PyTorch Hardware Acceleration in 2025

Next year may be the critical shift from NVIDIA CUDA towards AMD ROCm and other GPUs for ML compute; the 2025 State of Pytorch Hardware Acceleration Report has some interesting insights: PyTorch has become so ubiquitous that it can help provide a practical assessment on the state of maturity of hardware (GPU) accelerators. One interesting insight is that a key differentiator is not just the FLOPs, but largely the maturity of the software platform support on the end-to-end PyTorch 2.x runtime stack. It is clear that NVIDIA CUDA (ie H100/Blackwell) is still the operational gold standard for production training and serving because it has the most mature compiler path, widest kernel/operator coverage, and the least debugging/install friction (oh hi AMD!). It does seem that AMD ROCm has improved substantially and is now becoming a differentiator as it's cost-effective but anyone that has used it knows how painful the setup actually is; getting a smooth user experience is often easier said than done, and NVIDIA has a signficant advantage in both time/experience and developer ecosystem maturity. However as we approach 2026 there are quite a lot of options arising, such as Google TPUs, Apple Silicon (MPS), and many others that are being supported with low-level frameworks such as Vulkan (one of them we maintain under Vulkan Kompute!).

Anthropic MCP now in Linux Foundation

The need for a unified + scalable standard to bridge Agentic systems is only becoming more critical, which is why I am quite excited to see the move from Anthropic donating MCP to the Linux Foundation! We know that MCP as a protocol has quite a lot of flaws, not just in terms of functionality but also in terms of security, robustness, etc - however I am a believer that more often than not, a bad standard can be better than no standard. There are indeed many alternatives to MCP emerging, especially on how fast the domain is moving, but I do believe that it's likely easier to evolve towards a 2.0 MCP protocol within the backing of LF vs trying to standardise on something new that is not as widely adopted as a base. I have also been following how the standards for the protocol have been evolving through proposal / review processes for new features, and I keep getting surprised on the pace of innovation (some good, some not so good - but arguably in the right direction). This is certainly an interesting space to keep an eye, as it is clear it will continue to evolve at an accelerated pace!

The State of MLOps 2025 Survey 🔥

In 2025, the top 5 challenges organisations face in Production ML are:  ML System Monitoring 16% (+2% YoY), Access to Data 14% (+2% YoY), Data & ML Pipelines 13% (+2% YoY), Training/Experimentation Env Parity 12% (+1% YoY), ML Governance 8% (+2% YoY), and Security 8% (+6% YoY). It is super interesting to see the big jump on ML Security from 2% all the way to 8%, as well as the continued increase of ML Monitoring as the top 1 challenge, which reflects the awareness and maturity of organisations in understanding the implications and risks of production ML systems. It is also interesting to see how last year Gaps in ML Tooling was the top 2 on the list with 12% and this year it's down to 8% completely out of the top 5 which reflects how choice is no longer becoming the blocker, and instead it's the cohesive integration and robust productionisation. Finally it seems that also Engineering Talent is slowly declining with 8% this year vs 10% in 2024, which again shows that the skill gap is slowly closing. The top 5 challenges in production ML in 2025 do seem to reflect some of the qualitative trends that I see on our day-to-day. If you want to dive deeper you can access the full results here: https://ethical.institute/state-of-ml-2025

OpenAI Enterprise AI Report

OpenAI released their 2025 State of Enterprise AI Report, where they brought together responses from 9000 workers across 100 companies to assess the trends, challenges and opportunities: Some of the outcomes won't come as surprises, as one of the main takeaways was that people are moving away from ad-hoc prompting towards scalable / repeatable production workflows. It is no surprise they have also seen a surge in ChatGPT Enterprise message volume grew by 8x YoY, with significant increase using workflows via Projects/Custom GPTs rising ~19x YoY. One key insight next year will be the impact that Google will have on these numbers given their aggressive and accelerated momentum in the enterprise space. Other interesting insights: workers report measurable productivity gains - an OpenAI report likely would say so, however as we have seen there are contrasting views from domain to domain, so it will be interesting to explore as this matures in the year to come.

AI Eats the World 2025

Ben Evans releases his annual keynote on "AI Eats the World" for the 2025 edition with some interesting insights: We currently seeing investments forecasted on capex (compute) at nearly $400 billion in 2025 which as of today is disproportionately larger than the ROI so a major shift will be required next year. The landscape is also transitioning from model differentiation to model commoditization; namely performance on general benchmarks is converging, which means that the competitive "moat" moves toward traditional software advantages like distribution, proprietary data, and product UX. We can also assume that this "infinite interns" premise that LLMs theoretically provide will likely trigger a Jevons paradox, namely expanding the volume and complexity of work organizations can handle rather than merely replacing labor.

The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning.

✉️ Email, 🐦 Twitter, 💼 Linkedin