Tech • IA • Crypto
Why Silicon Valley Is Abandoning Nvidia?
9/10TL;DR
Nvidia posts record results but its stock dips as investors anticipate rising competition from custom AI chips led by Google and hyperscalers.
KEY POINTS
Record financial performance at Nvidia
Nvidia reported quarterly revenue of $81 billion, up 85% year-on-year, setting a new all-time high. Its data center division alone generated $75 billion, growing 92% and now exceeding the combined revenues of Intel, AMD, Qualcomm, and Broadcom. The company also boosted shareholder returns with a 25x dividend increase and authorized $80 billion in share buybacks, underscoring massive cash generation.
Stock declines despite strong results
Despite these figures, Nvidia’s stock fell after the announcement. Markets focused less on past performance and more on forward risks, particularly intensifying competition in AI hardware. The reaction reflects concern that Nvidia’s dominance may erode faster than previously expected.
Google escalates with new TPU chips
Google unveiled two new AI chips: TPU 8T for training and TPU 8i for inference. The company claims nearly 3x better performance per dollar for training and 80% gains for inference compared to prior generations. This signals a shift toward specialized infrastructure, splitting AI workloads into distinct industrial layers.
Shift toward custom AI chips accelerates
Major AI players including Anthropic, Meta, and OpenAI are increasingly adopting Google’s TPUs. At the same time, Amazon, Microsoft, and Meta are building their own chips. This trend reflects a strategic push to reduce reliance on Nvidia’s GPUs, which dominate but come at high cost.
Market share pressure emerges
Nvidia’s share of AI accelerators, near 90% in 2025, is projected to fall toward 70%. While still dominant, a potential 20-point drop in a short timeframe raises concerns about long-term pricing power and growth sustainability.
Custom chips outpace GPU growth
Specialized AI chips are growing 45% annually, compared to 16% for GPUs. This marks the first year custom silicon outpaces general-purpose accelerators, driven by efficiency gains and tailored architectures.
Economic advantage of TPUs
Companies switching to TPUs report major savings. One example showed inference costs dropping from $2.1 million to under $700,000 per month, a 65% reduction, with ROI achieved in 11 days. Such economics are increasingly difficult for competitors to ignore at scale.
Different strengths: training vs inference
GPUs remain dominant in research and model training due to flexibility. TPUs excel in inference, where workloads are predictable and efficiency matters most. By 2030, inference could represent 75% of total AI compute demand, strengthening the case for specialized chips.
Software ecosystem remains Nvidia’s moat
Nvidia’s CUDA platform, built over nearly two decades, remains a major barrier to switching. Migrating away often requires rewriting code and retraining teams. However, emerging tools and compatibility layers aim to reduce this friction, potentially weakening this advantage.
Cloud strategy shifts competitive dynamics
Unlike GPUs, TPUs are primarily accessed via Google Cloud, meaning adoption often shifts customers into Google’s ecosystem. When companies like Meta or OpenAI adopt TPUs, Nvidia not only loses demand, but Google gains long-term cloud revenue.
Nvidia adapts with specialization
Nvidia is responding by developing more specialized chips, including architectures optimized for specific inference tasks. This marks a shift from its traditional “universal platform” approach toward more targeted solutions.
Hidden winners: Broadcom and TSMC
Much of the value flows to suppliers like Broadcom, which designs chips for multiple players, and TSMC, which manufactures them. Broadcom’s AI revenue reached $8.4 billion, up 106%, with a pipeline exceeding $73 billion. TSMC’s advanced capacity is fully booked through 2028, reflecting demand far exceeding supply.
CONCLUSION
Nvidia remains dominant but faces structural competition from custom AI chips and cloud ecosystems, shifting the battle from raw performance to platform control and long-term infrastructure economics.