Nvidia commands 81% of the AI chip market, a grip forged over three and a half years according to IDC data. Billions pour into its GPUs. Hyperscalers like Microsoft and Amazon snap them up. Yet cracks appear. Alphabet looms largest among challengers—not AMD or Intel, but Google’s decade-long push into custom silicon. Its tensor processing units, or TPUs, started in 2015 for internal workloads. Now they scale massively. Ironwood, the seventh-generation TPU unveiled last November, delivers four times the performance per chip for training and inference over prior models. The Motley Fool spotlighted this as Nvidia’s top long-term risk.
Short punch. Google’s Axion CPUs join the fray. Arm-based, they promise twice the price-performance of Intel and AMD’s x86 chips. Ironwood pods link 9,216 chips, yielding 42.5 exaFLOPS—more than 24 times El Capitan, today’s fastest supercomputer. Bandwidth hits 9.6 Tb/s per chip. Nvidia’s Blackwell matches on raw specs like 4.5 petaFLOPS FP8 compute and 192 GB HBM. But Google edges in cluster scale and efficiency, roughly twice the performance per watt of older H100s. And supply? Google favors elite partners. Anthropic grabbed up to a million TPUs. Meta inked a multibillion-dollar deal, testing inference gains. Citadel Securities reports faster training. G42 chats up volumes. Los Angeles Times.
Inference shifts the battlefield. Nvidia’s GPUs rule training versatility, as CEO Jensen Huang insists. TPUs specialize. Jeff Dean, Google’s chief scientist, says demand for quick AI queries makes dedicated inference chips sensible. At Cloud Next, Alphabet unveiled TPU 8t for training—three times Ironwood’s compute, twice per watt—and TPU 8i for agents, packing 1,152 chips per pod with triple the SRAM at 384 MB each. Performance per dollar jumps 80%. No direct Nvidia benchmarks. Yet Gemini tops reasoning speed. CNBC.
Hyperscalers build their own. Amazon’s Trainium and Inferentia. Microsoft’s Maia. Meta tests everything. They rent Nvidia now—projected $1 trillion from Blackwell and Rubin in 2026-27. But in-house cuts costs long-term. Broadcom eyes $100 billion in AI ASICs next year. AMD targets $100 billion data center revenue by 2030. Contracts flow to rivals: OpenAI, Anthropic with AMD, Intel, Broadcom. Cerebras filed for IPO April 17, wafer-scale chips for OpenAI’s 750-megawatt deployment starting 2026—$20 billion over three years. Customers: Amazon, Meta, Mayo Clinic. Revenue hit $510 million in 2025, up 76%. The Motley Fool on Cerebras; Reuters.
Nvidia fights back. Builds racks for rivals’ chips. CUDA locks developers—hundreds of millions of GPUs trained them. But Google supports PyTorch on TPUs. China pivots too. Huawei’s Ascend runs DeepSeek V4 inference, CANN mimicking CUDA at 95% compatibility. U.S. export curbs birthed this stack. Jensen Huang warns: better a China hooked on Nvidia than Huawei-dominant. X posts echo: DeepSeek rewrites code, erodes moat. Forbes notes broader pressures.
Power grids bind all. Data centers crave gigawatts. Plumbers and electricians lag fabs. Nvidia’s scale unmatched—$193.7 billion data center revenue last fiscal year, 60% margins. But trajectory matters. Alphabet allocates TPUs amid shortages, prioritizing elites like Anthropic over startups. Demis Hassabis: high interest from labs. Supply echoes Nvidia’s woes. Amin Vahdat: internal-only risks isolation.
And so. Rivals nibble. Alphabet bites deepest, purpose-built for Google Cloud, Gemini, partners. Inference explodes—75% of compute by 2030, $255 billion market. TPUs shine there. Nvidia adapts with Grok inference buys. Market grows blistering. Room for winners. Yet Google’s vertical edge—chips, models, cloud—threatens most. Hyperscalers negotiate harder. Margins compress. Dominance frays.
WebProNews is an iEntry Publication