Summary

Kenny Zhang’s core framework: AI cloud and traditional cloud are “of the same origin but different lineage”—AI cloud is not a simple extension of traditional cloud but a paradigm reversal, a return to the mainframe era: traditional cloud’s core resource is bandwidth and high-frequency distribution, while Neocloud’s is the centralized management and optimization of hyperscale compute clusters. Neocloud rests on three structural forces: AI Labs and hyperscalers are both suppliers and competitors—Microsoft and OpenAI are the clearest example; Neoclouds have no downstream application layer and are therefore more neutral; and hyperscalers do not want to direct capital toward heavy infrastructure earning only 8%-10% ROIC versus 40%-50% in their core businesses.
Compute scarcity has flipped the strategic positions of OpenAI and Anthropic: Dario was publicly criticizing OpenAI last year for buying unlimited compute, while Anthropic’s annualized revenue has risen from roughly $9B at the end of last year to possibly $50B now and could exceed $100B by year-end, forcing it into “rigid procurement”—“you can’t even call it elastic procurement anymore.” 曹卿云 believes xAI has effectively exited the foundation-model race, with its H100/H200 capacity being monetized passively as a Neocloud selling to Anthropic; Kenny expects OpenAI and Anthropic to have roughly equal total compute supply once all of xAI’s capacity is online, while OpenAI’s early compute build-out leaves it “very relaxed” and gives it a chance to catch up.
CoreWeave and Nebius represent two viable business models: CoreWeave follows the traditional infrastructure playbook, using 5-6-year contracts to cover the commercial GPU cycle—“for every dollar spent, you can earn three dollars back over the next six years”—while its financing cost has fallen from 14.5% high-yield debt in 2023 to 6.2% investment-grade debt this year, taking levered IRR comfortably above 50%; Nebius uses 40%-50% customer prepayments to reduce its reliance on leverage, giving up roughly 20%-25% of contract pricing while still generating 35%-40% unlevered returns. Kenny is comparatively neutral on bare-metal-heavy models such as Oracle and some crypto-mining converts.
Kenny gave the same answer to the market’s two main objections—borrowing long against depreciating assets and the sudden margin collapse: the market is reading the accounts incorrectly. The debt is backed by the value of the entire contract, not the residual value of the GPUs; residual value is merely “icing on the cake.” CoreWeave’s margin fell from 12%-14% to 1%-2% because roughly 1GW is under construction while only about 800MW is actually generating rent; as that construction-to-rent ratio converges, margins should return to 12%-14% by year-end, reach roughly 25% by the end of 2028 and move toward 30%. Neoclouds currently trade at roughly 10x steady-state earnings despite potential 120%-130% CAGR over the next 3 years, versus traditional cloud at roughly 20-25x and about 30% growth—a very wide valuation gap.
The backlog debate—roughly $90B-plus, or more than $99B according to the show’s opening line—turns on a waterfall structure between revenue and capex: the more than $30B spent today corresponds to roughly $90B-$100B of cash recovery over the next 6 years, with revenue overtaking capex around 2029. The backlog is already highly diversified: Meta accounts for 35%, Microsoft 20%, OpenAI roughly 20%, and Anthropic and Jane Street 8%-10% each. Orders are difficult to cancel, and Nvidia has committed to taking back chips that cannot be deployed, so the financial risk exposure “is not as large as people imagine.”
The supply landscape is being rewritten by non-Nvidia chips such as TPU and Trainium: TPU shipments were just over 1M in 2025—Kenny gave examples of 1.1M, 1.2M or 1.3M—around 4M this year, and could reach 10M-12M next year, approaching Nvidia in scale. Trainium could reach 4M-5M next year. TPU growth is creating FluidStack, “the Nebius of the TPU world”; Kenny sees the future’s three major Neoclouds as CoreWeave, Nebius and FluidStack. Google’s partnership with Blackstone to build an independent TPU cloud is an effort to push TPU into a broader ecosystem, unlike Nvidia’s earlier DGX Cloud attempt, which found no takers.
The biggest future variable is not hardware but the PaaS layer, especially the convergence of storage and databases: models may interact directly with hardware, and as large-model responses get faster, “I don’t think it’s because inference is getting faster … it’s because many of the answers have already been stored.” Some legacy storage companies may struggle; Kenny specifically highlights private company Vast Data, whose backlog could increasingly include hyperscalers, AI compute companies and Neoclouds, making storage’s TAM “very large.”
The closing analogy is deliberately grand: AI is the next outlet for America’s expanding white-collar workforce, similar to the cost-cutting and productivity cycle that moved large amounts of US manufacturing to China after China joined the WTO in 2001—“chips are labor; they live inside data centers,” while hyperscalers and Neoclouds are the employers, which is why Nvidia calls them AI Factories. But the infrastructure lesson remains unchanged: “the owner of the first hotel is very likely to lose money or go bankrupt,” and “infrastructure is never a winner-take-all industry.” Over the next 2-3 years, “definitely more than half” of Neocloud companies with a scale above $1B will be eliminated—and that will be healthy and inevitable.

Deep dive

1. Three Computing Eras: AI Cloud Is a Return to the Mainframe Paradigm

Kenny’s historical periodization runs from the 1950s to the late 1990s, when hardware dominated and enterprises bought and operated their own systems. After 2000 came the virtualization era, driven by internet giants: AWS and Alibaba Cloud succeeded by externalizing the internal compute needed for Black Friday and Singles’ Day peaks. In the mid-2010s, a single CPU delivered roughly 300 TFLOPS while enterprise applications might require only 10-20 TFLOPS; bandwidth and high-frequency distribution became the core competitive resources of the cloud era.
The AI cloud era “is completely the reverse, returning to the era of the large computers of the 1950s and 1960s.” The objective is no longer high-frequency distribution but centralized compute management that keeps hyperscale clusters operating at high performance; demand from training, inference, Agents and enterprise inference is “exploding at the same time.”

2. Why Neoclouds Exist: Three Reasons for “Same Origin, Different Lineage”

Kenny’s framing is that traditional cloud and Neocloud are “of the same origin but different lineage.” The first reason is specialized expertise in managing and optimizing extremely large, densely packed compute. The second is the incentive structure: hyperscalers are “both suppliers and competitors” to AI Labs. OpenAI worried that Microsoft “couldn’t possibly tilt all its resources toward it,” so it backed neutral compute providers such as CoreWeave that have no downstream application business and are unlikely to build one.
The third reason is capital allocation. Kenny has followed data centers for 15 years and says their unlevered returns have generally been around 8%-10%, versus ROIC of 40%-50% in internet companies’ core businesses. There is little reason to commit resources to a low-return, capital-intensive sector that also requires specialist expertise. As a result, in data-center markets outside China, hyperscalers account for roughly 30% of customers, while third parties still provide around 60%-70% of power and cooling.
Responding to a report cited by 曹卿云, Kenny said the numbers were reasonable and likely durable: Transformer-based deep-learning advertising models at Meta and Google account for more than 20% of compute, LLMs such as Anthropic and OpenAI account for roughly 30%, and the rest goes to enterprise inference and consumer-facing demand.

3. OpenAI’s Foresight and Anthropic’s Rigidity: A Strategic Divide That Has Been Validated

Sam Altman believed from day one that compute demand would be enormous, with inference “very likely 10x training or even higher.” OpenAI therefore began making large compute purchases and investments in 2022 and 2023. Dario Amodei was still criticizing OpenAI in interviews last year for buying so much compute without limit.
Anthropic’s annualized revenue has risen from roughly $9B at the end of last year to possibly $50B now and could exceed $100B by year-end. The compute shortage is already constraining revenue growth: “you can’t even call it elastic procurement anymore; it’s rigid procurement now.”
Kenny said Anthropic’s revenue exceeded OpenAI’s at one point early this year, coming close to 2x its level. But from the end of April onward, OpenAI’s early compute build-out gives it “a chance to catch up with Anthropic.”
The account of xAI’s retreat and compute monetization came from 曹卿云, who said xAI’s H100/H200 fleet has effectively turned the company into a Neocloud selling to Anthropic, with little likelihood that it will continue developing the business. Kenny expects OpenAI and Anthropic to have roughly equal overall compute supply once all of xAI’s capacity is online.

4. The Two New-Cloud Camps: “Fourth-Cloud” Players and Bare-Metal Providers

The margin math is straightforward: AWS earns roughly 40% margins, versus 25%-30% for pure hardware. The extra 10-15 percentage points come from storage, distribution, bandwidth, databases and cloud-management PaaS. Those software services represent roughly 30%-35% of revenue and can generate 80%-90% gross margins. The CoreWeave-Nebius camp wants to add compute optimization, KV Cache deployment and retrieval, serverless RL and other PaaS capabilities on top of the hardware. CoreWeave’s storage and KV Cache revenue could reach roughly $300M annualized by year-end and potentially $1B or $2B in the future.
The ceiling is lower because the customers—LLM companies and hyperscalers—already have strong PaaS capabilities of their own. For a Neocloud, PaaS reaching 15%-20% of revenue would “already be very good,” putting overall margins in the low 30s; pure AI GPU hardware generates roughly 25% margins.
The other camp includes Oracle, SoftBank Energy, crypto-mining convert IREN and private company Crusoe, along with FluidStack. These companies are currently focused mainly on bare racks and bare-metal rentals, making them closer to hardware lessors and more exposed to pressure from pricing, returns and funding availability. Kenny’s view is “comparatively neutral.” Oracle may pursue both models. Later in the discussion, Kenny describes FluidStack as the TPU version of Nebius.

5. AI-Native Cloud: One Term, Three Interpretations

Kenny expects CoreWeave and Nebius to become “the fourth and fifth clouds,” with the eventual path leading to AI-native cloud—an end-to-end redesign of workflows for Agentic workflows. But the term means different things to different companies. DigitalOcean approaches it as a product philosophy and has launched roughly 15 or 20 products; Nvidia and its closely aligned partner CoreWeave use a broader definition centered on managing large AI compute clusters.
Hyperscalers largely avoid the term. Their view is that the underlying model has not changed; the B2B workflow has simply shifted from people using software, databases and storage to machines using them on people’s behalf: “it used to be mainly about people, and now it’s mainly about machines.”
Market share depends on both demand and resources. Whichever model captures the largest future market share will pull its upstream suppliers along with it. The outcome will also depend on downstream customers’ growth, access to a stable powered shell—including power and cooling infrastructure—and the ability to secure the most advanced and appropriate chips.

6. A Rapidly Changing Supply Map: TPU Scale-Up and the Future Big Three

The host offered 2025 revenue figures: CoreWeave generated $5.1B, or 22% of the Neocloud market; Nebius and Lambda each generated roughly $500M. Kenny expects the structure to change quickly. TPU shipments were just over 1M in 2025—he gave examples of 1.1M, 1.2M or 1.3M—could reach roughly 4M this year and, in his forecast, 10M-12M next year, approaching Nvidia in scale. Amazon Trainium could reach roughly 2M this year and 4M-5M or more next year.
CoreWeave, Nebius and Lambda are largely part of the Nvidia ecosystem, with chips and cluster deployments built around Nvidia. Kenny expects TPU growth to lift FluidStack: “you can think of FluidStack as the Nebius of the TPU world.” The future’s three major Neoclouds could be CoreWeave, Nebius and FluidStack. Amazon, with Trainium at the center of its strategy, will also hold a seat, although it is itself a hyperscaler.
Lambda is more programming-focused and could pursue a more vertical, regional and specialized Neocloud model.

7. The CoreWeave Model: Lock in Long Contracts, Add Leverage, Cut Financing Costs

The model is classic infrastructure finance. With a commercial GPU life cycle of roughly 5-6 years, CoreWeave locks that period to a single top-tier customer, including Anthropic, OpenAI and newer large compute buyers such as Jane Street and HRT. “For every dollar spent, you can earn roughly three dollars back over the next 6 years”; residual value is merely “icing on the cake.”
The financing curve reflects growing market acceptance. CoreWeave’s first loan in 2023 carried a 14.5% rate and was high-yield debt; this year’s new financing carries only 6.2% and is investment grade. Unlevered IRR is roughly 23%-25%. Three years ago, levered returns might have been 35%-40%, or even roughly 30%-35%; today, levered returns are “comfortably above 50%.”
Traditional infrastructure projects such as railways, ports and highways generate roughly 5%-6% unlevered returns and 10%-12% levered returns, with the better projects reaching 15%. Kenny’s former employer Brookfield manages roughly $2T in assets and is a century-old Canadian manager of commercial real estate and infrastructure assets that has long used this model of stable cash flow and predictable returns.

8. The Nebius Model: Use Prepayments to Reduce Leverage

The team came from the founders of Yandex, formerly Russia’s largest search engine. After selling Yandex and relocating to Amsterdam, they moved into Neocloud. Their DNA is “very internet-minded”: they are willing to take on compute-price volatility but do not want to take on balance-sheet leverage risk.
Several large Microsoft and Meta orders were discussed on earnings calls, with prepayments reaching 40% or even 50%. Nebius itself mainly issues company-level convertible bonds, so its overall leverage remains low.
The trade-off is pricing: Kenny estimates that contract prices are discounted by roughly 20%-25%, but the large upfront payments still allow 35%-40% unlevered returns. When the host asked whether that leaves room for more leverage, Kenny said “yes,” but expects Nebius not to add much and to continue using the 40%-50% prepayment model. He emphasized that all of the returns cited above are based on net cash flow, because EBITDA excludes depreciation.

9. The Host’s First Challenge: Does Long-Term Debt Against a Depreciating Asset Hold Up?

曹卿云 asked whether the model was “using an asset that depreciates to borrow long-term debt,” given that a new GPU generation arrives every 1-2 years and useful life is only a few years.
Kenny called that “a market misconception about this business model.” Data-center equipment and power systems depreciate over roughly 15-20 years. Traditional-cloud CPUs were initially depreciated over 5-6 years, and some hyperscalers have now extended that to 7-8 years, which is not dramatically different from AI cloud.
The key is that the debt “is not backed by the residual value of the GPU hardware; it is backed by the value of the entire contract.” CoreWeave and Nebius generally sign 5-6-year contracts that consume most of a chip’s commercial value. After that, every additional day of use produces another day of earnings—after electricity costs, almost 100% profit. Banks underwrite the contract, operations and cash flow as a whole, explaining the sharp decline in CoreWeave’s interest costs over the past 3 years.

10. The Margin Cliff: Construction Capacity Is Outrunning Rented Capacity

The public-market puzzle is familiar: CoreWeave’s margin fell from 12%-14% in the second and third quarters of last year to 8% in the fourth quarter, then “fell off a cliff to 1%-2%” in the first half of this year. Why are margins falling when spot GPU prices rose 40% in the first quarter? Kenny cited his former company GDS, where the numbers looked similarly poor at its 2016 IPO because capacity under construction exceeded the footprint already generating rental income.
The numbers explain the gap. CoreWeave has 1GW of energized capacity and a total power footprint of 3.5GW, with roughly 1GW likely under construction. Of the 1GW that is energized, only about 800MW is actually generating revenue because there is a 2-3-month lag between power activation, customer rack delivery and server deployment. Capacity under construction and energized capacity not yet generating revenue therefore drag down blended margins.
By year-end, CoreWeave will likely have 1.8-2GW generating rent while continuing to build roughly 1GW; the rented portion should earn about 25% margins and the overall business about 12%-14%. By the end of next year, roughly 3GW could be online and generating rent, with construction capacity equal to about 30% of the existing revenue base, lifting margins into the low 20s. With higher-margin PaaS such as KV Cache and RL, margins could reach roughly 25% by the end of 2028 and gradually approach 30%.

11. The Valuation Gap and Terminal Margins: A Discount Is Justified, but Not This Large

The current mismatch is stark. Traditional cloud trades at roughly 20-25x earnings with about 30% growth; Kenny estimates that Neoclouds trade at only about 10x steady-state earnings despite profit growth above 100% and a possible 120%-130% CAGR over the next 3 years. He expects the market to recognize in advance that “new cloud is not actually that different from old cloud,” with similar chip-depreciation and earnings models.
Over the long term, new-cloud valuations should still sit somewhat below traditional cloud. CoreWeave and Nebius spend roughly $45M-$50M of capex per MW, versus $15M-$20M for traditional cloud. With proprietary chips, white-box networking and bandwidth equipment, traditional cloud could reduce that to $10M-$12M per MW. Neoclouds require more capital and their margins will “never reach” the roughly 40% level of traditional cloud.
The end state may be convergence. Traditional cloud’s packageable PaaS functions could face pressure from AI applications, users doing their own Web coding and small software developers, pushing traditional-cloud margins down to 30%-35%. Neocloud margins could rise to 25%-30%. The two would become “relatively close, but still not equal.”

12. The $90B-Plus Backlog: A Defense Built on Waterfall Cash Flows and Diversification

The host highlighted the risk: CoreWeave currently generates roughly $5B of revenue, guides to $12B-$13B, and needs to spend $20B-$30B on capex; roughly two-thirds of 2025 revenue came from Microsoft, alongside a large volume of OpenAI orders. Kenny called this the market’s “biggest misconception” about CoreWeave: revenue has no direct relationship with current-period capex. The structure is a waterfall—spend $30B or even $35B this year, recover roughly $90B-$100B over the next 6 years, and revenue should exceed capex around 2029.
Unlike AI model companies, whose capex typically needs to convert into revenue within 6-12 months, CoreWeave and Nebius spread contract value over 5-6 years. As long as the customer has no serious financial problems and the platform continues delivering, cash flow keeps accumulating.
The forward backlog is already diversified: Meta accounts for 35%, Microsoft roughly 20%, OpenAI roughly 20%, Anthropic 8%-10% and Jane Street 8%-10%. These orders have financial binding force and are not easy for customers to cancel. Another factor often overlooked by investors is Nvidia’s commitment to take back a substantial portion of the chips CoreWeave purchased if they cannot be deployed. As a result, “the financial risk exposure is not as large as people imagine.”

13. DeepSeek Anxiety and the Semiconductor Cycle: Efficiency Is a Requirement, Not a Threat

On the risk that algorithmic efficiency could weaken demand, Kenny returned to the underlying cloud logic: “retail becomes wholesale.” Hyperscalers use scale, technical capabilities and core-business cash flow to lower compute costs, keep part of the benefit and pass part to customers, prompting customers to move more demand onto the cloud. DeepSeek is “a great company,” but improved compute efficiency is normal; efficiency gains and falling token prices are necessary conditions for software, equipment and edge demand to flourish.
This year’s spot pricing is “abnormal,” but the root cause is the supply side’s memory of the cycle. Suppliers expanded aggressively during Covid, then faced what may have been the worst semiconductor downturn of the past 20-30 years. SK Hynix, SanDisk and Kioxia were “almost on the verge of bankruptcy” in 2022 and 2023, only to enter what may be the biggest upcycle of the past 20-30 years. Companies are therefore reluctant to expand aggressively.
Kenny remains cautious on the view that capex will grow 50%-60% next year. Much of that increase could come from components, especially higher memory prices, without a proportional increase in silicon content. A healthy market would see silicon content rise 30%-40% while capex rises 50%, alongside another sharp decline in token costs; that is what would unleash genuine demand.

14. The Triangle: Microsoft Becomes Peripheral, While Nvidia Wants CoreWeave to Show the “Factory-Built Supercar”

Kenny said this was the most opinionated part of his view: Microsoft has “already become relatively peripheral” in the system. It originally served as the intermediary between OpenAI and CoreWeave, essentially providing credit enhancement. Infrastructure ratings are closely tied to the customer, and OpenAI had enormous negative cash flow at the time. Without Microsoft’s backing, CoreWeave’s pricing to OpenAI could have been extremely high. Today, OpenAI has “basically become independent of Microsoft’s system.”
Nvidia’s objective is to build its own hardware-infrastructure cloud ecosystem. Large hyperscalers and internet companies—including ByteDance, Meta and Alibaba—will increasingly develop their own chips, so Nvidia needs a moat within its own system. The traditional-cloud playbook is hardware white-boxing: reduce dependence on OEMs and bring Taiwanese ODMs into design and integration earlier. Hyperscalers are like someone who buys a GTR from Initial D and modifies it themselves, using hardware optimization to create differentiation.
Nvidia’s philosophy is the opposite: “you should buy my supercar and not modify anything.” CoreWeave uses Nvidia’s full networking stack in some clusters and focuses only on PaaS-level optimization, creating “the real inference and training capability Nvidia wants to show the world.”
The exchange includes Nvidia owning roughly 11% of CoreWeave, priority access to chips and servers, software and PaaS-tool support, and a commitment to take back chips that cannot be delivered or deployed.

15. Google x Blackstone’s New TPU Cloud: Sell the Engine Without Bundling the Car

The logic is an ecosystem flywheel. Nvidia’s ecosystem is stronger than AMD’s because its chips have been optimized and deployed by hyperscalers, model companies and DeepSeek. Google wants more people to use TPU so ecosystem participants can optimize it together. But Google Cloud inherited the traditional-cloud era’s technology and PaaS, which “may not necessarily be accepted by the new generation of AI-native companies.” Google is therefore partnering with Blackstone and other well-capitalized financial groups to push TPU independently into a broader market.
The mirror-image example is Nvidia’s earlier DGX Cloud attempt. Customers worried that “if your cloud develops too quickly, you may stop selling me chips in the future and sell me only compute,” so they rejected the model from day one. Nvidia later stopped talking much about it. Google’s structure gives the market a choice: “I may like your engine, like a BMW or Mercedes, but I may not like the design philosophy or interior of your car.”
Asked whether CoreWeave’s close Nvidia relationship could become a disadvantage, Kenny said that based on his conversations with management, “I don’t think they see themselves as completely exclusive to any one company,” although CoreWeave internally does believe Nvidia chips are the optimal solution.

16. The Railroad Lesson: The Assets Remain, the Companies Disappear, and Infrastructure Is Never Winner-Take-All

The host raised the 19th-century railroad analogy: a company may go bankrupt, but the infrastructure remains. Kenny agreed and extended it to all capital-intensive assets. The joke inside Brookfield is that “the owner of the first hotel is very likely to lose money or go bankrupt; the second owner generally breaks even; only the third owner starts making money.” In 2008-09, Australia’s largest infrastructure asset manager, Babcock & Brown, went bankrupt, and Brookfield bought roughly 20%-25% of Australia’s railway, port and infrastructure assets at a deep discount.
The structural conclusion is that “infrastructure is never winner-take-all, because that is almost impossible.” Competitive differences are not large enough; as long as capital is cheap and return expectations are low—even 2%-5% without leverage—a player can secure a position. AI chips, however, do not share the 40-50-year depreciation profile of railways. Their residual value may be low after 6 years, and the assets do not exist forever.
The shakeout is coming. The compute chain is currently “a hundred flowers blooming—there are already too many to count on two hands.” Over the next 2-3 years, “definitely more than half” of Neocloud companies with more than $1B in scale will be eliminated. Pressure will come from the three clouds, new entrants such as CoreWeave, Nebius and FluidStack, and whether downstream customers themselves survive. “But I think this is healthy and inevitable.”

17. The Biggest Future Variable Is the Storage Layer: The Answers Have Already Been Stored

Kenny moved past hardware and identified the real variable as PaaS. Differences at the IaaS layer have not been especially large: Microsoft favors smaller, lighter deployments with more nodes, while Amazon favors large data-center deployments. Hardware and physical-layer differentiation is limited; the competition is concentrated in PaaS.
Existing databases, visualization tools and cloud-management services are primarily designed for people, and may matter less to machines. Kenny expects “a major future direction to be the combination of the storage layer and the database layer.” Models may interact directly with hardware, while the storage layer runs through training, inference, GenAI workloads, enterprise workloads, consumer-facing workloads and potentially local deployments.
His evidence is that many questions may already have been asked tens of thousands of times and the underlying data sources are nearing saturation, while large-model answers are getting faster: “I don’t actually think it’s because inference is getting faster, because no new cards have really appeared. I think it’s because many of the answers have already been stored.” Whoever can optimize the storage of information and answers efficiently, frequently, at scale and low cost will address a “very large” TAM.
On investable names, Kenny said some traditional storage companies may struggle, including Dell’s EMC, NetApp and Pure Storage. He specifically highlighted private company Vast Data, which has deep experience in low-latency storage. Before the AI era, Vast Data’s main customers were high-frequency trading firms; today, many leading hyperscalers, AI compute companies and Neoclouds could become part of its backlog.

18. Digital Labor: AI Is the Next Outlet for America’s Expanding White-Collar Workforce

Kenny’s historical framework starts with US inflation reaching the low double digits or higher in the 1980s and a shortage of manufacturing labor. Manufacturing first moved to the Asian Tigers and other regions; after China joined the WTO in 2001, the US moved large amounts of manufacturing to China. From 1985 to today, the US has lost roughly 7M manufacturing jobs, while Asia added roughly 120M-150M; costs fell by about 60%. Over the same 30 years, the US added roughly 40M white-collar jobs. By 2010, the US services sector employed roughly 130M people, including about 50M core white-collar workers.
After 2012, the supply of white-collar labor also became insufficient. The second cost-cutting and productivity push came through cloud computing, software and SaaS. Cloud saved some costs but could not eliminate the existing 50M white-collar workers or their wage bill. SaaS completed the transition from on-premise software to recurring billing, with system lock-in and accumulated data creating stickiness while software companies continued to raise prices.
Kenny sees AI as America’s next solution to the expansion of knowledge labor. Starting next year, AI may replace some labor or even most labor; as its usefulness and reliability gain acceptance, the process could continue for years or even a decade. The digital-economy mapping is simple: “chips are labor; they live inside data centers,” while the employers are hyperscalers and Neoclouds that buy the chips and pay the data-center rent. That is why Nvidia calls them AI Factories. China experienced similar bottlenecks around water, power and labor in the 1990s and early 2000s; optical interconnect and other networking equipment play the role of highways and telecom companies. Chips do not sleep, but they require more electricity.
Kenny explicitly set the boundary of the theory: it is “based only on the capabilities of AI today.” What we use today may represent just 1% of AI’s future capabilities, or even less than 0.1%; we may still be seeing only the “first shoots” of AI’s development.

Inside Legora's Tech Stack: Why Token Maxing is Failing Enterprise Startups | Legora CTO Dan Loeb: The Lost Art of Short Selling, and Why Stock Picking is Back