Google is capitalizing on its latest AI model breakthrough to mount a full-scale challenge against NVIDIA's chip supremacy. The search giant has begun pitching its proprietary TPU chips to major clients like Meta, aiming to expand its alternative AI chip solution beyond Google Cloud rentals into broader markets.
Sources reveal Meta is negotiating with Google to deploy billions worth of TPUs in its data centers by 2027, alongside plans to rent Google Cloud chips next year. This potential deal could capture 10% of NVIDIA's annual revenue, generating billions in new income for Google—a target reportedly discussed internally among Google Cloud leadership.
The recent launch of Gemini 3, primarily trained on Google's TPUs and rivaling OpenAI's ChatGPT in performance, has triggered market reassessments of AI chip hierarchies. Alphabet's stock surged 6.3% to a record $318.58, narrowing its market cap gap with NVIDIA to $526 billion—the smallest since April 2023—while NVIDIA shares dropped nearly 10% this month.
NVIDIA CEO Jensen Huang has responded aggressively to the threat. After Google secured a deal to supply Anthropic with up to 1 million TPUs, Huang committed billions in investment to the AI startup while securing its continued GPU usage. Similar moves followed reports of OpenAI considering Google Cloud TPU rentals, with Huang negotiating a potential $100 billion investment.
Google's TPU Strategy Evolution: From Cloud Rentals to On-Premises Deployment
Having long rented its custom Tensor Processing Units (TPUs) through Google Cloud, the company now markets direct data center deployments to clients including Meta and major financial institutions. Google emphasizes that on-premises TPUs meet stricter security and compliance requirements, particularly for sensitive data processing. The company developed "TPU Command Center" software to ease chip adoption, positioning it against NVIDIA's dominant CUDA platform. While Google's Jax programming language remains niche, clients can operate TPUs using PyTorch-compatible tools without Jax expertise.
Since mid-2023, Google has actively courted smaller cloud providers typically reliant on NVIDIA chips, offering to host TPUs in their facilities. One confirmed partnership with London's Fluidstack includes a $3.2 billion financial backstop provision for data center leases.
Gemini 3 Breakthrough Resets Market Expectations
Gemini 3's launch—praised for surpassing ChatGPT in speed, reasoning, and cost-efficiency—has demonstrated TPUs' viability as NVIDIA alternatives. Though less flexible than GPUs, TPUs offer lower development costs and reduced power consumption at scale. Analysts note the model's TPU-based training challenges NVIDIA's stronghold in AI training clusters, with Meta reportedly discussing TPUs for new model development rather than just inference tasks.
D.A. Davidson estimates Google's DeepMind and TPU operations could be worth nearly $1 trillion independently. Alphabet's 68% year-to-date stock surge outpaces the "Magnificent Seven" index (22%) and Nasdaq Composite (18%), with TPU manufacturer Broadcom rising over 63%.
NVIDIA's Counteroffensive
Huang has moved to secure commitments from major AI players exploring TPU alternatives. Beyond the Anthropic deal, NVIDIA negotiated preliminary terms with OpenAI involving potential $100 billion investments for data center development and GPU leasing. While NVIDIA states its investments don't mandate GPU purchases, industry observers note Huang could preempt Google's TPU deals with key clients like Meta.
Huang acknowledged Google's chip progress in a recent podcast, stating respect was due given seven TPU generations. NVIDIA has made similar financial commitments to cloud partners like CoreWeave.
Market Restructuring Underway
Goldman Sachs traders describe Gemini 3 as a "disruptive model" reshaping AI investment ecosystems, potentially delaying competitors' product cycles and increasing capital expenditures. While NVIDIA's earnings remain strong, focus has shifted from its dominance. Analysts caution that Oracle and other hyperscalers invested heavily in NVIDIA cloud rentals may face pricing pressure from lower-cost TPU alternatives.
Even modest erosion of NVIDIA's AI advantage could trigger market ripple effects. Bernstein analysts emphasize the broader question isn't about winners or losers, but whether AI's economic opportunity proves sustainable—a scenario that would benefit all major players, or conversely, challenge them collectively.
Comments