AI's Architects Are Worried: Where the Cracks in the Industry's Foundation Lie

At the Milken Global Conference in Beverly Hills, five architects of the AI economy met: people who see the industry from every angle, from processor design to model deployment. Together they acknowledged that the AI industry has run into a series of systemic problems that may require decisive changes to its foundation.

Who sat at the table

The panel brought together experts from all the critical layers of the supply chain: chip manufacturing, software development, cloud infrastructure management, and academic circles doing fundamental AI research. This diversity is what made the conversation so candid: each participant could point to the bottlenecks visible from their own level.

Chips: a bottomless shortage

The first and most obvious problem is the shortage of advanced processors. Despite the efforts of NVIDIA, Intel, Samsung and other manufacturers, demand for high-performance GPUs remains many times higher than supply. Companies that would like to launch their own AI projects cannot do so because they lack access to the necessary chips.

Participants said this is not merely a manufacturing problem but a geopolitical and technological trap. Export controls restrict the distribution of hardware, chip prices remain prohibitively high, and delivery times are measured in months.

Energy: data centers at their limit

The second critical problem is electricity. Training large language models requires staggering amounts of energy: megawatts of load running around the clock. Traditional data centers are approaching the physical limits of cooling, and the power grid in many regions is not built for such loads.

Hence the idea of orbital data centers that participants mentioned: if servers are moved into space, where there is no atmosphere, overheating can be avoided. Solar panels work more efficiently in space, and energy costs fall.

• Server overheating requires gigawatts of cooling
• Local power grids are not ready for the load
• Traditional solutions are already inefficient

The architecture in question

But the most alarming point was the explicit admission that the current AI paradigm may be wrong or exhausted. In recent years the industry has grown through scaling: more parameters in models, more data, more computing power. That worked, but conference participants began saying that the returns curve is already heading down. A completely different architecture may be needed: not just bigger transformers, but something fundamentally new.

What this means

When five leading architects of the AI ecosystem simultaneously voice concerns about shortages, energy and a fundamental flaw in the approach, that is a signal. The era in which one could scale endlessly and count on exponential progress is ending. What lies ahead is an era of rethinking. The industry has to grow out of its childhood and become more systematic and efficient.