Avoiding Dependencies: Lessons From The Pandemic

Remember when Uber rides were $5?

For a brief window, it genuinely felt like the economics of getting around a city had been rewritten for good. A car showed up in minutes. The ride across town cost less than a coffee. It seemed almost too good to be true.

It was.

First the subsidies shrank. Then regulators got involved. Then it was only a matter of time before the underlying math of running a global transportation network caught up with the pricing. Rates climbed up and up. The convenience stuck around, but the cheap rides did not.

All this, mind, after a not-insubstantial amount of Taxi companies had been put out of business.

Does This Sound Familiar?

So, what’s the relevance of this today?

We talk a lot about how great AI is here at HyperscaleCEO, and for good reason: The current generation of AI tools is genuinely impressive. AI allows one employee to do the work of many, encourages rapid prototyping, and has fostered the emergence of new, smaller, AI-native companies that are nipping at the ankles of the bigger players who are still struggling to adapt.

The cost of all this, at least so far, has been low enough that adoption has outpaced almost any prior technology wave. 

Stanford's Institute for Human-Centered AI has tracked a dramatic expansion in organisational AI use as generative models moved into mainstream business workflows.

When something is both powerful and cheap, of course people rush in. The harder question is whether that pricing reflects reality, or whether someone else is currently picking up the tab.

Right now, a lot of the cost is being absorbed by vendors fighting for market share. The companies building these models are running some of the most expensive computing infrastructure ever assembled. The chips alone run tens of thousands of dollars per unit.

Powering and cooling the data centres that house them requires enormous and ongoing capital.

McKinsey and others have estimated that scaling global AI compute capacity over the next decade could require trillions of dollars in investment across the industry.

That bill has to land somewhere eventually.

Today, AI services are typically priced by usage, per request, per token, per inference call. On paper those unit costs look trivial. A single model interaction might cost a fraction of a penny. 

But that math changes quickly at scale. 

A product serving millions of users generates millions of model calls per day. Each new AI-powered feature layered into a product adds another line of recurring inference cost that compounds quietly in the background - Costs that are invisible to the end user during early experimentation. 

When a product, or an organization, becomes wholly AI dependent, those dependencies accumulate into a structural cost that grows every time usage does.

CEOs need to ask themselves: Are we making architectural decisions that assume today’s pricing is stable?

After all: We’ve seen this play out before.

Streaming services subsidised content to build audiences before raising prices. Cloud providers used aggressive discounting to move enterprises off on-premise infrastructure.

Once customers had reorganised themselves around the new platform, the pricing dynamics shifted from capturing markets to generating revenue. 

AI is unlikely to be exempt from that cycle. If compute demand continues to outpace supply, vendors adjust. If new capabilities such as larger context windows, real-time agents and multimodal reasoning require substantially more compute, they'll be priced accordingly.

The Necessity Of Flexibility

The more a company embeds a specific AI provider into its products, internal tools, and data pipelines, the harder it becomes to respond when economics change. 

Migrating to another provider, or bringing models in-house, can mean months of engineering work and significant operational disruption. You can end up structurally dependent on a supplier whose cost structure you don't control.

None of this is an argument for slowing down. 

The productivity gains from AI are real, and in many industries they're already competitively necessary. But there's a difference between experimentation and dependency, and the best leaders will treat them as two distinct phases.

Experimentation should answer a genuine question: does this actually improve the product or the efficiency of the business? Creating a possible dependency deserves the same scrutiny you'd apply to a long-term energy contract or outsourcing a critical manufacturing process. That means stress-testing the business model against scenarios where inference costs increase, or where the next generation of more capable models is priced several times higher than what you're using today. 

Most companies haven't run those numbers, because the current bills are small enough that nobody's worried yet.

It also means thinking carefully about architectural flexibility. 

Organisations that maintain control over their data pipelines, core orchestration layers, and embeddings will have real options when the market shifts. Organisations that build tightly coupled integrations around a single vendor may find that switching costs are far larger than they anticipated.

The Last Word

The lesson from the $5 Uber ride of old is such: Early pricing rarely reflects the long-term economics of operating a complex, capital-intensive technology network.

AI is following the same arc. The opportunities are real, but so are the temptations to treat current pricing as a permanent feature rather than a phase. 

CEOs who understand the difference will be far better positioned when the industry's actual cost structure catches up with what's being charged today.

Image Credit: Stockcake