CUDA lock-in is a real business risk, but it has moved. At the level most companies actually work, frameworks such as PyTorch, vLLM and SGLang, AMD's ROCm now runs mainstream models, and large buyers are hedging: OpenAI agreed in October 2025 to deploy 6 GW of AMD GPUs. The lock-in that remains sits lower down, in hand-written CUDA code, NVIDIA-only tools, and the maturity of advanced serving features. SemiAnalysis judged AMD competitive in several FP8 inference cases in February 2026, but more than six months behind when advanced distributed-serving techniques are combined. For an SME the risk is less a licence trap and more the time and expertise a less-travelled software stack demands. On workstations, NVIDIA's lead in tooling is wider still.
Where is the lock-in, layer by layer?
| Layer | NVIDIA | AMD (Nov 2026) | Lock-in level |
|---|---|---|---|
| Low-level kernels | CUDA 13.x, cuDNN, TensorRT-LLM | ROCm 10.1, HIP | High: hand-written CUDA must be ported |
| Frameworks | PyTorch CUDA builds | PyTorch ROCm builds (Linux) | Low to medium |
| Serving | vLLM, SGLang, TensorRT-LLM, Dynamo | vLLM (ROCm wheels), SGLang | Medium: features and test coverage lag |
| Translation layers | n/a | ZLUDA: llama.cpp works, PyTorch not robust | CUDA licence restricts translating CUDA output |
| Rack scale | GB300 NVL72 in production | Helios MI455X: AMD says volume in 2H 2026; SemiAnalysis expects production tokens in Q2 2027 | High at frontier scale |
| 32 GB workstation card | RTX 5090, RTX PRO range | Radeon AI PRO R9700 (US$1,299 launch) | Tooling gap wider (community reports) |
Source: AMD ROCm release history (10.1.0, 5 Oct 2026); Phoronix (27 Aug 2026; 13 Jan 2026); SemiAnalysis InferenceX v2 (16 Feb 2026); AMD and OpenAI announcement (6 Oct 2025); Club386 (Oct 2025). The "lock-in level" column is SourceIT's summary of these sources.
How far has AMD's software come?
A long way since 2024. In December 2024, SemiAnalysis's five-month training study described AMD's out-of-box experience as "very difficult to work with", and its analysts needed AMD engineers to fix bugs. AMD's chief executive, Lisa Su, acknowledged the gaps.
Since then AMD has moved to a six-week release cycle. ROCm 10.0, released in August 2026, jumped the version number to end confusion over 7.x releases and added production vLLM containers and support for local fine-tuning on Ryzen AI Max (Phoronix). By February 2026, SemiAnalysis found the MI355X competitive with NVIDIA's B200 on performance per total cost of ownership in several FP8 serving setups on SGLang.
The same report found the gaps that matter for production. When disaggregated serving, wide expert parallelism and FP4 are combined, AMD was "not competitive" with NVIDIA. vLLM's test suite then had no MI355X tests, and AMD's vLLM image was a fork of an older version. Results vary by model: on SemiAnalysis's live gpt-oss-120B comparison, B200 cost 22% less per million tokens at 104 tokens per second per user and 35% less at 171, and the page warns its data may be out of date.
Does the CUDA licence itself lock you in?
Partly. NVIDIA's CUDA licence says users "may not reverse engineer, decompile or disassemble any portion of the output generated using SDK elements for the purpose of translating such output artifacts to target a non-NVIDIA platform." Tom's Hardware reported in March 2024 that the clause had been in the online licence since 2021 and in installed documentation from CUDA 11.6. Recompiling source code for another platform, for example with AMD's HIPIFY tool, is a different activity from translating compiled output, but the licence doesn't explicitly address it. Ask a lawyer if this matters to you; this is not legal advice.
In practice the bigger lock-in is code and habit: custom CUDA kernels, NVIDIA-specific libraries, and teams whose tutorials, containers and troubleshooting all assume NVIDIA hardware.
If your team only calls PyTorch, vLLM or Ollama, switching is mostly a testing job. If it writes CUDA, switching is a porting project.
Is depending on NVIDIA a risk in itself?
Yes, in supply and price. NVIDIA's DGX Spark launched at US$3,999 and rose to US$4,699 in February 2026, citing constrained memory supplies; TechRadar reported US$6,950 for a 128 GB unit in October 2026, so current pricing should be checked. A single supplier gives buyers little leverage when supply tightens. That is part of why OpenAI's AMD agreement matters: very large buyers are deliberately keeping a second source.
For most SMEs, though, the bigger dependency is often not the GPU at all but the cloud API or model provider. Apple's MLX is another challenger to CUDA on the desktop, covered in The End of CUDA Dominance?, and NVIDIA's interconnect advantage is covered in Is NVLink NVIDIA's Biggest Competitive Advantage?. For a 2024 chip-level comparison, see NVIDIA H100 vs AMD MI300.
What about AMD on the desktop?
The gap is wider there. One developer's account of the Radeon AI PRO R9700 with ROCm 7.2 (February 2026) reported the card unsupported in stock Ollama, missing 4-bit kernels in vLLM that held speed to about 10 tokens per second, and a failed FP8 cache, concluding that "picking AMD and ROCm for AI is choosing hard mode." That is one user's experience, not a benchmark, and ROCm 10 has shipped since.
AMD's Ryzen AI Max+ 395 mini PCs with 128 GB are a more practical option. They sold for about US$3,449–4,349 in October 2026, and community testing put gpt-oss-120b at roughly 30–56 tokens per second depending on the software back end. NVIDIA's GB10 machines were reported at about 55–59 tokens per second on llama.cpp. Our post on smaller AI workstations compares the compact options.
How should a business manage the risk?
- Write to frameworks, not hardware. Keep custom CUDA to a minimum and record where it exists.
- Test your actual models on both platforms before a large purchase. Benchmarks vary widely by model and serving setup.
- Count staff time. A cheaper accelerator that needs weeks of tuning may cost more overall.
- Keep data and models portable: open model formats and containerised serving make a later switch easier.
- Revisit yearly. AMD's release pace means today's gap may not be next year's.
Buying in Singapore
On the NVIDIA side, the RTX PRO 4000 Blackwell 24 GB is S$5,640 and the RTX PRO 6000 Blackwell Workstation Edition 96 GB is S$29,100. For an AMD-based workstation platform, the AMD Ryzen Threadripper PRO 9975WX is S$6,604.46 (all listed 8 Nov 2026). We don't list AMD Instinct or Radeon PRO accelerators; ask us and we can check availability. Browse our NVIDIA graphics cards. Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can PyTorch run on AMD GPUs? Yes. PyTorch has ROCm builds for Linux, and vLLM and SGLang support AMD Instinct GPUs. Features and test coverage still lag NVIDIA, according to SemiAnalysis (February 2026).
Is AMD cheaper for AI inference? It depends on the model and setup. SemiAnalysis found the MI355X competitive on cost in several FP8 cases, but its gpt-oss-120B comparison showed NVIDIA's B200 22–35% cheaper per million tokens at 104–171 tokens per second per user.
Does the CUDA licence ban running CUDA code on AMD? It prohibits reverse engineering CUDA-generated output to target non-NVIDIA platforms. Recompiling source code is a separate question the licence doesn't explicitly address. This is not legal advice.
What version is ROCm on now? ROCm 10.1.0, released 5 October 2026. AMD jumped from 7.x to 10.0 in August 2026 and now releases about every six weeks.
Should an SME choose AMD to avoid lock-in? Only if your team can test and support it. For most small teams, the lower-risk path is NVIDIA hardware with portable framework-level code, reviewed each year.
