MOUNTAIN VIEW, CA – The three-year period of undisputed dominance enjoyed by NVIDIA as the "absolute monarch" of the burgeoning Artificial Intelligence (AI) industry is facing its most credible threat to date.
Search giant Google has successfully developed and launched its cutting-edge AI model, Gemini 3, using only its internally developed AI semiconductor design—the Tensor Processing Unit (TPU)—completely circumventing NVIDIA’s foundational GPU chips.
This technological feat has ignited a fierce market rivalry, characterized by aggressive customer poaching and a visible decline in NVIDIA’s stock valuation.
Since the launch of ChatGPT in November 2022, marking the beginning of the AI bubble, NVIDIA has transformed from a semiconductor manufacturer focused on gaming graphics to the world's leading company for creating the essential tools needed for AI. The company recently peaked at a market capitalization of approximately 750 trillion yen ($5 trillion).
NVIDIA's near-monopoly on high-performance chips capable of training and running advanced AI models has resulted in extraordinary financial performance, including a staggering gross margin of 73.4%.
This exceptional profit margin, considered nearly impossible for a manufacturing industry like semiconductors and difficult even in software, was realized because no other entity could create the necessary chips for training the most advanced AI models.
The high price paid for these chips was widely referred to within the industry as the "NVIDIA tax" or "protection money," which every AI player was forced to pay.
Eliminating the NVIDIA TaxGoogle has been developing its Tensor Processing Units (TPUs) internally for over a decade.
The TPU design is highly efficient at processing the repeated, high-efficiency matrix calculations (tensor processing) required by AI.
Although TPUs have long offered superior cost performance for inference (the process of responding to user queries), they were generally not viewed as true rivals to the GPU, especially for the intensive process of training (or learning) large models.
The successful deployment of the 7th generation TPU (code-named Ironwood) for the intense training of Gemini 3 has shattered this perception. The significance of this achievement is primarily financial:
Previously, if a company invested 100 billion yen in AI, approximately 70 billion yen (70%) went directly to NVIDIA.
Now, Google has demonstrated that similar results can potentially be achieved with an investment of only 30 billion yen using proprietary chips.
For NVIDIA, the existential fear was always that state-of-the-art AI could be created using non-NVIDIA chips, and this fear has now been realized.
Google is aggressively targeting NVIDIA’s massive customer base, with internal documents reportedly setting a target to capture 10% of NVIDIA’s market share. Given NVIDIA’s valuation, this target represents a potential loss of approximately 70 trillion yen for the GPU giant.
Grammarly is an AI-powered writing assistant that helps improve grammar, spelling, punctuation, and style in text.
Notion is an all-in-one workspace and AI-powered note-taking app that helps users create, manage, and collaborate on various types of content.
This market offensive is evident through several high-profile customer acquisitions:
Open AI, NVIDIA's single largest customer, was secretly utilizing Google's TPUs earlier in 2024 to save on the heavy "NVIDIA tax". When NVIDIA CEO Jensen Huang discovered this defection, he reportedly called OpenAI CEO Sam Altman immediately. This competitive threat is viewed by some analysts as the likely catalyst for the subsequent announcement of a massive 15 trillion yen cyclical transaction deal, which effectively locked in future purchases from NVIDIA.
Meta CEO Mark Zuckerberg is reportedly engaging in deals to rent and even embed Google’s TPU chips directly into Meta’s data centers, rather than merely accessing them through the Google Cloud platform. This move signals a significant escalation in the rivalry.
AI developer Anthropic, in which Google is an investor, is also expected to utilize 1 million TPUs in its operations, driven by this internal ecosystem.
These transactions confirm that the TPU is no longer a theoretical alternative but a genuine rival to the GPU in real-world business competition.
The immediate market impact of Google’s announcement and NVIDIA’s response indicates significant internal alarm:
• Market Reaction:
NVIDIA’s stock price has fallen approximately 90 trillion yen ($600 billion) from its recent peak.
• The Viral Tweet:
Following the Gemini 3 announcement, NVIDIA’s official account posted a congratulatory tweet directed at Google, immediately followed by an awkwardly boastful caveat: "Congratulations Google, but we will continue shipping GPUs to you." This perceived "tough-guy move" quickly went viral, becoming a meme depicting Jensen Huang crying behind a celebratory mask, suggesting the company was deeply rattled by the news.
• Refuting Critics:
The company’s CFO recently held an off-the-record meeting specifically for Wall Street analysts.
The sole purpose of this meeting was to publicly refute the claims of famous short-seller Michael Burry (who predicted the subprime mortgage crisis and has named NVIDIA as his primary short-selling target).
Sources indicate that NVIDIA’s defensive posture—providing a 7-page document refuting Burry’s specific claims about GPU longevity—demonstrates that the criticisms are "hitting home" and that the company is struggling to control the narrative.
While training new AI models currently accounts for about half of AI chip usage, this balance is expected to shift drastically.
As AI models become widely deployed and used by the general public (inference), that side of the business is projected to account for 90% of chip usage within five years.
This shift poses a major long-term threat to NVIDIA.
Because TPUs are inherently more cost-efficient for inference, they are perfectly positioned to seize market share as the industry moves from building new models to deploying existing ones.
Google’s current aggressive market position is a dramatic reversal from its state just three years prior.
Following the launch of ChatGPT in November 2022, Google reportedly descended into a state of panic, issuing a "Code Red" emergency alert.
Top engineers were reportedly reassigned, not to innovate, but to reverse-engineer competitors’ products, specifically analyzing ChatGPT’s inputs and outputs to determine how it functioned.
This panic led to employee turnover, with some engineers leaving because they felt they had not joined Google "to be a copycat".
Furthermore, the company was criticized for having a disorganized, confusing array of AI brands and a reputation for being excessively "woke" (liberal) in its outputs.
Despite this chaotic period, and despite the fact that OpenAI was initially founded in 2015 explicitly to prevent Google from monopolizing AI research, Google's technical foundation—having pioneered much of the original AI research—has allowed it to swiftly recover.
This comeback has culminated in Google outperforming NVIDIA year-to-date, marking a significant victory in the high-stakes AI chip war.
.png){kind=logo}
