AI Frenzy: Token Costs Soar, Memory Triples, CPUs Exhausted, and the Secret Mythos Model

1. Token Inflation and Cash Burn

A consulting firm’s AI token subscription surged from a few tens of thousands of dollars to $7 million annually—a 25‑fold increase that now consumes 28% of the company’s payroll and could soon exceed total salaries. An analyst spends $6,000 in tokens each day and, within three weeks, single‑handedly outperformed a ten‑year‑old, hundred‑person energy‑data team, mapping the entire U.S. power grid.

2. The Hidden "Mythos" Model

Anthropic’s internal project, codenamed Mythos , leapt from an L4 to an L6 capability in two months, achieving a 93.9% score on SWE‑bench and surpassing all existing models. The team deliberately withholds the full model for security reasons, releasing only a “capped” version called Opus 4.7. Despite the restraint, the product drives $35‑45 billion ARR and lifts gross margins from 30% to 72%.

3. Supply‑Chain Strain

DRAM: Demand has surged, pushing prices up 2‑3× while annual capacity growth remains at only 20‑30%, delaying additional supply until the end of 2027.

TSMC: Capital expenditures for 2028 are projected to exceed $100 billion as the foundry races to build new fabs, creating a “tail‑effect” that exhausts downstream equipment and EUV lithography from ASML.

CPU Shortage: Reinforcement‑learning training workloads now run entirely on CPUs; the more complex the task, the greater the CPU consumption, leading to a market‑wide shortage.

Materials: PCB copper foil, glass fiber, and laser components are sold out, indicating that AI is exhausting the entire electronics supply chain.

4. Industry Realignment

Execution costs have collapsed to near zero, making the combination of a good idea plus AI proficiency the most valuable asset. Traditional SaaS pricing (per‑seat) is being displaced by token‑ or task‑based billing. Programmers are re‑ranked: junior roles see demand plummet, while senior architects who can wield AI agents command premium rates; those unable to use AI agents are rapidly becoming obsolete.

5. China’s Cost Advantage

Chinese models such as DeepSeek, Tongyi, Kimi, and MiniMax achieve GPT‑4‑level performance at merely 1/600 of the cost, thanks to Ascend hardware and aggressive algorithmic optimizations (MLA attention, MoE architectures). The low‑price, open‑source stack forces global providers like OpenAI and Anthropic to lower their prices.

6. Forecasted Backlash

Dylan Patel predicts that within three months the United States will see large‑scale anti‑AI protests, driven by low public goodwill toward AI, high‑profile incidents such as a Molotov‑cocktail attack on Sam Altman’s house, and the perception that AI is stealing jobs without delivering clear societal benefits.