Gartner: $3.5 Trillion in AI Infrastructure at Risk of Depreciation by 2030

Gartner: $3.5 Trillion in AI Infrastructure at Risk of Write-Down by 2030

🔎 Gartner just did the math nobody wanted to do

On October 14, 2026, Gartner published a statement of a kind rarely seen in recent tech history: more than $3.5 trillion in AI infrastructure spending could be exposed to a write-down by the end of 2030, if the current pace of capex continues. This isn't a blogger chasing traffic — it's the firm that CIOs around the world pay to arbitrate their budgets.

The market didn't wait for the nuances. CoreWeave plunged as much as −17% in Wednesday's session, followed by Nebius, IREN, DigitalOcean, and much of the neocloud sector. Bloomberg picked up the alert the very next day under a blunt headline: "AI Race Puts $3.5 Trillion at Write-Down Risk, Gartner Warns".

Why is this warning coming out now? Because AI capex is set to reach nearly $1 trillion in 2026, with an additional $6 trillion projected between 2027 and 2030. Gartner asked the question everyone is avoiding: how much will these data centers have to generate to justify what they cost? The answer is a number no one will be able to look away from.

Key takeaways

• More than $3.5T exposed: AI infrastructure spending could be written down by the end of 2030, according to the Gartner note of October 14, 2026 — a projection, not a fact.
• $1T in 2026, $6T after that: Will Sommer (Senior Director Analyst, Gartner) calls the current capex pace "unsustainable".
• $8.5T in revenue needed vs. $3.6T expected: that cumulative effort would be required to cover capex, financing costs, and a 12% target ROIC.
• More than $4T in assets exposed to write-downs through 2032; the 12% ROIC would not be reached before the end of 2032 at the earliest.
• More than 80% of neoclouds would cease operating independently by 2030 — through acquisitions, partnerships, or bankruptcies.
• Market: CRWV down as much as −17% intraday, DigitalOcean −14%, Hut 8 −10%; CoreWeave's unsecured bonds in distress territory.
• Pushback: Mike Intrator (CoreWeave) calls the market "irrational", Ed Zitron sees it as validation, and optimists point out that AI technology itself is not the problem.

Recommended tools

Following this story requires monitoring stock prices, bonds, and analyst notes in parallel. Here are the tools I use for this type of monitoring:

Gartner's numbers: $8.5 trillion in revenue required, $3.6 trillion at the ceiling

No, with the current assumptions, the equation doesn't add up — and that's the entire point of the note. Will Sommer puts numbers on the mechanics: nearly $1 trillion in AI capex in 2026, then around $6 trillion cumulative between 2027 and 2030. That works out to roughly $7 trillion committed over five years.

For these investments to be profitable, Gartner calculates that $8.5 trillion in cumulative revenue would be required. That amount covers three line items: paying back the capex itself, financing costs, and a return on capital at the hurdle rate used — a 12% ROIC.

That's where the problem lies: enterprise and consumer revenue attributable to AI data centers would top out at $3.6 trillion. The gap exceeds $4.9 trillion. This isn't a matter of a few margin points; it's a chasm between what the machine costs and what it brings in.

"The lines don't cross until, at the earliest, the 2030s": the revenue and capex curves don't cross until the 2030s at the earliest, sums up Will Sommer via Fierce Network. In other words, the current generation of data centers risks becoming technologically obsolete before it is fully depreciated on the books.

A detail that often goes unmentioned makes the picture even more damning: more than half of AI revenue will go to providers with no significant infrastructure. Those who build the data centers, then, don't capture the majority of the value created by AI. All the risk on one side, a minority of the returns on the other.

My take: this argument is purely an accounting one, and that's precisely what makes it solid. This isn't a debate about the intelligence of the models — Gemini 3.1 Pro, GPT-5.5 and Claude Opus 4.7 keep improving quarter after quarter. The debate is about who pays for the steel, the concrete and the chips, and who pockets the revenue. On that question, arithmetic doesn't negotiate.

More than 80% of neoclouds would cease to exist as independents by 2030

Yes, Gartner projects that more than 80% of neoclouds will stop operating independently by 2030. An important clarification: "ceasing independent operations" can mean an acquisition, a partnership, or outright bankruptcy, notes Will Sommer via TechTarget. The watchword is not "disappearance" — it's "consolidation".

The sector has experienced what the analyst calls a "Cambrian explosion" of neoclouds — CoreWeave, Nebius, IREN and dozens of others — drawn in by GPU demand from overwhelmed hyperscalers and by euphoric valuations. This proliferation was a feature of the bull cycle. It becomes a breaking point when financing tightens.

Consolidation will leave "no more than a few" compute and services providers, according to Gartner. The AI-optimized IaaS market will indeed grow from $66 billion in 2027 to $174 billion in 2030 — real growth — but it would be insufficient to amortize the hardware before the current generation of chips is replaced.

This is exactly the scenario Groq anticipated when raising $650 million to pivot to the neocloud model. In a market that is closing up, either you reach critical mass, or you become a footnote. The window to choose is closing before our eyes.

The question TechTarget recommends CIOs ask is worth its weight in gold: how are these neoclouds capitalized? A provider backed by solid equity presents nothing like the same counterparty risk as a player financed by high-yield debt.

CoreWeave, Nebius, IREN: The Market Responded in a Single Session

Yes, the neocloud stocks plunged right at publication, proof that the market considered the note credible. According to Yahoo Finance/IBD, CoreWeave lost as much as 17% intraday on Wednesday — an "event"-type move, not an ordinary correction.

The detail, figure by figure (via TradingView/SeekingAlpha): CRWV −15% by midday, DigitalOcean −14%, Hut 8 −10%, NBIS and IREN −9%, Cipher and Applied Digital −8.7%. This isn't one stock faltering — it's an entire sector repricing within a few hours.

The most worrying signal doesn't come from the stocks but from CoreWeave's unsecured bonds, which entered distress territory. Comments from Chamath Palihapitiya and Elon Musk about the sector's debt added pressure to an already tense day. CoreWeave's valuation still stands above $36.5 billion: the market isn't saying the company is worth zero, it's saying its model is fragile.

The following night, the slide continued at a slower pace: CRWV −4%, NBIS −2.5%, IREN −1.5%, with retail investors, according to Yahoo Finance, eyeing the dip to buy. For the year 2026, the picture remains mixed anyway: CRWV −11%, IREN −22%, but NBIS +120%.

The reaction from the man concerned: Mike Intrator, CoreWeave's CEO, called the market reaction "irrational" in an internal memo seen by Bloomberg. It's also worth noting that higher-than-expected inflation was weighing on all of tech that day. Possible. But when a CEO calls the market irrational on the very day a note quantifies the fragility of his financing model, the simplest explanation is often the right one.

A Projection, Not a Finding: The Three Assumptions You Need to Know

None of this has happened yet — this is a Gartner projection based on explicit assumptions, not a record of bankruptcies. Three assumptions underpin the calculation, and you need to know them before declaring a crash or dismissing the warning.

First assumption: the 12% ROIC. It's a hurdle rate, a return requirement chosen by Gartner. If investors settle for 8%, the gap mechanically narrows. If the cost of capital keeps climbing with inflation, it widens. The verdict depends in part on a parameter no one controls.

Second assumption: the $3.6 trillion revenue cap. Gartner assumes that enterprise and consumer demand attributable to AI data centers will not take off beyond that level. The optimists bet the opposite: every model generation — GPT-5.5, Gemini 3 Pro Deep Think, DeepSeek V4 Pro — unlocks use cases that didn't exist twelve months earlier.

Third assumption: conditionality. The very title of the press release says "could be exposed" — might be exposed. And a write-down is not a liquidation: it's an accounting impairment. Microsoft or Google can write off assets without ceasing to exist; a heavily indebted neocloud, far less so.

Gartner also accompanies its press release with a note whose title speaks for itself: "Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032". More than $4 trillion in assets exposed to write-downs through 2032, and a 12% ROIC that wouldn't be reached before the end of 2032 at the earliest.

In the interest of transparency: Gartner sells consulting, and an alarmist note helps sell audits. But the figures in this note are arithmetic, not ideology — and to date, no one has published a rebuttal calculation as thorough. The burden of proof now lies with the vendors.

Why capex is exploding anyway: the global race for data centers

Because all the major players have reached the same conclusion: compute capacity is the limiting factor in the race for models. No GPUs, no training, no inference, no revenue. This conviction isn't irrational — it's simply shared by too many players at the same time.

The examples are piling up at a pace that's hard to keep up with. Microsoft, Meta and Google climb to $725 billion in AI capex — profitable companies that finance part of their investments through their cash flows, which changes everything compared to an indebted neocloud.

China is accelerating on its side: ByteDance is preparing $70 billion in AI capex for 2026, while Moonshot AI raises $2 billion to push Kimi K2.6, which dominates the open-weight rankings. The race is no longer just American, it's geopolitical.

Even space is becoming a server room: SpaceX and Reflection AI have sealed a $6.3 billion deal to turn Colossus 2 into a kind of AWS for AI. When Musk locks up orbital infrastructure, it's because everyone believes in compute scarcity — and no one wants to be the one who stops first.

This is where Gartner's reading becomes interesting: this logic holds for one or two players capturing the majority of revenue. It doesn't hold for an entire industry building the same capacity at the same time. If more than half of AI revenue goes to players without significant infrastructure, the data center race looks like a war of attrition whose winners have already been designated — and they're not necessarily the ones building the most.

The counterpoint: Ed Zitron, Krishna Rao, and those who reject the crash scenario

No, not everyone endorses the Gartner diagnosis — and the debate is more open than the headlines suggest. Ed Zitron, a longtime critic of the AI economy, sees it as a resounding validation of what he has been writing for years about the capex arms race and the absence of measurable return on investment.

On the other side, the optimists are not naive. Anthropic's CIO, Krishna Rao, has argued that their capex is not a bubble — the central argument being that inference demand is growing faster than available capacity, and that today's machines will be saturated by tomorrow's use cases.

An analysis relayed by ComputerWeekly offers the most useful nuance in this debate: one must distinguish the sustainability of the capital-intensive model from the AI technology itself. The models keep improving — Gemini 3.1 Pro (92), GPT-5.5 (91), Claude Opus 4.7 (90) on aggregate leaderboards — and usage demand is real. What is fragile is the financing structure, not the technology.

And the market context remains favorable: the AI bull market celebrates its fourth year, with Q3 expected at +31% even as strategists evoke overheating. You can have a growing sector AND structural overcapacity — telecoms in 2000 demonstrated this painfully.

My position: both camps are right, but about different things. AI demand will exist in 2030 — that's almost certain. Gartner's question is different: will the $8.5 trillion in revenue needed actually reach those who built the data centers? On this specific point, doubt is not only legitimate, but quantified.

The telecom precedent: what 2000 teaches us about 2026

Yes, this story has been told before — and it ended both ways at once. In the late 1990s, telecom operators dug up the ground to lay millions of kilometers of debt-financed fiber, betting on Internet demand that didn't yet exist at that scale.

The result: cascading bankruptcies, massive write-downs, accounting scandals. But also a frequently forgotten fact: the fiber laid at a loss in 2000 became the cheap infrastructure on which YouTube, Netflix, and the cloud built their empires in 2005-2010. The technology won. Many investors lost everything.

This is exactly the distinction made in the analysis relayed by ComputerWeekly: AI technology is not threatened by the Gartner note — it's the capital-intensive financing model that is. A $4 trillion write-down by 2032 wouldn't stop the models from improving — it would simply change who owns the machines and at what price they run.

The difference from 2000, and it's a major one: today's hyperscalers fund much of their capex from their cash flows, not from debt. The systemic risk is concentrated elsewhere — on the over-leveraged neoclouds and their creditors. That's where the Gartner note should be read with the most attention.

What This Means for CIOs and Compute Buyers

If you're buying GPU capacity, the Gartner note is a signal to cover your bases, not to change your strategy overnight. Three concrete actions are needed before your next contract renewal.

First, ask your vendors TechTarget's question: how are you capitalized? A neocloud backed by a hyperscaler or with solid equity doesn't carry the same risk as a player whose bonds are trading in distress territory. Demand the financial structure, not just the price per GPU-hour.

Next, eliminate vendor lock-in on a single compute provider. Multi-vendor architecture, data portability, service continuity clauses: if 80% of neoclouds change hands by 2030, your 2027 provider may go by a different name in 2030 — or belong to your competitor.

Finally, monitor the bond market as closely as the stock market. CoreWeave's bonds signaled the tension before most analyst commentary. Credit tells the truth before equities do, almost always.

A relatively good note to end on: the service isn't disappearing. AI-optimized IaaS grows from $66B (2027) to $174B (2030) according to Gartner. The demand will be there — it's the providers that will change names, not the needs.

❌ Common Mistakes

Mistake 1: Confusing write-downs with bankruptcy

A write-down is an accounting impairment of assets — servers, data centers — whose economic value has collapsed. A company can write off billions on its balance sheet and keep operating normally. The solution: read Gartner's figures as a risk of value destruction for shareholders and creditors, not as a list of bankruptcy filings.

Mistake 2: Taking the projection as prophecy — or dismissing it out of hand

"Could be exposed" is conditional: everything depends on the required ROIC, the cost of capital, and the demand trajectory. Ignoring the note on the grounds that "Gartner is often wrong" would be just as unrigorous as taking it as gospel. The solution: track the indicators that will settle the question — neocloud bonds, quarterly capex, published AI revenues.

Mistake 3: Building your entire business on a single neocloud

If more than 80% of neoclouds cease operating independently by 2030, your provider statistically has a four-in-five chance of changing hands or changing business model. The solution: multi-provider architecture, continuously exportable data, and contracts with continuity and exit clauses.

Mistake 4: Confusing criticism of capex with criticism of AI

This is the key point of the ComputerWeekly analysis: one can judge the financing unsustainable and still deeply believe in the technology. The solution: keep the two debates separate. Models will keep improving even if over-financed data centers consolidate — and that is probably what will happen.

❓ Frequently Asked Questions

What exactly does "write-down" mean in the Gartner note?

It is the accounting recognition of an asset's loss of value. A data center built for $10 billion can be written down well below that if the revenue it generates no longer covers its cost of financing. Gartner estimates that more than $4 trillion of AI assets are exposed to this risk through 2032.

Where do the $8.5 trillion in required revenues come from?

That is the cumulative amount required to cover the announced capex (nearly $1 trillion in 2026, then $6 trillion between 2027 and 2030), financing costs, and a 12% ROIC. Yet enterprise and consumer revenues attributable to AI data centers would cap out at $3.6 trillion, according to Gartner.

What are neoclouds?

AI compute providers that emerged after the hyperscalers — CoreWeave, Nebius, IREN — specializing in large-scale GPU rental. Gartner projects that more than 80% of them will cease independent operations by 2030, through acquisition, partnership, or bankruptcy.

Should you sell CoreWeave, Nebius, and IREN?

I don't give investment advice. The sourced facts: CRWV −11%, IREN −22%, and NBIS +120% since the start of 2026, CoreWeave bonds in distress territory, valuation maintained above $36.5 billion. Read the note, watch the credit, and decide according to your time horizon.

When will the situation become clearer?

According to Gartner, the 12% ROIC would not be reached before late 2032 at the earliest, and neocloud consolidation would be complete by 2030. The coming quarters — announced capex, published AI revenues, bond spreads — will provide the first signals of validation or invalidation.

Does this note call AI itself into question?

No. That is the distinction made by the analysis relayed by ComputerWeekly: the technology is progressing and the demand exists. What Gartner considers unsustainable is the pace of capex and the financing structure that accompanies it — not the value of AI.

✅ Conclusion

Gartner has just put a number on what critics have been murmuring for the past two years: more than $3.5 trillion in AI infrastructure is exposed to depreciation by 2030 if the $8.5 trillion in required revenue fails to materialize — and today, the trajectory shows no sign of it appearing. The ball is now in the hyperscalers' and neoclouds' court: prove that demand is keeping pace, or prepare for consolidation. To follow what comes next — stock prices, bonds, upcoming notes — keep Yahoo Finance bookmarked, and if you're building your own AI business on the side, start by laying solid, low-cost foundations with Hostinger.