Gavin Baker on AI Stock Selloff: Why the Data Still Looks Strong

Last updated: August 6, 2026, 3:46 a.m. ET (9:46 a.m. CEST)

The sharp decline in artificial-intelligence and semiconductor stocks during July 2026 created a striking split between public-market prices and the operating evidence coming from the companies building, selling, and using AI infrastructure. The Philadelphia Semiconductor Index entered bear-market territory in mid-July, high-profile AI names fell rapidly, and investors began questioning whether the industry had crossed from productive investment into a debt-financed capital cycle. Yet the same earnings season produced accelerating cloud growth, record data-center revenue, rising operating cash flow, expanding order backlogs, and fresh evidence that advanced memory and computing capacity remained scarce.

That contradiction is the central issue behind investor Gavin Baker’s argument that markets may be pricing the AI buildout incorrectly. Baker, the founding partner and chief investment officer of Atreides Management, described July as “2022 in a month”: a compressed cycle of narrative shocks, valuation compression, and indiscriminate selling. His case is not that every AI stock is cheap, that every data center will earn an adequate return, or that the industry is immune to a bust. It is narrower and more testable. He argues that the most important demand indicators—cloud growth, token consumption, GPU availability, memory pricing, and hyperscaler operating cash flow—have generally strengthened even as share prices weakened.

The available evidence partly supports that view. Microsoft reported 43% growth in Azure and other cloud services for the quarter ended June 30, 2026. Amazon Web Services grew 37%, its fastest expansion in 18 quarters. Alphabet said Google Cloud revenue rose 82%, while its backlog reached $514 billion. Nvidia’s most recently reported quarterly revenue rose 85% from a year earlier, with data-center revenue up 92%. Meta’s revenue grew 28%, and its operating cash flow reached $31.86 billion in the second quarter.

But the same evidence also explains the market’s fear. Meta generated only $784 million of free cash flow after $31.08 billion of capital expenditures and finance-lease payments. Alphabet produced negative quarterly free cash flow of $5.9 billion after $44.9 billion of capital spending. Amazon’s trailing-12-month free cash flow swung to an outflow of $7.6 billion as property-and-equipment purchases surged. Microsoft remained strongly free-cash-flow positive, but its quarterly capital expenditures reached $41 billion including finance leases. The buildout is producing revenue and operating cash flow, but it is consuming capital at a scale without precedent in the technology sector.

The correct interpretation is therefore not that the selloff was irrational or that the AI boom is certain to continue. It is that investors are debating the wrong level of the income statement. The bullish case focuses on demand, revenue growth, utilization, and operating cash flow. The bearish case focuses on capital intensity, financing costs, technological obsolescence, regulation, and the possibility that competition drives token prices down faster than usage grows. Both can be true at once.

Key Takeaways

• Main development: AI and semiconductor stocks sold off sharply in July even as major cloud and infrastructure companies reported accelerating demand and operating cash flow.
• Best evidence for the bullish case: Azure grew 43%, AWS grew 37%, Google Cloud grew 82%, and Nvidia’s latest data-center revenue increased 92% year over year.
• Best evidence for the skeptical case: Meta, Alphabet, and Amazon showed severe pressure on free cash flow as AI-related capital expenditures absorbed most or all of their operating cash generation.
• Most important market question: Whether installed compute can be repriced and utilized at levels high enough to finance future capacity without a destabilizing increase in debt.
• Most important nonfinancial risk: Regulation of data centers, electricity use, water, exports, and AI deployment could slow or redirect the buildout even if commercial demand remains strong.
• What comes next: Nvidia’s second-quarter fiscal 2027 results on August 26, 2026, will provide the next major public test of pricing, demand, margins, and Blackwell-to-Vera-Rubin execution.

What Happened in the July 2026 AI Stock Selloff

The selloff did not begin with one earnings miss or one definitive collapse in demand. It developed through a sequence of events that investors interpreted as evidence that the AI trade had become overcrowded, overfinanced, and vulnerable to cheaper alternatives. By July 17, Reuters reported that the Philadelphia Semiconductor Index had fallen 20% from its June record, meeting the conventional definition of a bear market. The decline spread beyond Nvidia to memory suppliers, semiconductor-equipment makers, networking companies, AI-cloud operators, and software businesses whose valuations depended on continued enthusiasm for artificial intelligence.

The first source of anxiety was competition from Chinese open-weight models. Moonshot AI’s Kimi K3 and Z.ai’s GLM-5.2 strengthened the argument that highly capable models could be offered at lower prices and with downloadable weights. For investors focused on the model layer, this looked like an attack on the economics of proprietary systems. If customers could obtain comparable performance from open models, the reasoning went, revenue and gross margins at the frontier labs might compress. That concern was amplified by public token indexes that appeared to flatten as usage shifted among providers.

The second source of anxiety was Meta’s reported plan to sell excess AI computing capacity through a cloud business. The initial interpretation was straightforward: if Meta had excess capacity, perhaps the industry had overbuilt. That reading ignored an alternative explanation. Meta could have seen an opportunity to monetize a portion of its installed base at attractive market prices while continuing to raise internal capacity. Its second-quarter results later showed that the company had not reduced capital spending. Instead, it reported $31.08 billion of quarterly capital expenditures and finance-lease principal payments, while raising or maintaining an aggressive infrastructure outlook.

The third source of anxiety was the credit market. Amazon, Alphabet, Meta, Oracle, and other hyperscale companies increased bond issuance as AI spending rose. Reuters calculated that Amazon, Alphabet, Meta, and Oracle had issued about $194 billion of bonds in 2026 through July 7, 79% more than their issuance during all of 2025. Corporate-bond spreads widened, credit-default-swap prices rose, and long-term Treasury yields increased. Even companies with enormous cash balances faced a higher cost of capital.

The fourth source of anxiety was China’s progress in semiconductor manufacturing equipment. Reuters reported on July 28 that China had begun producing its first domestically developed immersion deep-ultraviolet lithography machines. The tools were not described as immediate equivalents to ASML’s most advanced extreme-ultraviolet systems, and high-volume manufacturing reliability remained unproven. Even so, the report mattered because it suggested that export controls might delay rather than permanently prevent China from building a more independent semiconductor supply chain.

The fifth source of anxiety was simply valuation and positioning. AI-related stocks had delivered extraordinary gains, and many portfolios were exposed to the same group of securities. When investors began reducing risk, the selling became self-reinforcing. A fundamental concern about one part of the supply chain quickly became a general reassessment of every company associated with AI. The market did not wait for evidence that demand had declined. It priced the probability that demand, margins, or financing conditions could deteriorate.

That distinction matters. Markets discount future cash flows, not present demand alone. A company can report record revenue and still fall if investors believe the current economics represent a peak. Nvidia’s valuation debate illustrates the point. Its earnings have risen so quickly that the company’s forward price-to-earnings multiple compressed even when the stock price remained high. A lower multiple does not automatically mean the shares are undervalued. It can mean that the market expects earnings estimates to fall, margins to normalize, or capital intensity to migrate from customers back toward the supplier.

Baker’s argument is that the market moved too quickly from “risks are rising” to “the earnings are unsustainable.” The evidence supports the first statement. It has not yet proved the second.

The Central Thesis: Public Prices Versus Private Demand

Baker’s investment framework begins with a mismatch in visibility. Public investors can see the financial statements of Nvidia, Microsoft, Amazon, Alphabet, Meta, Micron, and other listed companies. They cannot see complete, audited operating data from private frontier-model developers, specialized inference clouds, AI-native software companies, or newly formed infrastructure providers. That creates a gap between where demand originates and where it is reported.

OpenAI, Anthropic, and other private labs purchase or reserve large amounts of computing capacity, but their financial disclosures are limited. Open-weight models generate usage across a fragmented group of providers, self-hosted deployments, and specialized inference platforms. Much of that activity does not appear in a single public-company revenue line. Public token trackers can capture a portion of API traffic, but they may miss private deployments, internal enterprise workloads, bundled services, and routing among models.

This is the “dark matter” problem in AI infrastructure. The gravitational effect can be observed through GPU shortages, cloud backlogs, memory contracts, networking demand, and data-center construction, even when the final application revenue is not fully visible. The risk is that investors confuse imperfect measurement with absent demand. The opposite risk is equally important: industry participants may mistake scarcity created by temporary bottlenecks for durable economic value.

The strongest version of Baker’s case contains four linked propositions. First, advanced compute remains scarce, and rental prices for high-end accelerators have not fallen as quickly as expected. Second, hyperscalers signed long-term contracts when compute prices were lower, so a portion of their installed capacity may be earning below current market rates. Third, as contracts reset and newer systems enter service, revenue and operating cash flow could rise faster than consensus forecasts. Fourth, that higher operating cash flow could finance much of the next phase of capital spending, reducing the need for destabilizing debt.

Each proposition is plausible, but none should be treated as established fact. GPU rental markets are opaque. A quoted hourly price can vary with chip generation, cluster size, networking, duration, power, software support, location, and credit quality. A short-term spot rental is not economically equivalent to a multiyear capacity reservation. A hyperscaler may have underpriced a contract, but it may also have accepted that price to secure a strategic customer or to improve utilization. Repricing can raise revenue, but customers can respond by optimizing workloads, adopting custom silicon, or shifting to cheaper models.

The most useful insight is not the claim that every contract will reprice upward. It is that the AI industry has moved from a chip-sales story to a utilization-and-cash-conversion story. The value of an installed GPU fleet depends on how many productive tokens it can generate, what customers will pay for those tokens, how efficiently the operator can serve them, and how long the hardware remains competitive. Those variables matter more than the headline number of GPUs purchased.

This also explains why operating cash flow has become a central metric. Capital expenditures reduce free cash flow immediately, but the resulting infrastructure can produce revenue over several years. A period of weak free cash flow is not necessarily evidence of poor economics if operating cash flow is accelerating and the assets earn attractive returns. It becomes dangerous when cash generation fails to catch up, when assets become obsolete faster than they depreciate, or when companies must continue spending merely to defend market share.

The Earnings Evidence: Demand Is Accelerating, but Cash Conversion Is Uneven

The latest earnings reports provide the most reliable test of the thesis because they separate narrative from audited or formally reported results. They show that AI demand is not a single market. It appears in cloud consumption, advertising performance, software subscriptions, model access, custom chips, networking, and infrastructure sales. The financial consequences differ sharply by company.

Microsoft: the clearest operating-cash-flow proof point

Microsoft’s quarter ended June 30, 2026, offered the strongest evidence that AI infrastructure can coexist with rising cash generation. Revenue reached $90.0 billion, up 18% year over year. Operating income rose 18% to $40.6 billion. Azure and other cloud services revenue increased 43%, and Microsoft Cloud revenue rose 27% to $59.3 billion. Commercial remaining performance obligations climbed 84% to $678 billion, giving the company unusually deep visibility into contracted future revenue.

Cash flow from operations reached $55.4 billion, up 30%. Free cash flow was $19.6 billion after capital expenditures of $41 billion including finance leases. That is the financial shape the bullish thesis requires: capital spending is enormous, but operating cash flow is growing fast enough to keep the company strongly free-cash-flow positive.

Microsoft is not a pure test of AI economics. Its cash engine includes Office, Windows, security, advertising, gaming, and other businesses. It also has strategic investments in private AI companies that can create accounting gains or losses. Still, Azure’s acceleration and the expansion of the company’s backlog are difficult to reconcile with a broad collapse in enterprise AI demand.

Amazon: cloud acceleration with negative trailing free cash flow

Amazon’s second-quarter results were simultaneously powerful and cautionary. Net sales increased 20% to $200.6 billion. AWS revenue rose 37% to $42.2 billion, its fastest growth in 18 quarters. AWS operating income increased to $16.6 billion from $10.2 billion a year earlier. Amazon also said its AI business and chips business had each exceeded annual revenue run rates of $25 billion.

The cash-flow picture was more difficult. Trailing-12-month operating cash flow rose 33% to $161.4 billion, but free cash flow fell to an outflow of $7.6 billion. Amazon attributed the deterioration primarily to a $66.1 billion year-over-year increase in purchases of property and equipment, net of proceeds and incentives, reflecting AI investment.

Amazon demonstrates why operating cash flow alone cannot settle the debate. The company is generating more cash from operations, but it is reinvesting even faster. That strategy can create enormous value if AWS demand remains strong and Trainium, networking, and data-center assets earn attractive returns. It can destroy value if pricing falls, utilization disappoints, or customers avoid lock-in by spreading workloads across providers.

Alphabet: extraordinary cloud growth and a quarterly free-cash-flow deficit

Alphabet reported consolidated revenue of $119.8 billion, up 24%, and operating income of $40.8 billion, up 30%. Google Cloud revenue increased 82% to $24.8 billion, while cloud operating income more than tripled to $8.8 billion. The cloud backlog reached $514 billion, up more than $50 billion sequentially. Alphabet also began recognizing revenue from TPU system sales to customer data centers, extending its business beyond rented cloud capacity.

Operating cash flow was $39.1 billion in the quarter. Capital expenditures were $44.9 billion, of which roughly 60% went to servers and 40% to data centers and networking. The result was negative quarterly free cash flow of $5.9 billion, although trailing-12-month free cash flow remained positive at $53.3 billion.

Alphabet’s numbers strengthen both sides of the argument. Cloud demand and profitability are accelerating rapidly. Yet the company spent more on capital assets during the quarter than it generated in operating cash. That does not prove overinvestment, but it places a heavy burden on future utilization and pricing.

Meta: strong revenue, almost no free cash flow

Meta’s second quarter was the most visible warning. Revenue rose 28% to $60.80 billion, and operating cash flow reached $31.86 billion. Capital expenditures, including principal payments on finance leases, were $31.08 billion. Free cash flow was only $784 million, down sharply from the prior year. Costs and expenses rose 55%, partly because of legal charges and severance, while operating income fell 8% and the operating margin narrowed to 31% from 43%.

Meta differs from Microsoft, Amazon, and Alphabet because it does not have a mature public-cloud business through which it can directly monetize excess capacity. Its core advertising products can benefit from AI through better recommendations, engagement, targeting, and creative tools, but the revenue connection is less transparent. A compute-resale business could create an additional path to monetization, yet it would also place Meta in competition with established cloud platforms.

Meta’s quarter therefore represents the key test of the “under-earning compute” thesis. If the company can convert its installed infrastructure into higher advertising revenue, subscription revenue, enterprise AI products, or external compute sales, the current free-cash-flow trough may be temporary. If it cannot, the spending will look increasingly like an arms race in which capital intensity rises faster than defensible profit.

Nvidia: record results before the next major test

Nvidia’s most recently reported quarter ended April 26, 2026. Revenue reached $81.6 billion, up 85% from a year earlier. Data-center revenue rose 92% to $75.2 billion. GAAP operating income increased 147% to $53.5 billion, and GAAP gross margin was 74.9%. The company guided to $91.0 billion of revenue, plus or minus 2%, for the quarter scheduled to be reported on August 26.

Those numbers show no evidence of a demand collapse at the supplier level. They also explain why investors worry about “over-earning.” Gross margins near 75%, extraordinary revenue growth, and a dominant market position attract competition. Customers are developing custom accelerators, open models are improving, and regulators are scrutinizing concentration. The market can believe that Nvidia remains the strongest company in the ecosystem while also expecting its long-term margins or market share to normalize.

All figures are reported company data in U.S. dollars. Free-cash-flow definitions vary by company and are not perfectly comparable.

Operating Cash Flow Is Improving—But It Is Not the Same as Return on Investment

One of the most important distinctions in the debate is the difference between operating cash flow and free cash flow. Operating cash flow measures cash generated by the company’s ordinary business activities before capital expenditures. Free cash flow generally subtracts purchases of property and equipment and, depending on the company, some lease-related payments. AI infrastructure makes that distinction unusually large because the assets are expensive and the investment cycle is front-loaded.

A hyperscaler can show accelerating operating cash flow while free cash flow declines. That does not mean the business is deteriorating. It means the company is choosing to reinvest a larger share of its cash generation. The investment can be rational if the future cash flows produced by the assets exceed the cost of capital and compensate for technological risk. It can be irrational if management overestimates demand, underestimates competition, or treats strategic fear as a substitute for financial discipline.

The bullish interpretation of the 2026 results is that the cash engine is already responding. Microsoft’s operating cash flow rose 30%. Amazon’s trailing operating cash flow rose 33%. Meta produced more than $31 billion in quarterly operating cash despite legal and severance charges. Alphabet generated $39.1 billion while cloud growth accelerated. These are not companies relying primarily on speculative debt or venture financing. They have large, profitable businesses capable of supporting substantial investment.

The skeptical interpretation is that aggregate spending is rising so quickly that even powerful cash engines may be insufficient. Capital expenditures are not only purchasing GPUs. They include data-center shells, land, power equipment, networking, storage, cooling, backup generation, grid connections, and long-lived assets that may take years to become productive. Construction delays can create periods in which cash is spent before revenue begins. Shortages can raise the cost of every component. The assets may also require continuing upgrades rather than a one-time investment.

Return on invested capital cannot be inferred from cloud revenue growth alone. A cloud provider may report higher revenue while earning a lower return if depreciation, energy, financing, and replacement costs rise faster. Conversely, reported free cash flow can understate long-term value creation during a period of unusually productive investment. The decisive evidence will emerge through margins, asset turnover, depreciation, contract pricing, utilization, and the durability of customer commitments.

That is why Baker’s claim that installed compute may be under-earning is so important. If older contracts reprice upward while the underlying assets remain useful, revenue can rise without an equivalent increase in capital spending. That would improve operating leverage and validate the buildout. If prices fall before contracts reset, or if newer chips make the installed base less competitive, the expected cash-flow acceleration may not arrive.

Investors should therefore resist two shortcuts. Weak free cash flow does not prove the AI buildout is a bubble, and strong operating cash flow does not prove the investments will earn adequate returns. The argument must be evaluated asset by asset and company by company.

Nvidia’s Valuation: Cheap Relative to Growth Is Not the Same as Cheap

The claim that Nvidia was trading at its lowest forward price-to-earnings multiple in a decade is central to the idea that the market expects a severe decline in earnings power. It also requires careful qualification. Forward multiples depend on the share price, the earnings-estimate source, the measurement date, and whether the calculation uses the next 12 months or a specific fiscal year. Different data services can produce meaningfully different numbers.

What can be established is that Nvidia’s earnings have grown much faster than its share price over important periods, causing its forward multiple to compress. Its first-quarter fiscal 2027 revenue rose 85%, while GAAP net income increased 211%. A company producing that rate of earnings growth can become less expensive on a forward-earnings basis even while its market capitalization remains enormous. At the close on August 5, 2026, Nvidia traded at $219.22, with a market capitalization of approximately $5.35 trillion and a trailing price-to-earnings ratio near 33 based on the market-data calculation available at that time.

A multiple in the low 20s or low 30s can appear modest relative to near-term growth, but the comparison is incomplete. The market is not valuing the next quarter in isolation. It is estimating how much of Nvidia’s current revenue, margin, and market share can persist once competition increases and the infrastructure cycle matures. A low forward multiple can signal undervaluation, or it can signal that analysts’ earnings forecasts are near a peak.

Nvidia’s current economics are exceptional. Its data-center business dwarfs its other segments. Its hardware is supported by CUDA, networking, systems, libraries, and an ecosystem that reduces deployment risk. The company can sell not only chips but integrated racks and increasingly complete AI-factory architectures. That breadth allows it to capture a larger share of the customer’s total infrastructure spending.

The bearish valuation case has several components. First, hyperscalers are developing their own accelerators, including Google’s TPUs and Amazon’s Trainium. Second, model efficiency can reduce the amount of compute required for a given task. Third, open-weight models and routing can move value away from premium proprietary tokens. Fourth, export restrictions can limit Nvidia’s addressable market. Fifth, customer concentration means a small group of buyers has significant negotiating power. Sixth, gross margins near 75% are likely to attract competition and regulatory scrutiny.

The bullish valuation case is that these risks are visible and already reflected in the multiple while demand continues to exceed supply. Custom chips do not necessarily eliminate Nvidia demand because the total market can expand faster than competitors gain share. Efficiency can lower cost per task and stimulate more usage. Open models still require compute. Export restrictions can delay revenue but can also strengthen demand in permitted markets. Customer concentration can be offset by the strategic importance of receiving Nvidia allocations on time.

The most defensible conclusion is that Nvidia’s valuation has become less demanding relative to reported growth, but not simple. The stock is pricing a future in which earnings growth slows materially. The argument between bulls and bears is about how quickly, why, and from what level.

GPU Prices, Contract Repricing, and the Hidden Economics of Installed Compute

Public investors often treat “the price of a GPU” as if it were one observable number. In practice, the market resembles commercial real estate more than a retail electronics store. A B200 cluster with high-speed networking, reliable power, specialized software, technical support, and a multiyear commitment is a different product from a single accelerator rented by the hour. Location, network topology, availability guarantees, contract length, and counterparty risk all affect the price.

Baker’s most striking evidence was anecdotal: companies that rented Blackwell capacity at rates in the mid-$2 range per GPU hour were later prepared to pay close to $4 for similar capacity. He also referred to inference providers expecting substantial increases when older contracts expire. These reports cannot be independently audited from public filings, and investors should not generalize from one negotiation. They are nevertheless consistent with broader evidence of scarcity: rising cloud backlogs, capacity constraints disclosed by Alphabet, large multiyear commitments at Amazon, and long-term memory agreements across the semiconductor supply chain.

The economics of contract repricing can be powerful. Suppose a cloud operator acquired GPUs and power under a multiyear customer contract priced when demand was uncertain. If the asset remains competitive and the contract expires into a tighter market, the operator can raise revenue without rebuilding the cluster. Depreciation and financing costs may be largely fixed, so much of the incremental revenue can flow through to operating profit and cash flow.

That mechanism is not guaranteed. The customer may move to a different provider, adopt a custom accelerator, optimize the model, or negotiate a lower rate in exchange for a longer commitment. New chip generations can make older systems less attractive. Electricity and memory costs can rise. The operator may need to upgrade networking or replace components. A headline increase in the hourly rate does not reveal the all-in return.

The age of the hardware is especially important. In traditional computing, older equipment usually becomes cheaper as newer systems offer better performance. AI has complicated that pattern because total demand has grown rapidly and because older GPUs can remain useful for inference, fine-tuning, data processing, and less latency-sensitive workloads. If a newer chip is scarce or expensive, an older system can retain value far longer than expected.

This is one reason the “air pocket” risk around Blackwell did not develop into a clear demand collapse. Customers could use older Hopper and Ampere systems while waiting for newer capacity. Newer systems could be deployed first for training or premium inference, with older fleets serving other workloads. The market became tiered rather than obsolete overnight.

Yet scarcity can also conceal poor capital allocation. If companies are willing to pay almost any price because losing access to compute threatens their competitive position, the market can sustain uneconomic investment for longer than a conventional cycle. The buyer may accept a low expected return because the alternative is strategic irrelevance. That game theory supports demand but does not prove that the eventual financial returns will be attractive.

The public indicators that matter are therefore not isolated hourly quotes. Investors should watch cloud gross margins, utilization disclosures, remaining performance obligations, depreciation, finance-lease commitments, and the relationship between incremental capital spending and incremental operating cash flow. Those figures reveal whether scarcity is becoming durable profit.

Open-Weight AI Is a Margin Threat and an Infrastructure Tailwind

The July selloff treated the rise of open-weight models as a negative for the entire AI ecosystem. That interpretation confuses the economics of the model layer with the economics of the infrastructure layer. Open models can pressure the prices and margins earned by frontier-model providers while increasing the number of tokens generated and the amount of compute consumed.

Kimi K3 illustrates the change. Moonshot AI described the model as a 2.8-trillion-parameter mixture-of-experts system with 104 billion activated parameters, native vision, and a one-million-token context window. Its published paper said the model remained behind the strongest proprietary systems overall but achieved frontier-level performance across coding, agentic work, reasoning, and vision. Full model weights allowed developers and infrastructure providers to customize and deploy the system without relying on one proprietary API.

Z.ai’s GLM family and Meta’s Muse Spark releases added to the competitive pressure. Meta introduced Muse Spark 1.1 in July as a multimodal reasoning model focused on agentic tasks, coding, computer use, and tool use. SpaceXAI launched Grok 4.5 for coding, knowledge work, and agentic tasks, pricing it below several premium competitors. These releases increased the number of capable models available to enterprises and developers.

For the frontier labs, the risk is clear. If a customer can route 40% or 60% of requests to a cheaper customized model, the premium provider loses volume or must reduce prices. Gross margins at the model layer can decline. The customer gains bargaining power. A software company can build proprietary data and workflow advantages instead of functioning as a thin interface over one external model.

For infrastructure providers, the effect can be positive. A cheaper token lowers the cost of experimentation and makes more applications viable. A company that reduces its average cost per task may run many more tasks. This is a version of the Jevons paradox: efficiency reduces the amount of a resource required for one unit of output, but total resource consumption rises because demand expands.

The relationship is not mechanical. A token from one model does not always require the same compute as a token from another. Model size, architecture, quantization, context length, batch size, hardware, software optimization, and latency requirements all matter. Some open models can generate a useful result with fewer or cheaper operations. Others may require more infrastructure because they are large or because self-hosted deployments run at lower utilization.

The important point is that lower model prices do not automatically imply lower infrastructure demand. If margin moves from the proprietary model provider to the customer and the cloud, aggregate compute consumption can continue to rise. The economic value can shift down the stack even when the industry’s total revenue grows.

Routing makes this shift practical. An application can send routine tasks to a customized open model, use a stronger model for planning or verification, and reserve premium frontier capacity for the hardest problems. The resulting system can produce better outcomes at a lower average cost. It can also consume more total tokens because the application performs more steps, checks more outputs, and runs more agents in parallel.

That pattern helps explain why model commoditization can be bullish for Nvidia, cloud platforms, memory suppliers, and networking companies. It broadens the set of users and reduces dependence on any one lab. It also creates a new risk: if efficiency improves faster than demand expands, infrastructure utilization can fall. The evidence in 2026 points to expansion, but the balance can change.

The Credit Risk Is Real Even When Companies Can Fund the Buildout

The most credible bearish argument in Baker’s discussion concerned credit. Technology bubbles become dangerous when capacity is financed with debt that requires repayment before the assets have proved their economics. The late-1990s telecommunications and internet buildouts offer the classic example: companies borrowed against optimistic demand assumptions, created excess capacity, and then faced a rapid collapse when pricing and utilization disappointed.

The 2026 AI cycle is different because the largest buyers generate extraordinary operating cash flow. Microsoft, Amazon, Alphabet, and Meta are not pre-revenue infrastructure startups. They can issue investment-grade debt, use cash on hand, or finance assets through leases. Their core businesses produce enough cash to absorb losses that would bankrupt a smaller operator.

That strength does not eliminate credit risk. The scale of the commitments is rising. Reuters reported that four major hyperscale issuers—Amazon, Alphabet, Meta, and Oracle—sold about $194 billion of bonds in 2026 through July 7, compared with roughly $108 billion during all of 2025. Investors demanded wider spreads, and credit-default-swap prices rose. Long-term real yields also increased, raising the discount rate applied to future infrastructure cash flows.

Higher financing costs affect the ecosystem unevenly. A hyperscaler with hundreds of billions of dollars in operating cash flow can continue spending. A neocloud, data-center developer, or startup may depend on project finance, customer guarantees, or asset-backed loans. If lenders become more cautious, the smaller provider may be unable to complete a project even when end-user demand exists.

This can strengthen the incumbents. Scarce credit restricts new supply, supporting rental prices for existing capacity. Nvidia GPUs are relatively financeable because they have an observable resale and rental market, broad software support, and many potential users. A custom accelerator or unproven architecture may be harder to finance because the lender has fewer options if the borrower defaults.

The same dynamic can create concentration risk. If only the largest companies can finance advanced infrastructure, the market becomes more dependent on a small number of buyers and providers. That can improve near-term pricing but increase regulatory pressure and systemic exposure. A change in spending by one company can affect memory suppliers, foundries, networking vendors, utilities, and construction markets.

Credit also changes management incentives. When a company funds a project from operating cash, it can tolerate a longer payback period. When a project carries debt, it must generate enough cash to service interest and principal on a schedule. That makes a temporary decline in utilization more dangerous. The project can be economically valuable over ten years and still fail financially in year two.

The key ratio is not total debt in isolation. Investors should compare debt and lease commitments with the cash flows generated by the assets, the duration of customer contracts, and the useful life of the equipment. A multiyear take-or-pay agreement can support financing. A speculative plan based on future spot prices is more fragile.

Baker’s bullish conclusion is that operating-cash-flow acceleration can remove hundreds of billions of dollars of potential credit demand. The latest results support the direction of that argument, particularly at Microsoft and Amazon. They do not prove that all announced projects can be funded internally. The buildout includes private labs, independent cloud operators, utilities, and data-center developers that do not share the same balance-sheet strength.

Memory Long-Term Agreements Are Changing the Semiconductor Cycle

High-bandwidth memory has become one of the most important constraints in AI infrastructure. GPUs and custom accelerators require enormous memory bandwidth to move model weights and intermediate data quickly enough to keep compute units productive. Adding theoretical processing power without enough memory can leave expensive accelerators underutilized.

Historically, memory was among the most cyclical parts of the semiconductor industry. Producers expanded capacity when prices were high, supply eventually exceeded demand, and prices collapsed. Customers benefited from short contracts and the ability to play suppliers against one another. The AI buildout is changing that relationship because access to advanced memory can determine whether a customer can ship an entire system.

Micron’s June 2026 presentation described 16 strategic customer agreements covering data-center, consumer, and automotive markets. The agreements generally run from 2026 through 2030 and include binding take-or-pay volume commitments. Micron said the signed agreements represented roughly 20% of its DRAM volume and one-third of its NAND volume over the period. Fourteen of the agreements had minimum cumulative revenue of approximately $100 billion, and expected customer deposits and related financial commitments totaled $22 billion.

The pricing structures are especially important. Micron said many agreements included floor and ceiling prices, while some used fixed prices or market-based terms. The floor protects the supplier from a conventional collapse, and the ceiling gives the customer some protection in a shortage. The contract converts a volatile spot relationship into a strategic supply partnership.

That structure supports Baker’s game-theory argument. A customer that breaks a long-term agreement to obtain a lower price during a temporary downturn risks losing future allocations when supply tightens. In a market where memory availability determines product shipments and AI capacity, the cost of being deprioritized can exceed the savings from renegotiation.

The argument should not be taken too far. Contracts can be amended, disputed, or economically impaired. A customer can redesign its systems, diversify suppliers, or reduce demand. New capacity can still create oversupply. Floor prices can protect supplier revenue while making the customer’s hardware less competitive. Long-term agreements reduce cyclicality, but they do not abolish it.

They also redistribute risk. The supplier gains visibility and can invest with more confidence. The customer secures allocation but may overpay if market prices fall. The financial system can lend against contracted cash flows, increasing the amount of capacity that can be built. The industry becomes more stable in one sense and more interconnected in another.

For Nvidia, reliable memory supply reinforces system-level control. The company’s products combine accelerators, memory, networking, CPUs, storage, and software. The performance of the whole system depends on coordinated access to each component. A competitor with a good chip design can still struggle if it lacks memory allocation, packaging, networking, software support, or financing.

Nvidia’s Moat Is Expanding From Chips to the Financing and System Layer

Nvidia’s strategic position is often described as a chip monopoly. That description is incomplete. The company’s advantage increasingly lies in coordinating a full system: accelerators, CPUs, networking, storage, software libraries, developer tools, model frameworks, supply relationships, and deployment partners. The more complex the AI factory becomes, the more valuable integration can be.

The Vera Rubin platform announced in March 2026 demonstrates the strategy. Nvidia described a system containing the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 networking, BlueField-4 data-processing units, Spectrum-6 Ethernet, and integrated Groq inference accelerators. The company said Rubin systems could train large mixture-of-experts models with one-fourth the number of GPUs required by Blackwell and deliver substantially higher inference throughput per watt.

Those are company performance claims, not independent guarantees. They nevertheless show how Nvidia is repositioning the competitive comparison. A rival cannot win merely by producing a chip with more theoretical operations per second. It must deliver a system that can be installed, financed, cooled, networked, programmed, and operated at high utilization.

This systems approach can increase revenue per gigawatt. Nvidia can sell more components into each data-center deployment and potentially participate in software, support, networking, and financing economics. It can also use equity investments and commercial partnerships to help expand the customer base for its products.

Baker described an emerging structure that resembles a credit wrapper with revenue participation above a pricing floor. Public details vary by transaction, and it would be inaccurate to describe every arrangement as vendor financing. Nvidia often does not lend the purchase price directly. Instead, its involvement, equity capital, hardware allocation, or commercial commitment can make a project more financeable for third-party lenders.

That model can align incentives. Nvidia benefits from more deployed capacity and may share in upside if rental prices remain strong. The operator obtains access to scarce hardware and financing. The lender gains comfort from the liquidity and broad usefulness of Nvidia systems. The customer receives capacity that might otherwise be unavailable.

It also creates risks. If Nvidia invests in customers that use the proceeds to expand purchases within its ecosystem, investors must distinguish genuine end demand from financially supported demand. Money is fungible even when contracts restrict its direct use. The company’s equity portfolio can create gains during a boom and losses during a downturn. Concentration across suppliers, customers, and financing partners can magnify shocks.

The strongest defense of the strategy is that Nvidia has unique visibility into demand. It works with frontier labs, cloud companies, enterprises, governments, and startups. It can observe model roadmaps, cluster requirements, and bottlenecks earlier than most investors. If management is willing to commit capital across the ecosystem, that is a meaningful signal of confidence—though not proof that the investments will earn attractive returns.

The strongest criticism is that a dominant supplier can extend a cycle by helping customers finance purchases that would not otherwise occur. The difference between ecosystem development and circular financing depends on transaction terms, independent customer demand, and cash generation. Investors should demand clear disclosure.

China’s DUV Progress Is Strategically Important but Not an Immediate ASML Replacement

China’s reported production of domestic immersion deep-ultraviolet lithography tools triggered another wave of semiconductor selling because it challenged a core assumption behind Western export controls. Advanced chip manufacturing depends on lithography equipment that transfers intricate patterns onto wafers. ASML dominates the most sophisticated systems, particularly extreme-ultraviolet lithography.

Reuters reported on July 28 that a Chinese state-backed group had begun producing domestic immersion DUV machines, with deliveries expected to leading Chinese chipmakers. The machines still required testing, and they were not described as matching ASML’s most advanced performance or reliability. The immediate commercial threat to ASML was therefore limited.

The strategic significance was larger. China had moved from research and prototypes toward production. Even a less advanced tool can support mature-node chips, memory, and repeated-patterning techniques. Domestic equipment can reduce vulnerability to tighter export restrictions and provide a platform for learning by doing.

Semiconductor equipment is not a field in which one successful prototype closes a 20-year gap. High-volume manufacturing requires precision, uptime, service, metrology, process integration, software, optics, and a supplier network that improves through thousands of production cycles. A machine that works in a laboratory is not automatically competitive in a fab where downtime can cost millions of dollars.

The market’s immediate reaction therefore looked larger than the near-term earnings effect. ASML’s order book was unlikely to disappear because of one report. U.S., Taiwanese, Korean, Japanese, and European manufacturers still depend on the company’s technology. China itself continued to purchase permitted foreign tools.

Investors were reacting to the direction of travel. Export controls create a powerful incentive for China to develop alternatives. State capital can support long development periods that private investors might reject. Progress in DUV, etching, deposition, metrology, and memory can gradually reduce foreign suppliers’ market share inside China.

For the AI-infrastructure thesis, the development cuts both ways. More Chinese semiconductor capacity can expand global compute supply and pressure prices over time. It can also create separate technology ecosystems, duplicated capital spending, and greater demand for equipment outside China as Western countries build resilient supply chains. Decoupling can be inefficient but capital intensive.

The correct conclusion is neither dismissal nor panic. China’s DUV progress is a real strategic milestone. It is not evidence that ASML’s advanced position will disappear in the next quarter or year.

Data-Center Regulation May Be the Most Underpriced Risk

Financial models generally assume that capital can be converted into energized computing capacity on a predictable schedule. Data centers make that assumption fragile. A company may secure GPUs and financing but still lack power, transmission, water, permits, turbines, transformers, labor, or community approval. These constraints can delay revenue while interest, lease, and construction costs continue to accumulate.

New York’s July 2026 moratorium made the political risk concrete. Governor Kathy Hochul signed an executive order temporarily pausing state environmental permits for new hyperscale data centers while regulators develop a statewide framework. The state said the review would examine effects on energy demand, water, air quality, and local communities. The administration also proposed changes to tax incentives and mechanisms requiring developers to contribute to grid expansion and community investment.

The order did not ban all data centers permanently. It created a pause of up to one year for projects requiring discretionary state permits, and it was aimed at establishing standards. Yet the market significance extended beyond New York. Other states and localities can adopt similar measures if residents believe data centers raise electricity bills, consume scarce water, or produce too few permanent jobs.

The industry’s response has often been too absolute. Supporters argue that new large customers can spread fixed grid costs across more electricity sales, finance generation, and lower average rates. Some academic work has found that U.S. data-center expansion modestly reduced average retail electricity rates during earlier periods, while warning that future supply constraints could reverse the effect. Opponents point to local transmission bottlenecks, capacity payments, water stress, noise, land use, and the risk that utilities build infrastructure for projects that are delayed or canceled.

Both outcomes are possible. A data center paired with dedicated generation, storage, transmission upgrades, and a contract protecting ratepayers can improve local economics. A speculative cluster added to an already constrained grid can raise costs and reliability risks. National averages do not resolve local conditions.

Water is similarly site specific. Some facilities rely heavily on evaporative cooling; others use closed-loop or air-cooled systems. The indirect water used in electricity generation can exceed on-site consumption. A project in a water-rich region has a different impact from one in a drought-prone area. Claims that data centers have “no environmental impact” are not supportable, but neither are generalized claims that every facility threatens local water supplies.

Employment claims also require precision. Data-center construction creates large numbers of skilled jobs, and ongoing operations require technicians, electricians, security, maintenance, and supply-chain support. The permanent on-site workforce can be small relative to the capital invested. Community benefits depend on tax revenue, local procurement, wage levels, infrastructure commitments, and whether the project shifts costs to residents.

Regulation can reduce supply and make existing compute more valuable, which may support the bullish pricing thesis. It can also prevent companies from deploying the capacity required to meet demand, slowing revenue and creating stranded inventory. The distributional effects matter: incumbents with permitted sites and power contracts gain an advantage, while new entrants face higher barriers.

The investment implication is that megawatts are not interchangeable. A gigawatt announced in a press release is not the same as a gigawatt under construction, energized, connected to hardware, and generating billable output. Investors should discount projects according to permit status, power certainty, construction progress, equipment delivery, and customer commitments.

SpaceX Adds a New Public-Market Test of the AI Infrastructure Thesis

SpaceX’s transition into a publicly traded company introduced another large-scale experiment in AI infrastructure. The company’s first public quarterly results arrived after the interview was recorded, making them an important subsequent development. Reuters reported that the results exceeded several expectations but that the shares fell sharply as investors focused on AI-related capital spending and negative free cash flow.

SpaceX differs from a conventional cloud provider. Its core businesses include launch services, Starlink connectivity, satellite manufacturing, and increasingly AI infrastructure through the combined SpaceXAI organization. That integration creates potential advantages in land, power procurement, engineering, network connectivity, and eventually orbital systems. It also makes the financial statements harder to interpret because profitable and capital-intensive projects coexist inside one company.

The reported quarter showed the same tension visible elsewhere in the sector. AI revenue was growing rapidly, and management cited new cloud contracts. Capital expenditures, however, reached a level that alarmed investors. The share-price decline demonstrated that public markets were unwilling to reward growth without a clearer path to cash conversion.

Baker’s bullish view of SpaceX centered on execution. The company had demonstrated an ability to bring large amounts of compute online quickly and at comparatively low cost. If it could repeat that performance, it might monetize infrastructure at attractive rates while using Starlink and launch cash flows to fund expansion. Its relationship with Nvidia and its ability to integrate AI products could improve utilization.

The skeptical case is substantial. Building data centers is different from launching rockets, even when some engineering disciplines overlap. The company must manage power, cooling, networking, software, customer service, and financial risk. Capital spending competes with Starship, Starlink satellites, launch infrastructure, and other projects. An organization known for ambitious timelines can create valuation risk when public investors expect quarterly evidence.

Orbital compute is the most speculative part of the story. The theoretical advantages include abundant solar energy, avoidance of terrestrial permitting, and the ability to process space-generated data before transmitting results to Earth. The disadvantages include launch cost, radiation, maintenance, thermal rejection, hardware replacement, network limits, and the difficulty of dissipating heat in a vacuum.

Independent analyses have reached cautious conclusions. Orbital inference for space-native data can make sense because processing images or sensor data in orbit reduces downlink volume. Frontier-model training for terrestrial customers is much harder because training requires dense, reliable networking and frequent movement of enormous datasets. Radiators, power systems, and replacement cycles can erase the apparent energy advantage.

SpaceX has unique advantages if orbital compute becomes viable: launch capacity, Starlink laser links, satellite manufacturing, and internal demand. Those strengths justify serious attention. They do not make orbital data centers a near-term substitute for terrestrial hyperscale campuses.

For investors, SpaceX is useful because it exposes the AI-buildout thesis to public-market discipline. The company must now demonstrate that rapid infrastructure deployment produces cash returns, not only technological milestones.

Efficiency Improvements Could Expand the Market—or Break the Capital Cycle

AI infrastructure demand is often modeled as a straight line from better models to more compute. The reality is more complicated. Algorithmic improvements can reduce the compute needed to train or serve a model. Better quantization, sparse architectures, caching, speculative decoding, routing, and specialized accelerators can lower cost per useful output. Continual learning and more sample-efficient training could reduce the need to retrain enormous models from scratch.

This creates a genuine threat to hardware demand. If a model can achieve the same capability with one-tenth the training compute and if total model development remains constant, accelerator demand falls. If inference can serve the same user activity with fewer memory transfers and lower precision, the installed fleet becomes more productive and fewer new systems are required.

The bullish response is that productivity increases usually expand the market. Lower costs enable more users, longer context windows, richer multimodal input, more agents, and higher-quality verification. The industry can move from occasional chatbot use to persistent systems that plan, monitor, simulate, code, negotiate, and operate tools. Each user can consume far more compute even as the cost of one task declines.

Which effect dominates depends on elasticity. If a 50% reduction in cost causes usage to more than double, total spending rises. If usage increases by only 20%, spending falls. Early AI markets have shown high elasticity because new capabilities create new applications, but elasticity can decline as the technology matures.

Specialized hardware adds another layer. Nvidia’s integration of SRAM-based inference accelerators through Groq illustrates how workloads can be divided among architectures. Prefill, attention, and feed-forward operations have different memory and compute characteristics. A system that routes each stage to the best hardware can increase throughput and reduce cost.

That can be positive for the total infrastructure market while changing who captures the value. Nvidia may sell or integrate multiple accelerator types. Memory suppliers can benefit from more complex hierarchies. Cloud providers can improve utilization. A standalone GPU competitor may find that the market shifts toward systems rather than chips.

The risk for investors is extrapolation. A company can report strong demand for today’s architecture just as a new system changes the cost curve. Long lead times mean suppliers must commit capital before they know which model architectures and workloads will dominate. Long-term agreements reduce financial volatility but can lock customers into hardware assumptions that age quickly.

The most important technical question is not whether AI becomes more efficient. It will. The question is whether the number and complexity of economically valuable tasks grow faster than efficiency improves. Public revenue and utilization data—not benchmark headlines—will answer it.

Where the Economic Value Must Ultimately Come From

Infrastructure spending cannot compound indefinitely without customers who earn or save money from AI. The cash paid to Nvidia, cloud providers, memory suppliers, and data-center operators must ultimately come from higher productivity, new revenue, labor substitution, consumer spending, government budgets, or a transfer of profit from another industry layer.

The most constructive scenario is productivity-led growth. Companies use AI to produce more output, improve products, reduce errors, speed research, and create services that were previously impossible. Revenue grows, wages can rise, and spending on compute becomes a share of a larger economic pie.

The more disruptive scenario is direct labor substitution. A company keeps output constant while reducing headcount and redirecting part of the savings to tokens and software. Compute spending can rise even if economic growth does not accelerate. The benefits accrue to owners, customers, and high-productivity workers, while displaced workers bear the adjustment cost.

AI-native startups provide early evidence of both effects. Some reach substantial revenue with smaller teams than earlier software companies. They spend a larger share of compensation-equivalent budgets on models and infrastructure. That can indicate superior productivity, but it can also reflect venture-funded experimentation before durable demand is proven.

Large incumbents are adopting AI more slowly and unevenly. Technology companies can integrate coding agents quickly. Banks, health systems, manufacturers, and regulated enterprises face data, security, governance, and workflow constraints. International adoption varies with regulation, language, cloud availability, and labor costs.

This staggered adoption supports a long runway. A small group of advanced users can create a compute shortage before the majority of businesses deploy agentic systems at scale. If adoption spreads, demand can remain strong for years. If advanced users discover that agents produce limited economic value outside narrow tasks, the shortage can disappear before mainstream adoption arrives.

Company disclosures are beginning to connect AI spending with business outcomes. Microsoft reported growth in paid Copilot seats and cloud commitments. Amazon said its AI and chip businesses exceeded large revenue run rates. Alphabet reported rapid cloud growth and new TPU sales. Meta linked AI to advertising performance and product engagement. These are meaningful signals, but they do not yet provide a comprehensive return-on-investment calculation.

Investors should ask what customers are buying, not only how much capacity providers are building. Usage that improves a measurable business process is more durable than promotional experimentation. Renewal rates, expansion within accounts, price realization, and gross retention will matter more than pilot announcements.

The Strongest Skeptical Case Against the AI Infrastructure Thesis

A serious skeptical argument does not require predicting that artificial intelligence will fail. It requires showing that too much capital can chase a transformative technology and still produce poor investment returns. Railroads, telecommunications networks, and the internet created enormous social value while bankrupting many investors.

Capital spending may be based on competitive fear rather than customer returns

Frontier labs and hyperscalers face a prisoner’s dilemma. If one company slows spending while competitors continue, it risks losing model quality, customers, talent, or strategic relevance. All companies therefore invest aggressively even if each privately doubts the industry’s eventual returns. Rational behavior at the company level can create overcapacity at the industry level.

Token prices may fall faster than usage grows

Competition among proprietary and open models is reducing prices. Routing and optimization can lower the average cost per task. If customer budgets remain fixed, more tokens do not necessarily produce more revenue. Infrastructure providers can face declining unit economics even while usage rises.

Hardware can depreciate economically faster than accounting schedules assume

A server may remain operational for six years but become commercially unattractive much sooner if a new generation offers dramatically lower cost per token. Accounting depreciation does not capture the full economic obsolescence. Operators can be forced to reduce prices on older fleets or spend more to upgrade networking and memory.

Power and construction constraints can produce stranded assets

Companies can order chips before sites are ready. Delays in transformers, turbines, permits, transmission, or cooling can leave expensive equipment idle. Conversely, developers can build shells and power infrastructure before securing enough accelerators or customers. Either mismatch reduces returns.

Debt and leases can turn a slowdown into a crisis

Investment-grade hyperscalers can survive weak utilization. Smaller operators with project debt, equipment financing, or lease obligations may not. A few defaults can cause lenders to retreat, reduce new supply, and force asset sales. Falling secondary-market prices can then weaken collateral values across the sector.

Regulation can alter the economics after capital is committed

New rules on energy, water, exports, privacy, model safety, copyright, or competition can reduce utilization or raise operating costs. A facility designed for one jurisdiction or customer may not be easily repurposed.

Demand data is incomplete and often promotional

Private companies disclose selected growth figures, annualized run rates, or token volumes without audited context. Infrastructure providers have incentives to emphasize shortages. Model developers have incentives to publicize benchmarks. Investors may be observing a true boom through noisy, self-interested data.

These risks make the bearish case credible even when current demand is strong. The key question is whether the market has priced them proportionately. A 40% or 60% decline in a speculative AI stock can be rational if its previous valuation assumed flawless execution. The same decline in a profitable infrastructure leader can be excessive if cash generation continues to accelerate.

What Would Prove the Bullish Thesis Wrong

A useful investment thesis needs observable conditions that can falsify it. The following developments would materially weaken the argument that the AI selloff has moved ahead of the fundamentals.

• Sustained declines in GPU rental prices: Not a temporary discount or a lower price on older chips, but broad and persistent declines across comparable high-end clusters.
• Cloud growth deceleration without margin improvement: Slower Azure, AWS, and Google Cloud growth combined with continued high capital spending would indicate weaker returns.
• Falling operating cash flow: If operating cash flow stops accelerating before capital expenditures normalize, the internal-financing argument deteriorates.
• Backlog cancellations or weaker commitments: A decline in remaining performance obligations, take-or-pay contracts, or customer deposits would challenge the scarcity narrative.
• Material excess inventory: Rising GPU, networking, or HBM inventories and lower supplier utilization would show that orders had moved ahead of real deployment.
• Large impairment charges: Write-downs of data centers, chips, or capitalized infrastructure would provide direct evidence that expected returns had fallen.
• Private-lab revenue plateaus: Verified slowing at OpenAI, Anthropic, SpaceXAI, and major open-model inference providers would weaken the demand source behind public infrastructure spending.
• Credit stress among neoclouds: Defaults, restructurings, covenant breaches, or emergency equity raises would show that project economics were not supporting debt.
• Regulatory barriers spreading across major markets: Moratoriums, energy restrictions, export controls, or model-use limitations could reduce the deployable market.
• Efficiency gains overwhelming demand growth: If token volumes rise but revenue and compute hours decline, the Jevons-style expansion thesis would fail.

What Investors Should Watch Next

The next phase of the debate will be decided by a small group of measurable indicators rather than by general enthusiasm or fear.

Nvidia’s August 26 report

Nvidia’s fiscal 2027 second-quarter results will test whether the company achieved its $91 billion revenue outlook, maintained gross margins near 75%, and continued converting Blackwell demand into shipments. Guidance will matter as much as the reported quarter. Investors will also look for evidence about China, networking, memory supply, and the Vera Rubin transition.

Hyperscaler capital-expenditure guidance

Spending plans will reveal whether the July selloff changed management behavior. Cuts would support the view that companies had overbuilt or that financing conditions had tightened. Continued increases would show confidence but could intensify free-cash-flow concerns.

Operating cash flow relative to capital spending

The ratio between incremental operating cash flow and incremental capital expenditures is the most direct financial test. Microsoft currently shows the best balance. Meta and Alphabet need to demonstrate that revenue and cash generation catch up with the infrastructure already purchased.

Cloud backlog and capacity constraints

Microsoft’s $678 billion commercial remaining performance obligation and Alphabet’s $514 billion cloud backlog indicate strong contracted demand. Investors should watch recognition schedules, cancellations, and whether companies continue using third-party capacity because internal supply is constrained.

Memory contract performance

Micron’s strategic customer agreements and SK Hynix’s partnerships will reveal whether long-term contracts genuinely reduce cyclicality. Deposit collection, volume commitments, pricing bands, and margins will show how risk is shared.

Credit spreads and project finance

Wider spreads do not automatically stop investment, but they raise required returns. Bond issuance, finance-lease growth, data-center project loans, and neocloud funding terms will indicate whether the industry can finance announced capacity.

Regulatory approvals and power delivery

Announcements should be separated from energized capacity. Investors should track environmental permits, interconnection agreements, turbine deliveries, transformer availability, and actual megawatts placed into service.

Open-model economics

The critical question is whether open-weight adoption expands the total market. Infrastructure revenue, token volume, customer spending, and provider margins should be evaluated together. More tokens at much lower prices can be positive or negative depending on elasticity.

Why the 2022 and Dot-Com Comparisons Are Useful—but Incomplete

Baker’s description of July as “2022 in a month” captures the speed and emotional character of the decline. In 2022, investors confronted rising inflation, aggressive monetary tightening, falling valuation multiples, and doubts about technology demand after the pandemic. The 2026 selloff compressed several versions of that fear into a shorter period: higher real yields, widening credit spreads, concern about overspending, and a rapid reassessment of long-duration growth assets.

The comparison is useful because valuation can fall even when revenue continues growing. A higher discount rate reduces the present value of future cash flows. Companies with profits expected many years in the future are especially sensitive. In 2026, however, the largest AI infrastructure companies were already generating substantial current earnings and operating cash flow. This was not simply a group of unprofitable software companies valued on distant possibilities.

The late-1990s internet comparison is also unavoidable. Telecommunications carriers and startups spent heavily on fiber, servers, and network capacity because internet traffic was certain to grow. They were correct about the technology and often wrong about timing, pricing, financing, and who would capture the value. Capacity expanded faster than monetization, prices collapsed, and debt turned an operating disappointment into insolvency.

AI infrastructure shares several features with that period. The technology is transformative. Capital spending is large and visible. Competitive fear encourages companies to build before demand is fully proven. Investors can confuse social importance with shareholder returns. Infrastructure assets can remain useful even after their owners fail.

The differences are equally important. Today’s largest spenders have profitable global platforms, strong balance sheets, and existing customers. Many AI capacity commitments are supported by cloud contracts, enterprise demand, or internal products rather than speculative consumer portals. The underlying hardware is more fungible across workloads than some specialized telecom assets were, and leading GPUs have active rental and resale markets.

Another comparison is the DeepSeek shock of early 2025. That episode showed how one model release could trigger a rapid reassessment of compute demand. The fear was that better algorithmic efficiency would reduce the need for expensive hardware. The subsequent expansion in AI infrastructure suggested that lower-cost models stimulated adoption and that frontier development continued to require large clusters. The lesson was not that efficiency never matters. It was that the first-order market reaction can ignore demand elasticity and the continuing race for better capability.

Historical comparisons become misleading when they are used as labels rather than analytical tools. The relevant questions are specific. Is capacity financed with debt that requires immediate cash returns? Are prices falling because supply exceeds demand? Are customers renewing contracts? Are suppliers accumulating inventory? Are assets being impaired? Are companies protecting market share with uneconomic spending? The answers determine whether 2026 resembles a healthy investment boom, a temporary valuation correction, or the beginning of a destructive capital cycle.

The current evidence points to a hybrid. There is no broad collapse in demand, and the largest platforms have financial resources that telecom-era startups lacked. There is also unmistakable overextension risk in the pace of capital commitments and the dependence on continued growth. The dot-com analogy should encourage discipline, not an automatic conclusion that the industry is repeating the same outcome.

How Market Structure Can Turn Small News Into Large Price Moves

The violence of the July selloff may also reflect changes in how information is processed. Quantitative strategies, passive funds, options hedging, thematic baskets, and AI-assisted research can cause many investors to respond to the same signal at nearly the same time. A report about one model, one supplier, or one financing transaction can affect an entire group of securities before analysts have separated direct exposure from narrative association.

This does not mean the market is irrational. Fast information processing is often useful. It becomes destabilizing when the same interpretation is embedded across portfolios and risk systems. If many investors own similar AI baskets and use similar models to classify news, a small reduction in expected demand can produce broad selling. Falling prices then trigger volatility controls, stop-losses, margin calls, and systematic de-risking.

The result is a compressed capital-market cycle. Stocks can price years of optimism and pessimism in weeks, while the physical supply chain moves on much slower timelines. A capacitor manufacturer, memory supplier, or data-center developer may experience little change in orders while its valuation doubles and then falls by half. Price signals become less reliable as contemporaneous measures of operating conditions.

For fundamental investors, this creates opportunity and danger. A broad selloff can make a strong company cheaper even when its earnings outlook improves. It can also tempt investors to dismiss market warnings because short-term narratives have been wrong before. The proper response is to identify which measurable assumptions changed and which did not.

In July, several assumptions changed: the cost of credit rose, regulatory risk became more visible, and China’s equipment progress challenged long-term market-share expectations. Other assumptions did not clearly change: cloud demand remained strong, high-end capacity remained constrained, and major suppliers continued reporting rapid growth. The price decline bundled these separate facts into one bearish trade.

That is why the selloff should be analyzed company by company. A speculative provider dependent on spot pricing and debt deserves a different valuation response from a diversified hyperscaler with contracted backlog. A memory supplier protected by take-or-pay agreements differs from a component manufacturer exposed to short-cycle orders. Nvidia’s integrated platform differs from a chip designer without software, allocation, or financing advantages.

Market structure can explain the speed of the move. It cannot determine whether the move was wrong. Earnings, cash flow, contracts, and utilization will eventually decide.

Frequently Asked Questions

Why did AI and semiconductor stocks fall in July 2026?

The selloff reflected several overlapping concerns: competition from lower-cost Chinese open-weight models, fears that Meta had excess compute, rising bond yields and credit spreads, China’s progress in domestic lithography equipment, high valuations, and crowded investor positioning. The semiconductor index fell into bear-market territory even though major cloud companies continued reporting strong demand.

Was the AI stock selloff caused by collapsing demand?

Public financial results did not show a broad demand collapse. Azure grew 43%, AWS grew 37%, Google Cloud grew 82%, and Nvidia’s latest reported data-center revenue rose 92%. The market was pricing the risk that future growth, margins, or returns on capital would be weaker than current results suggest.

Is Nvidia cheap after the selloff?

Nvidia’s forward earnings multiple compressed because profit growth outpaced the stock price, but “cheap” depends on whether current earnings and margins are sustainable. The shares can look inexpensive relative to near-term growth while remaining vulnerable if estimates decline or competition reduces long-term profitability.

What does “under-earning compute” mean?

It refers to installed GPU capacity operating under contracts signed at prices below the current market. If those contracts expire and reprice higher while the hardware remains competitive, the operator can increase revenue and cash flow without equivalent new capital spending.

Why can open-source or open-weight AI be good for Nvidia?

Open models can reduce model-provider margins and lower the cost per token, making more applications affordable. If usage grows faster than prices fall, total demand for GPUs, memory, networking, and cloud infrastructure can increase.

Why is free cash flow weak when AI demand is strong?

Companies are spending heavily on data centers, servers, networking, power, and leases. Operating cash flow can rise while free cash flow falls because capital expenditures are deducted after operating cash generation. The investment is justified only if the assets produce adequate future returns.

Which company currently shows the strongest AI cash-flow economics?

Among the largest hyperscalers, Microsoft’s June 2026 quarter showed the clearest balance: $55.4 billion of operating cash flow, $41 billion of capital expenditures including finance leases, and $19.6 billion of free cash flow. Comparisons remain imperfect because business mixes and free-cash-flow definitions differ.

What is the biggest financial risk to the AI buildout?

The biggest financial risk is that capital spending and debt grow faster than the cash flows produced by AI infrastructure. This would become especially dangerous if utilization or pricing declined while projects still carried fixed interest, lease, and depreciation costs.

What is the biggest regulatory risk?

Restrictions on data-center construction, electricity use, water, exports, privacy, copyright, or model deployment could delay capacity and raise costs. New York’s one-year permit pause for large data centers showed that infrastructure policy can change quickly.

Does China’s DUV progress threaten ASML now?

China’s move toward domestic immersion DUV production is strategically important, but the reported systems still require testing and do not immediately replace ASML’s most advanced EUV tools. The near-term earnings threat appears limited; the long-term competitive direction is more significant.

Can SpaceX become a major data-center company?

SpaceX has engineering, power-procurement, network, and launch advantages, and its AI business is growing. It must still prove that infrastructure spending produces durable free cash flow. Orbital compute may be useful first for processing space-generated data, while large-scale terrestrial AI training in orbit remains technically and economically uncertain.

What would signal that the AI infrastructure boom is ending?

Persistent declines in high-end GPU rental prices, cloud growth deceleration, weaker backlogs, rising inventory, asset impairments, neocloud defaults, reduced hyperscaler spending, and verified plateaus at major AI labs would provide stronger evidence than stock-price declines alone.

Final Assessment

The July 2026 AI stock selloff exposed a genuine contradiction. Public markets began pricing a capital-cycle downturn while the operating data continued to show expanding demand. The strongest cloud platforms accelerated. Nvidia’s reported data-center business remained extraordinary. Advanced memory moved toward long-term contracted supply. Companies continued to describe capacity constraints rather than excess availability.

The bullish interpretation is that the market has mistaken a free-cash-flow investment trough for a collapse in underlying economics. If installed compute reprices higher, utilization remains strong, and operating cash flow continues to grow, the largest companies can finance much of the buildout internally. Open-weight models can expand the user base, and lower costs can stimulate more token consumption. Scarcity in memory, power, and permitted sites can support the value of existing infrastructure.

The skeptical interpretation is not contradicted by current demand. The industry may be spending too much precisely because the competitive stakes are so high. Companies can build uneconomic capacity rather than risk falling behind. Token prices can decline, hardware can become obsolete, and regulation can delay projects. Credit can remain available until it suddenly is not. Social value and investor returns can diverge.

The decisive evidence will come from cash conversion. Revenue growth must translate into operating cash flow, and operating cash flow must eventually exceed the capital required to sustain the system. Microsoft has shown that pattern most clearly. Amazon, Alphabet, and Meta are at different points in the same test. Nvidia must prove that its growth, margins, and system-level advantage remain durable as customers and competitors build alternatives.

Baker’s central challenge to the market is therefore valid: a falling stock price is not evidence that demand has weakened. The latest filings show the opposite in several crucial businesses. But the market’s concern is also rational because the scale of investment has moved the debate beyond demand. The question is no longer whether AI is growing. It is whether the infrastructure required to support that growth will earn returns commensurate with its cost, risk, and speed of obsolescence.

For now, the data argues against declaring an AI demand bust. It does not justify declaring the capital cycle safe.

Sources

• Nvidia fiscal 2027 first-quarter financial results
• Nvidia Vera Rubin platform announcement
• Microsoft fiscal 2026 fourth-quarter results
• Amazon second-quarter 2026 results
• Alphabet second-quarter 2026 earnings call
• Meta second-quarter 2026 results
• Micron fiscal 2026 third-quarter results
• Micron strategic customer agreements presentation
• SK Hynix and Nvidia multiyear technology partnership
• Kimi K3 technical paper
• Meta Muse Spark 1.1 announcement
• SpaceXAI Grok 4.5 announcement
• New York hyperscale data-center moratorium
• Reuters report on the July semiconductor selloff
• Reuters report on Meta’s planned compute-resale business
• Reuters analysis of hyperscaler bond issuance
• Reuters report on China’s domestic immersion DUV production
• Reuters report on SpaceX’s first public quarterly results
• McKinsey interview on the case for orbital data centers
• Research on the cost and network limits of space-based AI compute
• Research on data centers and U.S. retail electricity rates

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.