AI Capex Boom: OpenAI and Anthropic Concentration Risk

Last updated: August 2, 2026, 12:15 p.m. ET

The most important question raised by the latest round of Big Tech earnings is no longer whether demand for artificial intelligence exists. Microsoft, Alphabet and Amazon have each reported extraordinary cloud growth, expanding backlogs and continuing shortages of computing capacity. The harder question is whether the revenue supporting hundreds of billions of dollars in data-center investment is broad, durable and profitable enough to justify the scale of the buildout.

That is the issue behind a sharply skeptical argument made by Ed Zitron, chief executive of EZPR and publisher of the technology newsletter Where’s Your Ed At, in a Bloomberg interview that spread widely after the latest technology earnings reports. Zitron contended that investors have been encouraged to view AI capital expenditure as a response to diversified enterprise demand, when a material share of the incremental cloud business may instead be tied to two privately held frontier-model developers: OpenAI and Anthropic.

The public evidence supports part of that concern, but not its broadest possible interpretation. OpenAI and Anthropic are unquestionably enormous customers, strategic partners and investment holdings for the largest cloud providers. Their contractual commitments can materially affect bookings, remaining performance obligations, infrastructure plans and even reported net income. Analyst estimates cited in the Bloomberg discussion suggest that the two companies could represent a very large share of future Google Cloud and Amazon Web Services revenue. Yet those estimates are not company disclosures, and the latest filings also show substantial demand outside the frontier-model laboratories.

Microsoft offered the clearest rebuttal to the idea that its cloud business is simply a circular financing machine. On its July 29 earnings call, the company said that nearly 90% of full-year Microsoft Cloud revenue came from customers outside frontier-model companies. It also said that all sequential growth in commercial remaining performance obligations during the quarter came from customers outside those companies. At the same time, Microsoft separately disclosed that excluding OpenAI would reduce its commercial backlog growth from 84% to 25%, demonstrating that one customer can be both non-dominant in recognized revenue and extremely important to future contracted commitments.

Alphabet and Amazon disclosed broad enterprise momentum, rapidly rising cloud operating income and hundreds of billions of dollars in backlog. Neither company, however, provided a customer-level revenue breakdown that would allow outsiders to independently confirm or reject the concentration estimates cited by Zitron. The result is a debate with unusually high stakes and incomplete public data: the bullish case is supported by real revenue and profit growth, while the skeptical case is strengthened by the size of private AI laboratories’ compute commitments, their continuing need for external capital and the pressure that infrastructure spending is placing on free cash flow.

Key Takeaways

• The central issue: OpenAI and Anthropic are major cloud customers whose commitments influence Azure, AWS and Google Cloud, but the hyperscalers do not publish enough customer-level data to measure the exact concentration independently.
• Microsoft’s evidence: Microsoft said nearly 90% of full-year cloud revenue came from customers outside frontier-model companies, even though excluding OpenAI would reduce commercial remaining-performance-obligation growth from 84% to 25%.
• Alphabet’s scale: Google Cloud revenue rose 82% year over year to $24.8 billion in the second quarter of 2026, while Alphabet raised full-year capital-expenditure guidance to $195 billion to $205 billion.
• Amazon’s scale: AWS revenue rose 37% to $42.2 billion, but Amazon’s trailing-12-month free cash flow fell to an outflow of $7.6 billion as property-and-equipment spending accelerated.
• The private-company risk: OpenAI and Anthropic are growing rapidly, but both rely on massive capital commitments and still disclose far less than public companies would be required to disclose.
• The balanced conclusion: The AI infrastructure boom is not proven to be a fiction, but customer concentration, circular commercial relationships, capital intensity and limited transparency are legitimate risks that investors should not dismiss.

The AI Capex Concentration Debate in Plain English

Cloud companies normally prefer a broad customer base. A diversified portfolio reduces the damage caused by one customer cutting spending, renegotiating a contract, delaying a project or failing financially. The same principle applies to a bank’s loan book, a semiconductor company’s customer list or a commercial landlord’s tenant base. Concentration is not automatically bad—large customers can be highly profitable and creditworthy—but it raises the importance of understanding the customer’s finances and the enforceability of its commitments.

The AI market complicates that familiar analysis because the cloud providers are not merely vendors. Microsoft has invested heavily in OpenAI and has an extensive licensing and commercial relationship with it. Amazon and Google have invested in Anthropic while also selling it computing capacity. Microsoft reported a $3.2 billion quarterly gain from an Anthropic investment even as the company competes with Amazon and Google for Anthropic-related workloads. Amazon reported $53.4 billion of non-operating pre-tax income in the second quarter, primarily from its Anthropic holdings. Alphabet also recorded very large unrealized investment gains in the quarter, though it did not attribute the full amount to one company.

Those overlapping relationships create several distinct forms of exposure that are often blended together in public discussion. The first is recognized cloud revenue: the amount a customer actually generates in a reporting period. The second is bookings or contracted commitments: agreements that may convert into revenue over several years. The third is equity exposure: changes in the estimated value of a private AI company can flow through a public company’s income statement. The fourth is capacity planning: a hyperscaler may build data centers, acquire chips or sign power agreements based on demand forecasts that include one or two very large counterparties.

A company can have limited current revenue concentration but still carry substantial future commitment risk. It can also record a large investment gain that has no immediate cash benefit. Conversely, a large cloud contract does not necessarily mean the customer is weak: a rapidly growing private company may have substantial cash, committed financing and strong revenue growth even while reporting accounting losses. Sound analysis therefore requires separating revenue, backlog, investment gains, capital spending and cash flow rather than compressing them into a single claim that the system is either healthy or fraudulent.

What Ed Zitron Argued

Zitron’s argument has three main parts. First, he says investors have overestimated the breadth of AI demand supporting hyperscaler capital expenditure. He cited UBS estimates that OpenAI and Anthropic could account for 27% of Google Cloud revenue in 2026 and more than 48% in 2027. He also cited Barclays estimates that the same two customers could represent approximately 13% of AWS revenue in 2026 and 18% in 2027.

Second, he argues that the economics of the frontier laboratories remain dependent on outside financing because their operating costs and infrastructure commitments exceed internally generated cash. In his view, that makes cloud growth less durable than it appears: the cloud vendors are recognizing revenue from customers whose ability to pay ultimately depends on continuing access to venture capital, strategic investment, debt markets or public equity.

Third, Zitron contends that the infrastructure pipeline is too large to be justified by the current revenue base. He referred to research tracking approximately 190 gigawatts of announced data-center capacity and used a rough revenue-per-megawatt assumption to illustrate the level of annual sales that would be required to support that buildout. The underlying concern is not simply whether OpenAI and Anthropic survive. It is whether enough end-user demand emerges across enterprises, consumers and governments to keep a vast installed base of expensive accelerators, networking equipment and power infrastructure economically productive.

The interview’s most memorable phrase—“everyone has been sold a lie”—is stronger than the public evidence can establish. There is no proof that Microsoft, Amazon or Alphabet have systematically misrepresented recognized revenue, and their audited results show real cloud sales, operating profit and cash generation. The more defensible version of the critique is narrower: investor presentations emphasize broad AI opportunity, while the companies disclose too little customer-level information for outsiders to determine how much of the incremental infrastructure demand is concentrated in a handful of heavily financed private laboratories.

The Analyst Estimates Are Important, but They Are Not Audited Facts

The UBS and Barclays figures deserve attention because they are unusually large. If two customers were to represent nearly half of a cloud division’s revenue, the commercial and valuation implications would be substantial. Yet the figures should be described for what they are: analyst estimates based on private contracts, industry checks, capacity assumptions and forecasts. They are not numbers that Alphabet or Amazon has confirmed in a filing.

This distinction matters because cloud-accounting models are complex. A contract can include reserved capacity, equipment sales, infrastructure services, model distribution, minimum commitments, credits, equity-linked arrangements and payments spread across several years. Some agreements may be take-or-pay, while others may depend on delivery milestones or usage. Hardware sold into a customer-controlled data center is economically different from cloud capacity rented by the hour, even if both are reported within the same segment.

Alphabet’s second-quarter results provide an example. Google Cloud revenue rose 82% to $24.8 billion, but the company said the quarter included the first recognition of revenue from sales of Tensor Processing Unit systems delivered to customer data centers. Management also said cloud growth accelerated meaningfully even after excluding TPU system sales. That disclosure confirms that hardware was not the sole driver, but it does not reveal which customers generated the hardware revenue or how recurring it will be.

Amazon similarly said that both OpenAI and Anthropic had made multi-year, multi-gigawatt commitments involving its Trainium chips. That is direct evidence of concentrated frontier-lab demand. Amazon also said hundreds of thousands of customers use Bedrock and that spending on the platform in the second quarter exceeded all previous quarters combined. Both facts can be true: the two largest laboratories can dominate the most capital-intensive commitments while a much broader population of enterprises generates smaller but rapidly growing usage.

Microsoft’s disclosures are the most informative because the company explicitly presented figures both including and excluding OpenAI. Commercial remaining performance obligation reached $678 billion and grew 84% year over year. Excluding OpenAI, growth was 25%. Commercial bookings grew 10% as reported but 18% when excluding OpenAI, because the timing of very large OpenAI commitments distorted the comparison. These numbers reveal dependence without proving domination: OpenAI is large enough to reshape growth rates, but the underlying commercial business still expanded at a healthy pace.

Recognized Revenue Is Not the Same as Backlog

Remaining performance obligations, often shortened to RPO, represent contracted revenue that has not yet been recognized. They are valuable because they offer visibility into future sales, but they are not cash in the bank and are not identical to a legally unconditional receivable. Timing, implementation, capacity availability and contract provisions can affect when revenue appears. A multi-year commitment can inflate backlog today while contributing only gradually to quarterly revenue.

That difference helps explain how Microsoft can say nearly 90% of cloud revenue came from customers outside frontier-model companies while also showing that OpenAI accounts for a large portion of backlog growth. OpenAI’s influence is strongest in the forward-looking contract base, not necessarily in the revenue already recorded during the fiscal year. Investors evaluating concentration should therefore ask two separate questions: how much current revenue comes from frontier labs, and how much future infrastructure is being built in anticipation of their contractual demand?

Backlog quality also depends on the counterparty. A commitment from a profitable investment-grade company with stable cash flow carries a different risk profile from a commitment made by a private company that expects years of negative cash flow. That does not make the second agreement worthless. It does mean that the customer’s financing plan, access to capital and ability to adjust spending are central to the cloud provider’s risk analysis.

Private AI companies have so far demonstrated exceptional access to capital. They have raised sums that would have been unimaginable for software startups only a few years ago. They also command strategic importance for cloud vendors, chipmakers, sovereign investors and governments. The danger is not that financing must stop completely. It is that the cost or terms of financing could change before the infrastructure produces enough end-user revenue to fund itself.

The Circularity Question

“Circular financing” is often used loosely in the AI debate. At its broadest, it describes an ecosystem in which a technology company invests in an AI laboratory, the laboratory uses part of that capital or associated credits to buy cloud services from the investor, and the investor reports both cloud revenue and changes in the value of its equity stake. The arrangement can look self-reinforcing: investment supports customer spending, customer spending supports cloud growth, cloud growth supports valuation, and valuation supports more investment.

Not every circular relationship is economically artificial. Strategic investments are common in technology. A platform owner may finance an ecosystem partner because the partner’s growth creates genuine demand, expands the platform and produces valuable intellectual property. If the AI laboratory sells useful services to independent customers and eventually generates sustainable gross profit, the cloud spending reflects real economic activity even when the vendor is also an investor.

The problem appears when outside capital is the primary source of the customer’s purchasing power and the customer cannot generate enough cash from end users to sustain its commitments. In that situation, the cloud provider is partly financing demand for its own infrastructure. Revenue can still be recognized correctly under accounting rules, but its quality and durability become more dependent on continued financing.

The public companies’ accounting adds another layer. Amazon’s $53.4 billion quarterly non-operating gain, primarily from Anthropic, helped lift net income to $62.6 billion. That gain was not AWS operating profit and did not represent a cash payment from customers. Microsoft’s $3.2 billion Anthropic gain also benefited quarterly results. Alphabet’s large unrealized securities gains had a similar effect on consolidated net income. Readers comparing headline earnings must separate these valuation changes from the operating economics of cloud services.

Investment gains can reverse. Private-market valuations depend on financing rounds, transaction prices, accounting methodologies and market conditions. A higher valuation can improve reported earnings without improving free cash flow; a lower valuation can reduce earnings without necessarily weakening the core cloud business. This is why operating income, cash flow and segment results are more useful than consolidated net income when evaluating whether AI infrastructure is paying for itself.

What Would Prove the Skeptical Case?

The skeptical case would become stronger if several developments occurred together. OpenAI or Anthropic could slow cloud spending, seek to renegotiate commitments or raise capital on materially worse terms. Cloud backlogs could keep rising while recognized revenue growth decelerated. Hyperscalers could continue increasing capital expenditure even as utilization weakened. Depreciation, energy costs and interest expense could rise faster than cloud gross profit. Data-center projects could be delayed after equipment had already been ordered, or completed facilities could operate below expected utilization.

Another warning would be a widening gap between reported earnings and cash flow. Amazon’s second-quarter results already illustrate why this matters. Operating cash flow increased strongly, but trailing-12-month free cash flow moved to a $7.6 billion outflow because property-and-equipment spending accelerated. Alphabet reported negative free cash flow of $5.9 billion in the second quarter after $44.9 billion of capital expenditure. Microsoft remained free-cash-flow positive, but quarterly free cash flow of $19.6 billion reflected the burden of $41 billion in capital expenditure.

None of those figures proves that the spending will fail. Infrastructure businesses often invest before revenue arrives. The question is whether the future returns exceed the cost of capital and compensate for technological obsolescence, construction risk and customer concentration. If utilization, pricing and end-user demand remain strong, today’s cash-flow pressure can produce future earnings. If not, the industry may discover that it built too much capacity too quickly.

What Would Prove the Bullish Case?

The bullish case would be strengthened by broader, more transparent monetization. Cloud providers could disclose that growth is increasingly driven by thousands of enterprises rather than a handful of frontier laboratories. AI application revenue could rise faster than infrastructure spending. Microsoft 365 Copilot, GitHub Copilot, Google’s Gemini products, Amazon Bedrock and first-party AI services could produce recurring revenue with improving margins. Customers could move from experiments into production workloads that persist through economic cycles.

There is already evidence pointing in that direction. Microsoft said Microsoft 365 Copilot surpassed 30 million paid seats. Amazon said Bedrock has hundreds of thousands of customers, and its AI business exceeded a $25 billion annual revenue run rate. Alphabet said its cloud backlog reached $514 billion and that the majority related to typical Google Cloud contracts across a broad customer mix. Google also reported growth in Search, subscriptions and advertising products that use AI internally rather than merely renting compute to outside laboratories.

Productivity evidence is also real, though narrower than the market narrative sometimes implies. Research published through the National Bureau of Economic Research found that a generative-AI assistant increased customer-support productivity by approximately 14% on average, with larger gains for less experienced workers. Other field experiments have found improvements in writing, software development and service operations. These studies do not establish economy-wide transformation, but they show that generative AI can create measurable value in specific workflows.

The decisive evidence will come from cash economics rather than demonstrations or user counts. Are customers willing to pay prices that cover inference, training, support, data, safety and distribution costs? Does usage remain after trial subsidies end? Do companies reduce other expenses or increase revenue enough to justify subscriptions and token consumption? Can cloud providers raise utilization without sacrificing margins? Those questions will determine whether the buildout resembles a productive infrastructure cycle or an overextended capital boom.

Microsoft: The Best Public Window Into OpenAI Dependence

Microsoft’s fiscal 2026 fourth-quarter results provide the strongest available evidence for both sides of the concentration debate. The company reported $90.0 billion of quarterly revenue, up 18% year over year, and $40.6 billion of operating income, also up 18%. Azure and other cloud-services revenue increased 43%, while annual Azure revenue surpassed $100 billion for the first time. Microsoft Cloud revenue reached $59.3 billion for the quarter and $214 billion for the full fiscal year.

Those are not the results of a business whose customer demand has vanished. Azure’s growth accelerated despite its enormous scale, Microsoft continued to report that demand exceeded available capacity, and management said new capacity was quickly monetized. The company added 31 data centers during the quarter, bringing the fiscal-year total to 88, and said it added another gigawatt of capacity in the quarter. It also reported that throughput for Copilot workloads had increased fourfold since the start of the year through optimization across silicon, systems and software.

Yet Microsoft’s OpenAI disclosures show why investors cannot treat all cloud growth as equally diversified. Commercial remaining performance obligation rose 84% to $678 billion. Excluding OpenAI, it rose 25%. The difference is enormous. It means OpenAI-related commitments are large enough to account for most of the reported growth rate in the contracted backlog, even though they do not account for most recognized cloud revenue.

Commercial bookings present the mirror image. Reported bookings grew 10%, or 11% in constant currency, including Azure commitments from OpenAI. Excluding OpenAI, bookings grew 18%. The lower reported rate reflects the timing of prior-year OpenAI contracts and demonstrates how one customer can make quarterly growth comparisons unusually volatile. Microsoft warned that significant OpenAI contracts signed in the previous year would continue to create fluctuations in bookings and RPO growth.

The company also said all sequential commercial RPO growth came from customers outside frontier-model companies and that nearly 90% of full-year cloud revenue came from customers outside that group. That is an unusually direct answer to the claim that Azure is built almost entirely on OpenAI. It does not eliminate concentration risk, because the future commitment base remains heavily influenced by OpenAI, but it shows that the recognized cloud business has a much broader foundation.

Microsoft’s Capital Expenditure and Free Cash Flow

Microsoft spent $41 billion in capital expenditure during the quarter, including finance leases. Approximately two-thirds went to shorter-lived assets, primarily CPUs and GPUs, while the remainder went to longer-lived assets such as data-center infrastructure. Cash paid for property and equipment was $35.8 billion. Operating cash flow reached $55.4 billion, while free cash flow was $19.6 billion.

The composition of spending matters. Servers and accelerators depreciate faster than land and buildings, and their economic lives can be shortened by rapid chip improvements. A data-center shell may remain useful for decades, but the GPUs inside it can lose relative performance within a few product cycles. Heavy investment in short-lived equipment therefore creates a higher near-term hurdle: the company must generate sufficient revenue and gross profit before the assets become technologically less competitive.

Microsoft also announced that it would extend the estimated useful lives of data centers and office buildings from 15 to 25 years. The company said the change would have only a minimal benefit to fiscal 2027 operating income, but it would alter the classification of some future leases. More leases are expected to be treated as operating leases rather than finance leases, reducing reported capital expenditure without necessarily reducing the underlying economic commitment. Microsoft adjusted its calendar 2026 capital-expenditure expectation to approximately $175 billion largely because of that classification change.

This is a reminder that capex comparisons require accounting care. A lower reported number can reflect lease classification rather than a reduction in construction or equipment commitments. Investors should examine cash payments, lease obligations and total contracted infrastructure exposure—not only the headline capex figure.

OpenAI’s Role in Microsoft’s Financial Statements

Microsoft’s investment in OpenAI affects reported net income separately from Azure revenue. For fiscal 2026, the company said net gains from OpenAI investments increased net income by approximately $5.0 billion and diluted earnings per share by $0.67. In fiscal 2025, OpenAI-related losses reduced net income by $3.6 billion and earnings per share by $0.49. The swing illustrates how a private investment can amplify changes in consolidated earnings even when the core operating business remains strong.

The commercial relationship has also evolved. Under the revised partnership announced in April 2026, Microsoft remained OpenAI’s primary cloud partner, OpenAI products were to ship first on Azure unless Microsoft could not or chose not to support the required capabilities, and Microsoft retained a non-exclusive license to OpenAI intellectual property through 2032. OpenAI gained greater flexibility to serve products through other cloud providers, while Microsoft stopped paying a revenue share to OpenAI.

That revised structure reduces exclusivity while preserving deep interdependence. OpenAI can diversify its compute suppliers, which may lower Azure concentration over time but also exposes Microsoft to more competition for a customer it helped create. Microsoft, meanwhile, has invested in its own models, chips and orchestration tools and has expanded relationships with other model providers. The company increasingly presents Azure as a neutral platform on which enterprises can use multiple models rather than a single-vendor extension of OpenAI.

For Microsoft shareholders, the practical risk is not simply that OpenAI might fail. A less dramatic outcome could still matter: OpenAI could shift workloads to lower-cost providers, develop more efficient models that require less compute, negotiate better pricing, or reduce the pace of capacity reservations. Any of those outcomes could lower the revenue generated per dollar of infrastructure. Conversely, OpenAI’s growth could continue to drive Azure utilization while Microsoft monetizes AI through Copilot, GitHub, security, data and database services. The relationship creates upside and concentration at the same time.

Alphabet: Explosive Cloud Growth, Record Spending and Limited Customer Detail

Alphabet’s second-quarter 2026 results were among the strongest cloud numbers ever reported by a business of its size. Consolidated revenue rose 24% to $119.8 billion, while operating income increased 30% to $40.8 billion. Google Cloud revenue climbed 82% to $24.8 billion. Cloud operating income more than tripled to $8.8 billion, and the segment’s operating margin increased from 20.7% a year earlier to 35.6%.

Those results weaken any simple claim that Google’s AI infrastructure is producing only low-quality or loss-making revenue. A 35.6% segment operating margin is substantial, even though the figure includes a mix of cloud infrastructure, platform services, productivity software and hardware sales. The division is no longer merely pursuing scale at the expense of profitability.

Alphabet also reported that Google Cloud backlog rose by more than $50 billion sequentially to $514 billion. Management said the increase reflected strong demand for enterprise AI offerings and that the majority of backlog related to typical Google Cloud contracts across a broad mix of customers. Just over half of the backlog was expected to convert into revenue over the following 24 months.

This statement is relevant to the UBS estimate cited by Zitron. If OpenAI and Anthropic were on course to produce nearly half of Google Cloud revenue in 2027, their contracts would presumably account for a material share of future capacity and backlog. Alphabet’s description of the majority as “typical” contracts across a broad customer mix suggests wider diversification, but the wording is not precise enough to calculate concentration. It does not reveal the size of the largest customers, the distinction between cloud services and TPU system sales, or the timing of the largest commitments.

The New Importance of TPU System Sales

Alphabet said it began recognizing revenue from TPU systems sold into customer data centers for the first time in the second quarter. Cost of revenue increased partly because of inventory costs associated with those sales. Management emphasized that cloud growth still accelerated meaningfully after excluding TPU system revenue, but the new category complicates year-over-year comparisons.

A TPU system sale can produce a large amount of revenue when equipment is delivered. That revenue may be less recurring than ongoing cloud usage, though it can also lead to future software, support and networking demand. The economics may differ from renting access to Google-owned infrastructure. Investors therefore need more information about gross margin, customer concentration and the expected cadence of system deliveries.

Anthropic’s April 2026 announcement provides important context. The company said it had signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online beginning in 2027. Anthropic had already expanded its use of Google Cloud TPUs in 2025. OpenAI, despite competing directly with Google’s Gemini products, also added Google Cloud capacity in a deal first reported in 2025.

These agreements confirm that both laboratories are significant Google infrastructure customers. They do not confirm the precise UBS percentages, but they make the estimates plausible enough to warrant scrutiny. A multi-gigawatt commitment can translate into tens of billions of dollars over time, especially when it includes accelerators, networking, storage, power and managed services.

Alphabet’s $195 Billion to $205 Billion Capex Plan

Alphabet spent $44.9 billion on capital expenditure during the second quarter, with approximately 60% of technical-infrastructure investment allocated to servers and 40% to data centers and networking equipment. It raised full-year 2026 capital-expenditure guidance to between $195 billion and $205 billion, up from a previous range of $180 billion to $190 billion. Management said the increase reflected accelerated capacity delivery in response to demand.

The spending pushed quarterly free cash flow to an outflow of $5.9 billion even though operating cash flow was $39.1 billion. Trailing-12-month free cash flow remained positive at $53.3 billion. Alphabet ended the quarter with $242.5 billion in cash and marketable securities and $98.2 billion of long-term debt, giving it substantial financial capacity to continue investing.

Financial capacity is not the same as investment quality. Alphabet can afford the buildout, but shareholders still need to know whether the incremental return will exceed the company’s cost of capital and alternative uses of cash. The market’s initial negative reaction after the company raised capex guidance showed that strong revenue growth does not eliminate concern about spending discipline.

Alphabet’s advantage is that AI can improve several existing businesses in addition to generating cloud revenue. Search, advertising, YouTube, subscriptions and workplace products all benefit from AI models and infrastructure. Google reported that Search and other advertising revenue increased 17% to $63.3 billion, YouTube advertising revenue rose 13% to $11.1 billion, and subscriptions, platforms and devices revenue increased 15% to $12.9 billion. Some of the same infrastructure used for external cloud customers also supports internal products with large existing revenue streams.

That internal demand makes Google less dependent on frontier laboratories than a pure infrastructure provider would be. It also makes return measurement harder. If an AI model improves ad relevance or user retention, the economic value may appear as higher advertising revenue rather than a separately labeled AI sale. The absence of a clean AI-revenue line does not mean there is no return, but it gives management wide discretion in describing the benefits.

Why Google’s Customer Concentration May Look Different From Microsoft’s

Google Cloud is smaller than Microsoft Cloud or AWS, so one or two huge customers can represent a larger percentage of segment revenue. A $10 billion customer is important to any provider, but it has a much larger proportional effect on a $100 billion business than on a $200 billion business. Google’s newer TPU system-sales model may also create lumpy revenue recognition that amplifies concentration in particular quarters or years.

At the same time, Google’s proprietary TPU architecture offers differentiation. If customers believe TPUs provide better price-performance for certain workloads, Google can win business that would otherwise go to Nvidia-based clusters on AWS or Azure. Anthropic’s expanding TPU commitment suggests the hardware has become strategically important. The risk is that Google may need to build enormous capacity for customers that are simultaneously using AWS, Azure and other providers, limiting Google’s pricing power.

The strongest conclusion supported by public evidence is therefore not that Google Cloud is dependent on two customers to the extent claimed by UBS. It is that Google has signed large, multi-year infrastructure relationships with both OpenAI and Anthropic, while its disclosures do not allow investors to test the concentration estimate. That information gap is itself material when capital expenditure is approaching $200 billion a year.

Amazon: AWS Acceleration, Anthropic Exposure and Negative Free Cash Flow

Amazon’s second-quarter results gave the AI-capex bulls their strongest market reaction of earnings week. Net sales rose 20% to $200.6 billion, operating income increased 43% to $27.5 billion, and AWS revenue jumped 37% to $42.2 billion—the segment’s fastest growth in 18 quarters. AWS operating income rose to $16.6 billion from $10.2 billion, implying an operating margin of approximately 39%.

Amazon said AWS reached a $169 billion annualized revenue run rate and that its AI business and chips business each exceeded a $25 billion annual revenue run rate. It described AI demand as broad and said hundreds of thousands of customers use Bedrock. The company’s strong cloud growth and profitability helped shares rise sharply after the report, even though Amazon increased its 2026 capital-spending plan to approximately $220 billion.

The market reaction does not settle the concentration debate, but it demonstrates what investors are willing to accept when revenue acceleration is visible. High capex became less alarming because AWS growth exceeded expectations and operating margins expanded. Alphabet, by contrast, experienced a more cautious response after raising spending guidance despite similarly strong cloud results. The difference shows that markets evaluate capex in relation to near-term revenue momentum, margin performance and management credibility rather than applying one fixed threshold.

Anthropic’s Central Role at AWS

Anthropic has designated AWS as its primary training and cloud provider for mission-critical workloads. In April 2026, Anthropic announced an expanded agreement involving up to five gigawatts of additional compute. The company said Trainium2 capacity was coming online in the second quarter and scaled Trainium3 capacity was expected later in the year. It also planned to use additional capacity to expand Claude inference through Amazon Bedrock in Asia and Europe.

Amazon’s earnings release added that OpenAI and Anthropic had made multi-year, multi-gigawatt commitments involving Trainium. Those disclosures support the basic premise behind Barclays’ concentration estimate: the two laboratories are not ordinary customers. They are anchor tenants for Amazon’s custom-chip strategy and some of its largest infrastructure projects.

The exact revenue contribution remains undisclosed. Barclays estimates cited publicly suggested that Anthropic and OpenAI could represent approximately 13% of AWS revenue in 2026 and 18% in 2027. At AWS’s current scale, those percentages would imply tens of billions of dollars. Amazon did not confirm the figures, and the company’s broad customer claims suggest the remainder of the business remains highly diversified.

There is also a distinction between customer concentration and model concentration. Enterprises may access Anthropic or OpenAI models through Bedrock or other AWS services, causing end-user spending to appear as AWS revenue even when the underlying model provider receives part of the economics. A cloud provider can therefore have thousands of enterprise customers while still depending heavily on one model ecosystem. The public disclosures do not fully reveal how revenue and costs are shared across those layers.

Amazon’s Anthropic Investment Gain

Amazon’s second-quarter net income rose to $62.6 billion, but the figure included $53.4 billion of non-operating pre-tax income primarily related to Anthropic. That gain reflects an increase in the estimated value of Amazon’s investment, not a comparable increase in cash generated by retail or AWS operations. Excluding it is necessary for understanding recurring performance.

The gain also illustrates a feedback loop in private markets. Strong revenue growth, high-profile funding rounds and strategic cloud commitments can raise Anthropic’s valuation. That higher valuation benefits the reported earnings of investors such as Amazon and Microsoft. The public companies’ rising earnings can strengthen market confidence in AI, which in turn supports capital formation for the private laboratory. The loop can be economically justified if Anthropic’s sales and margins continue to improve, but valuation gains should not be confused with operating cash returns.

Amazon’s AWS operating income is the better measure of current infrastructure economics. At $16.6 billion for the quarter, the segment produced substantial profit. The concern lies not in whether AWS is profitable—it clearly is—but in whether the incremental $220 billion investment program will earn comparable returns after accounting for depreciation, energy, financing and possible overcapacity.

Free Cash Flow Has Become the Key Amazon Counterpoint

Amazon’s operating cash flow increased 33% to $161.4 billion for the trailing 12 months. Free cash flow, however, fell from an inflow of $18.2 billion to an outflow of $7.6 billion. The company attributed the decline primarily to a $66.1 billion year-over-year increase in purchases of property and equipment, net of proceeds and incentives, reflecting AI investment.

This is precisely the trade-off at the center of the debate. Amazon’s underlying businesses are generating more operating cash than ever, yet the company is reinvesting so aggressively that free cash flow has turned negative. If AWS demand remains constrained by insufficient capacity and long-term contracts convert into high-margin revenue, the investment can be rational. If demand forecasts are too optimistic, the negative free-cash-flow period may last longer and returns may fall.

Amazon’s ability to build at scale is a competitive advantage. Smaller cloud and data-center providers often rely on debt, lease financing or customer prepayments. Amazon can fund much of its investment from operations and has access to public debt markets on favorable terms. That lowers financing risk but does not eliminate commercial risk. A well-capitalized company can still overbuild.

What Amazon’s Results Do and Do Not Show

The results show that AI-related cloud demand was accelerating through June 2026, not collapsing. They show that AWS remained highly profitable, that custom chips were gaining adoption and that customers had committed to substantial future capacity. They also show that Amazon’s capital intensity had risen enough to consume all free cash flow over the trailing year.

The results do not show how much AWS revenue came from OpenAI, Anthropic or other frontier laboratories. They do not reveal the profitability of those specific contracts, the degree to which customer commitments are backed by prepayments or guarantees, or how much of the planned infrastructure can be repurposed if demand changes. Those are the variables that would allow investors to test the strongest version of the concentration critique.

Comparing Microsoft, Alphabet and Amazon

Figures are reported by the companies for quarters ended June 30, 2026. Cloud definitions differ and are not directly comparable. Microsoft Cloud includes more than Azure; Google Cloud includes infrastructure, platform and workspace products; AWS is Amazon’s cloud segment.

The comparison reveals a paradox. All three businesses are generating strong cloud operating results, but all three are committing capital at a pace that has transformed their cash-flow profiles. Their balance sheets can absorb the spending, yet the magnitude raises the cost of being wrong. A normal software miscalculation can be corrected by reducing hiring or marketing. A data-center miscalculation leaves buildings, power contracts, chips and lease obligations that cannot be unwound quickly.

Microsoft has the best concentration disclosure, Alphabet has the fastest reported cloud growth, and Amazon has the clearest connection between frontier-lab commitments and custom-chip strategy. None offers a complete customer-level picture. That asymmetry between capital spending and disclosure is one reason the debate has become so intense.

OpenAI’s Financial Position: Rapid Revenue Growth and Heavy Spending

OpenAI is central to the debate because it is simultaneously a model developer, consumer product company, enterprise software supplier and one of the world’s largest buyers of computing capacity. Its revenue has grown at a pace rarely seen in enterprise technology, but its spending has grown even faster. Because the company remains private, outsiders depend on selective company statements, financing documents and original reporting rather than complete quarterly filings.

Financial Times reporting in June 2026, based on financial information reviewed by the publication, said OpenAI generated approximately $13 billion of revenue in 2025 and spent about $34 billion. The reported net loss was close to $39 billion, but that figure included a very large non-cash charge linked to the company’s previous investor structure. The operating picture was less extreme than the headline net loss but still deeply negative. Reporting on the same financials described an operating loss of roughly $20.9 billion and an adjusted cash-oriented loss closer to $8 billion after removing several non-cash items and credits.

This distinction is essential. A $39 billion accounting loss does not mean OpenAI consumed $39 billion of cash in one year. The non-cash restructuring charge reflected changes in investor rights and corporate arrangements rather than a comparable payment for chips or salaries. At the same time, the lower cash-oriented loss should not be used to imply that the business was near self-funding. Billions of dollars of annual cash consumption remain substantial, especially when future compute commitments are measured in hundreds of billions.

OpenAI’s revenue trajectory has continued to rise. Reuters reported in June 2025 that the company’s annualized revenue had reached $10 billion. By the end of 2025, SoftBank executives said annual recurring revenue exceeded $20 billion. More recent private-company reports have placed the run rate higher, but run-rate figures should not be confused with completed annual revenue. A company producing $4 billion in one month can describe a $48 billion annualized run rate even if it has not yet sustained that pace for a full year.

Why OpenAI Needs Multiple Cloud Providers

OpenAI originally relied heavily on Microsoft Azure. As demand increased, it added Oracle, CoreWeave, Google Cloud, AWS-linked capacity and other infrastructure arrangements. The diversification is commercially rational. No single provider can necessarily deliver every chip, power connection or data-center site at the required speed, and dependence on one vendor weakens negotiating leverage.

The expansion also creates a challenge for the hyperscalers. They may build capacity for a customer that can shift workloads among multiple providers. Model training can be technically difficult to move once a cluster is configured, but inference workloads and future projects can be distributed. If OpenAI secures better economics from custom chips, dedicated data-center operators or sovereign partners, the revenue retained by any one cloud provider may be lower than expected.

OpenAI’s revised Microsoft agreement formalized greater flexibility. Microsoft remained the primary cloud partner, but the relationship became less exclusive. The company’s Google Cloud deal, its CoreWeave contract and the broader Stargate infrastructure initiative demonstrate that OpenAI is assembling a portfolio of suppliers rather than building exclusively on Azure.

The $250 Billion Azure Commitment

In October 2025, Microsoft said OpenAI had contracted to purchase an incremental $250 billion of Azure services. The number is large enough to affect any analysis of Microsoft’s backlog. Yet its meaning depends on timing, conditions and utilization. Spread over ten years, $250 billion would average $25 billion a year; over five years, it would average $50 billion. Public announcements have not provided a complete payment schedule or all contractual protections.

A commitment of that size can be both a source of revenue visibility and a source of counterparty risk. Microsoft can plan construction around the agreement, but OpenAI must generate or raise enough capital to pay for the services. The contract may include mechanisms that protect Microsoft, such as minimum payments, deposits, collateral or staged capacity delivery, but the public cannot fully evaluate those terms.

The existence of the contract does not prove circularity by itself. OpenAI sells subscriptions, API access, enterprise products and licensing arrangements to independent users. The critical test is whether those end-user sales eventually cover the Azure expense and other costs. Until they do, outside capital remains part of the funding chain.

OpenAI’s IPO as a Financing Event

OpenAI confidentially filed for a U.S. initial public offering in June 2026, according to the company and Reuters. The filing did not establish a listing date. Reuters later reported, citing New York Times reporting, that advisers had discussed waiting until 2027 to pursue a valuation around $1 trillion rather than accepting a lower valuation for an earlier offering.

An IPO would not automatically solve the business model. It would convert OpenAI from a private company funded largely by strategic and institutional investors into a public company with access to a deeper equity market. It would also require substantially greater disclosure. A registration statement would be expected to contain audited financial statements, contractual obligations, risk factors, related-party transactions and customer information that could clarify many of the questions now debated through analyst estimates.

Delay is not necessarily a sign of crisis. Companies postpone listings for market conditions, regulatory preparation, litigation, accounting work or strategic reasons. OpenAI also said there were activities that might be easier to pursue while private. Still, the timing matters because a capital-intensive company benefits from raising equity before financing conditions become less favorable. A successful offering could strengthen the entire AI infrastructure chain; a weak reception could force investors to reassess private valuations and cloud commitments.

SoftBank and Crystal Intelligence

Zitron also highlighted revenue associated with SoftBank’s “Crystal Intelligence” initiative, arguing that more than $800 million of OpenAI’s 2025 revenue came from the strategic investor. SoftBank has publicly described Crystal Intelligence as an enterprise AI initiative built with OpenAI, and company executives have discussed its development. The available public information is not sufficient to independently evaluate the precise revenue recognition, implementation progress or economic value of the program.

The broader issue is related-party or strategic-partner revenue. Revenue from a shareholder can be legitimate, especially when the shareholder buys a real product at commercial terms. It deserves additional scrutiny because the relationship may influence contract size, timing or pricing. A future OpenAI registration statement would likely provide more detailed related-party disclosures than are currently available.

Anthropic: Faster Enterprise Momentum, Similar Capital Questions

Anthropic has become equally important to the infrastructure story. Its Claude models, especially coding and enterprise products, have driven rapid growth. Reuters reported in October 2025 that enterprise customers accounted for approximately 80% of Anthropic’s revenue and that the company expected a 2026 annualized revenue run rate between $20 billion and $26 billion. By spring 2026, later reporting suggested the company had moved well beyond those earlier targets.

Reuters Breakingviews reported in May 2026 that Anthropic generated $4.8 billion of revenue in the three months to March and expected approximately $10.9 billion in the June quarter. It also expected a positive adjusted operating result for that quarter, including model-training costs but excluding stock-based compensation. Those figures, if realized, would indicate that Anthropic’s commercial position had improved dramatically and that the company was closer to operating profitability than many critics assumed.

Adjusted profitability should still be interpreted cautiously. Excluding stock-based compensation can materially reduce reported expense at a fast-growing private technology company. Quarterly profitability can also be affected by contract timing, credits, infrastructure accounting and annual prepayments. The important point is that Anthropic is not merely a laboratory with negligible sales. It has become a major enterprise supplier with revenue measured in billions per quarter.

Anthropic’s Multi-Cloud Strategy

Anthropic’s infrastructure relationships are intentionally diversified. AWS remains its primary training and cloud provider for mission-critical workloads. Amazon has invested heavily in the company and developed large Trainium clusters for it. At the same time, Anthropic has used Google TPUs for years and expanded its Google-Broadcom agreement to multiple gigawatts of next-generation capacity expected to come online from 2027.

This strategy gives Anthropic access to different chip architectures and reduces dependence on Nvidia alone. It also gives the company leverage in negotiations and provides redundancy. From the cloud providers’ perspective, however, it means Anthropic’s total compute demand can be enormous without any one provider capturing the entire spending stream.

Amazon and Google may both benefit from Anthropic’s growth while competing to supply it. Microsoft’s reported investment gain adds another connection. The same private company can therefore affect the revenue, capex plans and investment income of three public hyperscalers. This is an extraordinary degree of financial interconnectedness for a company founded only a few years ago.

Anthropic’s Funding and Valuation

Anthropic has raised very large funding rounds at rapidly increasing valuations. Reuters reported in May 2026 that the company raised $65 billion at a post-money valuation of $965 billion. The financing gave Anthropic substantial resources for compute commitments and reduced immediate liquidity risk. It also raised the return expectations embedded in the valuation.

A high valuation can strengthen a company’s purchasing power because it enables more equity to be sold for a given percentage of ownership. It can also create fragility. If revenue growth slows or public-market multiples contract, later financing may require more dilution or a lower valuation. Because hyperscalers are building infrastructure around long-term demand forecasts, changes in private-market valuation can indirectly affect data-center plans.

Anthropic’s stronger enterprise mix may offer better economics than a consumer-heavy subscription business. Enterprise customers can sign annual contracts, integrate models into workflows and generate recurring API consumption. Coding tools can also produce unusually high token usage because agents read repositories, generate code, run tests and iterate. The cost side remains substantial, but the revenue may be more predictable than consumer subscriptions that experience variable engagement.

OpenAI and Anthropic Are Not the Same Risk

Public discussion often combines the two companies as though they were interchangeable. They share several characteristics: private ownership, high valuations, enormous compute needs and strategic relationships with cloud providers. Their customer bases, product mixes, capital structures and profitability trajectories differ.

OpenAI has a larger consumer brand, a broader product portfolio and deeper historical dependence on Microsoft. Anthropic appears more concentrated in enterprise and developer workloads and has structured primary relationships with both AWS and Google infrastructure. OpenAI’s consumer scale can produce massive subscription and advertising opportunities, but consumer inference costs can be difficult to control. Anthropic’s enterprise focus may support higher-value contracts, but reliance on coding and API use creates exposure to model competition and customer optimization.

For cloud investors, the distinction matters because the quality of the underlying customer demand may differ. A compute contract backed by prepaid enterprise revenue is not equivalent to one funded mainly by speculative equity. Without full financial statements, outsiders cannot map the two companies’ cloud commitments to their operating cash flow with confidence.

The Data-Center Pipeline: Announcements Are Not Operating Capacity

The 190-gigawatt figure discussed in the Bloomberg interview comes from a database maintained by Currence, previously known as Sightline Climate. Its February 2026 outlook tracked 190 gigawatts across 777 large data centers and AI factories larger than 50 megawatts that had been announced since 2024. The number represents a project pipeline, not completed capacity.

That difference is crucial. The same report said at least 16 gigawatts were scheduled to come online during 2026, but only about 5 gigawatts were under construction. Approximately 11 gigawatts remained at the announced stage despite typical build times of 12 to 18 months. The researchers estimated that 30% to 50% of the 2026 pipeline might be delayed.

Pipeline data can overstate future supply because developers announce projects before securing every permit, power connection, customer or financing package. Some projects compete for the same prospective tenants. Others are resized, moved or canceled. Adding the nameplate capacity of all announcements is useful for measuring ambition, but it should not be treated as a forecast that every megawatt will operate on schedule.

This weakens the simplest version of the revenue-sufficiency calculation. If only a portion of 190 gigawatts is built, the industry does not need to support the full implied revenue immediately. The timing also matters: capacity scheduled for the early 2030s should not be compared with 2026 revenue as though it must be filled next year.

Nevertheless, the pipeline remains economically important. Even a fraction of 190 gigawatts represents an infrastructure program comparable to the electricity demand of large countries. Projects require land, substations, transmission, generation, cooling, fiber and enormous quantities of servers. Commitments made today can create fixed costs and local power-system obligations for decades.

Power Is Becoming a Binding Constraint

AI infrastructure is constrained by more than money. The International Energy Agency projects global data-center electricity consumption to reach approximately 945 terawatt-hours by 2030 in its base case, roughly double the 2024 level. The agency expects data-center electricity demand to grow about 15% annually from 2024 to 2030, more than four times faster than electricity consumption in the rest of the economy.

In the United States, Lawrence Berkeley National Laboratory estimated that data centers consumed about 176 terawatt-hours in 2023, equal to approximately 4.4% of national electricity use. Depending on growth assumptions, the share could rise to between 6.7% and 12% by 2028. Those projections predate some of the largest 2026 spending plans, which means local constraints may be more important than national averages.

Data centers are geographically concentrated. Northern Virginia, Texas, Arizona, Oregon, Ohio and parts of the Midwest have become major development regions. A project can be financially attractive at the corporate level and still face transmission delays, transformer shortages, water concerns or community opposition. Utilities may need years to add generation and grid capacity, while AI companies want clusters operational within months.

These bottlenecks cut both ways. They can reduce overbuilding by slowing speculative projects, supporting the value of existing capacity. They can also raise project costs and delay revenue. A cloud provider that has ordered chips but lacks power cannot monetize the equipment efficiently. A laboratory that has promised customers additional model capacity may need to rent higher-cost third-party infrastructure while waiting for its own sites.

How Much Revenue Does a Gigawatt Need?

Converting power capacity into required revenue is necessarily approximate. A gigawatt describes electrical load, not the purchase price of a data center or the revenue it can produce. Economics vary by chip generation, utilization, cooling design, power cost, networking architecture, software mix, depreciation schedule and customer pricing.

A simple revenue-per-megawatt assumption can nevertheless illustrate scale. If a facility required $10 million to $12 million of annual revenue per megawatt to support its capital and operating costs, 100 gigawatts would imply $1.0 trillion to $1.2 trillion of annual revenue. At 190 gigawatts, the implied number would be far larger. But that calculation should not be treated as a forecast because the pipeline will be delivered over many years, not all projects will be built, and some capacity supports internal advertising, search, productivity and consumer services rather than direct cloud rental.

The more useful question is incremental. For each dollar of new infrastructure, how much additional gross profit can the owner generate over the asset’s economic life? Revenue alone is insufficient. A low-margin hardware sale can produce substantial revenue without an attractive return, while high-margin software layered on the same infrastructure can create far more value.

Utilization is critical. A cluster operating near capacity under a long-term contract can generate strong economics even if the customer base is concentrated. An underused cluster depreciates whether or not it is busy. Because accelerators improve quickly, idle capacity is especially costly. The providers’ claims that demand exceeds supply are therefore reassuring, but investors need to know whether shortages persist after the current wave of pre-booked frontier-lab demand is satisfied.

Depreciation Is the Slow-Moving Risk

Capital expenditure affects cash flow immediately, but much of its effect on reported earnings arrives through depreciation over subsequent years. A company can spend heavily today, report strong operating margins and then experience rising depreciation as assets enter service. Alphabet explicitly warned that higher technical-infrastructure investment would continue to pressure the income statement through depreciation and data-center operating costs.

The accounting life assigned to an asset may differ from its economic life. A building can remain useful for 25 years, while a server may become inefficient relative to new hardware after four or five. Companies routinely reuse older equipment for less demanding workloads, which extends value, but the resale market for highly specialized AI clusters may be limited.

Efficiency gains can offset depreciation. Better software can increase the number of tokens processed per chip, and new pricing models can align customer charges with value delivered. Microsoft said it increased Copilot throughput fourfold during the fiscal year, showing that physical capacity is not the only determinant of output. If optimization keeps pace with model demand, returns can improve even without constant hardware expansion.

The risk is that efficiency reduces the need for compute faster than demand increases. Cheaper models, smaller specialized systems, open-weight alternatives and improved inference techniques can lower the amount of expensive frontier hardware required for a given task. That would benefit AI users but could reduce the scarcity value of data-center capacity. Infrastructure providers are betting that falling unit costs will stimulate enough new usage to keep total demand rising.

Is Enterprise AI Demand Broad Enough?

The hyperscalers’ defense rests on a claim that is broader than the success of OpenAI or Anthropic: enterprises are adopting AI across software development, customer service, security, data analysis, marketing, research and internal operations. If that adoption becomes embedded in everyday workflows, the two frontier laboratories may function as early anchor tenants rather than the permanent foundation of the entire market.

There are reasons to take that possibility seriously. Microsoft 365 Copilot surpassed 30 million paid seats, and GitHub Copilot consumption increased after Microsoft shifted pricing toward usage and value. Amazon said Bedrock had hundreds of thousands of customers and that second-quarter spending on the service exceeded all previous quarters combined. Alphabet described demand across core Google Cloud infrastructure, AI solutions and enterprise products, while reporting a backlog that exceeded half a trillion dollars.

The scale of these disclosures is inconsistent with a market limited to a handful of research laboratories. A large bank using models to review documents, a retailer generating product content, a pharmaceutical company searching scientific literature and a software company deploying coding agents may each consume less compute than a frontier-model training run, but together they can create a durable base of recurring inference demand.

The challenge is that usage does not automatically produce attractive vendor economics. Enterprise customers are becoming more sophisticated about model routing, token costs and workload design. They can use smaller models for routine tasks, reserve expensive frontier models for difficult queries and switch among providers. That improves customer return on investment but can limit pricing power for model developers and cloud vendors.

Enterprise adoption also moves more slowly than consumer enthusiasm. Companies must address security, data governance, regulatory obligations, integration and employee training. A successful pilot may not become a production deployment. Production systems can require years of process redesign, and the largest gains may depend on changing organizational structures rather than simply adding a chatbot to existing work.

What Productivity Research Actually Shows

The best-known evidence comes from controlled or quasi-experimental studies of specific tasks. An NBER study of customer-support agents found that access to a generative-AI assistant increased productivity by approximately 14% on average, measured by issues resolved per hour. Less experienced workers benefited more than top performers. Other studies of professional writing found that participants completed tasks faster and, in many cases, produced higher-quality output.

More recent field research has found gains in software development, online retail and customer-service operations, but the effects vary by task and worker. AI can reduce time spent searching for information, drafting routine text or diagnosing common problems. It can also create rework when outputs are wrong, encourage overreliance or reduce quality for experienced employees who accept poor suggestions.

These findings support the proposition that AI has economic value. They do not establish that the value is large enough to justify every current valuation or data-center plan. A 14% productivity improvement in one workflow may be highly profitable for the employer, but only a portion of that value accrues to the model provider or cloud vendor. Competition can push prices down, and customers may retain most of the savings.

Economy-wide productivity statistics also lag technology adoption. Businesses need complementary investments in data, training and process redesign before new tools affect measured output. The history of general-purpose technologies suggests that productivity gains can appear years after capital spending begins. That delay supports the long-term bullish case, but it also creates room for overinvestment during the transition.

The commercial question is therefore not whether AI can improve productivity. It can. The question is how much of the improvement becomes recurring revenue for infrastructure owners after competition, efficiency gains and customer bargaining power are considered.

Revenue Quality Matters More Than Revenue Labels

“AI revenue” can refer to several economically different activities. A cloud provider may rent accelerators, sell chips, charge for storage and networking, license a workplace assistant, distribute a third-party model or use AI internally to improve advertising. These businesses have different margins, capital requirements and retention characteristics.

Infrastructure rental is capital-intensive but can be durable when customers reserve capacity. Hardware sales can accelerate recognized revenue while producing lower recurring value. Software subscriptions are typically less capital-intensive but can face seat saturation and discounting. Usage-based model access can grow quickly while exposing the vendor to volatile inference costs. Advertising improvements may be highly profitable but difficult to attribute directly to AI.

Investors should be cautious when companies aggregate these categories into one narrative. An annualized run rate from custom chips is not equivalent to software recurring revenue. A backlog dominated by long-term infrastructure commitments is not equivalent to cash already collected. A private-company valuation gain is not equivalent to operating profit. Each may be economically valuable, but they should not be compared without adjustment.

The latest results show that cloud operating margins remain strong. AWS produced approximately 39% operating margin, Google Cloud reported 35.6%, and Microsoft’s Intelligent Cloud operating margin was about 41%. Those figures indicate that the existing cloud businesses are not sacrificing all profitability for growth. The concern is whether incremental AI capacity earns similar margins after depreciation and power costs fully enter the income statement.

Why the Historical Comparisons Are Useful—and Dangerous

The AI buildout is frequently compared with railroads, electricity, telecommunications and the dot-com boom. Each analogy captures part of the story, but none provides a complete forecast.

The Railroad Analogy

Nineteenth-century railroad investment created essential infrastructure and transformed commerce. It also produced repeated bankruptcies because too many companies built overlapping routes with borrowed money. Society benefited even when original shareholders lost. The lesson is that a technology can be economically revolutionary while individual investments generate poor returns.

AI infrastructure could follow a similar pattern. Excess capacity might lower computing costs, encourage new applications and produce broad economic gains. The cloud providers that finance the buildout may still face weaker returns if competition drives prices below expectations. Unlike many railroad companies, however, Microsoft, Alphabet and Amazon have diversified, highly profitable businesses and can fund much of the spending internally.

The Telecom and Dot-Com Analogy

The late-1990s telecommunications boom produced enormous fiber investment based on expectations of internet traffic. Demand eventually arrived, but not quickly enough to protect many highly leveraged builders. Fiber laid during the boom later became valuable infrastructure acquired at distressed prices.

The parallel with AI is the possibility that demand forecasts are directionally correct but mistimed. If data-center capacity grows faster than monetization for several years, prices and asset values could fall even as AI usage continues rising. The key difference is that AI chips depreciate faster than fiber. An underused fiber route can remain valuable for decades; a cluster of accelerators can become less competitive within a few years.

The Electricity Analogy

Electricity became a general-purpose technology only after factories redesigned production around electric motors. Early installations often replaced steam engines without changing workflows, limiting productivity gains. The eventual transformation required organizational change.

Enterprise AI may follow the same path. Adding a chatbot to an old process can save minutes but rarely changes a company’s economics. Larger gains may require rebuilding workflows around agents, data access and automated decision support. That transition can create sustained demand, but it takes longer than software vendors’ sales cycles.

Why Enron Is a Poor Direct Comparison

Zitron invoked the phrase “smartest guys in the room,” associated with Enron, to caution against assuming that intelligent executives cannot make catastrophic mistakes. The general lesson is valid: reputation and technical brilliance do not eliminate incentive problems or groupthink. A direct Enron comparison, however, is not supported by the current evidence.

Enron involved deceptive accounting, hidden liabilities and fabricated economic performance. Microsoft, Alphabet and Amazon publish audited financial statements, report large operating cash flows and operate established businesses. The legitimate concern is capital allocation and disclosure, not evidence of comparable fraud. Using the analogy too literally obscures the real questions.

Market Reactions Show Investors Are Not Treating All AI Spending Alike

Alphabet’s shares fell in extended trading after it raised 2026 capex guidance, despite an 82% increase in Google Cloud revenue. Microsoft’s shares surged after Azure growth reached 43%, its outlook exceeded expectations and management emphasized broad demand outside frontier-model companies. Amazon then rose sharply after AWS growth accelerated to 37% and operating margins expanded, even as the company lifted its spending plan to approximately $220 billion.

The different reactions suggest that investors are applying a dynamic test. Spending is rewarded when it is accompanied by accelerating revenue, strong margins and evidence of capacity shortages. It is punished when the increase appears to outrun near-term cash generation or when management provides insufficient detail about returns.

That test can change quickly. A single quarter of strong cloud growth does not validate a multi-year investment cycle, and one quarter of negative free cash flow does not prove failure. Market prices reflect expectations about years of future cash flows, which means sentiment can reverse if growth, margins or financing conditions change.

The market also distinguishes among layers of the AI value chain. Chipmakers and memory suppliers can receive cash earlier in the cycle because they sell equipment to infrastructure builders. Cloud companies bear more utilization risk. Model developers bear training and inference costs while competing for users. Application companies can benefit from lower model prices but may struggle to build defensible products. Returns can migrate from one layer to another even when total AI demand remains strong.

Five Scenarios for the AI Infrastructure Cycle

1. Broad Monetization

In the most favorable scenario, enterprise AI adoption expands rapidly, consumer products sustain engagement and agents create new categories of paid usage. OpenAI and Anthropic grow into their commitments, while thousands of smaller customers fill additional capacity. Cloud revenue rises faster than depreciation, utilization remains high and free cash flow recovers after the construction peak.

This scenario does not require every application to succeed. It requires aggregate demand to grow fast enough that efficiency improvements and price competition are offset by new use cases. The hyperscalers’ existing distribution, security and data platforms would give them a strong position.

2. Productive Technology, Excess Infrastructure

AI creates genuine value, but the industry builds too much capacity. Compute prices fall, weaker data-center developers fail and hyperscalers write down or repurpose some equipment. End users benefit from cheaper services, while infrastructure returns disappoint. This resembles parts of the telecom cycle: the technology wins, but not every investor does.

3. Frontier-Lab Consolidation

OpenAI and Anthropic remain important, but one gains a decisive commercial advantage or merges with a larger platform. Cloud commitments are renegotiated, shifted or concentrated around the winner. Providers with the strongest contractual protection and most flexible infrastructure perform better. Customer concentration becomes visible as a financial risk even though AI demand continues.

4. Financing Shock

Public markets reject high private valuations, an IPO is delayed or priced below expectations, and strategic investors become more selective. Frontier labs reduce spending to preserve cash. Backlog growth slows, planned data centers are postponed and equipment prices weaken. The hyperscalers remain solvent but experience lower returns and potentially large investment losses.

5. Efficiency Shock

Model architectures improve so rapidly that equivalent performance requires far less compute. Small and open models handle most enterprise tasks, while only a narrow set of frontier workloads needs giant clusters. AI adoption rises, but total infrastructure revenue grows more slowly than expected. This scenario is good for users and application developers but difficult for owners of expensive capacity.

The actual outcome could combine elements of all five. Broad adoption may coexist with local overbuilding. Frontier labs may thrive while specific data-center projects fail. Efficiency gains can lower unit costs while total demand continues rising. The investment question is not binary; it is about which layer captures the value and how quickly.

The Disclosures Investors Still Need

The concentration debate would be easier to resolve if hyperscalers provided more consistent information. Customer privacy and competitive sensitivity limit disclosure, but companies already publish concentration data when it is material. The scale of the current commitments justifies more transparency.

• Top-customer revenue concentration: The share of cloud revenue generated by the largest customer and top five customers, even if names are withheld.
• Backlog concentration: The proportion of RPO or contract backlog tied to frontier-model companies.
• Contract protections: General descriptions of prepayments, minimum commitments, termination rights and guarantees for unusually large infrastructure agreements.
• AI infrastructure utilization: Capacity in service, reserved, under construction and awaiting power.
• Revenue mix: Separate information for accelerator rental, custom-chip sales, managed AI services and application subscriptions.
• Capital intensity: Cash capex, finance leases, operating-lease commitments and expected depreciation associated with AI infrastructure.
• Related-party economics: Clear disclosure when cloud customers are also major equity investments or strategic shareholders.

Microsoft has moved furthest by publishing several measures excluding OpenAI. Alphabet’s statement that most backlog consists of typical contracts across a broad customer mix is helpful but imprecise. Amazon’s disclosure of multi-gigawatt commitments is important, yet it does not quantify revenue concentration. More comparable reporting would improve market discipline without requiring companies to reveal commercially sensitive prices.

What Management Is Getting Right

The hyperscalers are responding to real demand signals. Their cloud divisions are growing rapidly, customers report capacity shortages and enterprise adoption is moving beyond experimentation in several areas. Custom chips can reduce dependence on Nvidia and improve price-performance. Building now may be necessary because power, permitting and construction lead times are long.

The companies also have strategic reasons to avoid underinvesting. A cloud provider that lacks capacity can lose workloads that remain on a competitor for years. AI services create demand for databases, storage, security, networking and workplace software. The value of winning a customer can therefore exceed the revenue from raw compute.

Microsoft, Alphabet and Amazon have unusually strong balance sheets and operating cash flows. They are not relying solely on speculative debt to fund the buildout. That gives them time to optimize, repurpose capacity and absorb temporary underutilization. Their scale also lets them negotiate power and equipment contracts that smaller operators cannot obtain.

What Management May Be Underestimating

Competition can destroy the economics of an otherwise successful market. If every hyperscaler builds aggressively and custom chips reduce switching costs, compute may become increasingly commoditized. Customers can route workloads to the cheapest acceptable model and provider. Falling prices stimulate demand but can delay returns on existing assets.

Management may also underestimate the speed of technological obsolescence. The useful life of a building says little about the useful life of the accelerator cluster inside it. New chips can improve performance per watt, making older hardware expensive to operate. Software optimization can extend asset life, but it can also reduce the amount of capacity customers need.

Customer concentration is another risk that management teams may view through strategic rather than credit-oriented lenses. OpenAI and Anthropic are not merely buyers; they are partners in the effort to define the next computing platform. That can encourage providers to accept terms or build capacity they would scrutinize more harshly for an ordinary customer.

Finally, executives may be responding to competitive fear. No chief executive wants to be blamed for missing a technology transition. When every rival is spending, reducing capex can appear more dangerous than maintaining it. This creates a collective-action problem: individually rational investments can produce industry-wide excess capacity.

What Happens Next

The next phase of the debate will be driven by disclosure rather than rhetoric. OpenAI’s potential IPO could provide the first comprehensive public view of its financial statements and contractual obligations. Anthropic’s financing and listing plans could do the same. Registration documents would allow investors to compare annual revenue, gross margin, cash burn, stock-based compensation and compute commitments directly.

For the hyperscalers, the key indicators will be cloud growth, operating margins, capital expenditure, depreciation and free cash flow. Microsoft’s quarterly reporting of measures excluding OpenAI will be especially important. Alphabet investors will watch how much of Google Cloud growth comes from recurring services rather than TPU system deliveries. Amazon investors will compare AWS acceleration with the duration of negative free cash flow.

Data-center execution will also matter. Delays can constrain supply and support pricing, but they can increase costs and postpone revenue. Power agreements, utility interconnections and financing packages will reveal which announced projects are likely to be built. The gap between announced gigawatts and construction-ready capacity is likely to remain large.

Enterprise customer behavior will provide the clearest long-term answer. Renewal rates, token consumption, paid-seat growth and documented productivity gains will show whether AI is becoming an operating necessity or remaining an optional experiment. The market does not need every company to deploy frontier models. It needs enough customers to pay for the infrastructure at prices that produce acceptable returns.

Frequently Asked Questions

Is the AI capex boom entirely dependent on OpenAI and Anthropic?

No. Public filings show broad cloud demand, and Microsoft said nearly 90% of full-year cloud revenue came from customers outside frontier-model companies. OpenAI and Anthropic are nevertheless unusually large customers whose commitments materially affect backlog and infrastructure planning.

Did Google confirm that OpenAI and Anthropic will generate 48% of Google Cloud revenue?

No. The figure cited in the Bloomberg discussion was an UBS estimate. Alphabet has not published a named-customer revenue breakdown confirming it.

Did Amazon confirm that OpenAI and Anthropic represent 18% of future AWS revenue?

No. The figure was attributed to Barclays research. Amazon confirmed multi-year, multi-gigawatt commitments from both companies but did not disclose their percentage of AWS revenue.

How dependent is Microsoft on OpenAI?

OpenAI has a major effect on Microsoft’s future contracted revenue. Microsoft’s commercial RPO grew 84% including OpenAI and 25% excluding it. Current recognized cloud revenue is much more diversified, according to management.

Are AWS, Azure and Google Cloud profitable?

Yes. The latest reported operating margins were approximately 39% for AWS, 41% for Microsoft’s Intelligent Cloud segment and 35.6% for Google Cloud. Segment definitions differ, so the figures are not perfectly comparable.

Why is free cash flow falling if cloud revenue is growing?

Capital expenditure is rising faster than operating cash flow in some periods. Data centers, servers, networking and power infrastructure require cash before all associated revenue is recognized. Amazon reported a trailing-12-month free-cash-flow outflow, and Alphabet reported negative quarterly free cash flow despite strong operating cash generation.

Is OpenAI profitable?

Available reporting indicates OpenAI remained unprofitable in 2025. Its net loss included large non-cash items, but the company also reported substantial operating and cash-oriented losses. Because it is private, complete current financial statements are not publicly available.

Is Anthropic profitable?

Reuters Breakingviews reported that Anthropic expected a positive adjusted operating result for the June 2026 quarter, including training costs but excluding stock-based compensation. That is not the same as GAAP net income or sustained free-cash-flow profitability.

What does “circular financing” mean in AI?

It refers to relationships in which a cloud provider invests in an AI company that uses financing or credits to buy the provider’s infrastructure. The arrangement can support real economic activity, but it creates questions about revenue quality when the customer depends on continuing external funding.

Does the 190-gigawatt data-center pipeline mean all that capacity will be built?

No. The figure tracks announced projects. Currence estimated that 30% to 50% of the capacity scheduled for 2026 could be delayed, and only a portion was under construction when the report was published.

What would signal that the AI infrastructure boom is overbuilt?

Warning signs would include slower cloud growth, falling utilization, contract renegotiations, persistent negative free cash flow, declining compute prices, project cancellations and impairment charges on data-center or server assets.

What would signal that the spending is justified?

Broad enterprise renewals, rising application revenue, improving cloud margins, sustained capacity shortages, recovering free cash flow and evidence that customers generate measurable returns from AI would support the investment case.

Final Assessment

Ed Zitron’s criticism identifies a real weakness in the public AI narrative: investors are being asked to evaluate unprecedented infrastructure spending without enough information about customer concentration, contract quality and the financial condition of the largest private buyers. OpenAI and Anthropic are not peripheral customers. Their commitments affect cloud backlogs, chip strategies, data-center construction and the investment income of several public companies.

The strongest version of the criticism goes beyond the evidence. The latest earnings do not show a cloud boom sustained only by artificial transactions between hyperscalers and two laboratories. Microsoft reported that nearly 90% of cloud revenue came from outside frontier-model companies. Alphabet and Amazon reported broad customer activity, accelerating cloud revenue and high segment operating margins. These are meaningful operating results, not merely private-market valuation gains.

The concern shifts from whether demand exists to whether returns will justify the capital. Amazon’s negative trailing free cash flow, Alphabet’s negative quarterly free cash flow and Microsoft’s enormous server spending show that the buildout has changed the financial profile of businesses once celebrated for low capital intensity. Their balance sheets can fund the transition, but the opportunity cost is substantial and the assets begin depreciating before the long-term revenue outcome is known.

The most important unresolved issue is the gap between current revenue diversification and future commitment concentration. A company can have thousands of customers and still build its marginal capacity around one or two anchor tenants. If OpenAI and Anthropic continue raising capital, growing revenue and paying for contracted compute, the hyperscalers may have secured the defining customers of a new platform. If financing tightens or model economics improve faster than demand expands, the same commitments can become evidence of overbuilding.

The latest quarter therefore supports neither complacency nor a declaration that the AI boom is a fabrication. It supports a more demanding standard. Cloud growth should be evaluated alongside cash capex, lease obligations, depreciation, utilization and customer concentration. Private-company run rates should be separated from completed annual revenue. Investment gains should be separated from operating profit. Backlog should be examined for counterparty quality rather than treated as guaranteed cash.

AI can be a transformative technology and an overextended investment cycle at the same time. The public companies are showing real commercial strength, while the private laboratories are consuming capital at a scale that remains historically unusual. The outcome will depend less on the number of announced gigawatts than on whether businesses and consumers ultimately pay enough for useful AI services to support them.

This article is provided for general informational purposes and does not constitute financial, investment, tax, or legal advice.

Sources

• Bloomberg interview: Zitron on AI infrastructure economics
• Microsoft fiscal 2026 fourth-quarter earnings release
• Microsoft fiscal 2026 fourth-quarter earnings call
• OpenAI: The next phase of the Microsoft–OpenAI partnership
• Microsoft: The next chapter of the Microsoft–OpenAI partnership
• Alphabet second-quarter 2026 earnings call
• Amazon second-quarter 2026 earnings release
• Anthropic: Expanded Amazon compute partnership
• Anthropic: Google and Broadcom compute partnership
• Reuters: OpenAI adds Google Cloud capacity
• Reuters: OpenAI confidentially files for a U.S. IPO
• Reuters: OpenAI weighs 2027 IPO timing
• Financial Times: OpenAI spending and 2025 financial results
• Reuters Breakingviews: Anthropic revenue growth and profitability outlook
• Reuters: Anthropic funding and valuation
• Currence 2026 Data Center Outlook
• International Energy Agency: Energy demand from AI
• Lawrence Berkeley National Laboratory: U.S. data-center electricity demand
• National Bureau of Economic Research: Generative AI at Work
• Reuters: AI investment and Big Tech free-cash-flow pressure
• Reuters: Alphabet cloud growth and increased capex guidance
• Reuters: Microsoft Azure growth and AI spending
• Reuters: Amazon AWS growth and capital-spending plan

Affiliate disclosure: Businessfinance.news may earn compensation from qualifying actions completed through selected links on this website, at no additional cost to the reader. Affiliate relationships do not influence our editorial reporting, analysis, or conclusions.