Your graphics processing unit (GPU) landlord likely will not survive the decade as a standalone firm. Gartner expects more than 80% of neocloud providers — the firms renting out artificial intelligence (AI) compute power — to stop operating on their own by 2030 [1]. These sellers need $8.5 trillion in sales by 2030 to pay for the data centers they plan to build [1]. But Gartner thinks buyers can pay no more than $3.6 trillion [1].

The big picture:

I keep coming back to one line in the Gartner note. It changes how to look at an hourly GPU rate. It is not just a price, but a share of someone else’s debt and equity [1].

The math is harsh. Owners need $8.5 trillion of revenue by 2030 to pay for planned builds, debt interest, and a 12% return on invested capital (ROIC) [1]. With AI data center revenue capped at $3.6 trillion, the market faces a $4.9 trillion gap [1].

The clock makes it worse. Gartner does not expect that 12% return before the end of 2032 at the earliest [1]. That puts more than $4 trillion of asset value at risk of a write-down through 2032 [1]. The chips get old faster than the servers pay for themselves.

Spending on AI-optimized infrastructure as a service (IaaS) — rented server gear — will grow from $66 billion in 2027 to $174 billion in 2030 [1]. A Gartner analyst notes this is too low for most sellers to earn their money back before the current chips hit the scrap heap [1].

Money, not silicon, may now be the harder limit. Hock Tan of Broadcom and Lip-Bu Tan of Intel told a Singapore crowd this month that cash limits will cap AI growth [3]. Lip-Bu Tan said the other part is that “in the longer run, people want to see the ROI” [3]. Both men still expect demand to beat supply through 2029 [3].

Parts cost more now, too. Gartner data shows PC shipments fell 22.4% in the third quarter [4]. Memory chip costs jumped fivefold, forcing hardware makers to pass the pain along [4].

Meanwhile, buying the hardware outright is getting easier. The International Data Corporation (IDC) ran a study paid for by the Open Compute Project (OCP) Foundation and raised its 2029 forecast for open-standard gear to roughly $800 billion [2]. That is 2.7 times the $295 billion it had forecast for the same year a year earlier [2].

By 2030, this open-standard gear will hit $892 billion [2]. AI hardware makes up about $628 billion of that total [2].

My read: These two reports describe the same server rack from opposite ends. The physical gear is turning into a standard catalog item growing 42% a year [2]. But the business of renting it out cannot cover its own cost of money [1].

By the numbers

  • $8.5 trillion — Revenue owners need: Total required by 2030 to cover planned builds, debt interest, and a 12% ROIC [1].
  • $3.6 trillion — Revenue market pays: Gartner’s cap on AI sales traced to data centers, leaving a $4.9 trillion gap [1].
  • 80% — Operators losing independence: The share of neocloud firms Gartner expects to stop operating on their own by 2030, under cost pressure and without a distinct business model [1].
  • 2.7x — Open-gear forecast jump: IDC’s revised 2029 guess for OCP-backed equipment, from $295 billion to about $800 billion in a year [2].

What I’d watch:

What strikes me is how little of this money risk shows up on a rate card. A buyer comparing dollars per GPU-hour looks at a number that assumes the seller’s funding works out [1].

  • Who holds the risk: The write-down risk sits with whoever owns the gear [1]. A long-term rental moves that risk to the buyer only if the contract says so, making exit clauses key.
  • The shape of the merge: Gartner’s warning that operators will stop standing alone covers being bought as well as being shut [1]. Which path a seller takes decides if a rental deal survives on its original terms.
  • Open racks outside hyperscale: OCP is pushing its open-hardware designs into enterprise, regional, and edge sites [2]. If this holds, standard AI gear gets much easier to buy directly, skipping the rental firms.
  • Spending guidance: Gartner’s forecast only holds if providers keep spending at their current pace, so the next round of guidance from the largest cloud providers is the test [1].

The catch

The honest limit here is that this is a guess about planned spending, not a booked loss. Gartner warns sellers “could be exposed to impairment” only if they keep up their current spending pace [1]. The largest cloud providers have other cash streams that can float weak AI returns for years [1].

I could be wrong about what this means for a buyer. Renting still keeps the leftover hardware risk off your own books.

Kevin Goldsmith makes this exact point in a recent essay [5]. AI “collapsed the first cost of building,” he writes, but “the cost of having made it” did not change [5]. His point applies perfectly to hardware: It all comes down to what you are willing to own [5].

The part I keep circling is the gap in time horizons. An engineering leader plans two to three years out, while the write-down risk runs through 2032. Between those dates sits your contract, and its terms decide who eats the gap.

At a glance

  • The Big Shift: Gartner predicts more than 80% of neocloud providers will stop operating on their own by 2030, because AI data center owners need $8.5 trillion to cover their building plans and the market can only pay $3.6 trillion [1].
  • Why It Matters: A rented GPU-hour is priced on a seller’s funding, meaning the buyer’s real risk is the seller’s balance sheet. At the same time, standard open-hardware sales are surging, offering a direct way to buy the same racks [1][2].
  • What I’d Watch: How money risk gets split between buyer and seller as weaker operators get bought or shut down [1].
  • Exit terms: Whether long-term deals let you leave early, which decides who holds the bag when a seller changes hands [1].
  • The shape of the merge: Being bought keeps a service agreement alive; being shut does not [1].
  • Open racks outside hyperscale: Whether OCP-backed hardware keeps taking market share in regional sites, making direct ownership a real choice [2].
  • The Catch: The write-down figure is a conditional guess, not a booked loss, and the biggest hyperscalers can float weak returns for years [1]. Renting still keeps aging hardware off your books, making this a core choice about what you actually want to own [5].

Related reading

Sources

[1] Gartner, “Gartner Says Over $3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by the End of 2030,” October 14, 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030 [2] IDC and the Open Compute Project Foundation, “IDC Raises 2029 Forecast for OCP-Recognized Infrastructure Spending by 2.7x to $800 Billion,” October 13, 2026 — https://www.prnewswire.co.uk/news-releases/idc-raises-2029-forecast-for-ocp-recognized-infrastructure-spending-by-2-7x-to-800-billion-302906503.html [3] Quartz, “Intel and Broadcom CEOs warn capital may bottleneck AI growth,” October 12, 2026 — https://qz.com/intel-broadcom-ceos-capital-constraint-ai-growth-101226 [4] Gartner, “Gartner Says Worldwide PC Shipments Declined 22.4% in Third Quarter of 2026 As Component Cost Inflation Raised ASPs,” October 12, 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-10-12-gartner-says-worldwide-pc-shipments-declined-22-point-4-percent-in-third-quarter-of-2026-as-component-cost-inflation-raised-asps [5] Kevin Goldsmith, “Build vs Buy When Building Just Got Cheap,” October 16, 2026 — https://kevingoldsmith.substack.com/p/build-vs-buy-when-building-just-got