Blowing one's own trumpet is never elegant nor wise, especially while awaiting the Dragon to depart. But, as we prepare to enter the Year of the Snake, let’s start by doing a bit of clumsy in this final piece of the Dragon Year: blow a trumpet and turn into a snake charmer.

The Invisible Dragon: Somehow, it Interested No One. And Still No One. And Then Everyone.

2024 ended with a flurry of AI announcements in the holiday season. (See our year-beginning article “GenAI did not take days off this Xmas” at https://geninnov.ai/s/XBzlRD.) Dozens of breakthroughs and new models were unveiled, vying for the spotlight. We had to create a top 10 to summarise, and to the surprise of many, DeepSeek, a Chinese model announced around Christmas, took the number one spot. In our view, the announcement was far more important than the buzz created by Google's quantum announcement, OpenAI's Christmas blitz, and many others that consumed popular media.

What surprised us even more was the lack of attention DeepSeek initially received outside tech circles. Its phenomenal achievements seemed to be invisible to the mainstream media. We were so taken aback that we dedicated an entire, long article to DeepSeek (“Seeking Deep” at https://geninnov.ai/s/fSKrWw), explaining its technical prowess in plain language. Yet, the deafening silence continued. Our third article of the year went into the efficiency-focused approach of Chinese models (“Breaking the AI Narrative” at https://geninnov.ai/s/YAaXZr), a theme we had mentioned repeatedly last year, too (“The Rise of Chinese Innovators” at https://geninnov.ai/s/LmjvqQ). We started to wonder if DeepSeek was destined to remain a hidden gem, lost in the sea of AI developments.

But then, almost a month after its release, the subject opened like an aging wine clearing the threshold. DeepSeek's true potential, initially obscured, now flowed freely, captivating the global media. Its capabilities became a global sensation, with articles appearing everywhere, from The Economist cover and CNBC commentaries to in-depth FT pieces and countless blogs. No single product or feature release has garnered this much attention in recent memory.

Of course, the response isn't entirely positive. Snide remarks abound, doubts linger about the model's claims, and some even call for Western policymakers to curb China's AI development. However, the primary reason for DeepSeek's prominence lies in the stark contrast it reveals. It highlights the differing approaches to AI development between China and the US, the cost advantages of Chinese models, and the implications of massive investments in projects like Stargate. This raises questions about the future of big-cap stocks reliant on ever-growing AI infrastructure spending.

The Dragon's Breath of Venture Capital

In the US, the cost of capital, especially venture capital for AI, is incredibly low. Cynics argue, not without reason, that this environment encourages excessive spending and a "burn rate," leading to inflated valuations in subsequent funding rounds. Add to this the prevalence of compute credits from hyperscalers and chip manufacturers (an issue we've highlighted previously—though we won't bore you anymore with links to past articles!), and you have a recipe for potential overinvestment. While there's certainly validity to these concerns, and investors should be wary of the "bigger is better" mentality, it's crucial to balance this perspective with the need for bold investments in AI innovation.

The Bloated Dragon's Indigestion: The Specter of Overinvestment

The Stargate announcement has ignited a debate: Are we witnessing a divergence in AI development strategies between China and the US? Could this lead to massive overinvestment in the US, similar to the early days of the Internet? India and various nations in the Middle East, for instance, have also announced large-scale plans for AI infrastructure. Given the levels of efficiency possible, as proven by DeepSeek, there are obvious questions about the scale of the compute demand.

However, it's important to remember that most compute resources are not consumed by model training but by inference—the actual use of models by end-users. As previously shown, inference can account for 70% to 90% of compute demand. While significant, DeepSeek's efficiency gains in training may not drastically impact overall compute needs. Furthermore, the increased accessibility and affordability of AI models will likely drive even greater usage globally, potentially offsetting any reductions in training costs. Therefore, it's premature to declare that the current infrastructure build-out is excessive or inefficient.

The Snake’s Coil: Decentralization and the Path from Large to Small

Behind the scenes, a fascinating trend is emerging. While the largest models from companies like Anthropic and OpenAI may not be showing radical improvements, they are advancing enough to justify their development. However, the industry talks suggest that these behemoth models are not intended for widespread use; their primary purpose is to facilitate the creation of smaller, more efficient models.

This approach makes sense. Large models are costly to run. Imagine asking a massive model like GPT-5 a simple question like "What is 2 plus 2?" The compute cost would be exorbitant compared to using a smaller, more specialized model. By leveraging the power of large models to develop better, smaller models through the creation of synthetic data and optimized matrices, the AI industry can achieve both innovation and efficiency.

In short, the path to the most efficient models may involve pursuing the large and expensive. 

Dragon Scales: Capex as Armor in the AI Race

As we've argued before (without the links, as promised above), capital expenditure (capex) is crucial for generative AI and even physical AI business development with monetizable moats. Unlike the internet era, where software and architectural innovation dominated, the future of AI lies in creating business models that leverage substantial capex to build unique and difficult-to-replicate capabilities. In the world of instant copiability, our long-running theme, anything that can be done with small efforts, is likely to be replicated and improved instantaneously. Business moats may not be only through large capex, but they are unlikely to be in ideas of products and features alone - a huge departure from the previous era. 

Concerns about waste and oversupply are natural when discussing large AI projects. However, these initiatives are not frivolous "dig and fill" endeavors; they possess real utility. Furthermore, unlike historical bubbles such as the railway or internet infrastructure booms, the gains in AI are likely to benefit those driving the innovation rather than just external parties. Companies making significant investments in AI are best positioned to achieve breakthroughs that will profoundly shape the field and unlock valuable monetization opportunities.

The Dragon's Rigidity vs. the Serpent's Slither: Adaptability in the AI Landscape

The risks associated with grand projects differ. While concerns about resource wastage and oversupply are valid, these risks in AI diverge from those observed in historical infrastructure projects. Unlike railways or internet cables, AI projects are not merely about establishing a physical foundation; they entail continuously evaluating ongoing innovations and quick adaptations to position rightly in the fast-evolving competitive landscape. 

The field is dynamic and unpredictable, with new breakthroughs and challenges appearing continuously. Companies that rigidly adhere to predetermined plans, particularly those involving significant long-term investments or satisfying the spending targets set by big bosses in public settings, risk doing things that are no longer needed. Those with large ambitions must ensure that flexibility is paramount in a field where nothing has been in line with expectations so far.  

The greatest lesson from DeepSeek is not the possible efficiencies but that innovations in the cognitive realm, aka GenAI, can come from any direction and in any form. It is good to have checkbooks open, but not if it forces everyone to think in only one direction.

The Dragon’s Departure, the Snake’s Arrival: A Race Against Time

The spoils will belong to those who attempt. As we wrote in an op-ed in India (titled “India Must Wake Up” at https://geninnov.ai/s/iNFvnh; this is behind a paywall, so those interested could also reach us directly for our note), not attempting will leave companies, communities, and countries behind in the AI race. The pace of innovation is such that catching up will become increasingly complex, leaving those who hesitate with high usage costs and limited influence.

The risk of falling behind in the AI race far outweighs the risk of wasted investment, a crucial point for corporate decision-makers and policymakers. While commentators can afford to simply focus on potential overinvestment risks and historical bubble parallels, those in decision-making positions must recognize the shift in risk calculus, symbolized by the image of prominent AI investors alongside President Trump. Inaction carries increasing risks, especially given the rapid and unexpected advancements in AI capabilities. While some may debate implementation hurdles or express disappointment with current abilities, the potential consequences of hesitation are far greater than the potential for wasted resources. The same risks exist for those who want innovations to settle down before deciding their path.

This brings us to our likely theme for the year: In4, or "Infusing Intelligence Into the Inanimate." As we highlighted in our last piece, efficiency gains are accelerating the adoption of AI, particularly in physical AI. This goes beyond robotics and driverless cars; it encompasses imbuing all sorts of objects with cognitive power for diverse applications. If the previous era was about connecting devices to the internet, the In4 era would be about infusing intelligence into the world around us. This is where the serpent's wisdom lies - the spoils will belong to those who can flexibly navigate this new landscape.

Gong Xi Fa Cai! 恭喜发财! Kung Hei Fat Choi! 恭喜發財!