Today’s AI prices are heavily subsidized, far below what it actually costs to deliver the services. The largest providers spend more to serve requests than they charge, losing significant amounts of money in the process. This is a deliberate strategy, funded by capital, to quickly gain market share and lock in customers before competitors catch up.

Read “The AI bill is coming. Who will pay?” at DN (norwegian only) >

This makes today’s pricing unstable. When capital needs to be repaid and investors demand returns, it’s reasonable to expect prices for the most powerful models to rise.

A pattern we’ve seen before

This is well known in cloud computing. Services are sold cheaply to win customers, and prices are raised once enough users are locked in. Google Workspace kept prices flat for ten years while taking customers from Microsoft, then raised them by around 20 percent in 2023, and more for some. Microsoft is now bundling AI into subscriptions, increasing prices by 5 to 43 percent from July 2026. And when Broadcom acquired VMware, customers reported increases between 800 and 1,500 percent, customers who couldn’t easily switch.

Read “Google hikes G Suite prices” at TechRadar >

Read “New Microsoft 365 pricing from July 1, 2026” at Microsoft >

Read “VMware price hikes? 800–1,500%, claim Euro customers” at The Register >

AI locks you in more tightly than traditional cloud

Lock-in is intentional and part of the business model. It’s free to move data into the cloud, but expensive to move it out, and large data volumes can cost tens of thousands of dollars to transfer. With AI, the lock-in becomes stronger. Your data, fine-tuning, agents, prompts, and integrations all become switching costs. The deeper AI is embedded in your workflows, the less negotiating power you have when prices change. The EU has recognized this and requires switching costs to be removed by 2027, but until then, many are stuck.

Read “Cloud Egress Fees Explained” at Backblaze >

The honest nuance

However, costs are also falling in some areas. The price per token, that is, per unit of text the AI reads or writes, has dropped dramatically, sometimes by orders of magnitude per year, as models become more efficient and competition increases. Gartner expects the cost of running models to continue declining toward 2030.

Read “LLM inference prices have fallen rapidly but unequally across tasks” at Epoch AI >

What AI will cost in the future is still uncertain. But the total bill is growing regardless, because usage is increasing faster than prices are falling. Agents consume many times more than a single query, and cheaper tiers are disappearing or being bundled into more expensive packages. You should therefore not base a three-year decision on today’s subsidized pricing.

Read “AI inference costs set to plunge: Gartner” at CIO Dive >

What you should do

Use AI, but build for changing prices. In practice, this means maintaining flexibility with multiple providers, using open models where possible, and structuring data and integrations so you can move without starting over. Set limits and visibility on usage so agents operate within controlled and transparent budgets. Measure cost per outcome rather than per token. A 2025 MIT study found that 95 percent of enterprise AI pilots delivered little measurable impact on the bottom line. Cheap AI without clear value is still wasted money.

AI is cheap today because someone else is paying for you to become dependent. Take advantage of it if you want, but build your solution so you still have choices when the bill becomes yours.