In the AI Gold Rush, the Only Ones Smiling Are the Lads Selling Shovels

AI-enabled revenue feels a lot like tulip bulbs in 1637.

Back then, everyone was buying tulips because everyone else was buying tulips. Didn’t matter what they were worth — just that prices kept going up. A single Semper Augustus bulb traded for more than a canal house in Amsterdam. Everyone knew it was irrational. Everyone kept buying.

Today it’s the same, but with data. Leaders hoard it by the petabyte. Engineers polish it, label it, model it. All in the hope it’ll bloom into “AI-enabled revenue.” The board wants it. The investors expect it. The competitors claim to have it. So the mandate comes down from on high: “We need an AI strategy.”

Which usually means shovelling junk data into shiny dashboards and calling it transformation.

But most of it’s weeds. No scent, no beauty, no buyers. Just a market high on hype and cloud credits.

The data hoarding instinct

The execs hoard data like prepper tins. Because data’s the new oil, apparently. Someone said it at Davos. It was on a slide. It must be true.

So they collect everything. Every click, every transaction, every sensor reading, every log line. Terabytes of it. Petabytes. Into the data lake it goes — which, as anyone who’s worked with one knows, rapidly becomes a data swamp. Ungoverned, undocumented, and untouched by anyone who understands what questions it could answer.

Then they throw it at engineers and say, “Go make it valuable.”

The engineering theatre

So the engineers graft. And they’re good at it — that’s not the problem. They build pipelines. Elegant ones. Spark jobs, Airflow DAGs, Kafka streams, dbt models. They train ML models on whatever data they can find. They build dashboards — beautiful dashboards, with charts that update in real time and metrics that go up and to the right.

It all looks clever. Feels important. The demo goes well. The exec sponsor nods approvingly. The quarterly review slide says “AI-driven insights platform — Phase 1 complete.”

But nobody asks the uncomfortable questions:

Is the data any good? Half of it is duplicated. A quarter is stale. The lineage is undocumented. The schema has drifted three times since anyone last checked. The ML model is trained on data that was cleaned by an intern who left six months ago and whose Jupyter notebook is saved somewhere on a laptop that’s been reimaged.

Do the insights matter? The dashboard shows that sales of Widget A correlate with temperature in the southeast region on Tuesdays. Fascinating. What does anyone do with that information? Nothing. Because the insight doesn’t connect to a decision anyone needs to make. It’s trivia dressed up as intelligence.

Do the customers care? This is the question nobody wants to ask. Is there a customer — internal or external — who will pay more, stay longer, or do something differently because of what you’ve built? If not, you haven’t created value. You’ve created cost.

The cloud bill reality check

Meanwhile, the cloud costs rise faster than a Blue Origin rocket — Jeff Bezos’ other vanity project.

The data lake is in S3. The compute is in EMR. The ML models run on SageMaker. The dashboards are in QuickSight. The real-time pipeline uses Kinesis. The data warehouse is Redshift. Or maybe Snowflake. Or Databricks. Or all three, because nobody made a decision and everyone picked their favourite.

Each of these services has a usage-based pricing model that seems cheap at prototype scale and becomes eye-watering at production scale. The monthly AWS bill is a number that makes the CFO’s eye twitch. But nobody can turn anything off because nobody knows what depends on what, and the sunk cost fallacy has taken hold: “We’ve spent this much already, we can’t stop now.”

The ROI calculation, if anyone dared to do one honestly, would be grim. The platform costs more to run than the value of the insights it produces. But the sunk cost is too high to admit, and the exec sponsor’s reputation is tied to the project’s success. So it continues. More data. More pipelines. More dashboards. More cost.

Who’s actually making money?

Here’s the uncomfortable truth about the AI gold rush: the value chain is inverted.

The companies building AI-powered products? Most of them are spending more on infrastructure than they’re generating in AI-enabled revenue. The unit economics don’t work yet for the majority of use cases. The technology is impressive, but the gap between “impressive demo” and “profitable product” is vast.

The companies selling the infrastructure to build AI-powered products? They’re having the best quarter in their history.

AWS, Azure, and GCP are selling compute, storage, and managed AI services at enormous scale — Gartner forecast worldwide public cloud spending at $723 billion in 2025, up 21.5% year-over-year. Snowflake and Databricks are selling data platforms. NVIDIA posted $130.5 billion in fiscal 2025 revenue, up 114% from the prior year, almost entirely on the back of AI data centre demand. Anthropic and OpenAI are selling API calls. Every one of these companies is growing rapidly — not because their customers are making money from AI, but because their customers are spending money on AI.

In the gold rush, the reliable money was never in the gold. It was in the picks, the shovels, and the Levi’s.

What actual AI value looks like

None of this means AI is worthless. It means most organisations are doing it backwards. They start with the technology — “we need an AI platform” — instead of starting with the value — “what decision would we make differently if we had this information?”

The organisations I’ve seen succeed with AI have a few things in common:

They start with a specific, measurable business outcome. Not “AI-driven insights.” Something concrete: “reduce customer churn by 5%,” “cut manual review time by 60%,” “predict equipment failure 48 hours before it happens.” A goal that connects directly to revenue or cost.

They work backwards from the decision. What information would someone need to make a better decision? Does that information exist? Can it be collected reliably? Only then do they think about models and pipelines. The technology serves the decision, not the other way round.

They treat data quality as a prerequisite, not an afterthought. Garbage in, garbage out isn’t a cliché — it’s the single most common failure mode in AI projects. The organisations that succeed invest in data governance, lineage, and quality before they invest in models. It’s less exciting than a neural network, but it’s where the value lives.

They kill projects that don’t produce value. This sounds obvious but almost never happens. The sunk cost fallacy, combined with the career risk of admitting a project failed, means zombie AI projects consume budget for years. The best organisations run time-boxed experiments with clear success criteria and shut down the ones that don’t work.

They do the ROI maths honestly. If the platform costs £500k a year to run and produces £200k in measurable value, that’s not a success story with room for optimisation. That’s a bad deal. Unless you can credibly project the value growing faster than the cost, it’s time to rethink the approach — or stop.

The shovel sellers will be fine

AWS will be fine. Azure will be fine. Snowflake, Databricks, NVIDIA — they’ll all be fine. They’re selling into a market with almost unlimited demand and very little price sensitivity, because the buyers are spending someone else’s money on a strategic imperative they don’t fully understand.

The question is whether you’ll be fine. Whether the money you’re spending on AI infrastructure is producing value that someone — a customer, a user, a business unit — actually cares about. Or whether you’re just buying shovels because everyone else is buying shovels.

If customers value the outcome less than it costs to make, it’s a bad deal. No amount of sophisticated engineering or beautiful dashboards changes the arithmetic.

Tulip mania ended when people realised the flowers weren’t worth the paper contracts they were written on. The same thing will happen here — when customers stop pretending that data insights are gold and start asking what they’re actually worth.

How many of today’s “AI-driven” businesses will look like tulip traders in five years? My guess: most of them. The ones that survive will be the ones who asked “what’s this actually worth to a customer?” before they built the pipeline — not after.

In the AI gold rush, the only ones guaranteed to be smiling are the lads selling shovels. Make sure you’re not just another customer.