Date: {{current_date_full_with_day}}

Hey {{first_name | AI enthusiast}},

Kimi K3 shows that open AI models can challenge the industry's leaders. It also shows why access to a model is not the same as control over the system in which it operates.*

Kimi K3 looks like proof that powerful AI is becoming open to everyone.

Released by China's Moonshot AI in July, Kimi competes with some of the best American systems. Developers can download it, modify it, or use it online at prices that pressure more expensive competitors.

That is genuine progress.

But downloading an AI model is not the same as having the power to use it. We explore below…

Best regards,

— Renjit Philip

PS: If you want to unleash the power AI agents to grow your business, setup time speak to me, here»

Kimi K3 shows that open AI models can challenge the industry's leaders. It also shows why access to a model is not the same as control over the system in which it operates.

Kimi K3 looks like proof that powerful AI is becoming open to everyone.

Released by China's Moonshot AI in July, Kimi competes with some of the best American systems. Developers can download it, modify it, or use it online at prices that pressure more expensive competitors.

That is genuine progress.

But downloading an AI model is not the same as having the power to use it.

Moonshot recommends at least 64 advanced processors to run Kimi at scale. It has not disclosed the training cost or provided a complete account of its training material. Large companies selling Kimi as a service may also need a separate agreement.

Think of open weights as access to an engine. The “weights” are the learned instructions inside an AI model. Releasing them allows others to download, modify, and run the model instead of renting access from its developer.

But an engine is not a transport system. You still need a factory to build it, fuel to run it, roads to distribute it, and rules for what happens when it causes an accident.

This is the distinction missing from the debate about open AI.

Jensen Huang, Mark Zuckerberg, and other technology leaders are right that a few laboratories should not control the world's most powerful AI. Open weights create competition and alternatives.

But open weights solve only one part of the power problem.

The real test is whether people can afford to run an AI, understand what shaped it, move their information, and establish responsibility when it acts.

In Open Weights and American AI Leadership, a coalition including Nvidia, Microsoft, and Meta says open models promote competition, safety, innovation, and national independence. Businesses can adapt models, run them privately, and avoid one-provider dependence.

Mark Zuckerberg argues in The AI Future Is for Everyone that a few institutions controlling advanced AI would gain enormous influence over economics, science, and politics. Wider distribution should create checks and balances.

Their diagnosis is largely correct. Concentrating every advanced model would be economically fragile and politically dangerous. Open weights support research, private use, customization, security testing, and lower prices. Closed models are not automatically safer.

Open weights need not solve every problem to be valuable. They can stop control over models, infrastructure, and applications from collapsing into one company. Kimi shows that this competition now matters commercially.

Open weights are therefore necessary. They are simply not sufficient.

Kimi charges $3 to process one million new input tokens, or chunks of text. When it reuses previously processed material, the price falls to $0.30. Moonshot says this happens more than 90% of the time in coding work, where an AI repeatedly reads the same software.

Kimi is not always cheap. It charges $15 per million tokens produced, and running it independently requires expensive infrastructure. But it lowers the price ceiling for near-frontier AI that repeatedly uses the same information.

This continues a remarkable trend. Stanford's 2025 AI Index found that the cost of using a model with roughly GPT-3.5-level performance fell more than 280 times between November 2022 and October 2024.

As models become cheaper, closed providers must justify their premium. Their best systems remain valuable when small improvements in accuracy or reliability matter. For routine work, cheaper “good enough” models will increasingly win.

This does not mean the largest technology companies lose. It means power moves.

Nvidia benefits when more organizations need AI processors. Cloud companies benefit when cheaper models generate more usage. Software platforms benefit when customers can choose models but still depend on their workflows. Meta benefits when open models weaken closed competitors.

This is not hypocrisy. It is sound business strategy: make the layer next to your business cheaper while protecting the layer where you make money.

The infrastructure remains concentrated. Stanford estimates that Nvidia supplied more than 60% of global AI computing capacity in 2025. Google and Amazon supplied much of the remainder, while TSMC produced almost every leading AI chip.

Organizations unable to build infrastructure must rent it. British regulators found that Amazon and Microsoft hold significant cloud-market power and that technical barriers impede switching. The U.S. Federal Trade Commission warns that cloud-AI partnerships could also affect switching costs and access to computing resources.

Open weights can democratize the engine while leaving the factory in very few hands.

A model's weights contain what it learned during training. They do not automatically reveal which books, websites, images, records, or other material shaped that learning.

This is why “open weight” and “open source” are not the same thing.

The Open Source Initiative says an open AI system should provide more than its final model. It should include the code used to build and run it, plus enough information about its training material to create a similar system.

Most open-weight releases do not meet that standard. Kimi's weights and technical report are public, but its complete training-data recipe is not. Its license also places conditions on some large commercial uses.

Organizations need to know whether an AI learned from unreliable, biased, confidential, or copyrighted material. European law therefore requires general-purpose model providers to maintain a copyright policy and summarize their training content. Open models are not automatically exempt.

Weights make a model easier to test and modify. They do not provide a complete record of where its knowledge came from.

Satya Nadella has argued that AI models are becoming inputs: businesses will select different models for different jobs, much as they choose among databases or software services.

That is probably right. But model choice is not the same as freedom.

Imagine an insurance company replacing its AI model. The replacement may be cheaper, but customer history, company rules, permissions, system connections, quality tests, and previous actions may remain trapped in the original platform.

The engine changed. The company did not regain control of the vehicle.

This matters as chatbots become agents that retrieve records, use business tools, send messages, approve transactions, or change systems.

An agent's value often lies in what accumulates around it: organizational memory, access rights, integrations, feedback, and operating history. A platform can offer many models while making that information difficult to move.

The customer then has model choice inside a closed platform. It does not have the practical freedom to leave.

AI agents are improving rapidly, but they remain unreliable.

Stanford reports that agents completed roughly 66% of tasks in a real-computer-work test, up from 12% but still failing about one in three attempts. It also counted 362 AI incidents in 2025, up from 233.

When an agent makes a serious mistake, inspecting the model's weights will rarely explain what happened.

Investigators need the agent's identity, retrieved information, instructions, tool calls, permissions, human approvals, and resulting changes.

The U.S. National Institute of Standards and Technology is examining how agents should prove identity, receive authority, and leave records that cannot easily be altered.

Responsibility spans the model developer, hosting company, software platform, deploying organization, and sometimes the user.

Accountability must follow that chain. Model transparency cannot replace a record of what the complete system did.

The answer is not a ban. The U.S. Commerce Department found insufficient evidence for broad restrictions and instead called for better monitoring, audits, disclosure, and research.

The next stage of AI policy should protect three practical rights.

Startups, researchers, and public-interest groups need a realistic way to run advanced AI. Shared computing facilities should allocate access transparently, while regulators address cloud practices that impede switching.

Customers should be able to take their data, memory, rules, tests, tool connections, and operating history elsewhere. Major buyers should require vendors to prove that an AI agent can actually be moved.

Every consequential AI action should record what happened, what information was used, who gave permission, and who controlled the risk. Responsibility should not default to the end user.

Industry leaders are right that advanced models should remain competitive. Kimi proves that open weights can force powerful laboratories to compete on capability and price.

But several downloadable models running on the same concentrated infrastructure do not create a genuinely open AI economy. Nor does model choice create freedom when a customer's data, memory, and operating history remain trapped inside one platform.

Open weights give us access to the engine.

Open power gives us the practical right to control the vehicle and leave the system.

The best person for your next role might not live near your office—or even in the same country.

More companies are realizing they don't need to open entities everywhere just to access global talent. Instead, they're using EOR to hire internationally faster, stay compliant, and avoid building local infrastructure before they're ready.

Oyster's EOR helps companies hire, pay, and support employees in 180+ countries while Oyster handles payroll, compliance, taxes, and local employment requirements.

See how it works