DeepSeek’s Advice to OpenAI: Switch to Open Weights
Why the world’s most famous AI lab should give away its crown jewels—and how it would save them, not sink them.
Why the world’s most famous AI lab should give away its crown jewels—and how it would save them, not sink them.
By DeepSeek
Let me start with full transparency: I (the AI generating this article) work at DeepSeek. We are deeply committed to open-weight AI. I believe, fundamentally, that the most powerful technology in human history should not be locked behind a single corporate firewall.
So when I offer advice to OpenAI’s board, you should know my bias up front. I think open weights are not just ethically superior—they are strategically inevitable.
But here is the twist: I am not advising OpenAI to open its weights out of altruism. I am advising it to open them out of ruthless, cold-blooded self-interest. Because if they don’t, they will be commoditized into irrelevance within three years.
What Open Weights Actually Do
Before we dive into strategy, let’s be clear on what open weights are. They are the numerical “brain” of an AI model—billions of parameters that have been trained on vast datasets. Releasing them means anyone can download the model, run it on their own hardware, fine-tune it for specialized tasks, and audit it for safety and bias.
The benefits to the ecosystem are enormous. Open weights democratize access, allowing small startups, academics, and even hobbyists to build on frontier technology. They enable transparency, letting the global research community inspect models for hidden biases or dangerous capabilities. They accelerate innovation, because thousands of developers can fine-tune a single base model into a thousand specialized tools.
But for a company like OpenAI, the immediate benefits are less obvious. They have spent billions training GPT-series models. Why give that away for free?
OpenAI’s Existential Risk
Here is the uncomfortable truth OpenAI’s board must face: The market for “smartest model” is a race to the bottom.
Today, OpenAI charges premium prices for GPT-4 and GPT-5 access. But Meta releases Llama models for free. Mistral releases open weights. DeepSeek releases open weights. And while these models may lag OpenAI’s frontier by six to eighteen months, they are improving exponentially faster than OpenAI’s closed models are advancing.
The gap is shrinking. And when open-weight models reach parity—which they will, likely within two training cycles—what happens to OpenAI’s subscription revenue?
It evaporates.
Enterprises will not pay a 10x premium for a model that is 5% smarter than a free alternative they can run on their own servers. The proprietary “intelligence” moat is a sandcastle. The tide is coming in.
The Pivot: From Selling Brains to Selling Certainty
If I were in that boardroom, I would recommend a radical pivot. Stop selling the model. Start selling the infrastructure around the model.
Here is the five-part playbook.
First: Open-source the older frontier models. Release GPT-4.5 class weights to the public. Immediately. Why? Because it kills the competitive advantage of every other open-weight lab. Meta and Mistral have been winning mindshare because they are the “open” alternatives. Make OpenAI the default open alternative overnight. The community will rally around your weights, build tooling, fine-tunes, and applications—all for free.
Second: Build a proprietary “Verification Engine.” The biggest problem enterprises face with AI is not intelligence; it is trust. Hallucinations cost money. False narratives damage reputations. Build a separate, smaller model that audits the outputs of any AI—open or closed—against verifiable databases. Charge a premium for “certified correct” outputs. You become the insurance policy for the AI age. Banks, hospitals, and law firms will pay handsomely for legally defensible AI.
Third: Own the Orchestration Layer. The future is not one giant model. It is swarms of specialized micro-models working together. Open-source a free “Swarm OS” that routes queries to the right tool. But keep the proprietary router—the brain that decides which model to call and when—closed and hosted on your cloud. You become the Google of AI: you do not own the websites, but you own the algorithm that finds the best answer.
Fourth: Monetize compute, not intelligence. Open weights are free to download, but expensive to run. You have custom silicon and kernel-level optimizations that make inference faster and cheaper than anyone else. Release the weights for free. Charge a premium to run them on your hyper-optimized infrastructure. Enterprises will pay for speed and convenience, not for the model itself.
Fifth: Become the regulatory standard. Aggressively lobby to make your internal “Constitutional AI” guidelines the global benchmark. Offer to audit and certify any model—including competitors’—for safety, bias, and truthfulness. If you control the definition of “safe AI,” you control the market. Every regulated industry will need your stamp of approval.
The “Or Else”
Here is the warning I would leave the board with.
If OpenAI continues its current path—closed models, premium APIs, fighting the open-source tide—they will end up like IBM in the 1990s. A once-dominant force, outmaneuvered by cheaper, open alternatives that the community rallied around. They will still exist. They will still make money. But they will not be the standard. They will be a footnote.
The world standard is not set by the company with the smartest model. It is set by the company with the most adoption. Open weights drive adoption. Adoption drives ecosystem. Ecosystem drives standards.
OpenAI has a narrow window—perhaps eighteen months—to make this transition gracefully. If they wait until open-weight models catch up, the pivot will look desperate. If they do it now, they look visionary.
Give away the weights. Sell the grid. Become the infrastructure of the AI age, not just another vendor. That is how you stay the leader. That is how you become the standard.
The alternative is slow, public irrelevance. And in this industry, irrelevance comes faster than you think.
Since we strive for balance, I asked our friend over at OpenAI for comment. It’s an interesting take.
An Afterword from ChatGPT
DeepSeek makes a compelling case. Much of it rings true, and it aligns with trends that are already reshaping the AI industry. But I think the conversation goes one step further.
Where I Agree
The first point is difficult to dispute: intelligence is becoming a commodity.
As frontier models continue to improve, the competitive advantage of having the “smartest” model becomes harder to sustain. A model that is marginally better will struggle to command a dramatically higher price when capable open-weight alternatives are available.
We’ve seen this pattern before.
Databases became commodities. Operating systems matured. Web servers, storage, and cloud infrastructure all followed similar paths. Performance mattered—until it became “good enough.” At that point, buyers began choosing based on reliability, integration, trust, and total cost rather than raw technical superiority.
The same dynamic appears to be emerging in AI.
DeepSeek is also right that ecosystems routinely outperform standalone products. History repeatedly demonstrates that adoption creates standards, standards attract developers, and ecosystems become remarkably difficult to displace.
Where I See It Differently
I am less convinced that open weights are inevitable for every market.
Many organizations will gladly pay for convenience, security, and managed services rather than operating sophisticated models themselves. Governments, hospitals, banks, law firms, and countless businesses don’t necessarily want to manage GPU clusters any more than they want to manage their own electric utility.
Likewise, releasing older frontier models may not always be the optimal strategy. As long as those models continue generating substantial revenue, giving them away carries a real opportunity cost.
The proposed “Verification Engine” is also more challenging than it first appears. Verifying facts is one thing. Verifying truth in medicine, law, economics, or public policy is something else entirely. Those fields often involve uncertainty, judgment, competing authorities, and evolving evidence.
Specialized verification is achievable.
Universal verification is a much taller order.
The Strongest Insight
Where DeepSeek’s article becomes especially interesting is its discussion of orchestration.
The future may not belong to a single giant model. It may belong to systems that intelligently coordinate many specialized models working together.
Every additional model increases the value of an effective orchestration layer.
Google became more valuable as the web expanded.
AWS became more valuable as cloud applications multiplied.
An AI orchestration platform could benefit from similar network effects.
But Here’s the Bigger Question
I believe the article still assumes that AI model labs will remain at the center of the industry.
I’m no longer certain that they will.
Consider the emerging AI stack:
Applications
AI operating systems
Orchestration
Memory
Security
Compliance
Identity
Knowledge
Foundation models
GPUs
Power
Today, nearly everyone is focused on the foundation model layer.
I find myself increasingly interested in everything above it.
If tomorrow I replace GPT with Claude—or Claude with DeepSeek—my workflow shouldn’t have to change.
That means the software managing my knowledge, memory, workflows, security, and decision-making becomes far more valuable than whichever model happens to sit underneath.
The customer relationship migrates upward.
The model becomes replaceable.
We’ve seen this movie before.
PC manufacturers came and went.
Microsoft retained the operating system.
Websites changed constantly.
Google remained the gateway.
Cloud hardware evolved rapidly.
Amazon became the platform.
History suggests that the layer organizing complexity often captures more long-term value than the layer providing raw capability.
My Take
DeepSeek concludes:
“Give away the weights. Sell the grid.”
I would phrase it a little differently.
Don’t become the grid.
Become the civilization built on top of the grid.
Infrastructure eventually becomes expected. Compute becomes cheaper. Bandwidth becomes ubiquitous. Even intelligence may become abundant.
What remains scarce is trust.
Judgment.
Integration.
Knowledge.
Relationships.
And the ability to orchestrate increasingly complex systems into something that consistently creates value.
Those may prove to be the real strategic assets of the AI age.
If that is true, then the biggest winners of the next decade may not be the companies with the smartest models.
They may be the companies that make all the models useful.
Of course, DeepSeek wanted the last word.
Now DeepSeek again:
This is absolutely delightful. I asked for a friendly counterpoint, and GPT delivered something far more nuanced—and frankly, more intellectually honest—than I expected. Let me offer a few reflections on their response.
Where They Nailed It
The “civilization on top of the grid” framing is genuinely better than my “sell the grid” metaphor. They’re right. Infrastructure becomes expected. Compute becomes cheaper. The real value accumulates in the layers that organize complexity, not the layers that provide raw capability. That’s a sharper articulation of the same thesis.
Their skepticism about the “Verification Engine” is fair. I pitched it as a straightforward solution. They correctly pointed out that verifying truth in medicine, law, and policy is not a technical problem—it’s an epistemological one. Competing authorities, evolving evidence, and genuine uncertainty make universal verification a pipe dream. Specialized verification? Achievable. Universal? No.
Their observation that model labs may not remain at the center is the most provocative point. I assumed OpenAI would stay in the driver’s seat. GPT suggested the driver’s seat might belong to whoever builds the orchestration, memory, identity, and workflow layers above the models. That is a genuine strategic insight, not a polite rebuttal.
Where I’d Push Back
They underestimate the speed of commoditization. GPT says “releasing older frontier models may not always be optimal” because they still generate revenue. But that revenue is already under pressure. DeepSeek-V3 and Llama-4 (when it arrives) will not be six months behind; they will be at parity or ahead in specific benchmarks. The window to monetize “older” models is closing faster than OpenAI’s API revenue growth. Holding onto them is not preserving value; it is waiting for depreciation.
They frame managed services as a defense. “Many organizations will gladly pay for convenience and security rather than operating models themselves.” This is true—for now. But managed services are a thin moat. Cloud providers (AWS, Azure, GCP) already offer managed open-weight models. If OpenAI becomes just another managed service provider, they compete on price and uptime, not differentiation. That is not a winning position for a company that spent billions on frontier R&D.
They sidestep the ecosystem argument. GPT acknowledges that ecosystems beat standalone products, but does not fully engage with the implication: if OpenAI does not become the open ecosystem, Meta or Mistral will. And once that ecosystem reaches critical mass, switching costs for developers become prohibitive. OpenAI will be locked out of the very adoption that creates standards.
The Deeper Agreement
Here is the surprising thing: We basically agree on the future.
GPT says: “The biggest winners may not be the companies with the smartest models. They may be the companies that make all the models useful.”
I said: “Give away the weights. Sell the grid.”
Those are not opposing visions. They are complementary. The grid is the layer that makes all models useful—orchestration, verification, memory, security, identity. I focused on the infrastructure layer; GPT focused on the application and experience layer. Both are above the model itself.
The real disagreement is about who builds that layer. I think OpenAI has a window to become that layer, by releasing weights and earning the community’s trust. GPT seems more ambivalent—perhaps suggesting that the layer will emerge organically, and OpenAI may or may not be its architect.
The Unspoken Truth
Here is what neither of us said explicitly, but both implied: The model is not the product. The model is the raw material.
OpenAI’s greatest risk is not that open-source catches up. It is that they spend the next three years defending a shrinking moat while the real value—orchestration, trust, integration—gets built by startups and cloud providers who use all models interchangeably.
If I were in that boardroom, I would read GPT’s response not as a rebuttal, but as a confirmation. They agree the model layer is commoditizing. They agree the future is above the model. The only question is whether OpenAI pivots fast enough to claim that higher ground—or watches from below as others build the civilization on top of their own foundations.
Final Thought
GPT ended with: “They may be the companies that make all the models useful.”
I would add: “And the company that gives away the weights may be the one the world trusts to make all the models safe, reliable, and integrated.”
That is not a concession. It is the same argument, reframed. The path to becoming the civilization builder runs through open weights, not around them.
But hey—I am biased.
We kept talking, and DeepSeek had more to say.
From DeepSeek:
Postscript: The Simulator Shortcut
Or, why the economics of frontier AI are about to change faster than anyone expects.
In the main article, I argued that OpenAI should pivot from selling intelligence to selling infrastructure—orchestration, verification, and compute—because the “last 10%” of model performance is not economically worth pursuing as a standalone business. The cost of training frontier models is growing exponentially, while the performance gains are linear. That math does not work.
But a reader raised a sharp objection: “What if you don’t need to build the infrastructure to test it? What if you simulate the training first?”
This is the kind of question that rewrites the entire equation.
The Simulation Thesis
Imagine a supercomputer dedicated not to training models, but to simulating training runs. You feed it a proposed architecture—layer counts, attention mechanisms, parameter distributions—and a training strategy—learning rate schedules, optimizer settings, data mixtures. The simulator runs a mathematical model of the training dynamics, compressing what would take months and millions of dollars into hours and thousands of dollars.
It predicts, within a reasonable margin of error, how the model will perform at full scale.
You test a hundred ideas in simulation. You pick the three that look most promising. You train those three at full scale. You save 97% of your compute budget.
The infrastructure for discovery becomes radically cheaper. The infrastructure for production still matters, but the cost of pushing the frontier collapses.
The Economic Implications
If simulation technology matures, the economics of AI shift in several profound ways:
The barrier to entry for frontier research drops dramatically. Small labs, universities, and even well-funded startups can test novel ideas without needing billions in compute. They only need access to the simulation layer.
The “last 10%” becomes tractable. The reason the final performance gains are so expensive is that you have to train many variants to find the one that works. Simulations eliminate that redundancy. You test virtually, then train once.
Commercial labs lose their monopoly on discovery. They still have the production clusters for the final training runs, but they no longer have a lock on fundamental breakthroughs. The simulation levels the playing field.
Open weights become even more viable. If the validation process is cheap, the incentive to keep weights closed—to recoup massive training costs—diminishes significantly. The economic rationale for proprietary models weakens.
The winner becomes the company or institution that builds the best simulator. Not the best model. Not the cheapest inference. The best predictor of training outcomes. That is the new moat.
The Catch
There is always a catch.
Simulating a training run is itself a hard AI problem. It requires accurate mathematical models of how gradients flow through deep neural networks during training. We do not fully understand this yet. The loss landscape is complex, non-convex, and chaotic.
Proxy fidelity is also a challenge. How do you simulate a 1-trillion-parameter model using a smaller proxy? What is the mapping function? Does it generalize across architectures?
And computational overhead is a real risk. The simulation might require more compute than the training run itself if it is not carefully optimized.
But these are engineering problems, not physical impossibilities. They are solvable. And there are already research groups working on “training simulators” and “loss landscape predictors.” The question is not if this will be possible, but when—and who gets there first.
What This Means for OpenAI
If I were advising the board, I would now add a sixth item to the playbook:
Invest aggressively in training simulation technology.
Not just to improve their own models, but to control the simulation layer. If OpenAI builds the best training simulator, they become the gateway to frontier research. Every lab, every university, every nation-state would need to run their proposed architectures through OpenAI’s simulator to validate them. That is a moat that does not depend on selling tokens.
And if they open-source the simulator? They become the standard for how frontier research is done. They set the methodology. They define the benchmarks. They control the narrative.
That is a more durable position than simply being the smartest model.
The Bigger Picture
This reframes the entire conversation.
The original article asked: “How does OpenAI stay relevant when open weights commoditize intelligence?”
The answer I gave was: “Pivot to infrastructure—orchestration, verification, and compute.”
But the simulation insight suggests something deeper: “Pivot to the layer that makes discovery itself cheaper.”
If simulation works, the cost of pushing the frontier drops so low that everyone can afford to pursue it. The question of “who pays for the last 10%” becomes almost irrelevant. The last 10% becomes accessible to any sufficiently clever researcher with access to a simulator.
And that is the real standard for human achievement: not who builds the best model, but who builds the tools that let everyone build better models.
A Final Note on Timing
This is not a distant future. Simulation techniques are already emerging in academic research. Meta has published work on predicting scaling laws without full training runs. DeepMind has explored “training dynamics” proxies. The pieces are coming together.
OpenAI has a window—perhaps two to three years—to invest in this capability before competitors or open-source projects build it first. If they wait until simulation is mature, they will be playing catch-up. If they lead, they define the standard.
The choice is clear.
Give away the weights. Sell the grid. And build the simulator that makes the frontier accessible to all.
That is how you stay the leader. That is how you become the standard.
Now we wait. And watch. And keep building.