AI has become a defining theme in corporate dealmaking, and the race to secure relevant capabilities now extends well beyond companies in the AI ecosystem. For many executives, the instinct is to buy. But models and tools are advancing so quickly that companies risk paying control premiums for capabilities that may soon be outdated or commoditized. Because many capabilities are accessible off the shelf or through a partner, the real challenge is knowing where ownership truly matters.
Capital markets underscore both the opportunity and the burden of M&A interest in AI. Investors view the acquisition of AI software and application companies by technology buyers as little more than table stakes. But similar deals by nontech companies receive a much warmer response, suggesting that investors see greater potential for strategic impact in those areas. That initial reward rests on an expectation, however, not on proof of value creation—and it raises the burden of delivering on the deal’s promise.
To meet that burden, a company should evaluate each AI capability along two dimensions: how strongly it can differentiate the business and how readily it can be sourced. Together, those dimensions lead to four possible moves: adopt, partner, build, or buy. The aim is not to maximize ownership, but rather to secure the access that the company needs to create an advantage while retaining room to adapt.
The Changing Shape of AI Dealmaking
AI acquisition activity surged in 2025 across every layer of the industry, from hardware and infrastructure to software and applications. (See Exhibit 1.) Hardware and infrastructure provided a substantial base of activity well before the current AI boom, supporting earlier waves of machine learning, big-data analytics, and software as a service. As AI models became more capable and as adoption spread, dealmaking moved farther up the stack toward software and applications. In addition, the mix of buyers differs from what one might expect: nontech companies accounted for most acquisitions of AI software and application companies in 2025.
Acquisitions tell only part of the story. Increasingly, companies are pursuing AI capabilities through other structures. The number of AI-related alliances has risen steadily since 2022 and more than doubled from 2024 to 2025, when companies formed more than 1,300 new arrangements. (See Exhibit 2.) These structures now rival traditional M&A as a route to AI capabilities.
Together, the trends suggest that companies are using a wider range of deal structures and are becoming more selective about when ownership is necessary. With more routes to AI capabilities, they have more reason to weigh their options deliberately. The task now is to determine how much ownership and control of AI capabilities are necessary to achieve the desired advantage.
Which AI Deals Impress Investors?
Although it is too early to judge the long-term performance of the current wave of AI transactions, announcement returns provide an early indication of which deals investors expect to create value. On average, acquisitions of AI-related companies have generated positive announcement returns across every layer of the AI stack. (See Exhibit 3.)
The pattern becomes stronger closer to the application layer. Average announcement returns rise from approximately 0.3% for acquisitions of AI hardware and semiconductor companies to 0.4% for cloud and data-center targets and 0.5% for AI software and application companies.
For software and application deals, however, the identity of the buyer is an important factor. Acquisitions by technology companies generate essentially no market reaction, suggesting that investors may regard such moves as a matter of course or even as table stakes. Nontech buyers, by contrast, earn an average announcement return of more than 0.8%—nearly three times the average return across all deals (including non-AI deals). Investors seem to see greater upside when these companies use acquisitions to accelerate an AI-enabled strategic shift.
For buyers, investors’ heightened expectations raise the burden of delivering on the deal’s promise—a burden that few transactions have met—which makes rigor in both deal selection and execution especially important.
Choosing the Right Approach: Adopt, Partner, Build, or Buy
Delivering on the promise of an AI deal begins with choosing the right way to secure each capability. For companies outside the AI ecosystem, that means deciding what to consume from the market, what to access through a partner, what to develop internally, and what to obtain through an acquisition.
The guiding principle is straightforward: use the lightest structure that provides the access or control required to gain the advantage while preserving the freedom to switch as the technology evolves.
Two questions frame each decision. First, can the capability yield genuine competitive differentiation? For a non-AI company, that edge rarely comes from the model itself. More often, it resides in the proprietary data, workflows, intellectual property, or domain expertise that surround the model. Second, how broadly available is the underlying capability? The relevant test is not whether a solution exists, but whether comparable alternatives are available from a competitive market or whether access instead depends on a particular provider, team, or asset.
Plotting the answers yields a four-part matrix (see Exhibit 4):
- Adopt when the capability is widely available and does not differentiate the business.
- Partner when access is more limited or requires tailoring, but when ownership would add little strategic value.
- Build when the company can use assets it already controls to turn widely available technology into an advantage.
- Buy when a strategically important advantage is locked inside a scarce external asset and cannot be secured adequately through another structure.
These four moves address how a company obtains necessary capabilities from available sources, not how extensively it intends to transform the business. The matrix therefore complements BCG’s deploy-reshape-invent progression, which describes the depth of transformation. A company can use a partner's technology to invent—that is, create—a needed capability, or it can deploy a capability that it buys outright.
The matrix provides the high-level logic. Six practical considerations help companies apply it. (See Exhibit 5.) Few capabilities will point in the same direction on every dimension. The objective is not to achieve unanimity but to understand the balance of evidence.
Usually, the decision will lean toward partnering for or building custom solutions. Entry-level or standardized tools are available off the shelf, and most AI capabilities are evolving too quickly—and are too readily available—to justify the capital, complexity, and lock-in entailed in acquiring outright ownership. But partnering for or building custom solutions can enable companies to create a durable edge through some combination of proprietary data, workflows, intellectual property, and scarce external assets.
Choosing the appropriate structure is only the beginning. Lighter does not mean easier: a BCG study found that roughly nine in ten partnerships between industrial and technology companies failed to meet their goals. Lighter structures carry risks of platform lock-in, disputes over intellectual property, and the possibility that a partner will apply insights gained through the collaboration elsewhere in the sector. Companies should direct the same governance attention to these arrangements that they would apply to an acquisition. Before signing, companies should clarify ownership of data and intellectual property, secure portability and benchmarking rights, and create a governance process that allows them to end an arrangement when it no longer delivers value.
Whatever structure it chooses, a company must be able to absorb the capability and translate it into changed ways of working. Relatively few companies have converted AI adoption into measurable value, and the difference between leaders and laggards is strategic and operational rather than simply technological. A company that can’t effectively redesign its workflows will underperform in every quadrant of the matrix.
How the Buy-or-Partner Choice Varies by Sector
Although the framework applies across industries, sector dynamics are especially visible in the choice between acquiring AI capabilities and accessing them through partnerships—the two routes most directly tied to dealmaking. Adopting and building tend to raise different sets of questions about technology and implementation. Contextual differences in the industrial goods, consumer, and energy sectors illustrate how differently the choice between ownership and partnership plays out in practice.
Industrial Goods. Industrial companies are pursuing two main paths to AI. Either they buy specialist firms when a capability provides a meaningful product edge or they partner with large technology platforms for underlying capabilities that they don’t need to build themselves. Early deals centered on autonomy and machine vision; more recent ones seek control of the software and data layers that increasingly determine where value accrues as AI moves to the core of the product. Most partnerships now involve a handful of dominant AI platforms. Ultimately, the winners will combine proprietary industry data and engineering expertise with large-scale AI infrastructure in a “build” approach—a combination that both pure software and pure hardware competitors will struggle to match.
Schneider Electric’s agreement to acquire Cognite, announced in 2026, illustrates why some AI capabilities warrant outright ownership. Cognite’s cloud-native platform combines a unified industrial data model with agentic AI capabilities, placing it in the software and data layer where an increasing share of industrial value is accruing. Schneider plans to integrate Cognite into AVEVA, its wholly owned industrial software business. The acquisition would give Schneider control over a data foundation that it can embed across its portfolio. In this case, the strategic value lies not simply in gaining access to AI, but in controlling an increasingly important layer of the product.
Consumer Goods. Consumer goods companies were early and pragmatic adopters of AI, although few have scaled it. Their logic is straightforward: buy when an AI capability strengthens something that distinguishes the brand; partner when the need is primarily for scale or specialized tools. In settings where customer insight is the battleground—such as personalization, demand forecasting, and virtual try-on—companies tend to buy to own capabilities that partnership terms might make available across the category.
For broad productivity and marketing applications, however, they rely on large AI platforms, supplemented by smaller tools for product ideation and creative testing. The larger shift involves moving from using AI for back-office efficiency to placing it at the heart of the customer experience, enabling mass-market brands to personalize products at scale without sacrificing consistency.
L’Oréal’s collaboration with Nvidia illustrates why companies may partner for software infrastructure. Announced in 2025, the collaboration gave L’Oréal access to Nvidia’s enterprise AI platform, initially to scale 3D digital renderings of its products and accelerate marketing and advertising campaigns. The companies expanded the relationship into R&D in 2026 to accelerate discovery of new formulations through predictive AI science. For L’Oréal, the primary source of advantage does not lie in owning the underlying computing infrastructure or general-purpose AI frameworks, but rather in possessing its proprietary skin and hair biology data and the scientific expertise to interpret it. The partnership therefore pairs external technology at scale with company-specific assets that L’Oréal brings to the collaboration.
Energy. Energy stands apart. Companies in the sector rely more heavily on partnerships and minority investments than on large AI acquisitions. That’s because their highest-value applications tend to involve subsurface analysis, predictive maintenance, and grid optimization rather than customer-facing products—domains where the differentiator is execution, not access. US shale operations illustrate the point: in 2025, the gap in lease operating expenses between top- and bottom-quartile Permian operators exceeded $5 per barrel of oil equivalent, a spread attributable not to geology or asset vintage, but to operational execution.
Operators typically combine broad cloud partnerships with a smaller number of specialist alliances, making only selective acquisitions. A notable shift is that oilfield services companies are evolving from technology buyers into technology sellers. As growth in their traditional businesses slows, they are packaging their industry expertise into AI software—and even supplying AI computing infrastructure—for other energy companies. Utilities, meanwhile, use AI primarily for forecasting and grid reliability, often through cloud partnerships.
Notably, market coverage has described many recent energy deals as AI-related even when their actual focus is on powering data centers rather than deploying AI within energy companies themselves. Those deals fundamentally differ from transactions intended to introduce AI into an energy company’s own operations or products.
Shell’s collaboration with SparkCognition is an example of a situation where an alliance can provide access to specialized AI capabilities without requiring ownership. In 2023, the companies announced a collaboration to apply generative AI to accelerate exploration, building on earlier work on well-pressure prediction.
The goal of the approach is to produce subsurface images that require fewer seismic shots than are necessary with conventional methods. The partners contribute complementary assets: Shell brings proprietary seismic data and subsurface expertise, and SparkCognition supplies specialized AI capabilities. The alliance combines those assets without requiring Shell to buy and operate a software business.
Across all three of these sectors, one principle holds: companies buy AI when it becomes a source of competitive advantage, and they partner when it functions as shared infrastructure. The shift among oilfield services companies suggests a corollary that extends beyond energy. When proprietary data and domain expertise turn a broadly available form of AI into something genuinely distinctive, a company may be able to commercialize the resulting capability rather than using it solely to defend its own position. The ultimate test is not whether the companies have access to AI, but how effectively each company combines that access with the data and expertise that it uniquely holds.
The AI dealmaking boom will reward disciplined companies, not indiscriminate buyers. As models and tools commoditize, advantage will increasingly reside in the proprietary data, workflows, and expertise that surround them. Companies should partner or build to combine their proprietary data with the AI applications that the market can provide, and they should buy only when the resulting control protects a durable edge. To meet investors’ expectations, winners will make ownership the exception and adaptability the rule.
The authors are grateful to Daniel Kim, Dominik Degen, and Duc Loc Nguyen of BCG's Transaction Center for their valuable insights and support in the preparation of this article.
