This project is impressive!
[Projects in Financing] GaN/SiC embedded power modules, fully active automotive suspension, automotive and energy DSP chips, high-end organic silicon new materials, AI acoustic applications, etc.
Source: Azat TV January 27, 2025
Key Points Overview
·According to reports, Meta is negotiating to lease and purchase billions of dollars worth of Google Tensor Processing Units (TPUs), indicating its intention to seek AI hardware alternatives outside of NVIDIA.
·Google’s TPU performs excellently in large AI model tasks but is limited to Google Cloud Platform; whereas NVIDIA GPUs are versatile across multiple industries and platforms.
·NVIDIA holds up to 90% of the data center chip market share, with rapid growth and significant lock-in effects in the developer ecosystem.
·Transitioning from NVIDIA’s CUDA technology stack to Google’s XLA environment is challenging, limiting the widespread application of TPUs.
·Meta’s move is part of a moderate diversification strategy rather than a complete departure from the NVIDIA platform.
Meta’s Direction: Ripples, Not Tsunamis
By the end of 2025, the AI hardware field is rife with speculation: one of the world’s largest hyperscale cloud computing companies, Meta, is reportedly in deep negotiations to lease and ultimately purchase billions of dollars worth of customized Google Tensor Processing Units (TPUs). This is no small shift — in recent years, NVIDIA GPUs have been the core pillar of nearly all major AI projects, and the industry’s reliance on them has become the norm, making Meta’s move a rare deviation. Following the news, Alphabet’s stock price rose, while NVIDIA experienced a brief decline.
However, as reported by Tom’s Hardware, NVIDIA’s response was both congratulatory and subtly provocative. The company reminded the public that Google is still procuring its chips and, in its own words, “leading the industry by a generation.” NVIDIA claims its platform can run all AI models in any scenario — whether in the cloud, data centers, edge devices, or local workstations. In contrast, Google’s TPU remains strictly tied to the Google Cloud Platform, with more limited applications focused on a single goal.
TPU vs. GPU: A Clash of Architectures
Google’s TPU is an Application-Specific Integrated Circuit (ASIC) designed to support high-throughput matrix operations for large language models. The flagship product, TPU v5p, features 95GB of HBM3 memory, with a peak throughput of over 450 teraflops in bfloat16. These chips are capable of large-scale deployment, with a pod accommodating nearly 9,000 units, manufactured by TSMC based on Google’s proprietary architecture, with Broadcom responsible for chip implementation.
Meanwhile, NVIDIA’s Hopper H100 GPU contains 80 billion transistors, 80GB of HBM3 memory, with FP8 performance reaching 4 petaflops. Its successor, Blackwell GB200, further enhances these figures and can seamlessly integrate with the Grace CPU to form a hybrid computing configuration. NVIDIA’s technology stack is highly versatile, deeply embedded in industry workflows, and supported by decades of developer investment in mature frameworks like CUDA, cuDNN, and TensorRT.
What are the Core Differences? TPUs excel in specific AI tasks within the Google ecosystem. In contrast, NVIDIA GPUs are “jack-of-all-trades,” supporting not only AI but also enabling scientific computing, simulation, and edge deployments across various industries from automotive to retail. This flexibility is central to NVIDIA’s dominance and a major barrier for Google in achieving its ambitions.
The Developer Dilemma: Ecosystem Lock-In Effects
Changing the underlying hardware for AI models is not as simple as swapping chips. Google’s TPU programming is optimized for JAX and TensorFlow through the XLA compiler stack. Although XLA offers performance portability, it requires developers to rewrite code, adapt to new performance bottlenecks, and sometimes even switch frameworks entirely. For companies like Meta that have invested heavily in JAX, migration is feasible but still fraught with resistance.
In contrast, NVIDIA’s CUDA-based technology stack is ubiquitous. It is the default choice for distributed training, mixed precision scheduling, and low-latency inference. The ecosystem is mature, with pre-trained models, commercial support, and a vast pool of developer talent. Stepping away from this “comfort zone” means taking on the risk of productivity and performance loss — few companies are willing to make that leap without strong incentives.
Secondary Resilience: Bargaining Power in a Changing Landscape
Why is Meta considering TPUs? The answer lies in strategic considerations. Major hyperscale cloud computing companies like Amazon Web Services (AWS) and Microsoft have their own custom chips (Trainium, Maia), and Google has been refining TPUs for nearly a decade. By exploring alternatives, Meta gains greater leverage in future hardware negotiations while also mitigating vendor lock-in risks. Even if TPUs are only used for overflow tasks or specific inference jobs, this potential migration allows hyperscale cloud computing companies to gain bargaining power.
Google Cloud executives estimate that deals like Meta’s could generate revenue equivalent to 10% of NVIDIA’s current annual revenue from data center business — a figure that, while substantial, remains speculative. Google has committed to providing up to 1 million TPUs to Anthropic and is actively promoting its technology stack to AI startups seeking alternatives to NVIDIA’s high-priced GPUs. However, the challenge lies not only in scale but also in whether TPUs can adapt to the complex and varied AI workflow demands outside of Google’s “walled garden.”
NVIDIA’s Next Steps: Vertical Integration and Market Expansion
While Google aggressively promotes its chips, NVIDIA has not been idle. The new Grace Blackwell architecture further solidifies NVIDIA’s advantage by unifying GPU and CPU memory, simplifying training and inference processes. Developers can train models in the cloud without rewriting code and deploy them to edge or enterprise environments. NVIDIA is also deeply penetrating verticals like robotics, manufacturing, and retail, where TPUs have yet to venture.
NVIDIA’s market share in the data center space is impressive. According to The Motley Fool, NVIDIA holds up to 90% market share, with data center business revenue reaching $51.2 billion in Q3 2026, far exceeding competitors like AMD. Its growth trajectory is rapid — the data center chip business grew 66% year-over-year, with next quarter revenue expected to reach $65 billion. Alphabet’s AI-driven advertising business is massive, but while Google Cloud is growing rapidly (with Q3 revenue of $15.15 billion, a 33.5% year-over-year increase), it still represents a small portion of the overall business.
Both NVIDIA and Alphabet are expected to perform well in 2025, with stock prices rising 30% and 58%, respectively. However, as Barron’s points out, this competition is only half the story. NVIDIA’s architecture is the core engine for generative AI and agent AI, while Alphabet leverages AI to defend its dominance in search and advertising, integrating new tools like Gemini and AI Overview.
Looking Ahead: Transformation or Continuation?
Can Google’s TPU shake NVIDIA’s position? The answer is unlikely in the short term. The inertia of the developer ecosystem, the versatility of NVIDIA’s technology stack, and the vast scale of its platform mean that TPUs will remain a niche market — even as it expands. Meta’s purported adoption behavior is a signal, not a major shift. It marks a trend towards diversification, with companies seeking bargaining power and hedging against dependency risks rather than a complete migration.
The challenges for Google are evident: achieving widespread adoption of TPUs. TPUs need to prove their value in complex and diverse AI workflows outside of Google’s controlled environment. For NVIDIA, the task is to continue innovating and reinforcing ecosystem lock-in effects, ensuring its platform remains indispensable as AI evolves.
Ultimately, the battle for AI hardware dominance is not a zero-sum game. It is a long and complex game — inertia, strategy, and ecosystems are as important as raw performance. While news headlines may sensationalize this competition, beneath the surface, it is the interplay of bargaining power, developer habits, and evolving workflows that will shape the next chapter of the industry.
Despite the growing interest in Google’s TPU, and Meta’s exploratory moves, NVIDIA’s position remains solid due to its unmatched versatility, entrenched developer ecosystem, and broad applicability across industries. Google’s progress is worth watching, but for now, this is merely a minor tremor in the AI hardware landscape — not an earthquake.