Prominent Industry Figures: AI Supply Chains are Highly Taut and Civil Unrest is Building Up

Dylan Patel, CEO of independent semiconductor and AI research institution SemiAnalysis, shed light on the core contradiction in today's AI industry during an in-depth conversation - demand is growing at an unexpected rate, while supply chain flashes red in almost every aspect.

Patel referred to the current evolution of models as "the greatest capability leap in the past two years". He took Anthropic as an example, stating the company originally planned for its models to achieve Level 4 (competent engineer) programming capabilities by the end of 2025, and Opus 4.6 essentially fulfilled this goal. However, the subsequent model internally code-named Mythos, has reached Level 6 in benchmarking tests, equivalent to a senior technical expert with years of experience. The leap between the two levels happened in just about two months.

Mythos is a model with significantly expanded training scale, effectively utilizing the computing power of approximately 100,000 NVIDIA Blackwell chips. This also indirectly confirms a key consensus in the industry remains valid: investing more computing resources will continue to enhance model capabilities.

"Severe Shortage" in the Supply Chain

Patel described a supply chain scenario where almost every line is in dire straits due to the leap in model capabilities.

Regarding the memory chip sector, he surmises that substantial new capacity will not be available until 2028, and DRAM prices may rise two to three times during this period, essentially eliminating some demand through pricing.

In the foundry side, he forecasts that TSMC's capital expenditure by 2028 may reach $100 billion, a number that sounds astonishing but one he believes is entirely realistic.

The demand for CPUs is surging, on the one hand, because the training environment for reinforcement learning depends on CPUs rather than GPUs, and on the other hand, AI-generated code ultimately needs to be deployed on CPU servers.

Additionally, there has been supply tension and prepayment scrambles for upstream niche materials such as PCB copper foil, glass fiber, and lasers.

The lifecycle of GPUs is also underestimated by the market. Patel pointed out that Hopper clusters in service for three to four years are renewing contracts for the next three to four years, actual service life is significantly longer than the previously assumed five years by the market.

"The economic value that the best models can create is growing at a rate beyond the ability of Token actual supply," Patel summarized the current industrial landscape in one sentence. This gap will continue to widen until every part of the supply chain begins to re-examine its own deserved profit share.

A Growing Populist Mood in the United States

Beyond the rapid advancement of technology and the capital frenzy, Patel issued a severe warning about the social mood in the U.S. triggered by AI.

He points out that as companies leverage AI to greatly increase efficiency and potentially lay off employees, the hostility of the American public towards AI is quickly reaching a critical point.

"I believe there will be large-scale protests against big model companies within three months," Patel said candidly in an interview. He cited recent extreme events, such as Sam Altman's house being hit by Molotov cocktails twice within a week, and people cheering in the news comments section, which is just the beginning. AI is now more unpopular than politicians.

He partially attributes this public anger to the PR strategy mistakes of U.S. AI company executives. He believes that the leaders of top U.S. AI companies (such as Sam Altman and Dario Amodei) lack charm when facing the public, and they are overly keen on discussing how AI will "change the entire world" and "automate all jobs".