Why AI Models Are Becoming Harder to Train: The $100 Billion Compute Problem

Training next-gen AI just hit a trillion-parameter wall—and the compute costs are bleeding even Big Tech dry.

Data Hunger Games

Today's models swallow datasets larger than entire internet archives from a decade ago. Each training run chews through petabytes—enough to make cloud providers wince at the electricity bill.

Hardware Bottlenecks

GPUs that powered the last AI boom now look like toy calculators. Custom chips help, but fabricating specialized silicon adds years and billions to development cycles.

The Diminishing Returns Trap

Throwing more data at models yields smaller performance bumps. Researchers now hunt for architectural breakthroughs—not just bigger datasets—to move the needle.

Regulatory Headwinds

Privacy laws limit data scraping. Export controls throttle chip access. Every new restriction adds friction to the global training pipeline.

Meanwhile, VCs still pitch 'AI disruption' like it's 2023—ignoring that training costs now exceed the GDP of small nations. Maybe they'll monetize those loss-generating models through crypto tokens next.

Elevate Your Investing Strategy:

• Take advantage of TipRanks Premium at 50% off! Unlock powerful investing tools, advanced data, and expert analyst insights to help you invest with confidence.

For example, a linguist who worked on OpenAI’s o3 model last year said that he used to find three or four tasks per week that the AI couldn’t solve. Now, working with the newer GPT-5, he’s only finding one or two. Other experts, like those in biology or chemistry, are still having some success finding things the model can’t do, but even that is becoming more difficult as models advance. In addition, the tasks themselves have become extremely complex.

For instance, a chemistry question asked the model to find a research paper using detailed molecular data, locate and cite documents, reformat computational structures, and analyze chemical similarities. Most people, especially those without advanced degrees, couldn’t even begin to answer it. As a result, this raises a new issue: as models get smarter, OpenAI and Anthropic will need even more highly specialized experts to train them. But convincing professionals with decades of experience, or even Nobel Prize winners, to spend time teaching these models may be tough.

Is MSFT Stock a Buy?

Turning to Wall Street, analysts have a Strong Buy consensus rating on MSFT stock based on 34 Buys and one Hold assigned in the last three months. In addition, the average MSFT price target of $625.78 per share implies 23.1% upside potential.