The New Rules of Partnerships · Chapter One
Randy McGraw · Chapter One, open to read
Two are better than one, because they have a good reward for their toil... a threefold cord is not quickly broken.
— Ecclesiastes 4:9–12
In the AI era, strategic advantage comes from assembling scarce technical and commercial assets faster than rivals. Partnerships — from short, binding commercial contracts to staged joint ventures, syndicated R&D and projects, and co-investments — are the primary mechanism to combine distribution, technical and commercial capabilities, data, models, talent and regulatory positioning at speed and scale. And these structures only succeed when commercial incentives, measurement, governance and economic, legal and tax engineering are designed together. That, in one paragraph, is the argument of this book. This chapter establishes the first half of it — the indispensability. The chapters that follow establish the second half — the design.
Private enterprises can no longer rely on standalone efforts to win and sustain market leadership. Partnering is now strategic, not optional. Rapid digital transformation and the rise of AI mean that critical capabilities — technical expertise, distribution channels, customer relationships, and operational scale — move faster and cross borders with unprecedented speed. Securing scope, distribution, know-how, new supplier and customer contracts, data access, or a defensible market position increasingly depends on timely, well-rationalized and well-structured partnerships.
Driven by changes in technology, today’s collaborations can take many forms: commercial contracts, joint ventures, IP sharing, earn-ins, project co-investment, and other new formats and structures. Each must align technology, execution assets, risk tolerance, commercial incentives, organizational culture, and governance. Success requires understanding success levers and their interconnections, adjacency to global competitive dynamics, anticipating platform and marketplace shifts, integrating data and workflows, and designing commercial and equity arrangements that convert transient advantages into durable value. Firms that move slowly, negotiate from incomplete information, or lack skills in dealcraft face lost opportunities, diluted returns, and exposure to competitors who are quicker to align.
These are large claims, and a book that asks its reader to reorganize a core strategic discipline owes that reader more than assertion. It owes evidence. So before we turn to how AI transforms the partnership landscape, we must first walk through the longstanding motives that explain why firms choose, and often need, to partner — motives that are evergreen — and test each one against the historical record. Six rationales; six cases; six measurable unlocks of value that no standalone effort could have produced.
The first and oldest rationale: local partners provide channels, customer relationships, regulatory legitimacy, and faster go-to-market rollouts. The door to a market is opened from the inside — and no case in industrial history demonstrates it more completely than Maruti Suzuki.
In 1981, the Government of India founded Maruti Udyog Limited to solve a problem decades of autarkic industrial policy had failed to crack: the absence of an affordable, modern car for a middle class that did not yet drive. In 1982 it signed a joint venture and license agreement with Suzuki Motor Corporation — then a modest Japanese maker of small cars, dwarfed by Toyota and Nissan at home and barely visible abroad. Company lore holds that Suzuki only pursued the deal after a senior director, on a domestic flight in India, read that rival Daihatsu was circling the same opportunity. Suzuki entered with a 26 percent stake. The first car, the Maruti 800, launched in December 1983 and drew 120,000 bookings in its first year, in a country where waiting lists for the incumbent Ambassador stretched for years. As results and trust compounded, Suzuki raised its stake to 40 percent in 1987, 50 percent in 1992, and ultimately a controlling majority above 56 percent after the government’s exit — trust converted into structure, one tranche at a time.
The scoreboard four decades on: Maruti Suzuki commands more than 40 percent of the Indian passenger vehicle market — a level of dominance no global automaker has achieved in any other major market — and India has become the center of gravity of the entire Suzuki enterprise. Maruti is Suzuki’s largest subsidiary, at times contributing roughly 39 percent of the parent’s global revenue and more than half of its worldwide unit sales. What did Suzuki supply that India lacked? Technology, product, and operating discipline. What did India supply that Suzuki could never have replicated centrally or internally? Localized market knowledge, regulatory cover, political legitimacy, and channel trust. Each party held the half the other needed — the defining geometry of every partnership this book will examine.
I know the grammar of the Suzuki story from the inside, because I spent four years writing a sentence in it myself. From 2019 to 2023 I led partnership development for Amazon Prime in Japan — one of Amazon’s highest-priority consumer growth markets, and one of its hardest. Japan’s e-commerce battlefield is dominated by two forces: Amazon, and Rakuten, the domestic challenger that had spent years deliberately constructing one of the most integrated loyalty ecosystems in the world — points flowing across e-commerce, travel, financial services, mobile and content, engineered so that the Rakuten relationship reinforces itself: spend anywhere, earn everywhere, stay forever. That was the competitive reality Amazon faced. And it was equally the reality NTT Docomo faced in its own market — a carrier watching a retailer build the points-driven customer retention infrastructure that telcos had historically owned. Our partnership was not born from a distribution conversation. It was born from a shared threat assessment — a point we will return to, because it is the answer to the first cause of partnership failure in Chapter Two.
What we built with NTT Docomo and KDDI was not one deal. It was four, constructed in sequence over several years — each one making the next possible. First, app preloads: Amazon’s properties pre-installed across the carriers’ full handset ranges — more than forty million devices, over twenty million new handsets a year — putting first contact on the home screen before the consumer had decided to look. Useful, but commercially shallow: no shared economics, small aligned incentive based on a decrement to the HVA value. Second, carrier billing: Prime’s ¥600 charge routed directly to the phone bill the customer was already paying — no new payment credential, no card entry, no visible moment of commitment. In a market where payment friction kills subscription adoption, this was not a payments feature; it was a flywheel ignition mechanism, and it had to be negotiated as a commercial term, not designed as a product feature. It also created the first shared, measurable, attributable outcome — the proof that funded the more ambitious conversation. Third, Prime subscription bundling itself — which those proofs unlocked — delivering more than five million Prime subscribers and over two million first-time Amazon customers, jointly acquired. And fourth, the d-point loyalty integration, which I initiated and took to MOU before moving on. Combined value: approximately $950 million — at the time, the two largest telco partnerships in Amazon Prime’s global history.
Set the two stories side by side and a principle emerges that this book will invoke repeatedly: sequence the commitment to match the evidence. Suzuki did not buy its Indian majority in 1982; it bought 26 percent, and let four decades of delivered results escalate the structure — 26 to 40 to 50 to control. We did not open with Prime bundling; we opened with preloads, and let each layer’s measured proof underwrite the next layer’s ask. The progression was not a roadmap presented at the outset — it was a discipline: you do not ask for the high-value structure until the low-value structure has produced proof. Staged commitment is how strangers become partners without either side betting the company on faith; it converts trust from a sentiment into a schedule. It is, as we shall see, also the front-side antidote to more than one of the failure causes ahead.
The second rationale: sharing CAPEX and OPEX for expensive infrastructure or large projects can reduce unit economics and investment risk. In industries where scale is destiny, pooling is not thrift — it is survival, and the Renault–Nissan Alliance is its canonical modern proof.
Formed in 1999 when Renault rescued a nearly bankrupt Nissan, the Alliance never merged its members; the two remained separate companies with separate brands, boards and shareholders. What they merged was cost structure. Through the Renault–Nissan Purchasing Organization, the partners moved from pooling 30 percent of their purchasing in 2001 to 100 percent by 2009, facing suppliers as a single buyer commanding the volumes of two global automakers, while converging engineering, manufacturing, logistics, common vehicle platforms and shared powertrains. Crucially, the Alliance measured and published the result every year: annualized synergies — cost reductions, cost avoidance and incremental revenue unavailable to either partner alone — climbed from €1.75 billion in 2011 to €2.69 billion in 2012, €5 billion in 2016, and €5.7 billion in 2017, the year the enlarged Renault–Nissan–Mitsubishi group sold 10.6 million vehicles and became the world’s largest automotive group by volume. Dwell on that number: €5.7 billion in a single year exceeds the total annual profit of most automakers. It was a phantom third company — a profit engine that existed only in the space between the partners.
The third rationale: combining complementary capabilities can avoid multi-year internal builds and capture first-mover advantage. And there are moments when speed is not a competitive variable but the entire game. No case makes the point like Pfizer and BioNTech.
In early 2020, BioNTech was a mid-sized German biotechnology firm with deep expertise in messenger-RNA science — a technology that had never yielded a licensed product anywhere in the world. Pfizer was a global pharmaceutical giant with unmatched capability in clinical trials, regulatory navigation, manufacturing scale-up and worldwide distribution — but no mRNA platform. Apart, each held half an answer to the deadliest pandemic in a century. Together, working critical aspects in parallel rather than in sequence, they compressed a process that historically consumes a decade into less than a year: roughly seven months from the first reported COVID-19 case to initiation of the pivotal trial, and eleven months to Emergency Use Authorization in the United States in December 2020.
The ledger of that speed is staggering on every axis. Three billion doses manufactured by the end of 2021. Peer-reviewed estimates that the vaccine’s first year of rollout helped avert some nine million symptomatic cases, seven hundred thousand hospitalizations, and more than one hundred ten thousand deaths. Commercially, Comirnaty posted peak sales of approximately $37.8 billion in 2022 — among the highest-revenue pharmaceutical launches in history — with gross profits split evenly between the partners. BioNTech, whose entire 2020 revenue was under half a billion euros, booked €19 billion in 2021, a thirty-eight-fold increase. Neither firm, acting alone, arrives in 2020. Perhaps neither arrives at all. When complementary capabilities are stitched together rather than built sequentially, time-to-market collapses — and in winner-take-most windows, the value of collapsing it is nearly unbounded.
The fourth rationale: distributing technical, commercial and financial downside for commercial experiments, R&D, or asset-heavy initiatives represents active de-risking. Some bets are simply too large and too uncertain for any single balance sheet — and in 2012, the future of the entire semiconductor industry was one of them.
Extreme ultraviolet (EUV) lithography — the technology required to keep shrinking transistors and thereby extend Moore’s Law — was a machine of nearly incomprehensible difficulty, years behind schedule and consuming R&D at a rate that threatened to outrun what its developer, the Dutch equipment maker ASML, could prudently fund. If EUV failed, the industry’s entire roadmap failed with it. ASML’s answer was a structure of remarkable ingenuity: the Customer Co-Investment Program, under which its three largest customers — Intel, TSMC and Samsung, themselves ferocious competitors — jointly committed €1.38 billion to next-generation R&D over five years and paid a further €3.85 billion for an aggregate 23 percent equity stake. Intel’s total commitment alone approached $4.1 billion. The design was as important as the dollars: the shares were non-voting, so no customer could steer the technology to private advantage; the program’s benefits were explicitly declared industry-wide; and the equity proceeds were returned to existing shareholders so the structure diluted no one. Three rivals who trusted one another with almost nothing agreed to trust a neutral supplier with the shared foundation on which all of them stood. A decade later, EUV machines were the sole gateway to leading-edge chips, ASML had become Europe’s most valuable technology company, and every AI accelerator now training the models of this book’s title descends, in part, from that shared underwriting of risk. Well-designed partnership structures make fundable what no rational actor would fund alone.
The fifth rationale: obtaining technology, manufacturing, specialized components, domain expertise, licenses, or channel relationships the firm lacks. In fast-moving technical domains the build-borrow-buy calculus tilts decisively toward borrow — because by the time an internal build matures, the frontier has moved. The defining case of the AI era is Microsoft and OpenAI.
In 2019, Microsoft’s leadership privately worried the company had fallen years behind in machine learning, particularly relative to Google. Rather than attempt a decade of catch-up research, Microsoft invested $1 billion in OpenAI, a small San Francisco laboratory with extraordinary research talent and no commercial infrastructure. The complementarity was almost geometric: OpenAI possessed frontier models and needed oceanic computing capacity and enterprise distribution; Microsoft possessed Azure, a global salesforce and the world’s most entrenched productivity software — and needed frontier AI. After ChatGPT reached one hundred million users within two months of its late-2022 launch — the fastest consumer adoption in history — Microsoft deepened its commitment to $13 billion, securing exclusive cloud-provider status and rights to embed OpenAI’s models across its product line. The returns compounded on multiple axes at once: analyses of company disclosures suggest roughly $30 billion in partnership-related revenue between 2023 and 2025, the bulk from Azure workloads, alongside Copilot products adopted by nearly seventy percent of the Fortune 500; and after OpenAI’s 2025 recapitalization, Microsoft’s stake was valued at approximately $135 billion — a tenfold paper return before counting a dollar of operating revenue. JPMorgan analysts suggested it might prove some of the best money ever deployed. Partnership is how a firm rents the frontier while the frontier is still worth renting.
The sixth rationale: on-the-job collaboration accelerates skill transfer, process maturity, and organizational adoption — far more effectively than vendor handoffs or one-way consultancy. Some knowledge cannot be bought, licensed or reverse-engineered; it can only be absorbed through joint work. The proof is a factory in Fremont, California.
In 1982, General Motors closed its Fremont assembly plant, by most measures the worst in the GM system: quality at the bottom of the league tables, absenteeism so chronic that managers overstaffed shifts to keep lines running, labor relations poisoned by wildcat strikes, cars towed off the line inoperable. Two years later the same plant, with substantially the same unionized workforce, was producing vehicles whose quality rivaled the best Toyota factories in Japan. What changed was the arrangement: in 1984 GM and Toyota opened New United Motor Manufacturing, Inc. — NUMMI — a fifty-fifty joint venture in which GM contributed the shuttered plant and Toyota agreed to run it. Each partner came to learn what it could not learn alone. GM needed to understand the Toyota Production System — then a near-mystical black box to Detroit, despite years of external benchmarking that had transferred nothing. Toyota needed to learn how to build cars in America, with American workers and unions, ahead of tightening import restrictions. Workers sent to Japan returned transformed; the turnaround was nearly immediate and, to Detroit’s establishment, nearly unbelievable.
NUMMI also teaches a harder lesson. As alliance scholar Benjamin Gomes-Casseres observed when the plant finally closed in 2010, the venture never made much money — its real product was learning, and Toyota harvested far more of it than GM did. Toyota converted its tuition into a North American manufacturing empire; GM cycled executives through Fremont for a quarter century and struggled to transplant what they saw into its own culture. Same partnership, same contract, same factory floor — radically different value captured. Partnership is not a transaction; it is a capability. The learning channel only delivers to organizations built to receive.
If the six rationales were static, partnership would merit a respected chapter in the strategy canon and nothing more. They are not static. The way AI changes product economics, customer behavior and competitive dynamics intensifies the need to partner along every one of the six dimensions simultaneously.
Run down the list again with AI-era eyes. Market access and distribution: AI deployments often require localized data, regulatory approvals and trusted incumbents — partners supply localized datasets, regulatory cover and channel trust that are hard to replicate centrally or internally. Resource pooling and cost sharing: model pretraining, fine-tuning and inference at scale demand large compute and labeled datasets; sharing these assets across partners materially lowers per-model cost and makes experiments economically viable — the EUV problem writ general. Speed to market: AI value accrues quickly; partnering lets firms stitch best-of-breed models, tooling and domain data to launch usable products in months rather than years. Risk sharing: AI projects have high technical failure rates and uncertain monetization; earn-outs, minority equity and JV structures align incentives and make high-risk bets fundable. Complementary capabilities: AI requires narrow, high-value inputs — proprietary datasets, specialized models, edge hardware, domain annotation — and partnerships provide these precise complements faster than internal build. And learning: joint development, secondments and embedded teams accelerate AI, MLOps, data governance and product integration learning far more effectively than any vendor handoff.
The amplification runs deeper still. Faster product development cycles mean time-to-market advantage compounds rapidly, and shorter product lifecycles create a cadence of continuous retraining and upgrade that partnerships and JV operations are better structured to support than one-off vendor contracts. Commoditization pressure on core features forces firms to combine unique data, domain expertise and channels via partners to sustain both reach and pricing power. Outcome-based commercial models — pay-for-performance and revenue share — require shared incentives and accounting certainty that only formal partnership frameworks enable. Ecosystem competition accelerates as startups and platform players assemble point solutions in weeks; incumbents need partner networks to match that breadth and speed without costly internal rewires. Upstream, hardware, data labeling, energy and compute provisioning have become strategic inputs, making supplier and infrastructure partnerships instruments of capacity security. And rising regulatory and stakeholder scrutiny makes neutral trustees, auditability and locally trusted partners de-risking factors for any commercial deployment.
Beyond amplifying existing drivers, AI creates discrete new commercial requirements that only formal partnership architectures can reliably deliver. AI increases the scarcity, specialization, and regulatory sensitivity of essential inputs — data, compute, models, talent — turning what were once advantageous options into strategic necessities that demand more frequent, deeper, and more structured partnerships. I have organized these requirements into a typology of ten AI-era imperatives. They are not a list to admire; they are a diagnostic instrument. For each imperative, the sections below give the reader four things: the principle itself; why it arises — the causal force that makes it structural rather than fashionable; the analytical test — the question a leadership team should ask to determine whether the imperative binds them; and the boundary conditions — when the imperative applies, and, just as importantly, when it does not. A typology without boundary conditions is a horoscope: vague enough to apply to everyone and therefore useful to no one. The discipline of asking when does this NOT apply to us is what converts a taxonomy into a toolkit — and it is the same discipline that will govern the failure-mode analysis of Chapter Two. Each imperative closes with the case evidence, because in a landscape moving this quickly, principles unanchored to deals are speculation.
The principle. Verifiable data lineage, auditable training pipelines and neutral custodianship are demanded by buyers and regulators; meeting that demand requires consortium governance, data trusts, licensing frameworks and escrow arrangements — in short, partnerships that engineer trust into the supply of data itself.
Why it arises. An AI system inherits the legal and reputational status of its inputs. A model trained on data of unknown or contested provenance is an unauditable model, and an unauditable model is an unsellable product in any regulated or enterprise market. The era of free scraping ended not by choice but by force — litigation, legislation and enterprise procurement standards converged on the same requirement: prove where it came from, and prove you had the right.
The analytical test. Can we demonstrate, to a regulator or an enterprise buyer, the origin of every critical input to our models — and our right to use it? If the answer is no for any input that matters, a provenance partnership is not optional; the only question is its form (license, trust, consortium, or escrow) and its price.
When it applies — and when it doesn’t. It binds hardest where the data is third-party, personal, creative, or contested, and where the buyer is an enterprise or the jurisdiction has an active AI or data regime. It binds weakly where a firm trains exclusively on data it wholly owns and generates — its own telemetry, its own transactions — or where the application is low-stakes and internal. The test is the data’s pedigree, not the model’s size.
The evidence. A licensed data supply chain has materialized in under three years. Google agreed in 2024 to pay Reddit roughly $60 million per year for structured, permissioned access to its content — and Reddit has since disclosed some $203 million in aggregate data-licensing agreements. OpenAI’s deal with News Corp was reported at up to $250 million over five years, and by mid-2026 OpenAI held roughly twenty verified publisher partnerships spanning more than 160 outlets in over twenty languages. The counterfactual proves the principle: publishers who did not partner litigated — the New York Times is suing OpenAI and Microsoft for billions, and a 2025 Munich ruling in the GEMA case strengthened rights holders’ hands across Europe. Partnership or litigation is now the binary; there is no third door marked “just take it.”
The principle. Co-development creates ambiguous ownership and downstream rights; partnerships must define licensing, revenue share and commercialization pathways ex ante to prevent costly disputes ex post.
Why it arises. AI collapses the neat boundaries on which IP law was built. When a model is trained on one party’s data, using another’s architecture, on a third’s compute, the resulting weights blend the contributions inseverably — a trained capability cannot be un-trained, returned, or partitioned the way a patent portfolio can. Default law was not written for assets that learn; whatever the contract does not allocate, litigation eventually will.
The analytical test. If this venture dissolved tomorrow, who walks away with the model, the data, the fine-tunes and the derivatives — and can each party keep operating? If the contract cannot answer that question cleanly today, the parties are carrying unpriced exposure that compounds with every training run.
When it applies — and when it doesn’t. It binds whenever value is co-created — joint development, contributed training data, shared fine-tuning, embedded teams. It binds far less in arm’s-length consumption: a firm calling a model API under standard terms faces vendor risk, not IP-allocation risk. The gradient is contribution: the more each party’s assets flow into the shared artifact, the more this imperative governs the deal.
The evidence. The most sophisticated partnership of the era is also its cautionary tale: Microsoft’s rights to OpenAI’s intellectual property were contractually bounded by the arrival of “AGI” — a trigger so ambiguous that the 2025 restructuring had to engineer an expert panel merely to verify whether it had occurred, a story Chapter Two will take up as an exhibit of weak dealcraft. The disciplines this imperative demands — explicit field-of-use licenses, fallback continuity rights, data escrow and trustee custody, hardened non-circumvention — are precisely the remedies Chapter Two prescribes under its second and third causes of failure.
The principle. Divergent export controls and national security rules force parallel commercial designs and neutral routing; partners enable compliant dual-market strategies that no single global architecture can deliver.
Why it arises. Advanced compute and frontier models have become instruments of national power, and their diffusion is now governed the way arms are governed. A firm serving markets on both sides of a control boundary cannot run one architecture; it must run two, with verified separation between them — and verification at that standard is not something a company can grant itself. It must be conferred by structure: partners, custodians, and frameworks that governments themselves accept.
The analytical test. Map every chip, model, dataset and workload in the venture against the control regimes of every jurisdiction it touches. If any asset is controlled in any market you serve — or plausibly will be within the life of the deal — the venture needs a two-track design, priced and governed from day one, not retrofitted after the first license denial.
When it applies — and when it doesn’t. It binds any venture with a cross-bloc footprint, advanced accelerators, frontier-scale models or sensitive data flows. It does not bind the single-jurisdiction venture built on commodity technology — though the boundary is migrating outward every year, and a deal designed without an answer to “what if this input becomes controlled?” is a deal with an unexamined assumption at its core.
The evidence. The defining case is Microsoft and G42. Alongside Microsoft’s $1.5 billion investment in the Emirati AI champion in April 2024, the two companies created a first-of-its-kind Intergovernmental Assurance Agreement — a binding framework between two private firms, developed in consultation with both the U.S. and UAE governments, setting enforceable standards across cybersecurity, export controls and technology transfer, data protection, responsible AI and know-your-customer practice. The price of admission was structural: G42 committed to divest Huawei technology, Microsoft’s president took a board seat, and G42 pioneered a “Common Operating Picture” giving U.S. partners real-time, cryptographically verified visibility over American chips deployed in the UAE. The architecture unlocked what money alone could not: the first U.S. export licenses for advanced AI chips to the Gulf, and a program that has grown into a $15.2 billion multi-year commitment. This is partnership functioning as an instrument of statecraft — private dealcraft absorbing treaty-like obligations — and it is the template every cross-bloc AI venture will now be measured against.
The principle. The AI stack is composable; partners let firms assemble models, tooling, datasets and inference infrastructure as modular building blocks rather than vertically integrate.
Why it arises. The layers of the stack — silicon, energy, data centers, models, data, applications, distribution — evolve at different speeds and demand wildly different capital intensities. Owning every layer means owning the depreciation and the obsolescence of every layer. Modularity converts fixed exposure into contractual flexibility: the firm re-composes its stack as layers commoditize, rather than writing down an empire.
The analytical test. For each layer of your stack, ask: is this a source of durable differentiation, or a utility? Own and defend the differentiators; partner the utilities — and re-run the test annually, because today’s differentiator is next year’s commodity. The failure is not choosing wrongly once; it is never re-asking.
When it applies — and when it doesn’t. It binds where a layer is fast-moving, capital-hungry, and available from multiple credible vendors. It does not bind — indeed it inverts — where the layer is the moat: proprietary domain data, a unique customer channel, a regulated license. Modularizing your moat is how incumbents wake up as commodities.
The evidence. The most consequential AI company on earth owns almost none of its stack. By late 2025 OpenAI had committed roughly $1.15 trillion in infrastructure spending across seven vendors — Broadcom ($350B), Oracle ($300B), Microsoft ($250B), Nvidia ($100B), AMD ($90B), Amazon Web Services ($38B) and CoreWeave ($22B) — composing chips, clouds and custom accelerators entirely through partnership contracts. The Oracle commitment alone, $300 billion over five years, is the largest single infrastructure contract in history. Whatever one concludes about the prudence of the sums, the architecture is unambiguous: at the frontier, the stack is assembled, not owned.
The principle. Cross-border AI and data work needs embedded legal, tax and accounting design to maintain compliance while preserving commercial economics.
Why it arises. In AI ventures, compliance is not a wrapper around the deal — it is load-bearing structure within it. Where data may reside, which entity may contract, how revenue is characterized for tax, which jurisdiction’s AI rules govern the model: each of these decisions moves margin. Compliance designed after the commercial terms routinely destroys the economics the commercial terms promised; the disciplines must be engineered concurrently or the venture ships with a defect in its keel.
The analytical test. Does our commercial design survive contact with every jurisdiction’s data, tax and AI regimes — and at what margin? The test is quantitative: model the compliance architecture into the unit economics before signature. If the venture only works assuming today’s rules persist, it does not work.
When it applies — and when it doesn’t. It binds every cross-border venture and every regulated vertical — health, finance, defense, critical infrastructure — and any deal whose data or models cross a boundary the parties do not control. It binds lightly in single-jurisdiction, unregulated applications; but even there, the tax characterization of AI revenue streams (license? service? royalty?) is unsettled enough to warrant design attention that most deal teams still skip.
The evidence. The pattern surfaces everywhere the previous imperatives’ cases have already taken us: the Microsoft–G42 IGAA is compliance engineering elevated to deal architecture; the GEMA ruling redrew training-data economics in Europe mid-deal; and the EU’s AI regime is doing to model deployment what data-protection law did to data flows a decade ago — turning legal design into a pricing input. The ventures that thrive treat regulatory engineering as a source of advantage: compliance others cannot match is market access others cannot buy.
The principle. Politically sensitive data or model sharing often requires neutral intermediaries and trustee services to build confidence between parties who could never extend it bilaterally.
Why it arises. Some of the most valuable combinations in the AI economy are between parties who are, and will remain, rivals — or between parties separated by politics no contract can bridge. Trust between them cannot be summoned; it must be manufactured, structurally: neutral custody, non-voting instruments, independent attestation, escrowed assets. The trustee is not a nicety; it is the enabling technology of otherwise-impossible deals.
The analytical test. Would each party hand this asset directly to the other? If the answer is no, the deal is not dead — it is a trustee design problem. Ask what neutral custody, verification or governance would let each party hand the asset to the structure instead. If no such structure can be designed, only then is the deal dead.
When it applies — and when it doesn’t. It binds when competitors pool assets, when data or models are politically sensitive, and when performance claims must be verified by a party neither side controls. It is unnecessary overhead in high-trust bilateral relationships over non-sensitive assets — and imposing trustee machinery where simple contract suffices adds cost and delay without adding confidence. The instrument must be matched to the trust deficit, not applied by reflex.
The evidence. ASML’s 2012 co-investment program remains the classic: three rivals funding a shared future through non-voting shares and industry-wide benefit, trust engineered where none existed bilaterally. The AI era has escalated the pattern to the level of nations: G42’s Common Operating Picture — cryptographic tracking of every workload on U.S.-origin clusters, visible to American policymakers — is a trustee mechanism between countries, embedded inside a commercial partnership. And the principle now reaches down into every AI-era commercial deal, through a discipline I learned the hard way structuring AI and cloud partnerships at Sertis: if your data partner controls the measurement environment, they also control the evidence base for your commercial entitlements. That is why I have become insistent that KPI measurement custody be addressed as a first-order term, not a side agreement resolved after close. You would not negotiate a revenue share without independent account reconciliation; you should not negotiate an AI-powered partnership without independent KPI attestation. Data escrow, model custody and neutral measurement — disciplines this book returns to repeatedly — are the trustee principle at deal scale.
The principle. In markets where the critical input is scarce and the capital requirement exceeds any single balance sheet, supply is secured by pairing long-term purchase commitments with financing — producing deal topologies in which supplier, customer, investor and investee collapse into the same counterparties.
Why it arises. When the constraint is physical — chips that cannot be fabricated fast enough, capacity that takes years to build — a purchase order secures nothing. What secures supply is making yourself structurally necessary to your supplier: their capital in your growth, your commitments in their order book. The loop is not an accounting trick; it is the rational response to scarcity plus capital intensity. But it imports a new class of risk: incentives that self-reference, valuations that feed on the loop’s own announcements, and systemic exposure if any link falters.
The analytical test. Trace each dollar around the loop and ask two questions. First: does the structure ultimately create real third-party revenue, or does it merely recycle the same capital between the parties while marking up the paper? Second: if one link in the chain falters, what cascades — and to whom? A circular structure whose incentives survive those two questions is a capacity-securing machine; one that does not is a correlated bet wearing a partnership’s clothing. In either case, the remedy is the same discipline this book teaches everywhere: independent measurement, staged commitments, and governance that no single node controls.
When it applies — and when it doesn’t. It applies when supply is genuinely scarce, dependence is genuinely mutual, and the capital requirement genuinely exceeds any one participant. It should not appear — and its appearance is a red flag rather than a necessity — in liquid, competitive supply markets, where circular structures serve mainly to flatter both parties’ reported demand.
The evidence. The 2025 AI capital cycle produced the largest circular architectures in business history. Nvidia agreed to invest up to $100 billion in OpenAI to finance a ten-gigawatt data center buildout — with OpenAI committing to fill those facilities with millions of Nvidia GPUs; in the event, Nvidia contributed $30 billion to OpenAI’s funding round while, by one analysis, every $10 billion Nvidia invests corresponds to roughly $35 billion of OpenAI spending on Nvidia chips. OpenAI simultaneously committed $250 billion in cloud purchases to its own investor Microsoft, structured a chip deal positioning it to become one of AMD’s largest shareholders, and late in the year Microsoft and Nvidia announced they would invest up to a combined $15 billion in Anthropic — Microsoft funding its partner’s chief rival, which in turn committed billions to Azure. Defenders call it a virtuous circle aligning suppliers, builders and customers against exploding demand; critics call it self-dealing with systemic risk. Both are describing the same object — which is exactly why the analytical test above, not the labels, must do the work. And the architecture is already propagating far beyond the AI majors, down the industrial supply chain the buildout pulls behind it. In July 2026, Prysmian — the world’s largest cable maker — signed a supply agreement with Molex worth up to €5.5 billion over ten years for optical fiber destined for AI data centers, with one detail that gives the whole game away: Molex paid €550 million in advance, before a single cable shipped. That is not procurement; it is capacity finance — the buyer has become the banker, underwriting the factories that will make what it intends to buy, because AI demand has inverted the scarcity: the scarce asset is no longer the customer’s order book but the supplier’s future capacity. Notice what the prepayment accomplishes as incentive design: Prysmian cannot take the money and under-invest, because the capacity is contractually spoken for; Molex cannot walk away from a decade of committed volume, because it has already paid for the factory. Each side has posted a bond against its own future defection — incentive alignment made physical, poured into concrete and drawn into glass. This is the doctrine I call economic engineering before governance, and it recurs throughout this book: get the incentive structure right and governance becomes light-touch; get it wrong and no steering committee can save you. Chapter Two returns to these circular structures as the probable failure case studies of the coming decade — the ones whose economic engineering was skipped.
The principle. Where traditional acquisition is blocked — by merger review, by valuation, by speed — AI capability is transferred instead through a partnership construct: hire the founding team and key talent, license the technology, and leave the corporate entity behind.
Why it arises. AI capability lives in two portable containers: people and model weights. Both can be transferred by contract — employment agreements and licenses — without a change of corporate control, and therefore, in many regimes, without merger review. When regulators made conventional AI acquisitions slow or impossible, dealcraft routed around the obstacle. The result is a genuinely new format on the partnership spectrum: more than a license, less than a merger, and governed by neither’s established playbook.
The analytical test. Ask first: what are we actually buying — the entity, or the capability? If the capability is severable (talent plus weights plus know-how), a license-and-hire structure can transfer it; if the value sits in customers, contracts, or the entity itself, it cannot. Then price the residue: the antitrust recharacterization risk (regulators have probed whether these deals are de facto mergers), the morale and litigation exposure of the stakeholders left behind, and the reputational cost of a structure widely perceived as leaving common shareholders and remaining employees stranded. A structure that transfers the capability but poisons the talent market against you has not created value.
When it applies — and when it doesn’t. It applies to talent-dense, pre-commercial targets whose value is team and technology, in environments where formal acquisition faces regulatory walls. It does not apply where the target’s value is its business rather than its brains — and it is increasingly hazardous even where it fits, as authorities on both sides of the Atlantic test whether the form should be reviewed as the substance it resembles.
The evidence. In March 2024 Microsoft hired Inflection AI’s co-founder Mustafa Suleyman — installing him as CEO of Microsoft AI — along with most of the company’s team, while paying roughly $650 million in licensing arrangements to a company it never acquired; Inflection, valued at $4 billion a year earlier, was left with a skeleton staff. Google followed with a $2.7 billion license that returned Character.AI’s founders; Amazon executed a comparable arrangement with Adept; and Meta paid $14.3 billion for 49 percent of Scale AI while hiring its chief executive — in aggregate, an estimated $40-plus billion deployed through license-and-acquihire structures in under two years. The Department of Justice has probed the Character.AI arrangement and sought information on the Inflection and Adept deals — a reminder that when partnership structures substitute for M&A, they eventually inherit M&A’s scrutiny.
The principle. AI has pulled power — and the land, water and grid capacity around it — out of the utility bill and into the deal sheet: gigawatt-scale compute is now secured through decades-long offtake agreements, co-development structures and direct equity in generation itself.
Why it arises. AI demand growth collides with the slowest-moving supply chain in the industrial economy. Power plants take a decade; grid interconnection queues run years; and no spot market can deliver a committed gigawatt in 2029. When an input is simultaneously essential, scarce and slow, procurement becomes partnership: only a long-duration, risk-sharing structure can underwrite the generation that does not yet exist.
The analytical test. Does our compute roadmap have an energy roadmap of equal contractual firmness? Match the duration and firmness of compute commitments against the duration and firmness of power commitments; any gap between them is unhedged exposure. Then interrogate the risk allocation: who bears development risk, regulatory risk, and completion risk on generation that has not been built — and is that allocation priced?
When it applies — and when it doesn’t. It binds any enterprise with gigawatt-class training or inference footprints and multi-year capacity commitments. It does not bind the inference-light business renting cloud capacity, whose energy exposure is embedded in the provider’s price — though such firms should understand that their providers’ energy partnerships now sit inside their own supply chain, and ask about them in diligence.
The evidence. Microsoft signed a twenty-year power purchase agreement with Constellation — reported at roughly $16 billion — to restart Three Mile Island’s 835-megawatt Unit 1, a retired reactor resurrected for a single customer’s data centers. Google committed to a 500-megawatt fleet of Kairos Power small modular reactors — the first corporate SMR fleet deal in the United States; Amazon invested $700 million in reactor developer X-energy alongside a $20 billion-plus nuclear-adjacent campus; Meta signed twenty years of output from a plant slated for retirement. By mid-2026, every major hyperscaler had at least one nuclear partnership, with thirteen announced projects committing nearly ten gigawatts. These are not electricity purchases; they are risk-sharing co-investments in generation — the fourth traditional rationale of this chapter, operating at civilizational scale.
The principle. Governments have entered the partnership landscape not merely as regulators but as partners, gatekeepers and co-investors simultaneously — and ventures at national scale now require a diplomacy track designed alongside the legal, tax and commercial tracks.
Why it arises. Nations have concluded that AI capability is strategic infrastructure — economically, militarily, culturally — and are acting accordingly: funding it from sovereign wealth, gating it with export controls, and demanding assurance in exchange for access. For the firm, this means the counterparty across the table is often, in substance, a state: the capital is sovereign, the consent required is governmental, and the deal’s durability depends on policy as much as performance.
The analytical test. Who must say yes besides the counterparty? Map every governmental stakeholder — on both sides — whose consent, license or acquiescence the venture requires, then price the assurance obligations (security frameworks, reporting, divestitures, local commitments) into the economics, and design explicitly for policy change: what happens to this structure when the government, or its priorities, turn over? A sovereign-adjacent deal that has not answered the transition question has not been diligenced.
When it applies — and when it doesn’t. It binds ventures involving national champions, sovereign capital, controlled technology or critical infrastructure — a category expanding fast enough that its boundary should be re-checked annually. It does not bind the purely commercial, single-jurisdiction transaction — but any AI venture of real scale should assume it is closer to this imperative than its deal team believes.
The evidence. Stargate — the joint venture among OpenAI, Oracle and SoftBank announced at the White House and slated to deploy $500 billion into U.S. AI infrastructure — is a private partnership consecrated as national policy. In the Gulf, the U.S. approved exports of chips equivalent to 35,000 of Nvidia’s most advanced processors each to the UAE’s G42 and Saudi Arabia’s Humain, contingent on rigorous security and reporting requirements — partnership terms set, in effect, by two governments. And the Microsoft–G42 architecture of Imperative 3 is this imperative’s fullest expression: a commercial deal whose real signatories included two nations. The lesson for the practitioner is uncomfortable but liberating: at the frontier, dealcraft and statecraft have merged, and the firms that staff for both will take the mandates the others cannot.
Abstract rationales become board decisions only when they cash out into concrete value, so it is worth being explicit about what, precisely, partnerships unlock. They unlock strategic and operational levers — accelerating time-to-market, scaling distribution, accessing technology and data, and de-risking investment to capture commercial value in the AI era. Concretely, though not exhaustively: faster time-to-market through access to proven products, engineering teams, distribution and regulatory approvals; expanded distribution through new channels, reseller networks and marketplace relationships that scale sales immediately; customer acquisition through shared customer bases, co-selling and white-label deals that lower acquisition cost; technology and IP access through licensed software, proprietary models, patents and curated datasets; operational capacity — manufacturing, fulfillment and service infrastructure that increases throughput without heavy capex; regulatory and compliance enablement that removes market-entry barriers; supply chain resilience through secured inputs and logistics capacity; cost sharing and risk mitigation across R&D, go-to-market spend and capex; local market expertise; talent and capability pooling without long hiring cycles; data and analytics scale — combined datasets, telemetry and labeling pipelines that improve model performance and monetization; brand and credibility lift; platform leverage that embeds a product into broader ecosystems with recurring revenue hooks; competitive deterrence and exclusivity that lock in advantaged routes to market and raise rivals’ barriers; and monetization multipliers — bundled offers, revenue-share models and shared go-to-market programs that accelerate unit economics and lifetime value.
Look back across the cases of this chapter and notice that every unlock on that list appears somewhere in the record — measured, attributed, and banked. Suzuki’s product multiplied by India’s market. Renault’s volume multiplied by Nissan’s volume. BioNTech’s science multiplied by Pfizer’s machine. Three chipmakers’ capital multiplied by ASML’s engineering. OpenAI’s models multiplied by Microsoft’s distribution. Toyota’s knowledge multiplied by GM’s assets. Aristotle’s ancient observation that the whole can exceed the sum of its parts is, in partnership economics, not a metaphor but a mechanism.
Notice, too, what the cases quietly insist upon: in every one, the economics, the legal structure, the governance and the incentives were engineered together, not sequentially. ASML’s non-voting shares were not a legal detail; they were the strategy. The fifty-fifty profit split between Pfizer and BioNTech was not an accounting footnote; it was the alignment engine. The G42 assurance framework was not compliance paperwork; it was the key that opened the market. Cross-border partnerships require concurrent legal, tax, accounting, technology and behavioral design — and failure to integrate these disciplines up front causes unexpected costs, regulatory blocking points and unworkable governance.
Together, the traditional rationales and the AI-era imperatives mean partnerships are no longer just an execution choice — they are an integral part of product strategy, commercial design and competitive defense. Partnerships have been permanently transformed from optional business vehicles into practical governance and execution mechanisms for building value in the modern business world. Which raises the question this chapter has earned and the next must answer: if partnerships are this valuable, why do roughly half of them fail — and what separates the ones that don’t?
Framing, rationales, AI magnification, AI-era imperatives and value unlocks adapted from the author’s practice thesis on Partnership Development (2025), from which core language is carried forward directly. The ten-imperative typology extends the author’s original six imperatives (provenance/auditability, IP allocation, two-track geostrategy, capability modularity, regulatory engineering, trusted consortiums) with four imperatives reflecting the 2024–2026 deal record (circular capital, license-and-acquihire, energy partnerships, sovereign AI).
Maruti Suzuki: company and JV histories (founding February 1981; JV and license agreement October 1982; Maruti 800 launch December 1983, 120,000 first-year bookings; the Daihatsu anecdote); Statista (market share above 40%; Suzuki majority stake ~56–58%); Business Standard (Maruti ~39% of Suzuki global revenue; over half of unit sales, FY23–FY24); stake escalation 26% → 40% (1987) → 50% (1992) → majority.
Renault–Nissan Alliance: Alliance/Nissan press releases — annualized synergies €1.75B (2011), €2.69B (2012), €5.0B (2016), €5.7B (2017); RNPO purchasing pooling 30% (2001) → 100% (2009); 10.6M vehicles in 2017.
Pfizer–BioNTech: Nature Biotechnology, "Delivering 3 billion doses of Comirnaty in 2021" (7 months to pivotal trial; 11 months to EUA; 3B doses by end-2021); first-year impact estimates (~9M cases, ~700K hospitalizations, 110K+ deaths averted); Pfizer SEC filings (50/50 gross profit split); peak Comirnaty sales ~$37.8B (2022); BioNTech 2021 revenue €19B (~38x).
ASML Customer Co-Investment Program: ASML press releases and SEC Form 6-K filings (2012) — Intel, TSMC, Samsung: €1.38B R&D over five years plus €3.85B for 23% non-voting equity; Intel total ~$4.1B; synthetic buyback; industry-wide benefit.
Microsoft–OpenAI: CNBC/Forbes ($1B 2019; $13B total; 2025–2026 renegotiation); Microsoft blog (Oct 2025): stake ~$135B, ~27% as-converted; ~$30B partnership-related revenue 2023–2025; JPMorgan commentary; ~70% Fortune 500 Copilot adoption.
NUMMI: HBS case "Knowledge Transfer: Toyota, NUMMI, and GM"; B. Gomes-Casseres, HBR (2009); Lean Enterprise Institute materials.
Imperative 1 evidence: Reuters/Engadget (Reddit–Google ~$60M/yr, Feb 2024); Reddit aggregate ~$203M in data-licensing disclosures; News Corp–OpenAI reported up to $250M over five years (WSJ); industry trackers (~20 OpenAI publisher deals, 160+ outlets, 20+ languages); NYT v. OpenAI/Microsoft (SDNY); GEMA v. OpenAI, Munich Regional Court (Nov 2025).
Imperative 3 evidence: Microsoft On the Issues (Nov 2025) and UAE Embassy materials — $1.5B G42 investment (April 2024); Intergovernmental Assurance Agreement scope; Huawei divestiture; Brad Smith board seat; Common Operating Picture cryptographic tracking; first U.S. export licenses; $15.2B program (2023–2029).
Imperative 4 evidence: analyses of OpenAI compute commitments — ~$1.15T across Broadcom ($350B), Oracle ($300B), Microsoft ($250B), Nvidia ($100B), AMD ($90B), AWS ($38B), CoreWeave ($22B); Oracle five-year $300B contract (SEC-disclosed; largest single AI infrastructure contract).
Imperative 7 evidence: Bloomberg, "AI Circular Deals" (Jan 2026); The Register/CoStar analyses — Nvidia up to $100B into OpenAI ($30B contributed to funding round); ~$35B of chip purchases per $10B invested (NewStreet Research); OpenAI $250B Microsoft cloud commitment; AMD shareholder structure; Microsoft+Nvidia up to $15B into Anthropic.
Imperative 8 evidence: NAAG and antitrust literature on reverse acquihires; Microsoft–Inflection ($650M licensing; Suleyman as CEO of Microsoft AI; $4B prior valuation); Google–Character.AI ($2.7B); Amazon–Adept; Meta–Scale AI ($14.3B for 49%); ~$40B+ aggregate; DOJ probe of the Character.AI arrangement (Bloomberg/Forbes).
Imperative 9 evidence: Microsoft–Constellation 20-year PPA (~$16B) for Three Mile Island Unit 1 restart (835 MW); Google–Kairos 500 MW SMR fleet; Amazon–X-energy $700M; Meta–Constellation Clinton 20-year PPA; industry trackers: 13 deals, ~9.8 GW committed by mid-2026.
Imperative 10 evidence: Stargate JV (OpenAI, Oracle, SoftBank; $500B; announced at the White House, Jan 2025); U.S. Commerce/BIS approvals — chips equivalent to 35,000 GB300 processors each for G42 and Humain, contingent on security and reporting requirements.
Prysmian–Molex (Imperative 7): contemporaneous reporting (July 2026) — supply agreement worth up to €5.5B over ten years for AI data-center optical fiber, including a €550M advance payment funding U.S. capacity expansion; analysis first published by the author at m2-ventures.com ("When the Buyer Becomes the Banker").
Author’s deals (Amazon–NTT Docomo/KDDI; Sertis): drawn from the author’s own record and published materials (m2-ventures.com; the author’s published Docomo analysis) — four sequenced agreements (app preloads across 40M+ devices, carrier billing, Prime bundling, d-point loyalty MOU); ~$950M combined deal value; 5M+ Prime subscribers and 2M+ first-time Amazon customers; the "sequence the commitment to match the evidence" and "front-loading the hard questions" disciplines, and the KPI measurement-custody doctrine, are the author’s own formulations.
© Randy McGraw. From The New Rules of Partnerships: Winning Through Collaboration in a World of Scarcity, Speed, and AI. Chapter One is shared here by the author; no reproduction or use without written permission.
← Books