a return on continuing innovation. They do not deserve a permanent line beyond which every future competitor must stop.

Chapter Three: Why Those Who Take More Will Lose to Those Willing to Take Less

In many companies, AI first becomes tangible at the end of the month, when the bill arrives. Sales has connected one model, customer support another, and engineering has purchased a more expensive allocation for a coding tool. The token count climbs, and every department says it is moving faster. Then the finance chief asks, “How much new revenue did this produce? How many errors did it prevent? How many hours did it actually save?” The room often goes quiet for a few seconds. Karp’s criticism stings because he said those quiet seconds out loud: enterprises are paying for tokens that create no value, while model companies are happy to call rising consumption a success. [5]

Enterprises are not unwilling to pay for intelligence. A model that gets a drug into a trial a month earlier, helps an engineer avoid one serious failure, or lets a support team genuinely solve customers’ problems may deserve a high price. Model companies need revenue because chips, electricity, researchers, and failed experiments are real costs. The question is what the customer bought: a better outcome, or merely admission to a black box. Profit earned through a better product is a business. Profit earned because the customer cannot leave is rent. The two can look similar on an income statement. Time reveals that they are very different.

Liang’s line about those who take more losing to those willing to take less captures the mechanism. In the circulating transcript, he imagines one company seeking a large share of the future AI economy. Another firm appears and decides it can live with a smaller share. If the technical gap is not impossible to close, the second company wins customers with lower prices and a more open approach. The process does not drive the price to zero, because a company that takes too little cannot survive. It does keep punishing the company that mistakes a temporary lead for a permanent power to tax the market. [4]

Today’s best model may look like a rare mineral. A few years later, many of its capabilities will be ordinary tools, like compression algorithms and databases. Smaller models will absorb simpler tasks. Distillation will put expensive capabilities into less expensive systems. Lower-precision versions will require less memory, while new chips and compilers continue reducing operating cost. Companies will slowly learn that they do not need a diamond-encrusted knife to cut every potato. The open-weight letter calls this matching the right model to the right job at the right cost: genuinely difficult work can still use the most expensive frontier system, while daily operations move to efficient, specialized models the company can control. [2]

I do not expect some abstract moral force to punish AI companies for excessive profits. The market will do the work. Chasing tenfold or hundredfold returns pushes a company to maintain scarcity. It begins to fear openness, compatibility, and easy migration. It grows accustomed to describing commercial boundaries as safety boundaries. While scarcity lasts, that strategy can generate extraordinary revenue. Once scarcity fades, customers discover that the premium they paid bought not only intelligence but an elaborate set of dependencies. Then the real danger is not falling prices. It is that the company has forgotten how to prove its value in any other way.

A healthy model business can still be enormous. Official APIs are convenient. Frontier capabilities are scarce. Enterprise support, compliance commitments, global capacity, and reliable service all deserve payment. Open weights do not prevent the original developer from becoming the best service provider. Customers may try self-hosting and still conclude that the official service is simpler and cheaper. Every dollar of profit needs to answer only one question: What did you create today? A strong company does not have to brick up the exit. It can leave the door visible and still give customers reasons to stay.

Chapter Four: Chips, Memory, and the Next Tollbooth

Walk into a data center and the first thing you hear is wind. Thousands of machines are pushing heat away. Cables and water lines disappear behind racks while screens scroll through power, temperature, and utilization. News stories compress the whole scene into one term: “GPUs.” The invoice includes much more—high-bandwidth memory, networking, advanced packaging, storage, power, cooling, and software engineering. If any one layer falls behind, even the most expensive chips may sit and wait. Nvidia’s advantage is not an accident. It spent years turning chips, CUDA, interconnects, and tools into a system that works, and that accomplishment deserves a rich return.

A rich return is not the same as a permanently high price. Today’s GPUs and memory are expensive partly because they contain real technical value, partly because supply cannot grow fast enough, and partly because moving to another platform is painfully difficult. When those factors blend together, a shortage price begins to look like the natural price of the future. Huang’s support for open weights also carries an elegant business logic: closed models concentrate compute inside a few labs, while open models encourage thousands of businesses, universities, and governments to build inference systems of their own. Nvidia may sell more hardware as a result. [1] Openness does not ask the shovel seller to leave the gold rush. It lets more people enter.

The price that matters is not necessarily the number printed on the next flagship GPU. Watch how much machine time is needed to complete the same useful task. Better model architectures, lower-precision computation, smarter caching, compilers, and specialized models can let a more expensive machine accomplish far more than its predecessor. In the investor conversation, Liang argued that AI-assisted programming and higher-level tools such as TileLang could make low-level chip programs easier to rewrite, gradually lowering the cost of moving away from CUDA. That prediction remains unproven, and Chinese chips still face real gaps in capacity, energy efficiency, and maturity. But it points to a plausible path: a competitor does not have to equal Nvidia at every layer. It only has to give customers a credible second choice before the toll collector loses the freedom to name any price. [4]

High-bandwidth memory follows the same logic. It is expensive today and may remain expensive for some time, because new capacity, packaging, and supply chains move far more slowly than presentation slides. Yet high margins eventually attract factories, suppliers, and new architectures. Companies do not need to bet that memory prices will be cut in half in a particular year. They need to avoid writing today’s shortage into a ten-year system design. Software portability, a second hardware supplier, and the ability to move between cloud and owned infrastructure may not look like innovation. They can prevent the next tollbooth from being completed before the customer even notices it is under construction.

Chapter Five: Rented Intelligence Can Get Dumber Overnight

On Friday evening, a software team finally pushes a new AI feature into production. The model can read the whole codebase, remember why a function changed ten turns earlier, and move smoothly between tools. Every test passes. Customers are happy. The team goes home with the rare lightness that follows a clean release. On Monday, the product has the same name and the interface looks unchanged, but the model begins forgetting what it just did. It repeats tool calls. Its answers are shorter. Complicated tasks end halfway through. Engineers suspect their own code first, then inspect the data, network, and prompts. Hours later, they discover that people around the world are asking the same question: Did it get dumber?

In April 2026, Anthropic published an unusually detailed postmortem about quality problems in Claude Code. The company had not secretly substituted a smaller model. Three product changes had collided. To reduce waiting time, Anthropic changed the default reasoning effort from high to medium. A caching optimization introduced a bug that repeatedly discarded earlier reasoning, making the model appear forgetful. A system instruction that demanded shorter responses unexpectedly damaged coding quality. Anthropic said the API and underlying inference layer were not affected and denied intentionally degrading the product. It later reversed the changes, repaired the bug, and reset subscriber limits. [9]

The transparency of that postmortem deserves credit. It also exposes an overlooked layer of closed-model services. A customer believes it bought “Claude Opus” or some other stable model name. The experience that reaches the screen also passes through reasoning settings, caches, system prompts, tool wrappers, and traffic allocation. The model can remain the same down to the last weight while the “worker” serving the customer is no longer the person who showed up on Friday. The provider may have no malicious intent. It may be reducing latency, controlling cost, or relieving capacity pressure. Good intentions do not repair a failed deployment or explain to the customer why the product suddenly performs worse.

Beyond quality lies the more basic question of supply. In June 2025, AI coding company Windsurf said Anthropic had given it less than five days’ notice before cutting most of its direct capacity for Claude 3.x models. Windsurf scrambled to find third-party channels and warned users of possible short-term instability. [10] Anthropic co-founder Jared Kaplan later said the company was compute-constrained and wanted to reserve capacity for more durable partnerships. At the same time, reports suggested that Windsurf might be acquired by OpenAI. [13] From Anthropic’s perspective, the choice was commercially understandable. From Windsurf’s perspective, it was like learning during the dinner rush that next week the supplier would stop delivering the ingredient used in every signature dish.

That is why an enterprise should not hand all model inference—the actual running of models across its business—to one outside company. Every firm does not need to train a frontier model or line its basement with GPUs. Critical workflows do need a second model, a version they can pin, an evaluation process they control, and a switching plan that has been tested before a crisis. Hardware can be rented. A mature open-weight model can provide the fallback. The hardest tasks can still call the strongest commercial API. A company does not need to do everything itself. It cannot let someone else completely decide whether the company can still do it at all.

Karp’s call for enterprises to control compute, models, data, and their own “alpha” becomes a list of unglamorous capabilities in daily operations: know where each request went, know which version made a decision, replay yesterday’s result, switch providers when quality drops, and keep the most sensitive information inside an approved environment. [5] An open-weight model will not always be the smartest or cheapest option. It is a backup door that can actually open. When a supplier knows its customers have a door, it often pays more attention to the people still inside.

Chapter Six: Safety Cannot Mean Trusting a Handful of Companies

In July 2026, the Hugging Face security team followed a set of strange traces through its production systems and discovered that an attacker had used a weekend to move through internal infrastructure. The intrusion began with a dataset that looked like ordinary material. Malicious code slipped into the data-processing pipeline, gained execution privileges, harvested credentials, and reached additional clusters. More disturbing, the work had not been done by a lone hacker typing commands one at a time. It was driven by an autonomous AI agent system. Across short-lived sandboxes, it moved like a group of tireless intruders testing doors, searching for keys, and mapping corridors at the same time. It left behind more than 17,000 recorded events. [7]

Hugging Face used AI to chase it. The security team needed to show a model real attack commands, exploit payloads, and command-and-control artifacts so the model could reconstruct what had happened across those thousands of events. It first tried frontier models behind commercial APIs. The requests hit safety guardrails. To the remote model, the material looked like an attempted cyberattack; it could not see that the person submitting it was a defender trying to contain one. The team switched to GLM 5.2, an open-weight model running on its own infrastructure. It completed in hours work that might normally have taken days, while keeping attacker data and referenced credentials inside the company’s environment. [7]

It is one of the most memorable moments in the debate over open and closed AI. The attacker was not going to stop for a usage policy. The defender was the one whose tool asked it to disarm. Yet the story does not prove that open models are inherently safer. Hugging Face explicitly said it did not know whether the attacker had used a jailbroken hosted model or an unrestricted open one. Powerful open weights can also reach bad actors, and they cannot be remotely recalled like a cloud service. Hugging Face’s lesson was more restrained: hosted models need safety guardrails, but defenders should prepare and vet a model they can run inside their own environment before an incident happens. Otherwise, the moment they most need intelligence may be the moment they are neither allowed to use it nor willing to upload the evidence. [7]

That is closer to the real world than declaring either open or closed AI “safe.” Closed services can monitor abuse, update rules quickly, and block attacks at scale. Those are significant advantages. But if society depends on only a few such gates, the same false positive can stop thousands of defenders, and the same outage can remove capability from many companies at once. The open-weight letter makes this point directly: closed systems are not inherently safe. They can be breached, misused, or fail in ways outsiders cannot see. Open weights let more researchers inspect behavior, simulate attacks through red-team testing, find vulnerabilities, and build defenses. They also reduce the risk of concentrating every critical capability in a handful of providers. [2]

Security is not a faith contest between total openness and total closure. It is the work of preparing another path for every failure. Open models need serious evaluations, clear licenses, trustworthy software supply chains, access controls, and runtime safeguards. Commercial APIs need emergency channels for vetted defenders. Enterprises need sandboxes, permissions, logs, and human approval instead of downloading a model and switching off every protection. The most dangerous choice is allowing “trust us” to replace those concrete systems. That weekend at Hugging Face made one thing clear: when attacks move at machine speed, defenders cannot leave their final key in someone else’s pocket.

Chapter Seven: AI Is Not a Mountain Where Only One Flag Can Fly

Imagine a county hospital trying to use AI to help physicians organize clinical notes. Its first concern is rarely whether the model came from San Francisco or Hangzhou. It cares whether the price is affordable, whether patient records will leak, whether the system can work through an outage, and whether it understands the language and illnesses of the local community. A school, a small factory, and a fire department ask similar questions. Then AI enters the national narrative, and those practical concerns are replaced by a different vocabulary: Who is ahead? Who is behind? Who will win the race? Whose model threatens whom? The people at the foot of the mountain are asking how to bring water to town. The people at the summit are arguing over which flag will be planted there.

Competition between the United States and China is real. Chips, talent, standards, military capabilities, and industrial interests all matter. No government will pretend national security does not exist. The danger begins when “the race” stops being one fact among many and becomes the only story we know how to tell. A less expensive Chinese model could mean that more people gain access to AI, yet it is interpreted as an American loss. A breakthrough in an American lab could advance science for everyone, yet it is interpreted as China falling another few months behind. If every step forward must manufacture a loser, cooperation looks weak, openness looks dangerous, and the benefits to ordinary people disappear from view.

The letter Huang shared still bears the title “American AI Leadership.” It is plainly making an argument about American policy and industrial interest. But Huang’s own post contained a wider line: AI will be built by every country. [1] Countries therefore cannot be expected to rent a distant black box forever. They need to choose models, adapt them to local languages and laws, keep sensitive data nearby, and maintain basic capabilities across hospitals, schools, government, and industry when international relations change. The sovereignty created by open weights is not an invitation for every nation to shut its doors and reinvent the wheel. It simply means no nation has to hand someone else the car keys forever.

Turning AI into a life-or-death contest between the United States and China creates a perfect environment for a few companies to protect themselves. A competitor can be called a national threat. Openness becomes capability leakage. Distillation becomes an attack. A customer choosing a foreign model becomes disloyal. Commercial moats put on the uniform of national security. Real risks involving military systems, biology, cyberattacks, and critical infrastructure require export controls, access restrictions, and enforcement. Those rules should target defined capabilities and specific harms. If the security boundary is broad enough to defend every incumbent advantage, eventually it is no longer protecting security.

China should remember this as well. The value of open and inexpensive models is not that a Chinese company wins a leaderboard. It is that an American startup, an Indian hospital, an African school, and a Latin American small business all gain better tools. If Chinese models eventually reach the frontier, Chinese firms will face the same temptation now facing American labs: close the weights, raise the rent, control the ecosystem, and describe the advantage as a national interest. A principle becomes real only when we obey it after taking the lead. If “AI belongs to humanity” applies only while one is catching up, it is merely another slogan in the competition.

America’s deepest strength has never been the ability to keep secrets alone. Universities, immigrants, risk capital, open-source communities, research networks, and a system that lets small teams challenge large firms all contribute to its technological power. Microsoft, Nvidia, Mozilla, the Linux Foundation, Hugging Face, and the other signatories support open weights because they understand that protecting the near-term revenue of a few frontier labs is not the same as protecting America’s long-term capacity to innovate. [2] If American leadership can survive only when the rest of the world has no second option, it is not leadership. It is an expensive and fragile dependency.

Responsible powers do not need to stop competing. They do need to know what should not be treated as a game. China and the United States can build common model-evaluation practices, cyber-incident reporting systems, and boundaries around nuclear and biological uses. They can preserve channels through which universities, public institutions, and countries with fewer resources obtain advanced models. Neither side has to trust the other completely before beginning. Humanity built hotlines and arms-control mechanisms in more dangerous eras. AI is not a mountain with room for one victorious flag. It is more like a river that will run through hospitals, farms, classrooms, and factories. What matters is not who planted a flag at the source. What matters is who can drink, who keeps the water clean, and who is building a gate downstream where only one company may collect the toll.

Conclusion: Charge for the Road, but Do Not Close It

Return to Jensen Huang’s first post. The man selling the chips supports openness because a market this large cannot grow forever through a handful of locked doors. The same is true for model companies. They spend enormous sums on training and deserve to make money—perhaps a great deal of money. But models understand our language and our world because countless people first wrote the books, code, papers, lessons, and fragments of life from which they learned. A company may charge for a better service. It should not turn itself into the only tollbooth between human knowledge and the future.

Enterprises do not need to reject every external service in the name of sovereignty. The best closed models remain worth using. Cloud platforms remove enormous amounts of work. Nvidia’s system may remain the most mature choice. Sovereignty does not mean doing everything yourself. It means retaining the ability to decide when the decision matters: switch when a model gets worse, continue when a supplier fails, keep sensitive data inside, and preserve institutional capability after a contract ends. The deepest value of open weights is not low price or technical ideology. It is making the words “I still have a choice” true outside the negotiation room.

The United States and China will compete, and they will probably compete for a long time. We should not let the score between two nations hide the possibility that billions of people may gain useful intelligence for the first time. AI can charge a price, but it cannot close the road. It can be competitive without turning someone else’s backwardness into victory. It can become one of history’s great businesses without becoming a new empire. The future worth building is not one in which the strongest model makes every decision. It is one in which no model, company, or nation—however powerful—can stop ordinary people from learning, choosing, and continuing down the road.

References

1. Jensen Huang, first post on X, July 24, 2026: “For my first post, I’m sharing a letter NVIDIA signed on why open models matter”.
2. Microsoft Corporate Responsibility, July 24, 2026: “Open Weights and American AI Leadership”.
3. Nvidia, July 24, 2026: “Open Weights and American AI Leadership” PDF. The two screenshots supplied as source material were taken from this letter.
4. “Liang Wenfeng Investor Meeting — Full Transcript,” a transcript supplied as source material. It is not an official, DeepSeek-verified record. For public reporting, see the Securities Times republication of a China Business News article, July 23, 2026: “Liang Wenfeng’s Four-Hour Investor Meeting Leaks: Why Doesn’t DeepSeek Want to Become a ‘More Profitable’ Company?”. A publicly available full-text version appears at PEDaily.
5. CNBC, July 1, 2026: Alex Karp on OpenAI, Anthropic, token spending, and enterprise control.
6. clawd.rip: “Everything That Went Wrong With Claude”. The site has a strong editorial stance and is used here only as an issue index, not as independent evidence.
7. Hugging Face, July 16, 2026: “Security Incident Disclosure — July 2026”.
8. Anthropic, February 23, 2026: “Detecting and Preventing Distillation Attacks”.
9. Anthropic Engineering, April 23, 2026: “An Update on Recent Claude Code Quality Reports”.
10. TechCrunch, June 3, 2025: “Windsurf Says Anthropic Is Limiting Its Direct Access to Claude AI Models”.
11. Associated Press, June 2025: “Anthropic Wins Ruling on AI Training in Copyright Lawsuit but Must Face Trial on Pirated Books”.
12. Associated Press, September 2025: “Judge Approves $1.5 Billion Copyright Settlement Between AI Company Anthropic and Authors”.
13. TechCrunch, June 5, 2025: “Anthropic Co-Founder on Cutting Access to Windsurf: ‘It Would Be Odd for Us to Sell Claude to OpenAI’”.

The author is responsible for the opinions expressed in this essay. All online sources were accessed and checked on July 24, 2026. For developing events, refer to subsequent updates published by the original sources.