AI Infrastructure At A Crossroads

By Ramin Farjadrad and Syrus Ziai

There is a big push to achieve greater scale, performance and sustainability to fuel the AI revolution. More speed, more memory bandwidth, less power — these are the holy grails.

Naturally, the one-two punch of StarGate and DeepSeek last week has raised many questions in our ecosystem and with our various stakeholders. Can DeepSeek be real? And if so, what does this mean for  investments in AI infrastructure like StarGate?

DeepSeek’s apparent breakthrough in optimizing AI training highlights an undoubtedly major shift. If the claims are true, we can now achieve powerful AI capabilities with significantly reduced computational costs. If model sizes could indeed be shrunk to cut training complexity by 10X or more, this would make AI more accessible beyond traditional data centers—fueling AI’s expansion into Edge devices and everyday applications.

DeepSeek leverages algorithms such as Mixture of Experts (MoE), which demand a lot of memory bandwidth and produce large amounts of temporary output token, which need to be stored in memory and read back. Eliyan’s die-to-memory technology with significantly higher bandwidth/power can directly address this market expansion.

As history has shown, once new technology becomes more accessible to the masses, applications and use cases explode. We have seen this story before with the birth and rise of the Internet.  The technologies that were created over the past two decades helped provide higher bandwidth Internet at lower costs, which helped dramatically expand the Internet market and the market cap of the companies delivering valuable products and services in that space, and to the overall economy. This expansion created many markets that did not exist before, and with that the demand for enabling hardware and software technologies exploded.

The next question is if we completely discount the impact of this dynamic in expanding the market, does this mean hyperscaler AI infrastructure spending will slow? Far from it.

The key distinction lies in scale—not just in model size, but in the sheer volume of data that AI must process for hyper-personalization.

Let’s compare:

• Current AI Models (e.g., ChatGPT, DeepSeek) are trained on nearly all publicly available human knowledge—books, articles, scientific papers, and structured datasets. This corpus, while vast, is relatively static and finite. Estimates suggest well over 90% (if not already ~99%) of today’s human knowledge has already been absorbed into AI models.
• Hyper-Personalized AI goes far beyond general knowledge. The next frontier is training AI on individual human behavioral data—a dataset growing exponentially across social media, emails, interactions, purchases, and even biometric inputs. This dataset isn’t just orders of magnitude bigger; it is constantly evolving, requiring continuous ingestion, processing, and retraining.

The difference in scale is staggering:

• The entire training corpus for general AI models may be measured in single-digit petabytes (PBs).
• The behavioral and personal data generated daily by billions of people could easily reach exabytes (EBs) or even zettabytes (ZBs) annually.

This shift from training AI on static, finite knowledge to dynamic, ever-expanding personal data is why hyperscalers will continue to invest heavily in AI infrastructure. While DeepSeek’s reported optimizations would make AI more efficient, the push for ultra-personalized AI requires orders of magnitude more compute power, storage, and high-performance AI hardware.

That’s why Hyperscalers, Open AI, StarGate and various other initiatives and entities are spending hundreds of billions of dollars on AI infrastructure, knowing well that they may not achieve positive ROI just by offering customers AI assistant apps or tools (this is especially true as the price of such AI assistant tools will surely drop significantly once the models that offer such capability become open source and more than two or three companies offer the same capabilities).

So while DeepSeek may reduce the compute hardware for the training by an order of magnitude, and allow training based on the existing knowledge base to be much faster (reducing from trillions of parameters to <100 billion), the characteristic dataset of all billions of humans online can be in hundreds to several thousands of trillions of parameters, and AI hardware is still far from being capable of training at that scale.

The bottom line? Efficiency gains won’t slow AI hardware demand—they will accelerate AI’s ubiquity while fueling an even larger demand for computational power. We are just at the beginning of an era where AI models aren’t just trained on what humanity has already discovered — but on who we are, individually, in real time.

Syrus Ziai is a co-founder of Eliyan.