NVIDIA’s $28 Billion Groq Deal: Jensen Huang’s “Inference” Moat
On December 20, NVIDIA announced that it would acquire AI inference chip unicorn Groq for $20 billion in a mix of cash and stock. This marks not only the largest single acquisition in NVIDIA’s history, but also a powerful rebuttal to the recent wave of “de-NVIDIA-ization.”
From “Training King” to “Full-Stack Dominance”In the second half of November, we saw Meta cozy up to Google’s TPU and Amazon aggressively promote its in-house Trainium chips. These giants are trying to bypass expensive GPUs and seek more cost-effective alternatives. Groq, with its distinctive LPU (Language Processing Unit) architecture, has long been regarded as the strongest challenger in the inference space. Its extremely high token generation speed and relatively low latency directly target NVIDIA GPUs’ weakness on the inference side. By acquiring Groq, NVIDIA eliminates a major potential threat in the inference market and integrates its technology into NVIDIA’s own CUDA ecosystem.
A Shortcut to Solving the “Physical Bottleneck” In the first half of November, we reported on the phenomenon of “dark datacenters” (idle data centers caused by power shortages). Groq’s architecture has an inherent advantage in performance per watt. Wall Street analysts believe that this move is not only about neutralizing a competitor, but also about rapidly acquiring a highly efficient inference architecture to cope with increasingly strict power constraints in data centers.
Market Reaction Following the announcement, NVIDIA’s stock surged 6%, approaching an all-time high. This indicates that the market agrees with Jensen Huang’s judgment: the future AI gold mine lies not in training ever-smarter models, but in enabling billions of users to use (inference) these models at low cost.
Some Side Notes.Before founding Groq, its founder and CEO Jonathan Ross worked at Google and was one of the key initiators of Google TPU. It’s fair to say he has an inside-out understanding of how to build high-performance AI inference architectures.
Another important point is that NVIDIA did not acquire Groq in a traditional, outright manner. Instead, it used a “hire + license” approach—bringing in the core team and licensing the IP on a non-exclusive basis—to obtain the company’s primary intellectual and technical assets. Groq will continue to exist as a company and keep operating Groq Cloud. This approach closely resembles Meta’s acquisition of Scale AI, Google’s acquisition of Windsurf, and Microsoft’s acquisition of Inflection, among others. On the one hand, it avoids taking on inefficient or low-value assets; on the other, it helps sidestep government scrutiny. A full acquisition would trigger antitrust review, which could drag on for a long time—or even be blocked altogether.
Meta Acquires Manus: A High-Profile Move into the Consumer AI Agents Market
On December 29–30, 2025, Meta officially announced its acquisition of Singapore-based AI startup Manus. This deal is one of Meta’s most important AI strategic moves at the end of the year. In Meta’s history, it is considered the third-largest acquisition, trailing only the WhatsApp acquisition and its strategic equity investment in Scale AI. While Meta did not disclose the exact price, multiple major media outlets estimate the deal value to be above $2 billion, with some suggesting it could be even higher. Manus had previously been valued at around $500 million in its last funding round, indicating a significant acquisition premium.
Manus’s core team comes from Chinese entrepreneur Xiao Hong and his company Butterfly Effect Pte. Ltd. Prior to developing Manus, Xiao Hong had already built experience in AI and automation tools through products such as Nightingale Technology, and later launched the browser AI plugin Monica, which accumulated a large user base. On March 6, 2025, Manus was officially released and was hailed by the media as one of the world’s first general-purpose AI agents. After launch, the product quickly attracted market attention, demonstrating capabilities in comprehensive task execution—far beyond simple Q&A or basic automation.
At its core, Manus is a “general-purpose AI agent”, with the goal of:
· Autonomously understanding target tasks with minimal human instructions
· Planning steps and executing complex tasks
· Enabling automated execution across a wide range of scenarios, from research, analysis, coding, and data processing to creative work and business workflows
This type of advanced agent system differs from traditional conversation-based chatbots. It places much greater emphasis on autonomy and task execution, proactively completing a sequence of actions rather than merely returning conversational responses.
Following its launch, Manus ignited the AI agents market and quickly attracted strong investor interest. In April 2025, Manus raised approximately $75 million in a Series B round at a valuation of around $500 million, with leading venture capital firms such as Benchmark participating. The round drew significant industry attention at the time and also triggered scrutiny from the U.S. Treasury Department regarding compliance issues related to China-linked investments.
Manus has millions of users, and its 2025 revenue reached $100 million. After the acquisition, Manus will continue to operate as an independent subscription-based product, with its apps and website services remaining unaffected. The company will continue to maintain its headquarters in Singapore. Manus founder and CEO Xiao Hong will join Meta as a Vice President, leading his team to contribute to Meta’s AI product development. Meta has clearly stated that the acquisition will sever Manus’s ties with China—including winding down China-related operations and relocating employees—to address regulatory and political concerns.
End-of-Year Review 2025: A Year of Dazzling Upheaval
With the completion of these two massive acquisitions, 2025 came to a highly symbolic close. Looking back, the year was defined by:
· Early-year frenzy: the eruption of “hundreds-model battles.”
· Mid-year bottlenecks: the emergence of power constraints and growing doubts about the Scaling Law.
· Year-end consolidation: GPT-5.2 stabilized the landscape, Gemini 3.0 made a strong breakthrough, while NVIDIA and Meta completed the final pieces of their strategies through acquisitions.
Looking ahead to 2026: startups in the middle layer of the AI stack will face an increasingly harsh survival environment. Being acquired or deeply specializing in vertical domains will become their main paths forward. Either they are absorbed by giants like Groq, achieving a “safe landing,” or they push a vertical niche—such as coding, like Anthropic—to the extreme. As for those “wrapper” companies that lack both a compute moat and unique scenario data, they are likely to quietly disappear amid the fierce competition of 2026.
