The AI Investment Cycle: Infrastructure Buildout and M&A

Introduction

The AI investment cycle is a front-loaded capital buildout: a small group of technology companies is spending cash, and increasingly borrowed money, on chips, data centers and power years before the revenue those assets are meant to earn arrives. It is the largest infrastructure buildout in technology history, inviting comparison with the fiber-optic boom of the late 1990s and the first cloud migration. A single figure captures the scale: Amazon, Alphabet, Microsoft and Meta alone guide to roughly $720-745 billion of capital expenditure in 2026. That spending generates deal flow across every layer of the stack, from semiconductors and data centers to power generation, GPU cloud providers and AI-native software. It also creates a financing gap between spending today and revenue later, which credit markets began testing in October 2026. For TMT investment bankers, the cycle is the dominant driver of M&A, equity, debt and potentially restructuring work, and knowing how it is funded matters as much as knowing what it builds.

Hyperscaler capex is the engine of the whole cycle. Alphabet and Amazon raised their 2026 outlooks with second-quarter results, Meta narrowed its range upward, and Microsoft changed how it reports; the table below shows each company's latest guidance in its own words and reporting period.

The sources are the companies themselves. Meta's second-quarter release narrowed its range to $130-145 billion including finance-lease principal; Andy Jassy told analysts Amazon now expects to spend $220 billion, citing higher memory costs; and Alphabet CFO Anat Ashkenazi lifted guidance to $195-205 billion as the Google Cloud backlog reached $514 billion. Microsoft's figure needs care: Amy Hood cut the calendar-2026 estimate to about $175 billion from about $190 billion purely because more future data center leases will be classified as operating leases, which sit outside reported capex, while spending plans stayed unchanged. Oracle, on a different fiscal calendar, guided to up to $95 billion, of which about $70 billion is its own spend and the rest is expected to be repaid by customers.

Two lessons follow for anyone comparing these companies. First, reported capex is a policy choice as much as a spending number: lease classification moved Microsoft's figure by about $15 billion without changing a single data center. Second, capital intensity has reached levels once associated with utilities and telecom carriers. Second-quarter capex ran at about 51% of revenue at Meta ($31.1 billion on $60.8 billion) and about 37% at Alphabet ($44.9 billion on $119.8 billion). Businesses that once converted most of their earnings into free cash flow now reinvest a large share of revenue, which is why the cycle has become a capital markets story as well as a technology story.

AI compute has turned data centers from a niche real estate category into one of the most active M&A sectors in technology, with several distinct deal categories. Hyperscalers acquire or build data centers directly, often partnering with developers and energy companies to co-develop large AI campuses that combine compute, power generation and cooling on a single site. Colocation providers (Equinix, Digital Realty, CyrusOne) lease space and power to hyperscalers and enterprises, and several have pivoted toward AI-optimized facilities with the higher power density (50-100+ kW per rack, compared with 5-15 kW for traditional enterprise workloads) that GPU clusters require. Specialized AI infrastructure companies (neoclouds) build GPU-dense facilities for training and inference. Each category attracts different buyers, from strategic acquirers to infrastructure-focused PE sponsors to sovereign wealth funds and infrastructure funds that treat data centers as long-lived real assets with contracted revenue, financed through the leases and project debt described in how data center financing works.

The power constraint has created an entirely new category of TMT-adjacent M&A: technology companies acquiring energy assets. Microsoft signed a long-term power agreement with Constellation to restart Three Mile Island's Unit 1 reactor, Amazon bought Talen Energy's data center campus beside the Susquehanna nuclear plant, and Google signed agreements with Kairos Power for small modular reactor supply. These transactions blur the line between technology and energy M&A and require TMT bankers to work with power sector colleagues on deal structuring, regulatory approvals (a reactor restart needs NRC sign-off) and the valuation of the power purchase agreements that underpin AI data center economics.

The scale of these deals keeps climbing. A consortium of the AI Infrastructure Partnership (AIP), Abu Dhabi's MGX and BlackRock's Global Infrastructure Partners agreed in October 2025 to buy Aligned Data Centers from Macquarie Asset Management, and the deal closed on July 21, 2026 at an enterprise value of about $40 billion, with a further $5 billion of growth capital committed. Aligned brings 51 campuses and more than 6.4 GW of operational and planned capacity, and the price is more than double the roughly $16 billion that Blackstone and its partners paid for AirTrunk in 2024, the previous record. The deal shows capital consolidating around the operators that control power and land, and it was AIP's first completed investment, a template for sovereign and infrastructure money entering AI at platform scale.

A new category of AI infrastructure companies has emerged to fill the gap between hyperscaler cloud services and the specialized compute needs of AI model developers. These neoclouds (CoreWeave, Lambda, Crusoe, IREN, Nebius) build and operate GPU-dense data centers optimized for AI training and inference, often at lower prices and with more flexible configurations than the major cloud providers.

Neocloud

A neocloud is a cloud provider built specifically to rent GPU computing capacity for AI training and inference, rather than offering the broad menu of storage, database and software services that general-purpose clouds such as AWS, Azure and Google Cloud sell. Neoclouds such as CoreWeave, Nebius, Lambda and Crusoe typically sign multi-year capacity contracts with AI labs and hyperscalers, then borrow against those contracts to buy GPUs and lease or build powered data centers. The model turns contracted demand into rapid growth, but it also leaves the provider with heavy debt and lease obligations that depend on a few large customers paying for years.

The neocloud model is a significant departure from the traditional cloud business, where hyperscalers built general-purpose infrastructure and customers ran diverse workloads. Neoclouds build for a single workload, secure long-term contracts with AI model developers and hyperscalers, and use that revenue visibility to raise large amounts of debt against future cash flows. CoreWeave is the largest, but it is not alone: Lambda focuses on AI research and development customers, Crusoe builds campuses around low-cost energy (including stranded natural gas), IREN converted a bitcoin-mining footprint into AI data centers, and Nebius (the former Yandex parent, now based in Amsterdam) runs AI clusters in Europe and the US. All share the same capital structure logic: aggregate GPU capacity, sell it on long-term contracts, and borrow against those contracts to expand.

That capital structure came under direct market pressure in October 2026. On October 14, CoreWeave shares fell more than 15% after a widely shared chart showed the yield on one of its bonds briefly passing 15%, and the sell-off spread to Oracle (down 8.8%), DigitalOcean (down more than 10%) and other AI infrastructure names; its unsecured notes due 2032 traded below 80 cents on the dollar, Reuters reported. CEO Mike Intrator told employees the revenue backlog more than supports the company's debt obligations. The numbers behind both sides of that argument are in CoreWeave's second-quarter 2026 results: a revenue backlog of about $104 billion at June 30, 2026 (remaining performance obligations plus other amounts the company expects to recognize, subject to delivering the service), against about $35 billion of recourse and non-recourse debt and about $16.5 billion of lease liabilities on the balance sheet, roughly $52 billion in all, with second-quarter net interest expense of $640 million on $2.6 billion of revenue.

The comparison with the overleveraged telecom carriers of 2001-2002 is now the question the market is pricing rather than a hypothetical. The difference is that CoreWeave's backlog is contracted, largely with Microsoft, Meta and OpenAI, whereas the fiber builders borrowed against traffic forecasts. The similarity is that the contracts pay only once capacity is built, powered and delivered, so the company must keep borrowing ahead of the cash, and its revenue rests on a handful of customers. Whether backlog converts to cash faster than debt comes due is exactly what lenders, bondholders and advisers now track for every neocloud.

The AI investment cycle is generating deal activity across every layer of the technology stack, from chips to applications. The common thread is that buyers are paying for scarce AI inputs (inference silicon, cloud security, proprietary data, research talent) rather than for scale or cost synergies, which changes how deals are structured as well as how they are valued.

The licensing structures matter for advisers because they move value without a change of control, which can reduce merger-review exposure but leaves the target company, its investors and its remaining employees in an unusual position. Bankers should be ready to structure non-traditional deals wherever the prize is a team or a model rather than a business.

The AI buildout requires front-loaded investment for compute, data centers and energy infrastructure, while the revenue from that investment arrives later. That timing mismatch, the financing gap, is reshaping how technology companies use capital markets, and it is funded very differently at the top and the bottom of the credit spectrum.

At the top, hyperscalers that once paid for everything from operating cash flow now borrow as well. Bank of America counted about $121 billion of hyperscaler bond issuance in 2025, more than four times the roughly $28 billion these five companies averaged over the previous five years, and the pace continued into 2026: Meta's second-quarter cash flow statement shows $24.9 billion of net long-term debt proceeds in the quarter alone, and Oracle paired its fiscal 2027 capex plan with plans to raise nearly $40 billion of debt and equity, including a $20 billion at-the-market share sale. How these deals are sized, rated and absorbed by insurers and other buyers is the subject of hyperscaler bonds and IG issuance; for this article the point is that investment-grade balance sheets let hyperscalers fund the gap at modest spreads. Further down, neoclouds rely on GPU-backed loans, project finance, high-yield bonds and vendor support, the layers mapped in how the AI boom is financed.

Remaining Performance Obligations (RPO)

Remaining performance obligations are the contracted revenue a company has not yet recognized: the value of signed customer contracts for goods or services still to be delivered, disclosed under US GAAP revenue recognition rules (ASC 606). Cloud and AI infrastructure companies report RPO, or a broader "backlog" built on it, to show investors how much future revenue is already under contract. RPO is not cash: it converts to revenue only as the company delivers capacity, and it can shrink if a customer renegotiates, defaults or terminates.

RPO has become the central metric of the financing gap because it is the collateral story behind the borrowing. Oracle reported $664 billion of RPO in its fiscal first quarter of 2027, and CoreWeave's $104 billion backlog sits behind its roughly $52 billion of debt and lease liabilities. Lenders and bond investors are effectively underwriting a conversion schedule: how quickly contracted revenue becomes cash once data centers are energized, compared with when interest and principal fall due.

For TMT investment bankers, the financing gap creates work across product groups. In debt capital markets, banks arrange investment-grade bonds for hyperscalers and leveraged loans, high-yield notes and structured project finance for neoclouds. CoreWeave's IPO in March 2025 (raising about $1.5 billion at a roughly $23 billion valuation) opened the equity capital markets to AI infrastructure exposure, and its October 2026 bond sell-off shows how quickly that appetite reprices. In restructuring and liability management, the same stress creates potential mandates: refinancing, covenant amendments and, if conversion falls behind, balance sheet repair for overleveraged infrastructure companies. The telecom cycle produced both lasting winners (Cisco, Juniper) and spectacular bankruptcies (WorldCom, Global Crossing, Adelphia), and advisers to AI infrastructure companies now have to price that same spread of outcomes.

The debate over whether the cycle is a sustainable technology transition or a speculative overbuild is itself part of the job. Bulls point to the transformative potential of the technology (comparable to electrification or the internet), enterprises' demonstrated willingness to pay for AI, cloud backlogs that keep rising, and the shift toward inference workloads that sustain compute demand even as training gets more efficient. Bears point to the concentration of AI revenue in a few foundation model companies (with some of it circular, flowing between hyperscalers and the labs they fund), the historical pattern of infrastructure overbuild, and efficiency breakthroughs such as DeepSeek's low-cost training results that could cut compute requirements. Advisory credibility requires holding both cases at once and structuring deals, valuations and financings that survive more than one scenario.

The AI investment cycle reaches every TMT coverage area, though the work looks different in each:

The sharpest valuation questions sit with the AI labs themselves. OpenAI's last priced round raised $122 billion at an $852 billion post-money valuation in March 2026, when the company said it was generating about $2 billion of revenue a month, and Anthropic's $65 billion Series H in May 2026 valued it at $965 billion post-money on run-rate revenue above $47 billion. Those prices imply roughly 20-35x run-rate revenue, multiples that rest on the belief that these companies are building category-defining platforms, far above where public software trades after the SaaS valuation reset. Both companies have since filed confidentially for IPOs, a story covered in the 2026 mega-IPO pipeline. For coverage bankers, the AI company valuation challenge is distinguishing genuine differentiation (proprietary models, unique data, defensible distribution) from AI washing, superficial AI features on traditional software that do not justify a premium.

The geographic dimension adds cross-border work. The US hosts most AI infrastructure investment, but Europe's data sovereignty rules create demand for local capacity, Gulf sovereign investors are deploying capital at platform scale (MGX's role in Aligned is one example), and Japan and South Korea are funding domestic compute. The CHIPS Act and its international equivalents add subsidies, creating mandates on public-private partnerships, government-backed financing and joint ventures between technology companies and sovereign investors.