Will expensive GPUs slow the progress of AI?

They say lightning doesn’t strike the same place twice. But software company Nvidia may disagree.

Nvidia originally became well known for making Graphics Processing Units (GPUs) for video games ...?

Until GPUs became all the rage with Bitcoin miners.?

Then, just as the smoke cleared from the Bitcoin frenzy, GPUs were in the spotlight again, as AI companies discovered they were perfect for training data-hungry Large Language Models (#LLMs).

This second lightning strike for GPUs has proved to be so big that some say it’s causing demand to outpace supply by a factor of 10.

It’s great news for Nvidia, which dominates the market. But not so great news for a rapidly growing number of AI companies that are emptying their wallets to pay for computing power.

And don’t look now, but China’s internet giants are also entering the fray—placing orders for $5 billion of Nvidia’s highly coveted chips.

When this much money chases a limited supply of hardware, it raises some tough questions.

Will expensive GPUs strangle AI progress just as it’s getting started? Will it put large companies and big budgets in control, and leave behind smaller businesses with less money?

These are important questions without easy answers. But there’s this: The tech sector is nothing if not innovative—and there are signs that a solution may be on the way.

Small yet mighty

Massive consumer-grade LLMs are getting all the attention now.?

But beneath the headlines, there’s a growing awareness that smaller LLMs can deliver impressive performance with a fraction of the resources.

We’re seeing these nimble, private LLMs beginning to flex their muscles in areas like finance and healthcare, where they’re focused on specific use cases and trained on specialized datasets.?

And private LLMs offer so much more value than just cost savings. They come with a host of other incredible benefits.

In the automotive industry for example, an LLM focused on domain-specific data could flag needed repairs and monitor battery health in electric vehicles—in real time.

There’s BloombergGPT for finance, and DialpadGPT for business conversations.

Already, we’re starting to see optimized LLMs supporting industry-specific use cases in more secure and efficient ways. It’s almost like a surgeon choosing a smaller scalpel to do a more precise job.?

Sometimes less really is more.