A growing number of AI companies are turning to custom-designed chips from vendors such as Cerebras and Marvell to power inference workloads, marking a significant departure from the GPU-dominated landscape, as revealed in a recent episode of The Register's Kettle podcast. Meanwhile, Microsoft has rolled out a long-overdue Windows 11 update allowing users to move the taskbar, among other enhancements, ending a five-year wait for the feature.
In a week that Register Systems Editor Tobias Mann described as 'quieter than usual' for AI bubble talk, the discussion instead turned to concrete hardware shifts. Mann highlighted that Cerebras, which has long taken an unconventional approach with its dinner-plate-sized chips, is now finding new traction for inference after earlier struggles with the compute demands of large language models.
'Cerebras has been around for quite a while, poking at the AI infrastructure problem longer than large language models have been the dominant force,' Mann said. Their chips pack 44 gigabytes of super-fast SRAM, avoiding the supply-constrained high-bandwidth memory that GPUs rely on. This has allowed them to serve tokens faster than any GPU stack, though they initially faced limitations for inference due to reliance on structured sparsity. Partnerships with AWS and AMD earlier this year have helped them overcome those hurdles.
Waymo, the Alphabet-owned self-driving taxi company, is also designing its own custom chips for its robotaxis rather than continuing to use off-the-shelf Intel silicon. This move reflects a broader trend among AI vendors seeking to optimize performance and reduce dependency on standard GPU suppliers, particularly Nvidia, which Cerebras CEO Andrew Feldman has labeled an 'AI arms dealer.'
On the consumer side, Microsoft has finally delivered a feature that users have requested for half a decade: the ability to move the taskbar in Windows 11. The update also grants more control over context menus and provides visibility into neural processing unit (NPU) activity in Task Manager. Register Microsoft Ecosystem Reporter Richard Speed weighed in on the changes, calling them 'belated but welcome.'
While the two stories may seem unrelated, they both underscore a push for greater customization—whether in the datacenter or on the desktop. As AI hardware diversifies beyond Nvidia's GPUs, and Microsoft catches up with user demands, the industry may be entering a phase of more tailored solutions across the stack.
Analysis
Why This Matters
- The shift to custom silicon could reduce reliance on Nvidia, which currently dominates the AI chip market with its GPUs.
- For enterprises, faster inference speeds from chips like Cerebras could lower costs and enable new applications, especially in real-time AI.
- Microsoft's Windows update, while small, signals a renewed focus on user experience after years of stagnation, potentially boosting adoption of Windows 11.
Background
GPUs, particularly Nvidia's, have been the backbone of AI training and inference since the deep learning boom of the 2010s. However, their reliance on high-bandwidth memory and general-purpose design has led to bottlenecks and supply constraints. Cerebras, founded in 2015, took a radical approach with a wafer-scale chip that integrates massive on-chip SRAM, eliminating the need for external memory. This gave them an edge in speed but limited compatibility with standard AI frameworks. Waymo, meanwhile, has been developing self-driving technology since 2009 and has increasingly relied on custom hardware to handle the unique demands of autonomous driving.
Microsoft's Windows 11 launched in 2021 with a centered taskbar that could not be moved to the bottom-left or top, angering many users. The company promised flexibility but only delivered it in 2026, alongside other overdue improvements like context menu customization.
Key Perspectives
[AI vendors (Cerebras, Marvell, Waymo)]: Custom silicon offers superior performance for specific workloads, reduces supply chain risk, and allows for architectural innovation beyond what GPUs provide.
[Nvidia]: The company maintains that its GPUs remain the most versatile and well-supported option for AI, and that its upcoming architectures will address efficiency concerns. It has not publicly commented on the custom chip trend.
[Microsoft users]: The taskbar update is a positive step, but many feel it should have been available at launch. The NPU visibility in Task Manager is seen as a useful addition for power users monitoring AI workloads.
What to Watch
- Adoption rates of Cerebras and other custom silicon by major cloud providers (AWS, Azure, Google Cloud).
- Nvidia's response: whether it will introduce more specialized inference chips or double down on its GPU ecosystem.
- Microsoft's next Windows 11 update: whether it will continue to address user feedback on other long-standing issues (e.g., File Explorer performance, system tray customization).