AI Token costs plummet: US industrial intelligence reaches a new inflection point

Why AI Token Costs Are About to Plummet

AI tokens—the basic unit of information processed by models and the standard for measuring AI usage and pricing—are about to see historic price drops. According to tests by SemiAnalysis, Nvidia's latest Blackwell GPU system (GB 300 NVL72) increases tokens generated per GPU per second from 90 to 6,000 compared to the previous Hopper system, a 65x improvement; and tokens generated per megawatt of power from 54,000 to 2.8 million, a 50x improvement. More critically, the cost per million tokens drops from $4.2 to $0.12, a 35x reduction.

This efficiency leap is not theoretical. Business Insider reports that Blackwell systems are being installed at scale and are expected to be fully operational by the second half of 2026. According to data from AI research firm SemiAnalysis, by then, new model training and inference will heavily rely on Blackwell, leading to an explosion of cheap tokens. OpenAI CEO Sam Altman has acknowledged that AI costs have become a "huge problem" and has promised to deliver more value. The token spending index from Silicon Data, a Silicon Valley data center monitoring organization, peaked at 2.06 at the end of May but fell to 1.75 on June 10, suggesting the downward trend has already begun.

Industry Restructuring: Who Benefits, Who Feels the Pressure?

• Benefiting Industries:
• AI model and application providers: Lower costs directly boost profit margins, while enabling price cuts to capture market share and stimulate downstream demand. This is especially beneficial for consumer-grade AI applications that rely on large-scale inference, such as programming assistants and customer service bots.
• Manufacturing and industrial automation: This is the largest potential beneficiary. The drop in token costs dramatically improves the economic viability of deploying AI in industrial settings. For example, using vision AI for product quality inspection, optimizing production scheduling with large models, and training predictive maintenance models based on real-time data—applications previously limited by high inference costs—can now be deployed at scale. U.S. manufacturers can leverage this to improve production line automation, reduce labor costs, and accelerate the construction of "smart factories."
• Data center infrastructure: Strong demand for GPU clusters will drive auxiliary industries such as liquid cooling systems, high-density racks, and power equipment, but it also puts pressure on the power grid.Industries Under Pressure:
• Traditional industrial software and automation suppliers: If AI can replace some traditional optimization algorithms or control logic at extremely low cost, traditional industrial software vendors face the risk of disruption.
• Power infrastructure: AI data centers are surging in energy consumption. The Blackwell system itself has higher power draw, and token usage may explode due to price drops, further driving up electricity demand. The aging U.S. power grid and slow approval processes will become more prominent, potentially constraining the pace of industrial AI expansion.

What does this mean for U.S. manufacturing?

The reshoring and intelligent transformation of U.S. manufacturing are entering a new phase. In the past few years, the CHIPS Act and the Inflation Reduction Act (IRA) have spurred construction of semiconductor and battery factories, but AI penetration in manufacturing remains low. One core barrier has been inference cost. Now, with token prices plummeting, a key cost obstacle for manufacturing AI applications has been removed.

• Specifically, small and medium-sized manufacturers will find it easier to afford AI tools, for example:
• Using AI vision systems to replace manual quality inspection and reduce return rates;
• Leveraging large models for equipment maintenance to minimize downtime;
• Optimizing supply chain logistics with AI to lower inventory costs.

A larger impact is that the U.S. may use this to narrow the gap with Japan and Germany in precision manufacturing, improving product quality and global competitiveness through AI-assisted processes.

What does this mean for the supply chain?

On one hand, the supply of Nvidia Blackwell becomes critical. TSMC's CoWoS packaging capacity remains tight, and Blackwell requires more advanced packaging technology. Insufficient supply could delay the pace of token price reductions. On the other hand, falling AI model costs may accelerate "nearshoring" decisions: U.S. factories, empowered by AI, improve cost efficiency, offsetting some of the labor cost advantages overseas, making supply chain reshoring more attractive.

Additionally, the data center equipment supply chain (liquid cooling, power supplies, cabinets) will benefit from the Blackwell deployment wave, with orders for related manufacturers such as Vertiv and Schneider Electric expected to grow.

What does this mean for corporate investment?

Corporate capital expenditure directions are shifting. Over the past two years, companies invested heavily in AI training infrastructure, buying H100 GPUs. Now, Blackwell's efficiency makes inference investment more attractive. From the second half of 2026 onward, companies are expected to accelerate the migration from Hopper to Blackwell, creating a new GPU procurement cycle. At the same time, electricity costs will account for a larger share of total costs, and companies may be more willing to invest in on-site renewable energy or sign long-term power purchase agreements (PPAs) to lock in electricity costs.

What does this mean for the next 5 years?Over the next 3-5 years, AI token costs may continue to decline, pushing industrial AI penetration from single digits to double digits. AI applications in U.S. manufacturing will move from "pilots" to "scale." This will have a dual impact on the labor market: on one hand, replacing some low-skilled inspection and sorting jobs; on the other hand, creating new positions such as AI operations and model fine-tuning.

However, the bottleneck has shifted from computing power to energy. Can the U.S. grid handle the dual load of data centers and industrial AI? Regulatory reforms in various states (such as Texas Governor Greg Abbott's proposal to limit the pass-through of data center electricity costs) will determine the pace of expansion. Similarly, the deployment of industrial AI also needs to address issues like data security and model reliability.

Conclusion: The plummeting token costs brought by Blackwell are not just a simple price war, but the true starting point of the U.S. industrial AI era. Investors should focus on industrial automation software, data center power equipment, and manufacturing companies with AI deployment capabilities.