The remarkable advancements made by the Chinese AI startup DeepSeek have shaken the global tech market, especially as they take the spotlight in discussing innovation in reinforcement learning and small model developmentThis breakthrough challenges the traditional notion that "substantial cash reserves create undeniable advantages," stirring conversations around the evolving dynamics between startup ingenuity and the might of established AI giants.

On January 29, in the midst of Silicon Valley's buzz surrounding DeepSeek's "Aha moment," both Meta and Microsoft released their quarterly financial reportsNotably, executives from both tech powerhouses referenced DeepSeek during their earnings callsSo far, Meta has not reduced its investment in AI, with plans to allocate between $60 to $65 billion this year aloneOn the other hand, Microsoft has indicated a downward trend in AI costs and an anticipated decline in capital expenditure growth for the fiscal year 2026 compared to 2025.

This shifting landscape highlights the strategic nature of investing in infrastructureMeta's CEO Mark Zuckerberg stated recently that it is too early to ascertain the long-term impact of DeepSeek's AI models on Meta and the broader tech industry.

As Zuckerberg elaborated, large companies that commit to significant investments in capital and infrastructure over time will develop a strategic advantageHe confidently suggested, "The ability to build such infrastructure will be a major advantage." While companies may require fewer computational resources to train powerful models eventually, the capability to maintain a vast number of servers might be essential for achieving superior intelligence and delivering high-quality services.

Zuckerberg further emphasized the extensive costs required to serve everyone effectivelyWith Meta's daily active user count surpassing 3.35 billion in the fourth quarter, the implications become even clearer.

Interestingly, Zuckerberg's perspective resonates with previous comments made by prominent figures such as OpenAI’s CEO Sam Altman and Meta’s Chief AI Scientist Yann LeCun

Advertisements

Altman noted that computational resources have become more critical than ever, while LeCun asserted that providing AI assistant services to billions still demands vast amounts of computational power.

DeepSeek has demonstrated that the funding and computational resources needed to build cutting-edge AI models may be less than previously estimatedNevertheless, Meta’s recent financial report still anticipates a $60 to $65 billion investment in AI this year, with total expenditure projected to exceed $114 billion by fiscal year 2025. Following these announcements, Meta's stock closed up by 2%.

Zuckerberg mentioned that Meta is still processing some "astonishing feats" attributed to DeepSeek, with aspirations of applying its advancements to Meta's own AI projects eventually.

"Meta will continue to learn from technologies like DeepSeek and optimize its own free and open-source AI model, Llama, to drive industry advancement," Zuckerberg stated.

On the same day, Microsoft also rolled out its latest quarterly reportThe company had previously committed to reserving $80 billion for AI data center expenditures within the fiscal yearMicrosoft's CEO Satya Nadella highlighted that this expenditure aims to overcome capacity constraints that hinder the company's AI utilization.

Nadella explained during an analyst call, "As AI becomes more efficient and widespread, we will witness exponential growth in demand." However, the company expects spending to gradually slow down beginning in fiscal year 2026, which starts July 1.

Microsoft's Chief Financial Officer Amy Hood described that the company's capital expenditures for this quarter and the next would remain around the second quarter’s $22.6 billion mark. "In fiscal year 2026, we anticipate continued investments driven by strong demand signalsYet, growth rates will be lower than those seen in fiscal year 2025, which ends in June," she remarked.

Amid these enormous expenditures in AI, both Meta and Microsoft face considerable scrutiny from the capital markets regarding their returns

Advertisements

In after-hours trading, Microsoft's stock dipped more than 5%, reflecting lower-than-expected growth in its Azure cloud division.

Analyst Daniel Newman from Futurum Group commented, "Due to these significant expenses, they need to ramp up revenue effortsI feel there's currently too much capital expenditure with insufficient consumer engagement concerning AI." Following the release of Microsoft's quarterly results, Nadella noted in an investors' call that DeepSeek indeed introduced some predictable innovations. "Everything has been commoditized, and it will gain widespread application," he mentioned.

Nadella further articulated that Microsoft had engaged in extensive collaborative efforts with OpenAI to broaden AI applications. "Simply releasing the best new model isn't sufficient; it must also be cost-effectiveIf the service is prohibitively expensive, that's unfavorable," he added.

DeepSeek's awe-inspiring AI model deserves attention not only for its technological capabilities but also for the innovative approach employed in training an AI model with limited human interventionThis "Aha moment" highlighted by DeepSeek could significantly reduce the development costs related to creating "killer" applications built on this technology.

DeepSeek revealed that when the model reassesses answers and adjusts processing times for different query resolutions, it experiences what they term the "Aha moment." The creators of DeepSeek noted in their research paper, "The 'Aha moment' powerfully reminds us that reinforcement learning could unlock new levels of intelligence in AI systems, paving the way for more autonomous and adaptive models in the future."

A technique to enhance chatbot responses involves "reinforcement learning from human feedback," an originality pioneered by OpenAI in the refinement of ChatGPTYet, this step is typically arduous, costly, and time-consuming, demanding the support of a team of human data annotatorsWhat sets DeepSeek apart is the automation of this last step using reinforcement learning.

Lewis Tunstall, a researcher at AI research firm Hugging Face, remarked, "The secret to achieving this seems to revolve around possessing a robust pre-trained model and excellent infrastructure to scale the reinforcement learning process effectively."

"The secret to intelligence is straightforward,” claimed Thomas Wolf, co-founder and Chief Scientist at Hugging Face. "DeepSeek's technology is comprehensible to others in the field, and I expect many teams will succeed in replicating their accomplishments."

Some AI experts hold the belief that DeepSeek showcases "genuine innovation." The company's approach is commendable for its transparency regarding technology development, as it openly releases detailed technical reports outlining how inference models were constructed.

Neil Lawrence, the first DeepMind machine learning professor at Cambridge University, reiterated that the models released by DeepSeek may only be "the tip of the iceberg." "History demonstrates that large companies struggle to innovate as they scale; rather, we often observe significant investments in computation superseding intellect-driven labor," he observed.

The analyst G Dan Hutcheson from TechInsights advised, "I don’t perceive DeepSeek as primarily a threat to NVIDIA but rather as a company that will significantly impact others trying to profit from selling AI products, such as OpenAI."

Previously, Pat Gelsinger, the former CEO of Intel, expressed on a LinkedIn post, "The market's reaction is misguided

Advertisements

Advertisements

Advertisements