APoV 002The Cloud, and the Cloud in the AI Era

Has AI turned the cloud into a better business?

This question has come up repeatedly over the past six months, prompting me to write down my thoughts. Answering it requires tackling two foundational questions first: what kind of business the cloud actually is, and how to define “good” before asking whether it has gotten “better.”

The Summer of Disconnect

In late July 2026, Alphabet, Microsoft, and Amazon reported their second-quarter earnings in quick succession. The top-line numbers were staggering. AWS generated $42.2 billion in revenue, up 37% year-over-year, its fastest pace in eighteen quarters. Azure grew 43%, crossing $100 billion in annualized revenue for the first time. Google Cloud posted $24.8 billion, surging 82%. By September, Oracle Cloud Infrastructure reported a 121% leap. Even more counterintuitively, operating margins expanded rather than compressed: AWS posted an operating margin of 39.4%, up 6.5 percentage points from a year earlier, while Google Cloud climbed from 20.7% to 35.6%.

The cash flow statements told an entirely different story. Over the trailing twelve months, Amazon's free cash flow swung to negative $7.6 billion, down from positive $18.2 billion a year prior. Its capital expenditures, spanning both logistics and compute infrastructure, reached $169 billion, exceeding AWS's total revenue over the same period. Alphabet burned $5.9 billion in quarterly free cash flow, raising nearly $50 billion net through fresh equity and mandatory convertible preferred stock. Oracle posted a $5.0 billion quarterly free cash flow deficit while issuing $20 billion in common stock. Capital poured into data centers and silicon, with combined 2026 capital expenditure budgets across Amazon, Alphabet, Microsoft, and Meta topping $700 billion. The fastest-growing, highest-margin engines on the income statement had become the heaviest cash drains on the cash flow statement.

Yet the market remains convinced this capital will be recouped quickly, anchored by two pieces of evidence. The first comes from Andy Jassy's remarks on Amazon's earnings call, noting that server and networking equipment takes an average of just under three years to break even, whereas AI compute capacity is being contracted for terms of at least five years. A three-year payback on a five-year contract leaves the remaining two years as locked-in profit. The second is the order book: remaining performance obligations (RPO) and cloud backlogs signed by customers across Microsoft, Oracle, Google Cloud, AWS, and CoreWeave now surpass $2.4 trillion in aggregate.

This paradox stems from a fundamental timing mismatch. Cash for chips and data centers goes out the door upfront, while rental cash flows trickle back over five to six years. Accounting depreciation spreads those massive capital outlays over time, flattering the income statement even as free cash flow bleeds red. As long as contracts outlast payback periods and counterparties stay solvent, the model works as a front-loaded capital relay.

Just four years ago, the cloud operated under an entirely different paradigm.

“Selling Compute Like a Utility”

Today's builders take cloud computing for granted, yet few examine how the industry arrived at its current form. A brief look back is worth our time, because tracing the evolution of the cloud is essential to understanding how AI is reshaping its core nature.

The idea of selling computing as a public utility predates the cloud itself. In 1961, computer scientist John McCarthy remarked at MIT's centennial celebration: “Computing may someday be organized as a public utility just as the telephone system is a public utility.” Paying only for capacity actually consumed was the original commercial blueprint. Time-sharing systems later pursued this vision, but the personal computer disrupted that trajectory: once desktop machines became cheap enough for every company to own outright, renting centralized compute lost its appeal.

McCarthy's prediction lay dormant for forty-five years. On March 14, 2006, Amazon launched S3 storage at fifteen cents per gigabyte per month; that August, it introduced EC2 to rent out compute cycles. The motivation was purely pragmatic. As Jassy later recalled, Amazon realized internally around 2003 that if building infrastructure was this painful for Amazon itself, outside companies were surely facing the same friction. That insight formed the cloud's foundational bargain: customers stopped paying upfront for hardware and switched to monthly rent, while cloud providers bought and operated the machines, absorbing the risk of idle capacity.

In the years that followed, other tech giants entered in ways that reflected their respective corporate DNA. Google launched App Engine in 2008, betting on a managed platform approach; Windows Azure made its commercial debut in 2010, leveraging Microsoft's deep enterprise distribution channels. For years, outside observers debated whether cloud computing was a viable high-margin business or a capital-intensive charity. The turning point arrived in April 2015, when Amazon broke out AWS financials for the first time, reporting $1.566 billion in quarterly revenue and $265 million in operating income, a 17% margin. Jeff Bezos noted in the earnings release that AWS was already a $5 billion business, and accelerating.

The ensuing decade was the golden era of conventional cloud computing. AWS's annual revenue surged sixteen-fold from $7.9 billion in 2015 to $128.7 billion in 2025; its operating margin expanded from 23.6% to 35.4%; meanwhile, storage pricing plunged roughly 85%. Surging revenue, plunging prices, and expanding margins happened simultaneously, proving that cloud computing profits not from retail markups, but from cost deflation driven by massive scale.

Google Cloud illustrated the flip side of this dynamic. It sustained heavy losses for years, turning its first quarterly operating profit only in the first quarter of 2023, when $7.45 billion in revenue generated a slim $191 million in operating income (versus a $706 million loss a year earlier). Below critical scale, the cloud is a bottomless cash sink.

Then, on November 30, 2022, ChatGPT launched. Over the next four years, the industry underwent a tectonic shift. Incremental demand rotated from millions of independent software workloads to a handful of frontier model labs demanding massive GPU clusters; billing evolved from hourly virtual machines to multi-year contracts measured in hundreds of megawatts. Google Cloud's operating margin rocketed from 11% to over 35% within two years. The sheer volume of AI demand catapulted cloud platforms past the scale threshold and into hyperdrive.

China's cloud ecosystem unfolded along a parallel timeline, but navigated a far rockier path. In 2008, Wang Jian joined Alibaba to lead development of the Apsara distributed operating system, which by 2013 scaled to single clusters of 5,000 servers. Yet in China's commercial environment, enterprise and government clients have long favored on-premises private deployments, willingness to pay for software remains structurally low, and price wars have never relented. In May 2024, Alibaba slashed API prices for its flagship Qwen models by up to 97%, triggering immediate matching cuts from ByteDance and Baidu, with several core models made entirely free. The financial results reflect this reality: for the quarter ended June 2026, Alibaba Cloud generated roughly $7.1 billion in revenue, up 45% year-over-year, yet its adjusted EBITA margin hovered around 12%, compared with AWS's operating margin of nearly 40%. While Alibaba has pledged 380 billion RMB over three years for AI and cloud infrastructure, under export controls that restrict access to advanced silicon, how much of that capital expenditure can be converted into effective compute, and ultimately into high-margin revenue, remains an open question.

What Kind of Business Is the Cloud?

At its core, the cloud is a machine that converts risk into rent. It takes specialized, capital-intensive hardware prone to rapid obsolescence and underutilization, and turns it into a predictable monthly bill. Where commercial banks perform maturity transformation, turning short-term demand deposits into long-term loans, cloud platforms perform asset and technology transformation, absorbing hardware obsolescence and utilization risk in exchange for steady rent.

Historically, this machine generated profit through four pillars:

1.Scale: Procuring hundreds of thousands of servers at a time unlocks steep volume discounts, justifies custom silicon design, and secures long-term power purchase agreements (PPAs).

2.Oversubscription: Much like a commercial gym, where memberships far outnumber the machines on the floor because patrons never show up all at once, traditional cloud platforms relied on millions of uncorrelated workloads. By aggregating peak loads through the law of large numbers, providers provisioned only for combined peak demand; the buffer saved by pooling those peaks formed their core profit pool.

3.Software and Services: High-margin proprietary layers, such as managed databases, access controls, and identity management, paired with the data gravity that makes switching costs prohibitive.

4.Capital and Balance Sheet: Fronting the capital for hardware and recovering it through monthly rent, earning a risk premium for absorbing idle capacity and technological obsolescence.

In the classical cloud era, the first three pillars held center stage while the capital dimension remained largely invisible. With millions of dispersed clients, no single customer could leave a data center sitting idle; utilization risk was diversified away to near zero, and high-margin software revenue masked the underlying balance-sheet intensity. For a decade, Wall Street valued cloud operators through the lens of asset-light, high-multiple software companies.

AI inverted this architecture completely. A cluster housing 100,000 cutting-edge accelerators often serves a single frontier-model tenant. Workloads run flat-out around the clock, eliminating temporal variance; gym-style oversubscription margins vanish. Frontier labs demand raw bare-metal throughput and networking fabric rather than managed middleware, compressing proprietary software margins. That leaves only scale and capital, pushing the balance sheet into the spotlight. Once the law of large numbers stops working, cloud providers can no longer buffer risk through elastic pooling; instead, they are forced to lock customers into five-year, take-or-pay commitments. The surge in multi-billion-dollar backlogs does not reflect a shift in customer preferences; it reflects an industry restructuring how unpoolable capital risk is allocated.

Oracle's recent disclosures illustrate this reality clearly. Management revealed that the vast majority of its newly signed contracts involve either customer prepayments or bring-your-own-hardware (BYOH) arrangements, requiring no incremental hardware capex from Oracle. Customers fund the silicon themselves, renting only Oracle's physical facilities, high-speed RDMA networking, and cluster operations. When an intermediary's balance sheet strains under heavy asset burdens, it strips out the financing layer, handing capital risk back to the customer and retreating into the role of a pure operator. This unbundles the cloud into its constituent parts: financing capital assets and running infrastructure operations are two fundamentally distinct businesses.

Did AI change the fundamental nature of the cloud? No. It simply stripped away the high-margin software veneer, laying bare the heavy-capital and energy infrastructure underneath. And once financing and operations decoupled, specialized players inevitably emerged to tackle a single layer.

Neocloud

Where traditional hyperscalers bundled all four pillars together, specialized challengers, including CoreWeave, Nebius, Lambda, and Crusoe, unbundled them. They skip managed relational databases and sprawling enterprise suites to optimize for a single objective: racking the latest GPUs at maximum speed, wiring up high-density networking and liquid cooling, and leasing out capacity under multi-year contracts. The industry calls them Neoclouds. Structurally, their closest analogue is not AWS, but aircraft lessors. They sign long-term leases, pledge the hardware as collateral to borrow heavily from institutional lenders, capture an interest spread, and underwrite the equipment's residual value over its useful life.

CoreWeave exemplifies this model. In the second quarter of 2026, it reported $2.575 billion in revenue, up 112% year-over-year. On a GAAP basis, however, it posted an operating loss of $49 million; interest expense alone reached $640 million, consuming a quarter of revenue and driving a net loss of $626 million. Capital expenditures hit $6.42 billion in the same quarter, or 2.5 times revenue. For the full year, projected capex stands near $38 billion against roughly $13 billion in expected revenue. Spending nearly three dollars on hard assets for every dollar of current-year revenue is virtually unprecedented in the history of the technology sector.

This debt-fueled velocity rests entirely on more than $100 billion in non-cancelable contract backlog. Lenders underwrite CoreWeave not on the strength of its own balance sheet, but on the creditworthiness of the counterparties on the other side of those contracts. In 2025, Microsoft alone accounted for 67% of CoreWeave's revenue; even after some diversification in early 2026, its top two customers still represented more than 60% of the total.

This points to the most common misconception about Neoclouds. They have never been true rivals to the hyperscalers; rather, they serve as external capacity warehouses and shadow balance sheets for them. When Microsoft leases capacity from CoreWeave, it converts what would have been massive upfront capital expenditures into operating lease expenses. Google went a step further, guaranteeing a $1.8 billion, ten-year lease for Fluidstack at TeraWulf, enabling the project to secure low-cost debt financing while earning Google roughly 8% in equity warrants. Here, the hyperscaler isn't leasing out data center floor space; it is lending out its corporate credit rating.

Between flattering adjusted EBITDA and true net profit lie the twin hurdles of depreciation and interest expense. For GPU lessors, those two line items define the business. As long as borrowing costs fall and older GPUs hold their rental value, the spread arbitrage works smoothly; but if compute supply outpaces demand or anchor tenants face a cash crunch, these highly levered balance sheets will be the first part of the value chain to feel the squeeze.

Has AI Made the Cloud a Better Business?

A great business generates a return on invested capital (ROIC) comfortably above its cost of capital, possesses a durable moat, and enjoys a long runway for reinvestment: in short, high returns, longevity, and compounding scale.

Benchmarked against the asset-light, high-retention, 25%-to-35% operating-margin profile of the traditional cloud era (2006–2022), the AI-era cloud can be evaluated across six dimensions.

Demand Is Larger, and Structurally Firmer

The explosion in top-line demand is obvious, but the structural shift in the nature of that demand matters even more. Traditional cloud demand came from pay-as-you-go consumption across millions of enterprises, broadly resilient, yet vulnerable to macro-driven spend optimization, as the 2023 cloud slowdown demonstrated. AI-era demand consists of multi-year, legally binding take-or-pay commitments. Backlogs have doubled within a few quarters, with hyperscalers repeatedly citing persistent capacity shortages on earnings calls.

Yet a critical question remains: is today's pricing power rooted in a durable technological moat, or is it an illusion created by temporary supply-chain bottlenecks? Current compute shortages are largely dictated by high-bandwidth memory (HBM) supply, advanced packaging capacity, and multi-year lead times for high-voltage transformers. Raw compute is fundamentally a fungible commodity; once upstream and downstream bottlenecks clear after 2027, pricing power propped up purely by hardware scarcity could face a sharp mean reversion.

From Millions of Small Accounts to Ultra-Concentrated Anchor Clients

Mega-clients bring unprecedented revenue visibility, but they cut both ways. First, extreme customer concentration wipes out oversubscription margins. Second, counterparty credit risk compounds rapidly. When a hyperscaler ties nearly half of its incremental backlog to a single frontier lab, circular financing loops emerge: cloud providers inject billions of dollars of equity into AI labs, which turn around and commit tens of billions of dollars back to the cloud provider in compute purchases. This vendor financing bears a striking resemblance to the late-1990s telecom bubble, when equipment makers lent money to carriers to buy their gear. It pads the order book, but tightly binds credit risk across the entire value chain.

Monetization Avenues Have Multiplied

This may be the most underappreciated shift of all. A decade ago, the cloud sold three basic primitives: compute hours, storage capacity, and hosted software. Today's monetization menu is far richer, spanning bare-metal GPU clusters, first-party token APIs, third-party model marketplaces, custom silicon, and even early-stage equity stakes swapped for compute credits.

Returns on capital vary dramatically depending on the path chosen. Think of a 1-gigawatt data center, housing hundreds of thousands of next-generation accelerators at a cost of tens of billions of dollars, as an industrial kitchen. There are roughly three ways to run it:

•Renting out the kitchen (GPU leasing): Wholesaling bare-metal compute hours yields an ROIC of roughly 30% after depreciation, power, and maintenance, with a static payback period of two to three years.

•Running the restaurant yourself (First-party model APIs): Owning both the compute and the model layer captures enterprise token value directly. Even if more than a third of capacity is dedicated to ongoing frontier research and pre-training that generates no direct revenue, using the remaining capacity for inference yields over 30% more total revenue than pure hardware leasing, pushing ROIC past 40%.

•Cooking as a tenant chef (Third-party model labs): Renting compute from a hyperscaler to sell tokens downstream forces a model lab to hand over 60% to 70% of its gross revenue to the landlord as rent. When price wars erupt in the token market, the rent stays fixed, and margins evaporate overnight.

The closer a provider sits to the final token delivered, the higher the economic yield per megawatt-hour. That explains why the entire industry is crowding toward the model and application layer. Hyperscalers with top-tier proprietary models naturally occupy the highest-return vantage point, while those without them must use strategic equity investments to lock in token distribution.

Margins Have Expanded

Early consensus assumed that upstream chip monopolies would capture the lion's share of AI profits, leaving cloud providers with thinner margins as mere hardware resellers. Actual financial results have shown a sharp divergence: AWS's operating margin is approaching 40% and Google Cloud's has climbed from single digits to over 35%, whereas players reliant on third-party merchant silicon and concentrated customers have seen gross margins come under pressure.

The primary driver of this divergence is custom silicon. Google's TPU and Amazon's Trainium were engineered not merely to top benchmark charts, but to internalize the hefty gross margins that would otherwise go to third-party chipmakers.

Still, an accounting caveat is in order. Today's pristine operating margins benefit from a depreciation lag. Hundreds of billions of dollars in annual capex are rapidly being capitalized onto balance sheets and depreciated over five-year schedules. As this wave of high-density AI infrastructure enters its peak depreciation window over the next two years, whether current operating margins can withstand the oncoming depreciation wall remains an open question.

Moats Have Shifted to the Physical World

Traditional cloud moats were built on sprawling software ecosystems and high switching costs. In the AI era, defensibility has shifted to harder physical constraints: power, custom silicon, and low-cost capital.

Power is the hardest constraint of all, and the bottleneck is not just power generation, but grid interconnection queues. In North America, connecting a multi-hundred-megawatt facility to the main transmission grid often takes four to seven years of permitting and engineering. Hyperscalers have begun buying up data center campuses directly adjacent to nuclear power plants, competing for the scarcest resource of all: land that already has high-voltage power hooked up. Custom silicon determines the depth of a provider's internal profit pool, while access to cheap capital serves as the entry ticket to a hundred-billion-dollar capex race.

In the inference phase, traditional scale economies return in a new form. The number of tokens a single GPU can generate per second depends heavily on batching, aggregating thousands of concurrent requests to run simultaneously. Whichever platform commands the largest enterprise customer base and call volume achieves the highest hardware utilization and lowest unit cost.

There is, however, a potential wildcard on the edge: on-device AI. If lightweight agent routing, routine filtering, and local context processing are increasingly handled by smartphones, PCs, and local NPU chips, cloud inference volume could fall short of linear projections, blunting the throughput advantage of massive centralized data centers.

From Pay-As-You-Go to Delayed Gratification

The ultimate measure of business quality is ROIC, which equals profit margin multiplied by asset turnover. AI has boosted the former while crushing the latter: across the major cloud providers, net property, plant, and equipment (PP&E) has surged more than 40% within a single year, while company-wide revenue growth has hovered in the 20–30% range, dragging down asset turnover.

Making this math work over a full cycle depends entirely on stable compute rental rates. If a shift in supply and demand drives rental prices down by 20% a year, project-level ROIC plunges from roughly 30% into the mid-teens. That is why five-year take-or-pay contracts have become the industry's lifeline: they convert asset price-deflation risk into counterparty credit risk. Isolated cases of prior-generation GPUs securing leases through 2029 mostly reflect discounted pricing for low-intensity edge workloads; lacking native support for low-precision formats, older hardware simply cannot sustain optimistic long-term residual values.

Taken together across all six dimensions, AI has transformed cloud computing into a heavy industry, one with firmer demand, richer monetization avenues, higher physical barriers to entry, and thicker reported margins, but also a much heavier balance sheet and far greater customer concentration. It looks better today largely because acute supply shortages confer pricing power and non-cancelable contracts provide a buffer. It looks worse because an agile, asset-light software business has morphed into a capital-intensive, leveraged infrastructure marathon. On longevity and reinvestment runway, the AI cloud is stronger than before; on capital returns, the unit economics of individual projects pencil out, but the full-cycle bill has yet to come due.

What's Next?

Will the Cloud Become Even More Essential?

Almost certainly yes, because the traditional cloud's two greatest strengths are coming back to the fore.

The first is concurrency orchestration. Model training is a single-tenant affair, whereas agentic inference requires managing unpredictable traffic spikes across millions of end users, returning to classic cloud multi-tenancy. Because AI agents must read enterprise databases, orchestrate workflows across systems, and enforce strict permission boundaries, the data security and enterprise networking layers that hyperscalers have spent decades building become indispensable once again.

The second is the rise of open-weight models. When enterprises use proprietary commercial models, they effectively pay two separate bills: a software licensing fee for the model and a consumption fee for the underlying compute. Open-weight models eliminate the software tax, collapsing the cost structure into a single compute bill. The cheaper and more commoditized raw models become, the more economic value migrates to the two ends of the stack: physical compute at the bottom, and enterprise data governance and workflow context at the top. Hyperscalers control both.

Will Overcapacity Emerge?

The most common historical analogy is the 2000 telecom fiber bubble, but there is a critical difference. Dot-com telecom players laid millions of miles of uncontracted “dark fiber” on speculation, whereas nearly all of today's AI data centers break ground only after legally binding multi-year contracts are signed. That makes an abrupt, industry-wide glut unlikely in the near term; instead, it shifts the risk to a different question: whether the customers who signed those contracts will survive long enough to honor them.

Rather than tracking aggregate FLOPs, three specific signals are worth watching: first, whether general enterprise backlog keeps growing smoothly once affiliated AI labs are stripped out; second, what actual renewal rates look like when initial leases on prior-generation GPUs expire; and third, whether the operating cash flows of frontier labs can scale fast enough to cover their tens of billions of dollars in purchase commitments. Future overcapacity is far more likely to show up as a structural mismatch in specific hardware generations than as a sudden, across-the-board halt.

Will Rising Costs Squeeze Operators?

Upstream supply-chain inflation is real, with rising prices for high-bandwidth memory and thermal components already forcing upward revisions to capex budgets.

Yet cost inflation cuts in opposite directions depending on the player. For cash-rich hyperscalers, rising hardware costs actually raise barriers to entry, and their market power allows them to pass those costs through to customers. For highly levered Neoclouds, however, hardware inflation combined with heavy debt service directly eats into their thin financing spreads. Rising input costs will not derail the industry as a whole, but they will act as a ruthless scalpel, carving out marginal players.

Will Frontier Labs Bypass the Cloud?

Efforts to bypass traditional cloud providers are already underway, but it is crucial to distinguish what can be bypassed from what cannot.

When frontier labs work directly with chip design houses or build custom inference silicon in-house, they are bypassing the hyperscalers' hardware markup. When they turn to third-party developers for turnkey builds or bring-your-own-hardware hosting, they are bypassing balance-sheet asset ownership. Yet at the operational and commercial layers, keeping clusters of hundreds of thousands of liquid-cooled GPUs running reliably, and selling AI into the core workflows of global enterprises, they cannot easily bypass the cloud giants.

This cleanly unbundles the cloud into three distinct layers:

•The Capital Layer: Who funds and owns the physical hardware. Sovereign wealth funds, infrastructure syndicates, tech giants, and private credit funds are all competing to price this capital.

•The Operations Layer: Who can keep ultra-dense clusters of hundreds of thousands of accelerators running reliably at scale, representing the core engineering craft of cloud operators.

•The Distribution Layer: Who can embed models and intelligence into the procurement pipelines, security frameworks, and daily workflows of global enterprises, representing the deepest moat built by incumbents over decades of enterprise IT dominance.

Frontier labs can bypass the intermediary's hardware markup and asset ownership, but they cannot bypass enterprise distribution. The strongest business model remains custom silicon paired with global enterprise distribution; the most fragile is leveraged hardware brokering without a diversified customer base.

From a Software Lens Back to an Infrastructure Lens

For a decade, the market valued the cloud through a software lens, prizing asset-light balance sheets, high gross margins, and sticky recurring revenue. AI has yanked the industry back to the logic of heavy infrastructure, forcing investors to underwrite contract duration, capital intensity, counterparty credit risk, and hardware residual values.

Viewed through a software lens, today's cloud looks like a worse business: it is too asset-heavy and consumes too much cash. Viewed through a utility infrastructure lens, however, it reveals a far grander ambition: using a heavy-asset skeleton of power plants, steel, and silicon to earn software-like profit margins.

Whether this model holds up over the long run comes down to two conditions. First, contract durations must outlast the true economic payback period of the underlying hardware; second, the model and application ecosystem must prove its commercial viability before capital markets lose patience. The first depends on the engineering and commercial discipline of the cloud providers; the second depends on the real economic value created by the age of AI.