I’ve been managing computer transformations for over 40 years. In the late 70s I was writing BASIC, assembly, and Fortran routines on early terminals. In the mid-1980s I was a front-line Apple Mac desktop publishing (DTP) evangelist and technical author, showing traditional typesetters how PostScript, Illustrator, PageMaker, and Motorola 68000 chips were about to tear apart their legacy workflows overnight. In the mid-90s I was part of the open web revolution, watching proprietary mainframes and walled gardens crumble under the weight of HTML, Linux, and open standards.
I’ve witnessed all of the major paradigm shifts in computing from desktop publishing to enterprise security and cloud architectures, and I know the pattern when history repeats itself.
The historical development of computing platforms in the last forty years is a clear four-decade arc. In the 1980s, it was proprietary phototypesetting rigs until Macintosh PostScript and desktop publishing came along and flipped the whole thing on its head. In the 1990s, proprietary mainframes were traded for the open web and the LAMP stack. The first AI boom had created a new CUDA GPU monopoly by the 2010s. In the 2020s, such a monopoly is challenged by domain-specific ASICs, open-source frameworks, open acceleration architectures such as Huawei’s CANN, Google TPUs, and domestic immersion DUV lithography.
If you pay attention to the mainstream tech media and the Silicon Valley hype, you’d think that today, the CUDA software lock-in and GPU hardware clusters of Nvidia are an unassailable monopoly, a permanent tollbooth on the road to AGI.
They are wrong. The tide is changing. The stranglehold of Nvidia on the AI hardware and software market is being stealthily, systematically, and inevitably dismantled. The primary engine behind the change is a massive influx of open-source AI from the East.
CUDA Moat leaks
Nvidia didn’t print fast silicon and magically become dominant in valuation. Its real lock-in was CUDA, the proprietary parallel computing platform and API introduced in 2006. If you were a developer developing deep neural networks, for 15 years CUDA was the only frictionless road in town.
Software history proprietary lock-in only lasts until the abstraction layer is above it.
Graphics programmers in the 1980s programmed directly against the proprietary display hardware. Open standards such as OpenGL and PostScript abstracted the hardware away. Today, high-level frameworks like PyTorch, Triton, and JAX occupy the space between the developer and the metal. AI engineers today do not write low-level CUDA; they write high-level tensor math.
Meanwhile, global hyperscalers are moving strategically to Application-Specific Integrated Circuits (ASICs) designed for matrix math:
- Google’s TPUs (Tensor Processing Units) power large frontier LLM workloads without a single line of CUDA code.
- AWS Trainium and Inferentia are gaining share in enterprise model training and inference.
- Meta and Microsoft are rolling out bespoke silicon (MTIA and Maia) at data center scale.
If model workloads compile via open layers or run on custom cloud ASICs, Nvidia’s proprietary software advantage dives.
The Sanction Paradox and Eastern Silicon Autonomy
You don’t prevent hyper-capable nations from locking you out of their ecosystem; you accelerate their self-sufficiency. This is strategic technology management.
The Western sanctions aimed to prevent mainland Chinese developers from obtaining top-of-the-line GPUs like the H100 or Blackwell architectures. Rather, these export controls were the ultimate catalyst for technological self-reliance. The application of these sanctions led to a complete hardware decoupling. This decoupling was split into two tracks: one for the rapid development of domestic DUV immersion lithography hardware production and the other for the development of open software acceleration frameworks such as Huawei’s CANN and OpenAI’s Triton. These two directions directly resulted in smarter, highly efficient, open AI architectures such as DeepSeek.
For years the Asian tech giants have been better off simply buying the hardware from Nvidia and building on CUDA. Overnight, the sanctions removed that option, forcing an entire ecosystem to adapt and align domestic hardware with open-source software:
Software Abstraction: Open Acceleration Frameworks
Huawei has developed CANN (Compute Architecture for Neural Networks) to replace CUDA, as part of its Ascend line of AI processors. Rather than shutting it down, Huawei took a page from classic open-source history and open-sourced CANN. Now local chipmakers, such as Cambricon and Hygon, are coalescing around open acceleration frameworks that are natively decoupled from American APIs.
Hardware & Lithography: Immersion DUV Realities
Skeptics said eastern foundries could never produce competitive silicon without access to the most advanced Extreme Ultraviolet (EUV) lithography of ASML.
SMIC and other companies are using the most advanced multi-patterning techniques on local Deep Ultraviolet (DUV) immersion lithography tools and chiplet design platforms to ramp up the fabrication of high-density AI accelerators without foreign supply chains.
China’s Global Dominance of Open-Source AI
As Western tech giants moved towards closed-model architectures, subscription paywalls, and API lock-in, mainland Chinese developers made a bold, strategic decision: to double down on open-source AI.
In software distribution, open source always wins the developers’ mindshare. Metrics from the Hugging Face platform show that models from Chinese labs, including Alibaba’s Qwen, DeepSeek, Kimi, and Zhipu GLM, now make up a massive share of global model downloads and fine-tuning derivatives around the world:
- The DeepSeek Breakthrough Instead of brute computation, labs like DeepSeek focused on structural efficiency, creating Multi-head Latent Attention (MLA) and open Mixture-of-Experts (MoE) architectures that equal closed frontier models with a fraction of the compute and inference cost.
- By releasing model weights under permissive open licenses (like Apache 2.0), Alibaba’s Qwen has given rise to hundreds of thousands of downstream open-source fine-tunes across enterprise, coding, and multimodal applications globally.
By opening up the model weights, training papers and execution code to global developers, China has established open-source AI as the global standard. Developers around the world can host, fine-tune, and run state-of-the-art models on their own infrastructure, thereby completely avoiding proprietary US cloud APIs and CUDA dependencies.
The Age of Open Source
In the 1980s I spent a lot of time convincing corporate leaders that proprietary phototypesetting units were dinosaurs waiting for the tar pit and that open PostScript DTP was the future, running on graphical Macs that even clerical administrators could use. In the 1990s I worked on the proof that open web protocols would replace closed enterprise networks.
The history of AI hardware and software is the same curve.
Today, Nvidia still is a great company, but the ground is shifting under it. The era of GPU monoculture is over. The West has hyperscaler ASICs, and the East has an independent, open-source-driven silicon ecosystem.
In the history of technology, open standards, accessible algorithms, and flexible silicon always win in the long run. That revolution is taking place now.
###
Dr Seamus Phan is head of content at Microwire.news (aka microwire.info), a content outreach and amplification platform for news, events, brief product and service reviews, commentaries, and analyses in the relevant industries. Part of McGallen & Bolden Group initiative. Copyrights belong to the respective authors/owners and the service is not responsible for the content presented.
