A look at the evolution of AI chips and where they are headed, as companies like Google, Amazon, Graphcore, and Cerebras look to challenge Nvidia's dominance
NVIDIA's GPUs dominate AI chips. But a raft of startups say new architecture is needed for the fast-evolving AI field Tweets: @jbonne5 , @cerebrassystems , @wireduk , @vickiturk , and @sub8u Tweets: Jean-Marc Bonnefous / @jbonne5 : The race is on for better AI chips, from gaming to mining to machine learning at the edge https://www.wired.co.uk/... Cerebras Systems / @cerebrassystems : There is “GPU impossible work” in #AI made possible by the WSE-2 from Cerebras Systems. Our CEO @andrewdfeldman spoke with @njkobie @WiredUK to discuss this work and the evolving AI chip industry. #machinelearning https://www.wired.co.uk/... @wireduk : NVIDIA's GPUs dominate AI chips. But a raft of startups say new architecture is needed for the fast-evolving AI field https://www.wired.co.uk/... Vicki Turk / @vickiturk : “There's an apocryphal story about how NVIDIA pivoted from games and graphics hardware to dominate AI chips - and it involves cats...” Great (and accessible!) piece by @njkobie on AI chips and how hardware is racing to stay ahead as AI accelerates https://www.wired.co.uk/... Subrahmanyam Kvj / @sub8u : Cats are behind most good things in tech. Good read on Nvidia, and the future of AI chips. https://www.wired.co.uk/... https://twitter.com/...
Context & Ripple Effects
This Wired UK feature sits mid-way through an arc that began with the 2018 chips arms race, when AI workloads first pulled tech giants and a wave of startups into custom silicon. Three years on, the challengers named here — Google, Amazon, Graphcore, Cerebras — have each staked out a different attack: hyperscalers building in-house chips for their clouds, and Cerebras claiming its wafer-scale WSE-2 unlocks 'GPU impossible' AI work.
What makes the story durable is what the later coverage shows: by 2023 Nvidia's ~$10K A100 had become the default tool of generative AI with an estimated 95% machine-learning GPU share, and by mid-2024 its data center revenue hit $22.6B in a single quarter with 90%+ data-center GPU share. The startup architecture argument was real; displacing the incumbent proved far harder than pitching it.
First-order effects
- Cerebras and Graphcore must convert architectural claims like the WSE-2's 'GPU impossible' positioning into paying workloads against a CUDA-anchored incumbent, while Google and Amazon deploy homegrown chips inside their own clouds to cut dependence on Nvidia supply.
Second-order effects
- Nvidia responds by widening the platform beyond the GPU itself — the DGX GH200 supercomputer push and data center networking tools bundle silicon with systems software — while AMD and Intel fight over a distant second place rather than the crown.
- Cloud buyers gain leverage: as rivals including Amazon's and Google's cloud units scale alternatives, pricing and availability pressure lands back on Nvidia's dominant franchise.
Third-order effects
- If the pattern holds, AI compute structurally splits into two camps — vertically integrated hyperscalers designing their own silicon versus Nvidia's full-stack platform — while independent chip startups face the monetization gap the later coverage documents, leaving the merchant market thinner than the 2018 arms race promised.
The trend: AI compute is consolidating around platform lock-in and hyperscaler vertical integration, squeezing the independent chip architectures that were meant to unseat the GPU.