Nvidia still holds 81% of the AI chip market, according to IDC, a position built on three and a half years of relentless GPU sales to hyperscalers like Microsoft and Amazon. But the most credible threat isn’t AMD or Intel—it’s Google’s decade-long investment in custom silicon. Its Tensor Processing Units (TPUs), first deployed internally in 2015, have scaled into a serious alternative. The seventh-generation Ironwood, announced last November, delivers four times the per-chip performance for both training and inference over its predecessor. Google’s Axion CPUs, built on Arm, claim twice the price-performance of x86 chips from Intel and AMD.
Ironwood pods link 9,216 chips to achieve 42.5 exaFLOPS—over 24 times the output of El Capitan, the current fastest supercomputer—with 9.6 Tb/s bandwidth per chip. Nvidia’s Blackwell matches on raw specs like 4.5 petaFLOPS FP8 and 192 GB HBM, but Google pulls ahead in cluster efficiency, offering roughly twice the performance per watt of older H100s. Supply is tight, though; Google prioritizes elite partners. Anthropic secured up to a million TPUs. Meta signed a multibillion-dollar deal for inference testing. Citadel Securities reports faster training cycles. G42 is in volume discussions.
The real battleground is inference. Nvidia’s GPUs remain the default for training versatility, but TPUs are purpose-built for the task. Jeff Dean, Google’s chief scientist, argues that the surge in quick AI queries makes dedicated inference chips a necessity. At Cloud Next, Google unveiled TPU 8t for training—three times Ironwood’s compute and double the per-watt efficiency—and TPU 8i for agent workloads, with 1,152 chips per pod and triple the SRAM at 384 MB each. Performance per dollar jumps 80%. Google didn’t release direct Nvidia comparisons, but Gemini now leads in reasoning speed benchmarks.
Hyperscalers are increasingly building their own chips—Amazon’s Trainium and Inferentia, Microsoft’s Maia, Meta’s in-house tests. They still rent Nvidia hardware (projected $1 trillion from Blackwell and Rubin in 2026–2027), but custom silicon cuts long-term costs. Broadcom eyes $100 billion in AI ASICs next year. AMD targets $100 billion in data center revenue by 2030. Contracts are flowing to rivals: OpenAI and Anthropic now work with AMD, Intel, and Broadcom. Cerebras filed for IPO on April 17, offering wafer-scale chips for OpenAI’s 750-megawatt deployment starting 2026—a $20 billion, three-year deal. Customers include Amazon, Meta, and Mayo Clinic. Revenue hit $510 million in 2025, up 76%.
Nvidia isn’t standing still. It’s building racks for competitors’ chips. Its CUDA ecosystem locks in developers—hundreds of millions of GPUs have trained on it. But Google now supports PyTorch natively on TPUs. China is pivoting too: Huawei’s Ascend runs DeepSeek V4 inference, with CANN achieving 95% CUDA compatibility. U.S. export controls accelerated this shift. Jensen Huang has warned that a China dependent on Nvidia is preferable to one running on Huawei. DeepSeek is rewriting code to erode the CUDA moat.
Power constraints bind everyone. Data centers need gigawatts; infrastructure can’t keep pace with fab output. Nvidia’s scale is unmatched—$193.7 billion in data center revenue last fiscal year, 60% margins. But trajectory matters. Google allocates TPUs selectively amid shortages, prioritizing Anthropic over startups. Demis Hassabis notes high interest from labs. Amin Vahdat warns that internal-only use risks isolation.
Rivals are gaining ground. Google bites deepest, with purpose-built chips for Google Cloud, Gemini, and partners. Inference will represent 75% of compute by 2030—a $255 billion market—and TPUs are optimized for it. Nvidia is adapting with Grok inference acquisitions. The market is growing fast enough for multiple winners, but Google’s vertical integration—chips, models, cloud—poses the most direct threat. Hyperscalers are negotiating harder, margins are compressing, and the once-unassailable grip is starting to fray.
Source: Webpronews