Hook
On a nondescript earnings call in late spring, Jensen Huang said something that should have sent a chill down the spine of every closed-model proponent in Silicon Valley. The Nvidia CEO, in his characteristic leather-jacket casualness, declared that open models are the key to AI growth. Not proprietary APIs. Not gated frontier systems. Open.
I read the transcript three times, searching for the nuance. There was none. It was a clean, unambiguous endorsement of a paradigm that, five years ago, would have been considered commercial heresy from the world's most valuable hardware company. But here's what the mainstream coverage missed: this wasn't a philosophical statement about AI democratization. It was a positioning move in a war that has nothing to do with models and everything to do with who controls the pipes.
Truth is immutable, unlike the price action. And the truth is that Nvidia's embrace of open models is the most calculated "altruistic" stance in tech history.
Context
To understand why this matters, you have to understand the tectonic shift happening beneath the AI industry's feet. For the past three years, the narrative has been simple: frontier models require frontier compute, and frontier compute means Nvidia. OpenAI, Anthropic, Google—they all came to the same conclusion: buy as many H100s as humanly possible, then buy more. This created the "scaling law" economics that pushed Nvidia's market cap past $3 trillion and made it the most valuable company on Earth.
But the ground is shifting. Meta's Llama 3, released with open weights, achieved performance that benchmarked within spitting distance of GPT-4. DeepSeek-V3, a Chinese model with a 671B parameter MoE architecture, matched or exceeded closed models on mathematical reasoning and code generation. The performance gap between open and closed models—once estimated at 20-30%—has compressed to something like 5-15% in certain domains.
Here's what most analysts don't tell you: Nvidia's data center revenue hit $47.5 billion in fiscal 2024, a 217% year-over-year increase. That growth was powered by training runs for frontier models. But the next wave of growth won't come from training. It will come from inference—the process of running trained models to generate outputs. And inference is where open models change everything.
Core
The "Selling Shovels" Strategy, Revisited
I've spent years in the crypto education space, and I've seen this play before. In blockchain, the "sell shovels" strategy meant selling infrastructure while staying neutral on which protocols would win. It worked for mining hardware companies, it worked for node providers, and it's working for Nvidia now.
But there's a nuance that the market is missing. Nvidia's embrace of open models isn't just about expanding total addressable market—though that's part of it. It's about expanding the geography of compute demand. Closed models concentrate inference in a few hyperscale data centers. Open models, by their nature, distribute inference across thousands of enterprise deployments, edge devices, and even consumer hardware.
I've audited enough smart contracts to know that when you reduce the barrier to entry, you don't just increase adoption—you create entirely new categories of usage. The same logic applies here. Open models mean a mid-sized insurance company in Ohio can deploy Llama-3-70B on its own infrastructure for claims processing. That's a GPU purchase Nvidia would never have gotten if the only option was paying for GPT-4 API calls.
The math is compelling. According to IDC projections, AI inference compute demand will surpass training demand by 2025. Nvidia's product line—from the H200/B200 for training, the L40S for inference, the L4 for edge, down to Jetson for endpoint devices—looks like a deliberate map of the open-model deployment landscape. TensorRT-LLM, Nvidia's inference optimization engine, already supports Llama, Mistral, and DeepSeek. NIM, their microservices platform, provides turnkey containers for open models.
This isn't neutral infrastructure. This is a land grab.
The Hidden Economics of Open Weights
But here's the insight that should keep both Nvidia bulls and bears awake at night. The open model ecosystem creates a paradox for Nvidia's margins. Open weights mean enterprises can run models on whatever hardware they choose. They can quantize models to 4-bit precision, dramatically reducing memory requirements. A model that needed an H100 at full precision might run acceptably on an L40S with aggressive quantization. Or even on AMD hardware.
I've watched this pattern before, in a different context. In 2020, I wrote extensively about how DeFi protocols were optimizing gas usage to run on cheaper L2s. The result? A flood of new users, but also a compression of fees at the base layer. The same dynamic is emerging here. Open models will expand the total number of GPU deployments—but they'll also push a significant portion of demand toward mid-tier hardware.
Nvidia's gross margins, currently around 75%, are the envy of the semiconductor industry. The question is whether open models, by commoditizing the model layer, will eventually commoditize the compute layer. If every enterprise is running the same open-source Llama variant on whatever GPU they can get cheapest, Nvidia's pricing power erodes.
The counter-argument—and it's a strong one—is that Nvidia's software moat, the CUDA ecosystem, locks in developers even when the hardware becomes more fungible. I've seen this play out in crypto: the chains with the best developer tooling win, regardless of raw performance. CUDA has over 4 million developers. That's a switching cost that AMD's ROCm hasn't been able to crack.
The Selective Openness Contradiction
Here's what bothers me ethically. Nvidia is preaching open models while maintaining a closed stack. CUDA is proprietary. TensorRT-LLM is proprietary. NIM is proprietary. The hardware architectures are proprietary. Nvidia wants the model layer open because that maximizes the number of customers. But Nvidia wants everything else closed because that maximizes margins.
This is the "selective open" strategy, and it's not unique to Nvidia—we see it throughout the tech industry. In the blockchain world, we call this "open-source washing." A project releases its frontend code but keeps the backend proprietary, then claims to be decentralized.
I'm not saying Nvidia is being deceptive. They're being rational. But when Jensen Huang talks about open models democratizing AI, I can't help but hear echoes of the arguments used to justify corporate blockchain consortia—the ones I spent 2017 declining to advise. The language of openness, deployed in service of centralization.
The Real Strategic Target: OpenAI
There's another dimension to this that the financial press has largely missed. Nvidia is OpenAI's key compute supplier. But OpenAI is developing its own chips with TSMC, presumably to reduce dependence on Nvidia. Anthropic is investing billions in its own inference infrastructure.
By publicly championing open models, Nvidia is signaling to the market: "You don't have to be locked into a single API provider. You can build your own AI infrastructure." That's a direct shot across OpenAI's bow. It's Nvidia saying to enterprise customers: "Buy GPUs from me, run open models, and you won't be held hostage by API pricing or model deprecation."
This is the same playbook Nvidia ran with CUDA in the 2000s. Make the ecosystem open and accessible, get developers hooked, then monetize the platform layer. The question is whether the model layer becomes so commoditized that the value shifts entirely to the compute layer—which would be a massive win for Nvidia.
Based on my experience auditing Tezos's consensus mechanism in 2017, I've learned to look for where the real lock-in lives. It's rarely where the marketing says it is. With Tezos, the governance mechanism was the lock-in. With Nvidia, it's CUDA and the software stack—not the open models they're endorsing.
Contrarian
Let me play devil's advocate against my own analysis. The open-model narrative could be a trap for Nvidia, not a strategy.
Consider this scenario: Open models continue to improve and eventually match or exceed closed models across all benchmarks. The model layer becomes fully commoditized. Enterprises deploy open models on whatever hardware is cheapest. Cloud providers like AWS and Azure, already building their own AI chips (Trainium, Maia), offer inference at razor-thin margins. In this world, Nvidia's high-end GPUs become a smaller slice of the market, and their margins compress.
The deeper risk is in the software layer. If open models run well on PyTorch—which they do—and if the optimization tools (vLLM, TensorRT-LLM) improve their support for non-Nvidia hardware, then CUDA's exclusivity erodes. AMD's ROCm has been making steady progress, and the open model ecosystem gives them a perfect testbed for catching up.
I see a parallel in the crypto world. When smart contract platforms opened up their execution environments to be EVM-compatible, they gained adoption but lost differentiation. Everything became a commodity. The same could happen to Nvidia's hardware if the software layer becomes truly portable.
And there's a regulatory angle that concerns me. Open models are harder to govern. Once weights are public, they can't be un-published. If an open model is used for a major cyberattack or biosecurity incident, Nvidia—as the key compute provider—could face reputational and legal exposure. The "tech neutrality" defense works in theory, but in practice, when something goes wrong, the public looks for the most recognizable name to blame. That's Nvidia.
Takeaway
The open-source movement in AI is not a charitable endeavor. It's a structural shift in where value accumulates in the AI stack—and Nvidia's public endorsement is a signal that they've calculated where they stand to win.
For those of us who've lived through the crypto cycles, this feels familiar. The decentralization narrative, deployed by centralized powers to expand their reach. The rhetoric of openness, masking the reality of infrastructure control. The promise of democratization, delivered through the gatekeepers of compute.
But I've also seen this play out differently than the cynics expect. The open model ecosystem, like open blockchain networks, creates genuine opportunities for innovation at the edges. The question is whether Nvidia will be the benevolent infrastructure provider or the dominant gatekeeper of the next era.
Truth is immutable, unlike the price action. And the truth is still being written.