The big story out of this week's Hot Chips conference wasn't that Nvidia's grip on AI chips is loosening. It's that its biggest customers are pouring more money than ever into building their own.
According to a research note from BNP Paribas Equity Research senior analyst Karl Ackerman, presentations from Meta Platforms, Inc. (META) and Alphabet Inc. (GOOGL)'s Google highlighted a growing industry shift: hyperscalers are expanding custom silicon not to replace Nvidia overnight, but to run AI workloads more efficiently and lower the cost of inference as demand scales.
Meta, Google and Amazon Build More of Their Own AI Infrastructure
The announcements showed just how quickly the largest cloud companies are broadening their AI hardware strategies.
- Meta detailed the next phase of its Meta Training and Inference Accelerator (MTIA) roadmap. Its upcoming MTIA 400 chip is designed to support both recommendation systems and generative AI models, delivering 3 PFLOPS of FP16 compute—roughly five times the performance of its predecessor—while expanding high-bandwidth memory capacity by 33% to 288 GB.
- Google, meanwhile, focused on the networking infrastructure behind its latest TPU platform. The company said its Virgo Network can connect 134,000 TPU 8t chips with up to 47 petabits per second of non-blocking bandwidth, enabling more than 1.6 exaFLOPS of compute while maintaining near-linear scaling. The architecture is designed to eventually support up to one million TPU chips across multiple data centers.
- Amazon, too, has been investing in proprietary AI processors through its Trainium and Inferentia families, reflecting a broader industry push toward custom silicon alongside merchant GPUs.
Custom Chips Are About AI Economics, Not Replacing Nvidia
Ackerman's broader point is that hyperscalers are increasingly optimizing their own AI infrastructure as inference—the process of serving AI models to users at scale—comes to account for a larger share of AI spending.
That helps explain why Meta and Google continue to invest heavily in proprietary chips even as they remain among Nvidia's largest customers.
For investors, that distinction matters. Nvidia continues to dominate the market for cutting-edge AI training, but the economics of running AI services are encouraging the world's largest technology companies to own more of the underlying hardware stack wherever it makes financial and technical sense.
Investment Takeaway
The AI chip race is evolving from a contest over raw computing power into one focused on infrastructure economics. Nvidia remains the industry's dominant supplier, but Meta, Google and Amazon are steadily building more AI capabilities in-house to improve efficiency and reduce long-term costs.
Investors should watch custom silicon less as an immediate threat to Nvidia and more as a measure of how hyperscalers plan to balance dependence on the AI leader with greater control over their own infrastructure.
