Related Articles
Stay Up To Date
Something has gone wrong, check that all fields have been filled in correctly. If you have adblock, disable it.
The form was sent successfully
What Matters
The Point
The AI infrastructure buildout is entering a more complex, higher-risk phase. The cost and availability of capital, the pace of data center construction, and the industry’s ability to monetize AI will increasingly shape investment and returns, raising the bar for management teams and investors alike.
The Proof
Demand still outstrips supply, with cloud revenue accelerating, capacity largely committed through 2027, and no meaningful build-up of GPU inventory. But the sources of risk have shifted: capital spending now outpaces operating cash flow at some hyperscalers, financing is growing more complex, and construction faces new political and physical constraints.
The Result
In this phase, we believe financial strength, competitive position, and the ability to earn attractive returns will increasingly distinguish the businesses that lead from those that lag. That reinforces the value of an active approach grounded in fundamental research, and the discipline to tell genuine signal from routine noise.
The AI infrastructure buildout is entering a more complex phase. For much of the cycle, the largest technology companies funded investment primarily with cash generated by their existing businesses. Today, capital spending is outpacing operating cash flow at some hyperscalers, while model developers and specialized cloud providers depend heavily on outside financing. At the same time, political opposition to data center development has introduced uncertainty around the pace at which capacity can be built.
These developments have contributed to a sharp repricing of AI-exposed companies, as investors question whether the scale and duration of investment will generate adequate returns. Yet the market’s concerns stand in contrast to strong operating results from cloud providers, accelerating AI usage, and multiyear commitments to secure data center components.
In our view, these pressures mark a transition to a higher-risk phase in which the cost and availability of capital, the pace of data center construction, and the industry’s ability to monetize AI will increasingly shape investment and returns. The bar has risen for management teams and investors alike.
We examine the strength of demand, the constraints on capacity, the financing required to support the buildout, and the prospects for earning a return on that investment.
Is Demand Keeping Pace?
Cloud growth is accelerating. Aggregate public cloud revenue at the hyperscalers grew about 50 percent year over year in the third quarter, the latest in a steady run of accelerating quarters, led by Alphabet’s cloud business at 103 percent, and Microsoft’s Azure at 48 percent. Customers are buying capacity faster than it can be built. Amazon describes AWS capacity as effectively sold out for 2026 and largely committed for 2027, Microsoft says demand continues to exceed available supply, and the three largest cloud providers now report more than $1.7 trillion in combined revenue backlog.
Usage is growing even faster. Token volumes, a measure of activity on the platforms tracked, have risen about fivefold over the past six months and roughly 50 percent in the past four weeks alone.1 Cheaper models have contributed to that surge by making new applications economical. The effect on demand for computing capacity depends on both growth in usage and improvements in the amount of compute required for each task.
Rental rates also point to tight supply. Rates for NVIDIA’s B200 have risen about 27.5 percent over six months, while rates for the older H100 have risen about 14 percent.1 SpaceX says new compute at current pricing pays for itself in less than a year, and its comments suggest that falling build costs and better monetization could shorten that period to about six months in 2027. These measures indicate strong current demand, although they do not establish the useful life or long-term value of the hardware.
We do not yet see some of the typical signs of a cycle peak. There is no meaningful build-up of GPU inventories in the supply chain, and lead times for components are still lengthening rather than shrinking. SK hynix has committed more than half of its 2027 and 2028 production capacity under long-term agreements, while ASML Holding says its planned supply for both years will not fully meet demand.
Can the Capacity Be Built?
The first phase of the AI infrastructure buildout was defined largely by the availability of advanced GPUs. As we discussed in Beyond GPUs, growing AI workloads have since expanded demand for memory, networking, and other components required to support increasingly complex systems.
The constraint is now extending further, into the physical infrastructure required to house and power those systems. A GPU cannot generate revenue until it is installed in a data center with sufficient electricity, cooling, networking, and supporting equipment. Building that capacity requires available land, permits, skilled labor, electrical infrastructure, and time.
These constraints can affect the pace of investment even when demand remains strong. If data center shells or power connections take longer to deliver, customers may be unable to deploy chips as quickly as expected. That could push out orders or create temporary imbalances in parts of the supply chain, even while the broader market remains short of usable computing capacity.
The concern is whether the industry can continue adding capacity at the pace investors previously anticipated. At the peak of optimism, the market expected new data center capacity added in 2028 to grow by more than 30 percent. Investors now worry it could be closer to flat with 2027, raising the possibility that chip production runs ahead of deployment. JPMorgan estimates that 55 percent of the 96 GW of new capacity planned for delivery in 2028 could face a “pause” or additional “conditions,” based on incoming governors’ campaign promises.2
The midterm elections added a source of uncertainty. Voters backed restrictions in all but one of more than 40 local ballot measures across 10 states, and economists expect new governors in Texas, Georgia, and Michigan to impose conditional moratoria while state legislation is drafted. States where new governors are expected to restrict development account for roughly 45 percent of planned U.S. capacity, with Texas alone representing about 25 percent. Texas’ incoming governor has called for a pause until the legislature acts in 2027, though she has since said she is not opposed to data centers.
We expect tougher rules rather than a lasting halt to development. Restrictions may include reduced tax incentives, new water and grid-connection requirements, and community benefits. Our conversations with builders and policy specialists in Texas suggest that projects with permits and tax agreements in hand are expected to be protected, with new rules likely by the second quarter of 2027. Under that scenario, the principal effect would be a permitting delay of several months.
Restrictions could also change where and how capacity is built. New projects may shift to other states, remote areas, or overseas, while developers may turn more quickly to on-site power. A pause in new Texas starts could leave a gap in 2028 and make the path of construction less smooth. We expect these constraints to slow the pace of the buildout, with a more limited effect on the long-term level of aggregate investment.
Can It Be Financed?
The AI buildout is requiring a growing amount of capital. Hyperscaler capital spending is on track to reach roughly $800 billion this year, and consensus expects it to rise to approximately $1.3 trillion in 2027. As investment has increased, so has borrowing. Hyperscalers have issued approximately $320 billion of debt globally through August.3
The ability to support this investment varies considerably across the industry. Our estimates suggest that over the next three fiscal years, capital spending will exceed operating cash flow after dividends for the four largest hyperscalers, leaving them reliant on external funding. By comparison, we estimate that the 13 other infrastructure builders will require significantly more.4 The hyperscalers’ cash generation and stronger balance sheets should give them greater flexibility to sustain investment if credit conditions tighten.
This distinction matters for the pace of the buildout. Hyperscalers have accounted for roughly 65 percent of net new data center capacity in recent years. The remaining capacity includes projects developed by companies with greater reliance on outside financing. If lenders become less willing to fund those projects, some investment could be delayed or require greater financial support from hyperscalers, model developers, and chip suppliers.
Financing is also becoming more complex. Private credit and suppliers are playing a growing role, as illustrated by an Apollo-led financing for SpaceX’s chip purchases and arrangements involving chip vendors and model developers. These structures can help bring capacity online, but they also create additional links between suppliers, lenders, and customers. As a result, the creditworthiness of each participant, along with borrowing costs, contract terms, and customer concentration, becomes more important to the broader ecosystem.
NVIDIA has itself become a significant source of financing for the buildout. Through supply and capacity commitments, loan guarantees, and equity stakes, the company has helped fund AI labs, cloud providers, and startups across the ecosystem, and its web of commitments has grown substantially over the past year. Following the departure of CEO Jensen Huang, NVIDIA has signaled a more disciplined and selective approach, applying tighter scrutiny to non-investment-grade specialized cloud providers and to new guarantees. A more cautious posture from one of the ecosystem’s largest backers would place greater weight on other sources of capital.
The intersection of these pressures is most visible among more leveraged infrastructure providers. CoreWeave’s bonds traded at distressed levels and its credit default swaps topped 1,000 basis points, despite a backlog of more than $104 billion, much of it with investment-grade customers. Credit spreads have also widened for some larger companies with significant data center exposure. At the same time, broader credit spreads remain tight, and investors have absorbed a wave of large investment-grade debt offerings.
Will It Pay?
AI revenue is growing rapidly. We estimate that the industry will generate approximately $400 billion in AI revenue in 2026. Anthropic’s IPO filing showed first-half revenue of $16.3 billion, up roughly 1,300 percent from a year earlier, and an adjusted operating margin of 19 percent in the second quarter. OpenAI has reportedly told investors it expects annualized revenue of at least $70 billion by year-end, supported largely by enterprise demand.5 Alphabet and Meta Platforms are also using AI to improve engagement and advertising targeting within their existing businesses.
These results provide evidence that AI is being monetized, but the scale of the infrastructure buildout raises the bar for future growth. The question is not only how much revenue AI can generate, but also whether that revenue can support the cost of building, operating, and eventually replacing the infrastructure required to deliver it.
Sustaining investment will likely require adoption to broaden beyond the applications driving growth today. Coding has emerged as one of the clearest uses for agentic AI. Models can write and test software, developers are comfortable adopting new tools, and results can often be evaluated quickly. We believe the opportunity remains significant as models become more capable and usage expands beyond professional developers.
The broader opportunity includes customer service, cybersecurity, research, and other knowledge work. AI spending remains a small share of enterprise technology and labor costs, leaving room for growth if companies can demonstrate gains in productivity, revenue, or cost savings. Adoption will depend on model capabilities and on businesses’ ability to integrate AI into workflows, provide it with the right context, and manage security concerns.
Continued progress in model capabilities could expand those use cases. Advances in training, reasoning, and the tools that allow models to interact with software have made AI useful for a wider range of tasks. Agents that can complete multistep work show how improvements can create demand that was difficult to anticipate, although the timing and commercial impact of future breakthroughs remain uncertain.
Revenue momentum across the leading model developers has been uneven, raising questions about how quickly revenue can grow relative to planned investment. Anthropic’s growth decelerated in the third quarter, with revenue rising about 47 percent from the prior quarter, down from 144 percent in the second, according to its amended public filings.6 OpenAI, by contrast, reportedly reaccelerated, with roughly 50 percent sequential growth and an annualized revenue run rate of about $50 billion as of the end of September.7 Anthropic’s subsequent disclosures eased some of these concerns, showing that its run rate rose approximately $10 billion in October, following the September release of Claude Opus.8 Competition can also redistribute spending among providers, so a gain for one model developer may not translate into faster growth for the industry as a whole.
Falling prices add another dimension to the return equation. Consumption-weighted token prices have roughly halved over six months on the platforms tracked, even as token volumes have grown about fivefold. Lower prices can make new uses economical and encourage customers to use AI more frequently. For providers, the benefit depends on whether greater usage and lower costs can offset the decline in revenue per token.
Competition from open-weight models is contributing to price pressure as their capabilities improve. These models may be well suited to simpler tasks, while more capable models could retain an advantage in complex work where accuracy and performance have greater economic value. For customers, the relevant measure is the cost of completing a task successfully, not simply the price per token. Data from third-party platforms captures only a portion of overall AI activity and may not reflect the market as a whole.
Our Take
The AI opportunity remains large, and current demand for computing capacity continues to exceed supply. In our view, recent concerns mark a transition to a more complex phase in which the pace of investment will depend increasingly on the availability of powered capacity, access to financing, and growth in AI revenue.
These factors are connected. Construction delays can slow chip deployment even when customers want more computing capacity. Tighter financing conditions can delay projects before there is evidence of weaker end demand. At the same time, the scale of planned investment raises the importance of broadening adoption and translating usage into sustainable revenue and cash flow.
What matters is less whether investment slows than how abruptly. An abrupt slowdown in deployment, combined with financing constraints, could impair the revenue growth expectations that have sustained the cycle. A more modest slowdown need not diminish the long-term opportunity. It could instead give adoption and monetization more time to catch up with infrastructure spending and reduce the outside capital required in a given year. Because advances in AI tend to arrive unevenly and unpredictably, we expect periodic bouts of volatility to be a recurring feature of the cycle rather than a sign of its breakdown. Any such adjustment would nevertheless have different consequences across the AI value chain.
Businesses with available computing capacity may benefit from prolonged scarcity, while delays in bringing new data centers online could create temporary imbalances for hardware suppliers if production outpaces deployment. Infrastructure providers that depend heavily on external financing may face greater pressure if borrowing costs rise or access to capital narrows.
The first phase of the buildout broadly rewarded providers of the physical infrastructure that AI depends on. In the next phase, we believe financial strength, competitive position, and the ability to earn attractive returns on investment will become more important differentiators. Opportunities may also broaden beyond infrastructure to businesses using AI to improve products and services. As the cycle evolves, we will continue to distinguish temporary disruptions from changes that could impair a company’s long-term economics.
1 https://www.citadelsecurities.com/news-and-insights/global-macro-strategy/tokenomics-2-0/
2 https://privatebank.jpmorgan.com/apac/en/insights/latest-and-featured/eotm/now-you-can-see-everything
3 J.P. Morgan, Eye on the Market, “Online US Economy and Iran War Tracker,” September 1, 2026.
4 Based on company filings and Sands Capital estimates; aggregate hyperscaler free cash flow turned negative in 2026 as capital spending exceeded operating cash flow.
5 Bloomberg News, “‘Black Box’ Revenue From Anthropic and OpenAI Stumps Traders,” October 9, 2026
6 Anthropic PBC, amended Form S-1 (Amendment No. 2), filed with the U.S. Securities and Exchange Commission, November 10, 2026
7 OpenAI third-quarter 2026 results as presented to investors and reported by Bloomberg, October 2026.
8 Anthropic PBC, amended Form S-1 (Amendment No. 2), November 10, 2026, reporting run-rate revenue of approximately $81 billion as of October 31, 2026.
Disclosures:
The views expressed are the opinion of Sands Capital and are not intended as a forecast, a guarantee of future results, investment recommendations or an offer to buy or sell any securities. The views expressed were current as of the date indicated and are subject to change.
This material may contain forward-looking statements, which are subject to uncertainty and contingencies outside of Sands Capital’s control. No reliance should be placed on these forward-looking statements.
There is no guarantee that future growth projections will accurately predict the actual growth experience of any of the companies involved or cause a portfolio to outperform its benchmark or index.
Consensus earnings estimates are subject to revision and may differ materially from actual results.
References to specific issuers or securities are provided for illustrative purposes only and are not intended as recommendations to purchase or sell such securities. Sands Capital may own, may have owned, or may not currently own any securities of the issuers referenced and, if such securities are owned, no representation is being made that they will continue to be held. It should not be assumed that any investment in the securities of the issuers referenced was or will be profitable.
As of November 19, 2026, Anthropic was held in certain Sands Capital strategies and in private funds managed by Sands Capital Alternatives, LLC, an affiliate of Sands Capital Management, LLC. OpenAI is held in private funds managed by Sands Capital Alternatives, LLC.
All investments are subject to market risk, including the possible loss of principal.
Recent tariff announcements may add to this risk, creating additional economic uncertainty and potentially affecting the value of certain investments. Tariffs can impact various sectors differently, leading to changes in market dynamics and investment performance.
Investments in companies utilizing artificial intelligence (AI) involve additional risks. AI technologies are rapidly evolving and may be subject to regulatory uncertainty, ethical concerns, and operational challenges. There is no assurance that AI-driven strategies will be successful, and companies may face risks related to data quality, model accuracy, cybersecurity, and unintended outcomes, which could adversely affect business results and investment performance.
Companies that manufacture AI infrastructure are subject to risks including rapid technological change, competitive pressures, supply chain disruptions, regulatory developments, customer concentration, and fluctuations in demand for AI-related products and services, which may adversely affect financial performance and valuations.
References to the “firm”, “we” or “our” are references to Sands Capital. Sands Capital refers to the combination of Sands Capital Management, LLC, Sands Capital Alternatives, LLC and Sands Capital Horizons, LLC. All three firms are registered investment advisers with the United States Securities and Exchange Commission in accordance with the Investment Advisers Act of 1940. The three registered investment advisers share certain personnel, office space, and other resources.
This communication is for informational purposes only and does not constitute an offer, invitation, or recommendation to buy, sell, subscribe for, or issue any securities. The material is based on information that we consider correct, and any estimates, opinions, conclusions, or recommendations contained in this communication are reasonably held or made at the time of compilation. However, no warranty is made as to the accuracy or reliability of any estimates, opinions, conclusions, or recommendations. It should not be construed as investment, legal, or tax advice and may not be reproduced or distributed to any person.

