IBM Chairman and CEO Arvind Krishna did not say that every AI data center is guaranteed to lose money. In a December 2025 appearance on The Verge’s Decoder podcast, he used rough “napkin math” to question whether the current, infrastructure-heavy race toward artificial general intelligence can earn an adequate return at today’s costs.
Krishna estimated that a fully built and equipped 1-gigawatt AI data center could cost about $80 billion. Applied to roughly 100 gigawatts of announced or discussed capacity, that implies approximately $8 trillion in capital spending. He then used an annual financing burden of about 10%—or $800 billion—to illustrate the scale of profit that would be needed merely to cover interest or the broader cost of capital. Those figures are estimates, not an audited industry total or an IBM forecast.
As an Amazon Associate I earn from qualifying purchases.
What Arvind Krishna actually argued
Krishna’s comments were about the economics of infrastructure being built in pursuit of increasingly capable AI systems, particularly an artificial-general-intelligence race requiring enormous computing clusters. Reports on the podcast discussion attribute the following rough framework to him:
- A 1-gigawatt AI data center could cost approximately $80 billion when the facility and computing equipment are included.
- Roughly 100 gigawatts of announced or discussed capacity would imply about $8 trillion of capital expenditure.
- A 10% annual financing or cost-of-capital assumption on $8 trillion would equal about $800 billion per year.
Krishna’s conclusion—reported as “there’s no way you’re going to get a return on that”—was conditional on current costs and assumptions. It is best understood as a warning about the scale, timing and financing of the infrastructure race, not as proof that AI applications or every data-center project are unprofitable.
#1 Best Overall
The remarks were reported by Tom’s Hardware and Data Center Dynamics. The available coverage attributes the figures to Decoder; it does not turn them into a formal IBM projection.
How the $8 trillion calculation works
| Assumption | Illustrative figure | What it means |
|---|---|---|
| Cost per 1 GW | $80 billion | Krishna’s approximate estimate for construction plus IT equipment |
| Capacity discussed | About 100 GW | Announced or discussed capacity, not necessarily operating capacity |
| Implied capital spending | $8 trillion | 100 × $80 billion |
| Annual financing burden | About $800 billion | Roughly 10% of the implied capital base |
The arithmetic is straightforward: 100 multiplied by $80 billion equals $8 trillion, and 10% of $8 trillion is $800 billion. The uncertainty lies in every major input.
The $80 billion figure is not a universal price for a 1-GW data center. A gigawatt measures power capacity, not a standardized facility size or cost. A project’s economics depend on whether the figure refers to total facility power or IT load, how much capacity is actually commissioned, and whether the campus is designed mainly for training, inference or a mixture of workloads.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Costs may include land, buildings, substations, transmission upgrades, backup power, cooling systems, networking, storage, servers and accelerators. They may also be counted at different stages, creating a risk of comparing a planned power envelope with a fully equipped campus.
Likewise, approximately 100 GW of announcements is not the same as 100 GW that has been financed, built, powered and filled with operating equipment. Projects can be delayed, redesigned, canceled or measured using different definitions. The resulting $8 trillion is therefore an extrapolation from assumptions—not a confirmed total of industry spending.
Why profitability is difficult at this scale
Accelerators can lose economic value before they fail
AI chips may remain physically usable for years while becoming economically less competitive much sooner. A newer accelerator can deliver more performance per watt or a lower cost per token, reducing the value of older equipment even when it still works.
That creates a difficult depreciation problem. An operator must recover the cost of an accelerator through revenue before a newer generation makes it unattractive to customers. Hardware replacement can also become a recurring capital requirement rather than a one-time investment.
Free tools Windows power users keep installed
One-click scans. No signup required.
High utilization is essential
A data center carries large fixed costs even when its equipment is idle. Operators need sustained utilization to spread construction, power, cooling, staffing and financing costs across enough billable compute.
Utilization is not guaranteed. Training demand can be episodic, with large projects followed by quieter periods. Inference demand may vary by time, customer and application. Customers may reserve capacity but use only part of it, while model efficiency improvements can reduce the amount of compute required for a given result.
Long-term contracts can make revenue more predictable, but concentration creates another risk: a facility dependent on a small number of customers may be exposed if those customers cancel, renegotiate or shift to another model or provider.
Power and cooling are part of the investment
The economics depend on obtaining reliable electricity at a manageable cost. A project may require utility interconnection, substations, transmission upgrades, backup generation and long-term power contracts. Electricity prices can vary, and local permitting or environmental rules can affect both construction schedules and operating costs.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
High-density AI systems also require sophisticated cooling. Depending on the design, operators may face costs for liquid-cooling equipment, water management, maintenance and specialized facilities. A site with cheap theoretical power is not necessarily economical if it cannot secure grid capacity or operate reliably at the required density.
Revenue may accrue to different parts of the AI economy
AI-related revenue is distributed across the stack. Chip makers, cloud providers, model developers, data-center landlords, power and equipment suppliers, software companies and end customers may all benefit from the same buildout. Growth in one layer does not automatically mean that every layer earns an attractive return on invested capital.
A cloud provider may generate revenue from compute, model APIs, databases, security, productivity software and enterprise contracts. A model developer may pay for infrastructure while capturing a different portion of customer revenue. A data-center owner may receive a stable lease payment but not participate fully in the upside of the AI applications operating inside the facility.
Interest is not the same as total profitability
Krishna’s $800 billion figure is best treated as an illustrative financing calculation. It is not necessarily based on a disclosed loan with a 10% coupon.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallA complete investment model would distinguish among:
- Debt interest: cash payments to lenders.
- Equity return requirements: the return investors expect for taking risk.
- Depreciation: the accounting and economic cost of equipment and facilities wearing out or becoming obsolete.
- Operating expenses: power, cooling, maintenance, staff, networking, insurance and administration.
- Taxes and fees: obligations that reduce cash available to investors.
- Replacement capital: spending required to refresh accelerators and other equipment.
A lower interest rate would reduce the annual financing burden, but it would not eliminate depreciation, energy costs or hardware replacement. Conversely, a high equity hurdle rate could require a return greater than the interest-only figure.
Why companies may keep spending anyway
Weak standalone data-center margins do not necessarily make the investment irrational. A hyperscaler may accept modest direct returns from infrastructure if the capacity supports more valuable businesses elsewhere.
Strategic reasons to continue spending include:
- Securing scarce chips, power and data-center capacity before competitors do.
- Protecting access to proprietary models and infrastructure.
- Building enough capacity for future inference demand.
- Using AI infrastructure to sell cloud, software, advertising, database and security services.
- Preventing a rival from gaining a technological or distribution advantage.
- Supporting national-security, sovereign-computing or strategic-industry goals.
- Expecting lower cost per computation as hardware, software and model techniques improve.
This is why “can the facility earn a return?” and “can the company earn a return on the ecosystem it enables?” are different questions. A company may tolerate low direct infrastructure margins if the investment improves customer retention, cloud usage or software sales.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →The bullish case: what could improve the economics
Krishna’s warning is not a fixed outcome. The economics could improve if the industry achieves several changes at once:
- More efficient models: Quantization, sparsity, distillation and better architectures can reduce the compute required for a task.
- Higher utilization: Better workload scheduling and a broader mix of training and inference can keep equipment busy.
- Specialized accelerators: Purpose-built systems may lower energy use and cost per unit of output.
- Longer useful lives: Older chips may remain valuable for less demanding inference and enterprise workloads.
- Higher-value applications: Businesses may pay premium prices for reliable AI that produces measurable savings or revenue.
- Improved power economics: Cheaper electricity, better power density and more efficient cooling can reduce operating costs.
- Falling hardware prices: Increased supply and competition could lower the cost of future capacity.
- Facility flexibility: Some sites may support conventional cloud workloads if frontier-model demand slows.
The central bullish assumption is that AI revenue and utilization will eventually catch up with infrastructure spending. That could happen through higher prices, more customers, more intensive usage, or major efficiency gains. It is an investment thesis, not an established fact.
What IBM’s position adds to the debate
IBM is not a neutral observer. It sells infrastructure, software and services and competes in enterprise technology, while its strategy emphasizes hybrid cloud, AI, infrastructure and consulting. IBM’s 2025 annual report presents AI as a major strategic platform, with an emphasis on enterprise deployment, governance and cost effectiveness.
That means Krishna’s comments should be read both as an economic argument and as a statement from a company with its own commercial interests. IBM may benefit if customers favor efficient enterprise AI and hybrid deployments rather than a winner-take-all race for the largest possible training clusters. That incentive does not make the arithmetic wrong, but it is relevant context when weighing the warning.
IBM has also described AI infrastructure as incremental to existing server demand rather than a simple replacement for all other technology spending. Its second-quarter 2025 infrastructure commentary described pressure shifting among hardware categories as customers adjusted their investments.
What IBM’s July 2026 results showed
Later IBM updates provide context, but they do not prove or disprove Krishna’s original calculation. In a July 14, 2026 investor update, IBM said customers had shifted some quarterly capital spending toward servers, storage and memory. The company reported second-quarter revenue of $17.2 billion and said infrastructure revenue fell 7%.
IBM’s finalized second-quarter results continued to describe AI as a structural business shift while highlighting the investment requirements and risks. IBM also pointed to strength in parts of its distributed-infrastructure business.
The mixed picture matters. Demand for infrastructure can remain strong while spending shifts away from other technology purchases or puts pressure on vendors exposed to older product cycles. Strong infrastructure demand is not identical to high returns for every owner, supplier or customer.
Is Krishna saying AI is a bubble?
Not necessarily. His argument is narrower than a claim that AI has no commercial value.
These are separate questions:
- Can AI applications generate useful revenue or savings?
- Can individual AI companies earn sustainable margins?
- Can cloud providers profit from inference?
- Can hyperscalers earn an adequate return on infrastructure?
- Can an AGI race justify ever-larger training clusters?
A positive answer to the first question does not guarantee a positive answer to the last. Enterprise AI, smaller models, specialized systems and AI-enabled software may be profitable even if an infrastructure race based on continual increases in scale produces disappointing returns.
What investors and business owners should watch
The most useful indicators are not headline spending totals alone. They include:
- AI-cloud utilization and whether customers actually consume reserved capacity.
- Revenue per accelerator, per rack or per megawatt.
- Hyperscaler capital expenditure compared with operating cash flow and free cash flow.
- Useful-life assumptions and evidence of accelerated hardware depreciation.
- Data-center leasing rates and the length and quality of customer contracts.
- Power prices, interconnection delays and cooling costs.
- AI application revenue, customer retention and willingness to pay.
- Delayed, canceled or repurposed projects.
- Evidence that newer models deliver more output with materially less compute.
For a personal investor, rising AI revenue or chip demand should not automatically be treated as proof of attractive returns across the industry. The relevant question is whether companies are generating enough incremental cash flow to justify the capital being deployed.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteWhat this means for companies that need AI compute
The infrastructure debate also has a practical implication: most businesses do not need to build a data center. They can use public cloud, dedicated hosting, colocation, on-premises systems or a hybrid model.
- Public cloud: Low upfront commitment and fast access, but potentially high long-term costs for predictable, intensive workloads.
- Dedicated hosting or colocation: More control and potentially better economics for stable demand, with longer contracts and greater management requirements.
- On-premises clusters: Maximum control and possible advantages at scale, but the customer absorbs hardware-obsolescence, power, cooling and staffing risks.
- Smaller or specialized models: Often a better match for measurable enterprise tasks than frontier-model training.
- Hybrid deployment: Keeps sensitive or predictable workloads close to the business while using cloud capacity for bursts.
Cloud GPU and accelerator services from providers such as Amazon Web Services, Microsoft Azure, Google Cloud, IBM Cloud and NVIDIA DGX Cloud avoid the upfront construction cost. They do not eliminate the underlying economics: sustained, intensive usage can make rented compute more expensive than owned infrastructure over time.
Pricing varies by accelerator, region, reservation term, storage, networking, data transfer and support. Individuals and small businesses should not interpret Krishna’s comments as a reason to purchase enterprise GPU infrastructure. A cloud GPU may be a poor fit when usage is sporadic, an existing model API is sufficient, or data-transfer costs dominate.
The fairest interpretation of the warning
“The AI data-center industry will lose money” is too broad a conclusion from Krishna’s remarks. The more defensible interpretation is that the current buildout may be overextended if companies assume that enormous infrastructure spending will automatically be recovered through future AI revenue.
Krishna’s calculation highlights a genuine tension: the more capital-intensive the buildout becomes, the more utilization, pricing power, efficiency and application revenue must improve. Companies can rationally spend ahead of demand for strategic reasons, but strategic value is not the same as a demonstrated financial return.
The $8 trillion figure should therefore be read as a warning scenario, not a verified bill. Whether the buildout becomes profitable will depend on how much capacity is actually built, how quickly chips become obsolete, how reliably facilities run, what customers pay, and whether AI services create enough economic value to support the infrastructure beneath them.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
