AsianDadEnergy's Substack Podcast
Every few weeks, a new headline appears claiming that AI is becoming unsustainably expensive. Companies are reportedly burning through millions of dollars a month on AI subscriptions. Some organizations are discovering that employees can consume astonishing quantities of tokens when every workflow becomes an AI workflow. There are even stories circulating of teams whose AI bills now exceed the cost of the employees using the tools. For many people, these stories confirm a suspicion they have been quietly nurturing all along. Maybe AI isn’t economically viable. Maybe the entire thing is a bubble. Maybe the costs will eventually become so overwhelming that widespread adoption simply won’t happen. I understand why people feel that way. There are enormous emotional and financial stakes involved. Entire industries are being disrupted. Careers are being questioned. Businesses are being reshaped. Some people stand to make fortunes while others worry about becoming obsolete. But after spending decades in the technology industry and watching wave after wave of disruption unfold, I have reached a different conclusion: AI is already cost effective today. And it is almost certainly going to become dramatically cheaper tomorrow. The Cost Problem Is Already Being Solved Most discussions about AI costs focus on the wrong question. People look at today’s prices and assume those prices are fixed. They aren’t. The history of technology is a history of making expensive things cheap. Computers were once available only to governments and large corporations. Cell phones were luxury items. Internet access was expensive. Data storage cost a fortune. Then competition arrived. Innovation happened. Costs collapsed. AI is following the same pattern. What makes the current moment particularly interesting is that many of the cost reducing innovations are not hypothetical future breakthroughs. They already exist. They are being deployed right now. How Sanctions Accidentally Accelerated AI Innovation One of the more fascinating developments of the past several years has been the unintended consequences of the technological competition between the United States and China. The United States has attempted to restrict China’s access to advanced semiconductors and semiconductor manufacturing equipment. The logic is straightforward: limit access to compute and you limit AI development. But resource constraints often produce innovation. When organizations cannot simply throw more hardware at a problem, they are forced to become more efficient. And efficiency is exactly where many Chinese AI companies have focused their efforts. The result has been a wave of innovations aimed not at creating the absolute most powerful models, but at creating models that are nearly as capable while being dramatically cheaper to train and operate. That distinction matters. Because in the real world, economics often wins. Smarter Models, Not Just Bigger Models For years, the dominant strategy in AI was simple. Build bigger models. Use more GPUs. Consume more electricity. Spend more money. But there is another path. Instead of scaling everything upward, researchers can improve the underlying architecture itself. A good example is the evolution of attention mechanisms inside large language models. Architectural improvements can dramatically reduce the amount of computation required during training while maintaining similar performance. In some cases, these optimizations reduce training costs by multiples rather than percentages. That is a profound shift. If a model can achieve similar results using half, a third, or even a fifth of the resources, the economics change completely. Why Reinvent the Wheel? Another powerful cost reduction technique is model distillation. Imagine spending years building an expert employee. Now imagine being able to transfer much of that expertise into a new employee at a fraction of the cost. That is essentially what distillation does. Instead of starting from scratch, newer models learn from the outputs and behaviors of existing advanced models. The result is significantly lower training costs and dramatically reduced data requirements. From a business perspective, this is incredibly attractive. Why spend one hundred million dollars recreating knowledge that already exists when you can acquire much of it for a small fraction of the cost? The Hidden Battle: Inferencing Costs Most people focus on training costs because the numbers are eye catching. But for many businesses, the bigger issue is inferencing. Inferencing is simply the process of asking an AI model a question and receiving an answer. Every prompt requires computation. Every computation costs money. And when millions of users are interacting with AI systems, those costs add up quickly. This is where some of the most important innovations are occurring. Techniques such as mixture of experts allow models to activate only the portions of the neural network necessary for a specific task. Instead of powering the entire machine every time, only the relevant specialists are called into action. The result can be reductions in inferencing costs ranging from significant to dramatic. When multiplied across billions of requests, the savings become enormous. Hardware Matters Too Software is only half the story. Hardware innovation is equally important. Because Chinese companies have limited access to the most advanced Chip manufacturing technologies, they have increasingly focused on specialized chips designed for specific workloads. These application specific chips may not be as versatile as cutting edge GPUs, but versatility is not always the goal. Efficiency is. When optimized for AI inferencing, specialized hardware can often deliver surprisingly competitive performance at substantially lower costs. Again, the pattern repeats. Constraints force optimization. Optimization reduces costs. Reduced costs accelerate adoption. The Energy Advantage There is another factor that rarely receives enough attention. Electricity. Training and running AI models requires enormous amounts of power. Power costs are not fixed. They vary dramatically depending on geography and infrastructure. As renewable energy continues to become cheaper, the cost of operating AI systems falls alongside it. This creates another powerful deflationary force acting on AI economics. Even if the models themselves never improved, cheaper energy alone would lower operating costs over time. But the models are improving. And the hardware is improving. And the software is improving. All at the same time. What Happens When American AI Companies Start Optimizing? The most interesting part of this story may be what has not happened yet. Many American AI companies have operated in an environment of abundant capital. When investor funding seems unlimited, optimization is often less urgent. Speed matters more than efficiency. Growth matters more than profitability. But markets eventually change. Investors begin asking harder questions. Profitability becomes important. Efficiency becomes important. And suddenly all of the techniques that were previously ignored become very attractive. If Chinese companies can reduce costs dramatically through architectural improvements, distillation, specialized hardware, and operational efficiency, there is nothing preventing American companies from doing the same. In fact, competitive pressure almost guarantees that they will. The Real Question Could the AI bubble pop? Of course. Every technological revolution creates bubbles. Money will be made. Money will be lost. Speculators will speculate. Some companies will fail spectacularly. That part is normal. But bubbles and underlying technology are not the same thing. The railroad bubble burst. Railroads did not disappear. The dot com crash happened. The digital economy kept growing. The real question is not whether investors will overpay for AI companies. The real question is whether the cost of using AI will continue falling. Looking at the technologies already available today, my answer is yes. Decisively yes. The future of AI may be many things. But “too expensive to survive” is not the outcome I would bet on. Get full access to AsianDadEnergy's Newsletter at asiandadenergy.substack.com/subscribe [https://asiandadenergy.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4]
34 Folgen
Kommentare
0Sei die erste Person, die kommentiert
Melde dich jetzt an und werde Teil der AsianDadEnergy's Substack Podcast-Community!