AI costs are rising, but proving value is still a challenge
It’s all happening far more quickly than I thought it would.
I wrote about the impending AI cost crisis back in April. We’d all had access to some of the world’s most advanced technology for years at a fraction of what it cost to build and run. It was only a matter of time until its creators cranked up the prices in search of a return on their investment. And then businesses would face tough decisions about their spiralling use of LLMs.
The price hikes came, and the questions around AI’s value began. It feels like an industry-wide equivalent of the challenge that keeps appearing online: “If AI is so powerful, where are all the successful AI-built products?”
Microsoft began cancelling its Claude Code licences. Nvidia’s vice president of applied deep learning said AI ran “far beyond the costs” of its employees. And Uber capped its employees at $1,500 per month in each of its AI tools, just a week after its COO said it was “very hard to draw a line” between AI spending and the desired output: useful features in its products.
The trend has not gone unnoticed by AI leaders. OpenAI CEO Sam Altman said that while AI costs “never came up” earlier this year, now it’s “a huge issue” – which isn’t exactly a reassuring way to talk about your customers when you estimate that you’ll spend $600 billion on compute by 2030.
Weighing code against value
But while AI rightfully grabs the headlines, I think the issue is connected to a broader problem the tech industry has been skirting for years: the question of the changing relationship between code and actual business value.
One defence of increased spending on AI might be productivity gains – isn’t the extra spend worth it if we can get a lot more done? But that equation doesn’t factor in what is getting done and how valuable it is.
How much has a product like Uber, or Amazon, or Netflix actually changed in the last three years? I’m not talking about small tweaks to the UI – I mean new features that either attract new customers or convince existing users to spend more. I’m certainly struggling to name many examples (and I’m not counting user-facing AI integrations, which rarely add much value).
The AI value equation is much easier for start-ups building something new than for established businesses making small UI tweaks.
For start-ups, the equation is straightforward – AI-assisted coding might enable a small team to build a product that would otherwise take a decade and tens of hires. But the value proposition becomes murkier for larger businesses with established products – when every developer is spending tokens, they risk multiplying their costs without a clear link to revenue.
As AI costs come under scrutiny, I expect much more pressure on teams to demonstrate that spend is tied to revenue. That might mean tighter caps, more usage reporting with costs attributed to specific projects, or more businesses leveraging internally hosted – though less powerful – models.
AI is too useful to disappear completely – used in the right contexts, it can be a force multiplier. But the days of businesses throwing unlimited tokens at a wall to see what sticks could be numbered as the bills start to rise.
