I've had this conversation before

I was having drinks with a friend a couple weeks ago when he started venting about his AI bill. Not the model, he loves the model. His team ships faster than they ever have. What’s killing him is the invoice. Every week he’s digging through usage logs trying to figure out which feature is burning through tokens like a space heater left on in July.

I laughed, and not because it’s funny. I laughed because I’ve had this exact conversation before, word for word, about a decade ago. Back then the subject was the cloud. Companies were racing to get off their own servers and onto AWS or Azure, thrilled with the flexibility, and then the first real bill landed and somebody in finance had a very bad Monday. My memory of that era is that CTOs were blowing way past budget, sometimes by what felt like multiples, not percentages. I can’t verify a specific number close to what I remember, so take that as my recollection and not a stat. Funny how none of those CTOs kept the receipts. But the actual research from that period backs up the shape of it: RightScale’s State of the Cloud surveys found that companies were wasting roughly 35 percent of their cloud spend, year after year, because nobody had gone back and right-sized what they’d built.

Same pattern. Different receipt.

The line everyone quotes and nobody can source

Somebody at that same table said “history doesn’t repeat, but it rhymes,” and I nodded like I knew where it came from. I didn’t, so I checked, and turns out almost nobody does. It’s usually pinned on Mark Twain, but there’s no evidence he ever said it. The earliest real match, according to Quote Investigator, is psychoanalyst Theodor Reik, writing in 1965: “It has been said that history repeats itself. This is perhaps not quite correct; it merely rhymes.” Twain got credited five years later, in 1970, by a writer who was pretty sure he’d read it somewhere and just went with it. Which feels like the whole theme of this post: nobody checked before they repeated something that sounded true.

Same shift, new line item

Here’s what’s actually rhyming. The cloud didn’t just move where your servers lived, it moved how you paid for compute. You went from capex (buy the hardware, own it, depreciate it over years whether you use it or not) to opex (rent exactly what you use, pay for it monthly, scale it up or down on demand). That shift accelerated hard starting around 2014, when Microsoft and Google jumped into the public cloud market and prices started falling fast. AWS’s revenue grew 70 percent in 2015 and 55 percent in 2016, according to NBER’s research on the period. That’s not gradual adoption. That’s a stampede.

AI is doing the same thing to compute again, one layer up. You used to need to own the model, or at least own the GPUs to run one. Now you pay per token, per call, per output, and the meter runs whether or not anyone checks it. It’s capex to opex for intelligence itself. And the stampede this time is bigger and faster. Gartner is now forecasting worldwide AI spending to hit $2.59 trillion in 2026, up 47 percent year over year. And the FinOps world, the discipline that exists specifically to clean up cloud cost messes, has noticed. Two years ago, 31 percent of FinOps teams were managing AI spend. Today it’s 98 percent.

My friend at the bar wasn’t wrong to be thrilled with what the model can do. He was just standing exactly where every cloud-migrating CTO stood in 2015: adopt first, ask questions never, get surprised later.

The optimization phase, not the panic phase

Here’s the part that should actually make him feel better. The cloud didn’t stay chaotic. FinOps became a real discipline, teams learned to right-size instances, schedule things to shut off when idle, and stop paying for capacity they weren’t using.

The same playbook is showing up for AI, and it’s grounded in real technique, not vibes. A January 2026 research paper testing prompt caching across OpenAI, Anthropic, and Google found that caching just the system prompt, and keeping dynamic content out of the cache, cut API costs by 41 to 80 percent. Add in the basics: don’t send your priciest model a question a cheaper one can answer, batch what doesn’t need to run in real time, and actually look at your usage logs instead of waiting for finance to forward you the invoice.

None of that is glamorous. It’s the same unglamorous work that made cloud spend survivable. It’s just wearing a new coat.

I texted my friend a few days later and told him what I’d found. He said something like “so we’re just doing the AWS thing again.” Pretty much. History isn’t repeating itself here. It’s rhyming, once you can actually track down who said that first.