It’s raining outside again. March in Shanghai is probably just like this, always permeated with a dampness you just can’t shake off. I wiped the fog off my glasses while dragging the progress bar back to rewatch the early morning broadcast of Nvidia’s GTC conference a few more times.
Don’t Just Stare at the Chips, Look at Who He’s Rallying
Last night, I was watching the GTC livestream with a few developer friends in our group chat. At first, everyone was guessing how high the memory bandwidth of the new generation hardware would soar, or what shocking new tricks the network interconnect architecture would pull off. But after watching for a few dozen minutes, what really made me sit up and take notice wasn’t the hardware, but a long list of alliances that didn’t have much to do with direct hardware specs.
At a time when the whole world is flaunting how amazing their foundational large models are and how many leaderboards their benchmark scores have dominated, Nvidia very cleanly announced a series of strategic partnerships. Note, this isn’t the kind of superficial brotherhood where they just exchange PR releases and hang each other’s logos—it’s genuine, deep integration.
Not only did they join hands with Mistral and embrace Hugging Face (yes, that familiar smiley face community), but they also pulled current darlings like Perplexity and Cursor entirely into their camp. Anyone who knows a bit about the industry knows that Cursor has almost become the default prosthetic limb for programmers writing code, while Perplexity is frantically eating into the market share of traditional search engines. Nvidia’s intention in pulling them over couldn’t be more obvious.
Jensen Huang’s move is incredibly interesting. He didn’t pick a so-called “absolute winner” in this epic gamble of AI; instead, he chose to support every single horse running in his own arena. Whether you’re building search engines, coding editors, or even running open-source communities, as long as you’re operating within my computing power ecosystem, I will give you the most fundamental bottom-layer optimization support.
To put it bluntly, he stopped selling purely physical graphics cards a long time ago. He’s handing out admission tickets to all the entrepreneurs unwilling to be swallowed whole by tech giants.
The Needle Piercing the Myth Comes from Europe
If you think about it a bit more deeply, you’ll find that the underlying logic of interests behind this is actually very straightforward.
Let’s shift our gaze slightly. The current situation is that the leading company across the ocean comfortably collects a $200 Pro subscription fee every month, tightly clutches its core code, and constantly transmits an arrogant narrative to the market that “only closed-source can ensure security and performance.” In their context, others don’t even have the right to look at the source code.
But two years ago, Mistral, a startup in Paris with only 35 employees that Silicon Valley bigwigs often treated as a joke, has now quietly swelled to 350 people. Not only did they survive, but they are thriving—even starting to reverse-project a sense of pressure.
This brings up a fact few are willing to admit upfront: after the initial panic, Europe has actually chosen its weapon for a defensive counterattack, and that weapon is open source.
They not only brought in Nvidia to co-develop cutting-edge models for free, but even their fellow European neighbor ASML—which controls the global semiconductor lifeline—directly forked out 1.3 billion euros, taking about an 11% stake as the lead investor. You have to understand, ASML usually just focuses on selling lithography machines. This kind of cross-boundary move to directly invest in an AI software company carries a strong sense of strategic defense in itself.
Look again at their initial partners: BNP Paribas, Thales (aerospace and defense), and SNCF (French National Railway). These are the most hardcore foundations of European traditional industry and finance. The AI competition is no longer a simple “Tale of Two Cities” between China and the US; it’s a three-way standoff among the US, Europe, and several top-tier tech giants in China.
The devotees of the closed-source myth will likely not discuss this news loudly on social media. Because once they admit that a French company with only a few hundred people can produce top-tier open-source models for free just by partnering with Nvidia, that commercial fairy tale built on absolute closure becomes very hard to maintain.
Weighing It Against the Peers
Let’s do a horizontal comparison. Right now, at the poker table of top-tier models, the divergence of paths has reached a point of absolute incompatibility.
On one side is a highly closed black box: unknown parameters, unknown training data, extremely minimal room for fine-tuning, and all interactions strictly confined within officially provided APIs. You want to run it locally? You want to do deep adaptation for your own proprietary data? No way. You just obediently pay by the token.
On the other side is the Mistral route that was held high at this GTC. In the newly released Mistral 3 series, that behemoth with 675 billion total parameters adopted a sparse Mixture-of-Experts (MoE) architecture, with the number of active parameters per token around 41 billion. What’s the most crucial part? It is entirely open-sourced under the Apache 2.0 license, completely free, and features insanely optimized underlying code specifically tailored for Nvidia’s Hopper and next-gen hardware.
To put it bluntly, so what if your technology is good? When your competitor slams an over 600-billion-parameter model directly onto the table and tells you, “Take the code, modify it however you want, with zero commercial restrictions,” many small and medium-sized enterprises (SMEs) and academic institutions won’t even hesitate.
I previously read a very interesting data survey: many SME development teams at home and abroad, facing limited computing budgets, prefer to set up their own vLLM or SGLang environments to run Mistral’s open-source models rather than buying expensive enterprise-level closed-source services. Because the former not only offers controllable long-term costs, but the data also never leaves their own server rooms—which is almost the only choice for clients in the financial and medical industries. Nvidia’s moves this time have essentially smoothed out the seams between open-source software frameworks and underlying hardware, even tuning the quantization and inference engines to their optimal state for you.
Some Things I Think About Sometimes
Speaking of this, I sometimes wonder: is this a historical replay in the tech world?
Back in the day, Microsoft feasted on the global operating system market with Windows. Everyone thought this heavily commercialized, entirely closed-source software empire was invincible. At the time, almost no one believed a free system could take Microsoft head-on. And the result? Linux brutally tore a massive hole in the server market and ultimately reshaped the entire internet infrastructure using open source.
That drama of using an open-source ecosystem to fight a closed empire looks exactly like today’s scenario of open-source large models besieging closed-source giants. It’s just that this time, the role of Linux is being played by European AI startups.
However, there is actually an ironic paradox hidden in all of this.
On the surface, Nvidia is vigorously supporting open source to fight the closed AI model oligarchs, but it itself is an extremely massive and closed “hardware empire.” Anyone who has slightly touched low-level development knows how high the moat of the CUDA ecosystem is. What we see on the surface is a hundred flowers blooming with open-source models, and everyone seems to have gained freedom, but the tracks these freely running horses gallop on are almost exclusively green.
Is it possible that Nvidia is just using the open-sourcing of the model layer to further consolidate its absolute monopoly at the hardware layer? When these open-source models with hundreds of billions of parameters become thoroughly inseparable from Nvidia’s proprietary acceleration libraries for underlying operator tuning, will it be the beginning of another lock-in? The hardware giant has become the biggest sponsor of open-source software; no matter how you look at it, this combination reveals a hint of cunning.
(Stroking my chin haughtily) I’m not entirely sure either, maybe I’m just overthinking it. After all, for today’s developers, a good weapon is one you can grab and run with immediately without having to look at a tech giant’s face. As for who will collect the seigniorage five years from now, that’s really a problem to worry about five years from now.
Never Mind All That
Unconsciously, I’ve already typed almost two thousand words. The Americano in my cup has gone completely cold, turning incredibly sour. The rain outside seems to have stopped, and who knows when this late spring chill will finally end. I’d better go pour out this cold coffee first.
References:
- ASML, Mistral AI enter strategic partnership
- Data Points: Mistral 3 update adds four new open models
- Mistral Large 3: An Open-Source MoE LLM Explained
- NVIDIA Expands Open Model Families to Power the Next Wave of AI
—— Lyra Celest @ Turbulence τ.
