All articles tagged with #tpus
Google’s edge stems from owning the full AI compute stack—massive data centers, millions of chips and servers, a global fiber network, and custom TPUs—that enables faster, more scalable AI delivery than competitors, a combination that has helped lift its stock and could widen the gap as compute becomes the deciding factor in AI success.
Google Cloud chief Thomas Kurian argues that owning the full AI stack—in-house chips (TPUs), Gemini models, and DeepMind-driven tech—will help Google close the gap with AWS and Azure. The company previewed newer TPU generations, touted its self-developed hardware and software as a differentiator, and expanded its Chips–to–Models ecosystem with a $40 billion Anthropic investment. Google says this approach reduces Nvidia dependency, supports faster, cheaper AI deployments, and underpins a forecast of rapid cloud revenue growth (around $70 billion this year) with heavy capex of about $185 billion.
In a detailed Stratechery interview, Thomas Kurian argues Google Cloud is shifting from pure AI models to an enterprise-ready agent platform built on Gemini, enabling real‑world, multi–step automation across large organizations. He highlights tightly integrated infrastructure (chips, GPUs/TPUs, data context, cybersecurity) and a Knowledge Catalog to map company data for reliable, grounded reasoning, with customers like Citi, Comcast, and Walmart already deploying complex agent workflows. Google’s strategy emphasizes a multi‑cloud, cross‑cloud lakehouse approach that lets data stay where it is while Gemini analyzes it, plus expanding an ecosystem of third‑party models and partners. New hardware and security offerings (TPU 8t/8i, Wiz integration) are framed as foundations to scale, monetize across labs, SaaS, and infrastructure, and broaden Google Cloud’s enterprise footprint without closed ecosystem constraints.
Google introduced the TPU 8t and TPU 8i—distinct chips for AI training and inference—claiming 3x compute, 10x faster storage, and 80% better performance-per-dollar, along with a custom Axiom CPU, as Alphabet seeks more cost-efficient cloud AI while continuing to rely on Nvidia GPUs for heavy lifting.
Google has signed a multiyear, multibillion-dollar deal to rent its Tensor Processing Units (TPUs) to Meta for AI model development, signaling a push to scale TPU-based infrastructure and potentially threaten Nvidia's dominance; Meta is also discussing buying TPUs next year, while Google pursues investment-firm–backed joint ventures to lease TPUs to other customers.
Nvidia still dominates AI compute with GPUs and faster growth, while Alphabet offers a niche TPU alternative and steadier, diversified cash flows from Google Cloud and other non‑AI businesses. Nvidia appears cheaper on forward earnings with higher upside, but Alphabet provides resilience and a balance of AI exposure; both are solid AI plays, with Nvidia carrying AI spending risk and Alphabet offering steadiness.
Anthropic is expanding its use of Google Cloud's TPU chips, planning to access up to one million TPUs by 2026 to train and serve future generations of Claude AI models, enhancing its AI research and deployment capabilities.
The article predicts that by 2035, Alphabet could become the Nvidia of quantum computing, leveraging its DeepMind research, custom hardware (TPUs), and quantum programming framework (Cirq) to build a dominant AI and quantum ecosystem, potentially transforming its market valuation and industry role.
Google has revealed its TPU-based supercomputer, TPU v4, which it claims is faster and more efficient than Nvidia's A100 chips. The system, which has been running since 2020, was used to train Google's PaLM model over 50 days. While Nvidia dominates the market for AI model training and deployment, Google has been designing and deploying AI chips called Tensor Processing Units since 2016. The power requirements of AI are also a boon to cloud providers like Google, Microsoft, and Amazon, which can rent out computer processing by the hour.