In the spring of 2023, at the peak of the ChatGPT frenzy, Amazon's name was conspicuously absent. OpenAI unveiled GPT-4, Google launched Bard, and Meta open-sourced Llama. And Amazon? Without a single large language model of its own, it bundled other companies' models on top of AWS and called the platform Bedrock. Around the tech industry, people said it openly: Amazon had already lost the AI war.

But starting in late 2025, the picture began to change. AWS's growth rate steepened again, and Amazon's homegrown AI chip, Trainium, has been quietly filling large data centers. Why is a company once dismissed as a straggler holding up so well? To find the answer, you have to start with where the AI industry's center of gravity has moved.

The Center of Gravity Has Shifted

The AI industry has two clearly distinct phases. First came the training era, a fight over who could build the biggest, most powerful model. Training a model with hundreds of billions of parameters demands enormous compute and world-class researchers. Every time GPT-4, Gemini, or Llama arrived, headlines erupted, and investors' attention fixated on that race.

The next phase is the inference era. Now the core challenge is running already-trained models in production, serving hundreds of millions of users reliably, and driving costs down to something sustainable. What separates winners is less which model is more powerful than how cheaply and quickly you can run it. Infrastructure operations, latency optimization, cost efficiency, and global distribution become the battleground.

Since 2025, the center of gravity has unmistakably tilted from training toward inference. AI startups are focused less on building new models than on wiring existing ones into products, and corporate AI spending is flowing toward API calls and infrastructure rather than model development. Amid this transition, everything AWS spent twenty years building began to shine again: a platform that has handled companies' most critical workloads across dozens of regions worldwide. Delivering AI services reliably and at scale was a problem Amazon already knew how to solve.

How to Hold On Until Your Game Begins

One way to understand Amazon is to look at how often the company has made choices that looked foolish in the short term.

When AWS launched in 2006, the reception was icy. The very idea of an online bookstore renting out servers struck people as absurd. When the Kindle appeared, it wasn't even clear an e-book market existed. In Prime's early years, the cost of fast shipping far exceeded membership revenue. By every short-term metric, these were all losses. And one by one, they became Amazon's greatest assets.

These choices share a common logic: work backward from what customers will eventually want, and invest in the capabilities that make it possible before worrying about near-term profit. AWS was built on the judgment that companies would eventually ditch their own servers and move to the cloud. Prime was built on the calculation that if customers stopped thinking about shipping costs, their purchasing behavior would change. These were decisions driven by principle, not prediction.

The same pattern repeated during the AI transition. In 2022 and 2023, while OpenAI monopolized the headlines, Amazon kept developing its Trainium chips and quietly expanded the Bedrock platform. It gave up the news cycle of the training-era race and instead laid the groundwork for the inference era. Amazon didn't gain its edge in AI by suddenly pulling ahead. It gained it by continuing to do what it was already good at.

One caveat is worth pausing on. Amazon did not accurately predict how AI would unfold. There's no evidence that in 2022 it was certain an inference era was coming. What it did was something of a different order: it kept investing, on the conviction that its core strengths — operating infrastructure at scale, driving down costs, and maintaining the trust of enterprise customers — would stay relevant over the long run.

What You Need to Work in the Inference Layer

Now let's bring this story into the context of Korea's solo entrepreneurs.

People weighing AI adoption often ask the same questions: "Which model is the best?" "Which tool should I use?" That is training-era thinking — chasing which model tops the benchmarks, which company built the most powerful AI. In that race, a one-person business has no way to win.

The inference-era competition happens somewhere else entirely. The battleground is how well you connect already-trained models to your own business context. The advantage goes to the person who has tested, in their own work, whether attaching AI to a task actually saves time, actually improves customer service, or merely looks impressive. That capacity for verification doesn't come from technical knowledge. It comes from knowing your own work in concrete detail.

When people discuss what remains for humans in the AI era, one view holds that attitude and habits of judgment matter more than technical skill — that deciding which capabilities to build, and staying steady as trends shift, is something only a person can do. Amazon's story proves that idea at the corporate level: an organization that kept doing its own work, while the news cycle pointed at OpenAI, until its own game began. That is what durability actually is.

There are questions worth asking yourself. If the AI tools you use have changed over the past year, did you replace them out of better judgment, or because you followed the news? Is there an area of your business where you actually use AI repeatedly? Are you investing now in capabilities that will still matter three years from now? If these questions feel uncomfortable, it may be a sign you haven't yet started asking inference-era questions.

Amazon didn't win the AI race. It simply kept doing what it needed to do until its own game began. That habit is now showing up under the name of durability. What Korea's solo entrepreneurs and small teams should take from this picture isn't a technology choice — it's that habit.