A $523 Billion Backlog May Not Be What It Seems

• Oracle (ORCL) holds a $523B backlog but shares fell 23.25% YTD as capex reached $20.54B. Nvidia (NVDA) posted Q4 revenue of $68.13B, up 73% YoY. Microsoft (MSFT) spent $29.88B in capex with $625B in RPO.

• OpenAI walked away from Oracle’s Stargate data center expansion because power infrastructure delays mean Nvidia’s current Blackwell chips will be outdated by the time the facility opens, with next-gen Vera Rubin available instead.

• The analyst who called NVIDIA in 2010 just named his top 10 AI stocks. Get them here FREE.

“The chip cycle is moving so fast that even the biggest infrastructure deals can’t keep up.” That line from CNBC’s Deirdre Bosa captures the core tension now rattling AI infrastructure investors. OpenAI reportedly walked away from expanding its flagship Stargate data center in Abilene, Texas, because power won’t be ready for at least a year, and by then, OpenAI doesn’t want the current generation Blackwell chips because NVIDIA’s next-gen Vera Rubin will be available. The practical consequence: Oracle committed the debt, secured the site, ordered the hardware, and the customer said the chips will be dated before the building is even ready.

For investors, this is the moment where a financial concept called the capex-to-revenue lag becomes very real. Understanding it is the difference between seeing Oracle’s $523 billion backlog as a treasure chest and seeing it as a liability waiting to be tested.

The CNBC framing is correct, and investors who dismiss it are taking on more risk than they may realize. Oracle has made an enormous bet on infrastructure at exactly the moment the chip cycle is accelerating faster than power grids and construction timelines can follow. Oracle’s capital expenditures in Q1 FY2026 alone reached $8.5 billion, consuming more than 100% of operating cash flow and pushing free cash flow to negative $362 million. In Q2, capex for the first half of FY2026 totaled $20.54 billion. That is real cash leaving the building to build infrastructure that must eventually be filled with paying customers running current-generation hardware.

READ: The analyst who called NVIDIA in 2010 just named his top 10 AI stocks

The financial mechanic at work here is straightforward. When a company signs a cloud infrastructure contract, it books a Remaining Performance Obligation (RPO), which is committed revenue that hasn’t been recognized yet because the service hasn’t been delivered. Oracle’s RPO surged 438% to $523 billion in Q2 FY2026. That number sounds extraordinary. But Oracle’s quarterly revenue is roughly $16 billion, which means the company is sitting on years’ worth of contracted work. The question is whether those contracts hold when the hardware they were built around becomes a generation behind.

Microsoft is spending aggressively too. Microsoft’s Q2 FY2026 capital expenditures reached $29.88 billion, nearly doubling year-over-year. But Microsoft has a critical buffer Oracle lacks. Microsoft’s quarterly revenue is $81.27 billion, giving it roughly five times the revenue base to absorb capex shocks. If one hyperscaler customer delays or redirects a contract, Microsoft can absorb it. Oracle cannot afford that. It doesn’t have the cushion that the other mega-caps have.

NVIDIA’s financial results make clear why infrastructure builders are scrambling to stay current. NVIDIA reported Q4 FY2026 revenue of $68.13 billion, up 73% year-over-year — a growth rate that reflects a chip market moving faster than any single infrastructure buildout can track. The data center segment was the engine of that growth, with networking revenue surging 263% year-over-year as hyperscalers raced to build out AI capacity. NVIDIA reported Q4 FY2026 revenue of $68.13 billion, up 73% year-over-year, with data center networking revenue alone surging 263% YoY. CEO Jensen Huang described the moment plainly: “Grace Blackwell with NVLink is the king of inference today, delivering an order-of-magnitude lower cost per token, and Vera Rubin will extend that leadership even further.” The message to infrastructure builders is clear: the window for any given chip generation is shrinking.

The implication for infrastructure builders is that a data center designed around Blackwell GPUs today may be commercially inferior by the time it opens. Power infrastructure, which is the binding constraint in Abilene, takes 12 to 18 months to provision. Chip generations are turning over faster than that. OpenAI’s decision to redirect capacity to newer sites where Vera Rubin can be deployed from day one is a rational response to this mismatch. For Oracle, it means a site that absorbed billions in upfront commitment may need to be re-negotiated or re-tenanted.

Oracle has supply commitments of $95.2 billion in total, which reflects just how deep into the hardware cycle the company has committed. That figure is not a backlog of revenue. It is a forward obligation to pay for chips and infrastructure regardless of whether the end customers show up on schedule.

The Stargate situation illustrates a structural divide in AI infrastructure investing. Companies with diversified revenue streams and massive cash generation can treat a delayed or renegotiated contract as a rounding error. Companies whose growth thesis depends on a small number of anchor customers absorbing massive new capacity are in a different position entirely.

An investor who bought Oracle at the start of 2026 is sitting on a year-to-date decline of 23.25%, with shares trading around $149.77. That drop reflects the market repricing exactly the risk Bosa described: if anchor tenants redirect capacity to newer sites with newer chips, the revenue recognition timeline stretches further out while capex obligations remain fixed.

Despite the selloff, analysts maintain a consensus price target of $257.29 with 32 buy ratings, suggesting the long-term thesis remains intact. The bull case assumes Oracle successfully backfills capacity and holds its existing contracts — a meaningful execution hurdle given the chip cycle dynamics now in play.

Microsoft, by contrast, is down 15.56% year-to-date but carries a commercial RPO of $625 billion spread across Azure, Microsoft 365, and dozens of enterprise products. A single infrastructure contract renegotiation does not move the needle for Microsoft the way it does for Oracle.

The Stargate situation creates a clear framework for separating AI infrastructure companies by their ability to absorb timing mismatches between capex commitments and revenue recognition.

For Oracle specifically, the key metrics to track are not the RPO headline number, which will continue to look impressive, but rather the rate at which RPO converts into recognized revenue. IaaS revenue grew 68% year-over-year in Q2 FY2026 to $4.08 billion, which is the real-time signal of whether committed deals are actually flowing through. If that growth rate decelerates while capex stays elevated, free cash flow pressure will intensify.

NVIDIA reads this situation differently. OpenAI walking away from Blackwell-era capacity to wait for Vera Rubin means demand for next-generation chips is pulling forward. NVIDIA guided Q1 FY2027 revenue to approximately $78 billion, and the Vera Rubin cycle has not yet contributed meaningfully to that figure. The risk for NVIDIA investors is different: the stock has pulled back 3.49% year-to-date despite the earnings strength, reflecting broader market caution rather than any fundamental deterioration.

The single most important thing to understand from this story is that in a chip cycle this fast, the companies bearing the most infrastructure execution risk are those whose revenue base is too small to absorb a timing mismatch between when deals are signed and when hardware is ready to deploy. Oracle is building the right infrastructure for the right market. Whether it can do so fast enough, and with the right chips at the right moment, is the question the market is now pricing in.

Wall Street is pouring billions into AI, but most investors are buying the wrong stocks. The analyst who first identified NVIDIA as a buy back in 2010 — before its 28,000% run — has just pinpointed 10 new AI companies he believes could deliver outsized returns from here. One dominates a $100 billion equipment market. Another is solving the single biggest bottleneck holding back AI data centers. A third is a pure-play on an optical networking market set to quadruple. Most investors haven’t heard of half these names. Get the free list of all 10 stocks here.