The Illusion of Certainty: RPOs, Oracle, and the Dot-Com Ghosts

Every cycle invents its own way of bottling the future animal spirits into today’s stock price. And on that topic, I’m reading the new biography on John Malone and he says the most important investing lesson is to always ask “what if not?”, to look past the upside and imagine the downside. That discipline is what keeps castles from turning into rubble and from ruining your personal net worth. And it’s the question too few investors ask when the market bottles all of the future into a single number. In 1999 it was “eyeballs. In 2007 it was “house prices never go down.” In 2025 it’s a sterile accounting line called Remaining Performance Obligations (RPO), a footnote that’s now treated like a stone tablet given from on high.

Oracle just told the world it’s current quarter was a miss, but that its RPO surged 359% year-over-year to $455 billion, thanks to a $300B OpenAI contract. And what’s funny is the market didn’t punish them for missing revenue estimates. Instead, it bid the stock up, rallying 40% and making Larry Ellison the richest man in the world. In the process algos, pods, or investors were effectively saying, I’m happy to pay today for numbers that won’t materialize until 2030, if at all. Because AI revenue means stock has to go up.

And that’s the problem, an RPO is not prophecy, it’s a ledger of promises that are enforceable, yes, but subject to the same human frailties that blew up backlogs in the dot-com bust and blew up the housing market in 2007.

I am not bearish on the sector. But I wanted to share a quick bit of research I did since I absolutely thought RPO was Run Pass Option, and not a means of calcuating future revenue. So treat this as a bit of knowledge, a bit of caution, and a bit of “what if not”.

The first thing to realize is that GAAP’s ASC 606 for RPO’s is real. This is a standardized generally accpeted accounting practice, so this is most certainly not fraud. This accounting allows for only non-cancellable contracts or those with penalties attached to make it into RPO. That’s why Oracle could book OpenAI’s $300B deal, on paper, because it’s enforceable. But binding ≠ certain. Companies warn in their 10-K’s that RPO can be hit by customer bankruptcies, regulatory shifts, and other risks.

In the late 1990s, telecom equipment makers like Lucent, Nortel, and Tellabs bragged about record backlogs, treating them as proof of unstoppable demand. On paper, those contracts looked binding, yet when the bubble burst, capital spending collapsed, orders were canceled, and revenues evaporated. Lucent went from beating Wall Street expectations 14 straight quarters to losing $64 billion in market cap in a single day when its backlog failed to convert. Nortel and Tellabs followed the same path, showing that even the most “locked-in” backlog can vanish when customers falter and credit dries up.

The reality for Oracle is that of the $455 billion, only about a third of it is set to convert in the next 12 months, the rest is a promissory note. And this is not a comment on the future direction of Oracle’s stock, only a caution that when you begin to pay higher and higher prices for things further and further into the future … things can go wrong.

Oracle’s deal with OpenAI is, at its heart, very simple: Oracle builds and runs giant computer factories, massive data centers filled with specialized chips, racks of servers, cooling towers, and the electricity hookups to power it all. OpenAI pays Oracle for access to that horsepower so it can keep training and running bigger AI models. Instead of OpenAI building those facilities itself, it signs a contract to rent the capacity in advance. Oracle keeps the lights on and the machines running; OpenAI gets guaranteed compute. To be clear, OpenAI and none of the huge numbers you see right now actually exist as cash in the bank. These are promises built off future cash flow models using assumptions on monetization, adoption, and things like energy prices.

The headline number is mind-numbing: 4.5 gigawatts of compute capacity, equal to four nuclear power plants, enough electricity to serve several million homes for a full year. And while it’s exciting to think about, the reality is that scale ≠ profitability. For those numbers to materialize, a lot has to go right: billions in financing must keep flowing, regulators must approve new power hookups, utilities must deliver grid capacity, hardware suppliers must ship millions of cutting-edge GPUs, and customer demand must keep rising fast enough to fill all that capacity. Breaks in the chain mean these promises become IOU’s that can be renogiated.

Academic research also shows that firms with large backlog swings are significantly more likely to restate revenue. Why? Because backlog invites opportunism. Companies in bull markets push for eye popping long deals with huge headline numbers, trumpet the RPO, and let the market mistake it for destiny. And why not? There is quite literally little to no downside. But when a backlog number gets too big, the temptation is to smooth earnings with aggressive recognition or hide softness by pointing to future promises.

There’s actually some academic research out there about this. A study of 64,000 firm-years of Compustat data, by Barber, Hollie, and Park found that order backlog is a double-edged sword. Yes, it can moderate short-term weakness but when backlog balloons relative to sales, future earnings are often worse, not better. The culprit is congestion, basically a high backlog-to-sales ratio is less a sign of vibrant demand than of bottlenecks: production stretched, supply chains clogged, operations unable to convert orders into revenue efficiently.

And one of the reasons why the finance world is skeptical in this case is that over the years, Oracle and Larry Ellison have faced questions about accounting and disclosure practices tied to the company’s rapid growth. In the early 1990s, Oracle encouraged customers to buy licenses up front, a strategy that boosted near-term revenue but later forced the company to restate earnings when future demand fell short. A decade later, investors raised concerns during the rollout of Oracle’s Suite 11i software, leading to lawsuits alleging the company had overstated sales and downplayed internal challenges; those cases were heavily litigated and ultimately dismissed for lack of proof that Oracle deliberately misled shareholders. And more recently, in 2016, a former finance manager filed a complaint suggesting Oracle’s cloud revenue disclosures were overly optimistic, claims the company denied. While some of these episodes reflect the scrutiny that comes with managing a global enterprise at scale, they also underscore why backlog and revenue metrics deserve close attention from investors. That history doesn’t make Oracle unique (or bad), many fast growing firms have faced scrutiny, but it’s a reminder that even the giants can push the limits of disclosure when incentives are strong.

The lesson is simple: when the crowd is willing to pay today for certainty decades away, it’s not discipline, it’s faith.

Oracle’s half-trillion-dollar RPO is staggering, but it rests on energy that hasn’t been built, customers that must raise capital to consume it, and timelines vulnerable to every macro gust. And the research is blunt: big backlogs increase the odds of accounting problems, not decrease them.

And for the million dollar question, or should I say $455 billion question, does any of this matter?

For now the flows into AI are unstoppable. The revenue growth is real. Concentrated, yes, but real none the less. Companies in the ecosystem have generated huge amounts of cash and are spending it on this next wave of the future.

The AI trade has unfolded in three powerful phases, each building on the last, and Oracle has just proven why it sits at the crossroads. Phase 1 was the chips, the raw horsepower of NVIDIA, AMD, Broadcom, and Micron that turned silicon into the new oil. Phase 2 is the infrastructure buildout, Microsoft, Amazon, Google, and especially Oracle, which has reinvented itself from legacy laggard into a hyperscaler with blistering cloud growth. Its breakout has become the clearest signal that the AI capex wave is institutionalized, pulling in Arista, Vertiv, and Super Micro to provide the racks, cooling, and networking muscle. And the future alpha sits in Phase 3, where platforms and applications like ServiceNow, Adobe, Snowflake, CrowdStrike, and HubSpot monetize usage, charging every time a token is processed or a workflow is automated. Each phase has pulled real cash into the system but history says it’s the application layer where faith and fundamentals smash together. Can this promise turn into something real and productive? I think so, but I guess we’ll see.

The lesson from all this is simple though: when the crowd is willing to pay today for certainty decades away, it isn’t discipline, it’s faith. Lucent, Nortel, and Tellabs once sold investors on backlogs that looked unshakable, until they weren’t. Real estate brokers in 2007 convinced people that home prices could only go one direction, up. And now Oracle’s $455 billion RPO may prove real, but it rests on energy not yet built and customers still scrambling for cash. Investors aren’t just buying compute, they’re buying castles in the sky. And castles in the sky have a way of looking solid, right up until they vanish.

Malone’s question lingers: what if not? What if the backlog slips, the power isn’t built, or the demand curve bends lower? Asking that simple question forces us to see RPOs for what they are not prophecy, but possibility. And possibility always carries risk.

Until next time,

Victaurs