Meta made a significant announcement this past Monday, revealing its acquisition of the Chinese AI agent startup Manus. This deal represents a major win for Manus's investors, including Benchmark Capital, ZhenFund, and Hongshan Capital—all of which participated in a funding round just last April that valued the startup at a mere $500 million. According to reports, the acquisition price paid by Meta exceeded $2 billion.
This investment, which yielded massive returns in just eight months, continues Meta's aggressive acquisition spree as the company seeks to fundamentally reshape its artificial intelligence business architecture. However, whether this particular transaction will ultimately enable Meta to stage a successful turnaround in its AI endeavors remains an open question.
The AI industry's exuberant party continued unabated throughout 2025, with venture capital and top research talent continuing to pour in, and compute-recycling financing deals becoming increasingly commonplace. Despite this, signs of party fatigue have begun to emerge in the market.
For instance, there have been several recent announcements of delays in data center construction projects, with expectations that such postponements will become more frequent next year, even as major AI firms continue to race, spending tens of billions of dollars to build out the computational infrastructure needed to power their models. Concurrently, the commanding lead once held by OpenAI in the AI chatbot market has narrowed dramatically; the performance of models from the current top three AI developers—OpenAI, Anthropic, and Google—are now largely on par, reigniting debates about the potential commoditization of large language models.
Compounding these challenges, core enterprise customers of AI models, such as Salesforce and Microsoft, have encountered difficulties in selling their own AI-powered products. Meanwhile, discussions about a potential AI bubble persist like an unwelcome guest at the party, refusing to go away.
Here is a rundown of the ten most impressive core trends that defined the artificial intelligence landscape in 2025.
✅ January: DeepSeek Sends Shockwaves Through the Industry The year began with a major tremor when GaoChi Capital, a previously obscure Chinese hedge fund, released an open-source large language model. This model claimed performance rivaling or even surpassing the top offerings from OpenAI, Anthropic, and Meta, while also asserting a drastically reduced training cost. The news sent shockwaves through Silicon Valley, with many fearing severe repercussions for AI developers, venture capital firms, NVIDIA, and cloud providers, and suggesting China had pulled ahead in the AI race.
However, the reality proved more complex. It was later revealed that the actual training cost for GaoChi's DeepSeek model was significantly higher than initially claimed. Despite this, the event shattered the absolute confidence many developers had in US AI technological supremacy, and the ongoing popularity of Chinese open-source models underscored China's formidable competitiveness in the AI arena.
✅ The Rise of Reinforcement Learning The reinforcement learning technique employed by DeepSeek rapidly gained traction across the AI industry. This AI development method involves rewarding models for achieving set objectives and penalizing them for rule violations. Subsequently, leading AI labs adopted this technique to optimize model performance, applying it across diverse scenarios including code generation, Excel spreadsheet creation, and medical consultations.
This trend also spurred the rise of reinforcement learning simulation environments—virtual settings that mimic various applications, allowing AI models to conduct practical training. For example, an executive at Anthropic stated the company plans to invest $10 billion over the next year to build such simulation environments.
✅ AI Applications Achieve Scale and Significant Revenue A central question in 2024 was whether any companies besides AI model developers, cloud providers, and NVIDIA could actually turn a profit from AI. In 2025, the answer became clearer.
According to data, over 25 AI application startups have now achieved an annualized revenue run-rate of at least $100 million, a notable accomplishment. The key question for 2026 becomes: Can these companies actually achieve profitability?
✅ Meta's AI Ambitions Stumble 2025 was a particularly challenging year for Meta. Its next-generation Llama 4 model, released in April, was widely criticized by developers. In June, CEO Mark Zuckerberg announced a staggering $14.3 billion investment in data annotation company Scale AI, a move that stunned the tech world, primarily aimed at acquiring the company's CEO, Wang Yalun, and a cohort of key talent to revamp Meta's AI division.
However, this massive investment has so far yielded minimal returns. Meta's new AI team has only managed to launch Vibes, an AI video application that was poorly received, all while undergoing multiple organizational restructurings and experiencing a continued exodus of core talent. Reports suggest Meta plans to release a new generation of text, image, and video models next year, hoping this new lineup can reverse this year's disappointing performance.
✅ Google's Strong Comeback in AI When OpenAI launched ChatGPT in 2022, Google was slow to react and spent subsequent years labeled as an AI laggard. In 2025, Google finally got back on track.
Throughout the year, Google released a series of well-received models. Its November launch of Gemini 3.0 was widely acclaimed—the model achieved a major breakthrough in code generation and was the first AI model to overcome the thorny "pretraining scaling bottleneck" (a development significant enough to trigger a "code red" emergency response at OpenAI following its release). Although the user base for Google's Gemini chatbot still trails far behind ChatGPT, it is catching up rapidly. Whether Google can maintain this momentum in 2026 will be a key area to watch.
✅ The Compute-Recycling Financing Frenzy If an AI lab hasn't secured financing from Microsoft, NVIDIA, or Amazon, only to use those funds to purchase chips or compute services from these very same companies, it is practically an outlier in the current industry.
This compute-recycling financing trend is arguably one of the most enduring features of the current AI boom, dating back to Microsoft's initial investment in OpenAI in 2019. AI labs like OpenAI and Anthropic have found this model to be an efficient way to secure funding for compute—their single largest cost.
✅ The Trump Administration Emerges as an AI 'Ally' Following his inauguration at the beginning of the year, President Trump introduced several policy measures favorable to the AI industry, including an executive order prohibiting states from enacting AI regulations and expediting the approval process for data center projects.
These actions were not coincidental: tech companies invested significant time and resources in courting Trump, for instance by contributing to his inaugural fund. (In contrast, some AI firms, like Anthropic, have maintained a distance from the Trump administration.)
✅ Unfulfilled Promises in AI Robotics In 2024, venture capitalists poured billions into robotics startups that promised large language models would lead to practical robots. This vision largely failed to materialize, with most robots still prone to frequent, basic operational errors.
Norway's 1X Technologies launched its Neo home robot, touted as capable of assisting with household chores and among the first such robots available for in-home testing—provided one can accept its $20,000 price tag and allow remote operators real-time visibility into their home.
✅ Growing Skepticism and the New Research Focus on 'Continuous Learning' Despite top AI labs generating billions in revenue, leading AI researchers are expressing increasing skepticism about whether current techniques can achieve Artificial General Intelligence (AGI)—AI that performs at human levels across most tasks.
Researchers, including former OpenAI Chief Scientist Ilya Sutskever, argue that achieving true AGI requires developing AI that can learn in real-time from real-world scenarios, much like humans do—a concept known as 'continuous learning'. While such technology does not yet exist, nearly every major AI lab is investing heavily in its development. Success in creating 'continuous learning' AI would be transformative for the industry, as it would require far less data and computational power than current models (a development that would undoubtedly be negative for NVIDIA).
✅ AI Labs Prepare for Public Markets Throughout 2025, leading AI developers like OpenAI, Anthropic, and xAI continued to raise capital at staggering valuations. In recent months, both OpenAI and Anthropic have signaled plans to go public within the next few years.
Both companies have strong motivations for an IPO: their businesses are capital-intensive, and they aim to capitalize on the current market's bullish sentiment toward the AI sector. A successful public listing would offer retail investors a chance to participate in the AI boom; however, they would also bear the losses if an AI bubble were to burst.
Other Industry Developments Andrej Karpathy's 180-Degree Turn As recently as October this year, OpenAI co-founder Andrej Karpathy publicly criticized existing large language models, particularly questioning their utility in code generation—one of the most hyped applications for conversational AI. In a podcast interview, he stated that AI programming tools were only marginally useful for code completion and writing boilerplate.
Clearly, something changed. Last week, Karpathy reversed his position on the social platform X, vigorously endorsing AI programming technology. He posted that the software engineering industry is undergoing a "disruptive restructuring," where the core work of programmers is diminishing, shifting more toward integrating various AI tools. He added, "I've realized that by properly integrating the AI tools released over the past year, my productivity could increase tenfold; failing to capitalize on this advantage is clearly a skill issue."
It's not entirely clear which specific programming tools or models Karpathy was referring to, but it's likely Anthropic's Claude Code—which he praised in a post earlier this month. Karpathy further noted that Anthropic's Opus 4.5 model, released in November, along with several new models launched last month, have achieved significant breakthroughs in code generation.
