Escalation Without Exit: Martin Shubik’s Dollar Auction and the Management Logic of the Global AI Infrastructure Race

In the last month, one of the clearest global management and technology trends has been the intensification of the artificial intelligence infrastructure race. Recent reporting and analysis show three connected developments: first, the International Monetary Fund has warned that growing interdependence and circular financing structures in the AI sector can create systemic financial vulnerabilities; second, BloombergNEF has described AI data-center building as advancing at full speed, with very large lease commitments and capital spending; third, the International Energy Agency has reported that electricity demand from data centers surged strongly in 2025, with AI-focused facilities growing even faster than the broader sector. Together, these developments suggest that competition in AI is no longer only about better models or better products. It is increasingly about scale, sunk cost, strategic commitment, and fear of being left behind.

This trend can be understood in a particularly useful way through Martin Shubik’s famous dollar auction. The dollar auction is a game in which participants bid for one dollar, but with a crucial twist: not only the highest bidder must pay, but also the second-highest bidder. Because the second bidder faces a sure loss if they stop, both parties are tempted to keep bidding beyond the value of the prize. The outcome often looks irrational from the outside, yet each next move appears temporarily rational to the people involved. The example is widely used in game theory to explain escalation of commitment, sunk-cost reasoning, competitive overreaction, and the difficulty of exit once losses have already been incurred. These concepts are old, but they have become newly relevant in a month when the AI economy has shown clear signs of strategic acceleration across finance, infrastructure, and energy systems.

This article argues that Shubik’s dollar auction offers more than a classroom metaphor. It provides a serious analytical framework for understanding current technology competition. The AI race is not identical to a dollar auction, and the analogy should not be pushed too far. Yet the structure is strikingly similar in several ways. Firms invest heavily in chips, cloud capacity, data-center space, power agreements, talent, and distribution. Once these investments are made, withdrawal becomes expensive. Competitors observe each other’s commitments and interpret them as signals. Markets reward participation in the race but may punish hesitation. Managers may continue to invest not only because expected future returns remain attractive, but also because the cost of stopping appears worse than the cost of continuing. In such conditions, escalation can become self-reinforcing.

The value of this framework is practical as well as theoretical. In management, the greatest errors often do not arise from ignorance but from distorted persistence. Organizations sometimes fail not because they invest, but because they lose the ability to distinguish disciplined commitment from trapped commitment. The difference is central. Disciplined commitment is based on fresh information, revised assumptions, and clear strategic logic. Trapped commitment is based on prior expenditures, prestige concerns, competitive anxiety, or fear of public retreat. The dollar auction shows how easily the second logic can hide inside the first.

At first sight, the current AI boom may not seem like a case of escalation. Many arguments for investment are real and serious. AI capabilities continue to improve. Enterprises are still exploring valuable use cases. Governments increasingly treat compute infrastructure as strategic. The OECD has also emphasized the growing importance of measuring and governing AI compute, while recent OECD work on digital measurement underlines the policy relevance of critical digital infrastructure such as data centers and cloud systems. In other words, there are solid reasons to invest.

However, the presence of real opportunity does not eliminate escalation dynamics. In fact, it can intensify them. The most dangerous auctions are often those in which the prize is genuinely valuable. The important question is not whether AI matters. It clearly does. The question is whether some actors are now investing under conditions in which the logic of continuation is being shaped by sunk costs, signaling pressure, and fear of strategic inferiority. The IMF’s recent warning is important precisely because it draws attention to interconnected funding, concentration, and amplification channels in the AI economy. Those are the kinds of conditions under which a rational contest can begin to resemble a strategic trap.

To understand why, it is useful to return to the mechanics of Shubik’s model. The dollar auction starts with a simple incentive: win the dollar at a price lower than one dollar. Early bids are easy to justify. But as the price rises close to one dollar, a new logic appears. If one bidder has already committed 90 cents and is about to lose, bidding one dollar and ten cents may appear preferable to accepting a loss of 90 cents. The other bidder then faces the same logic. What began as a small gamble becomes an escalating contest. Importantly, escalation is not driven by madness. It is driven by local rationality under cumulative commitment. Each individual choice may be defensible, even when the sequence as a whole is destructive.

This is exactly why the model remains relevant for management. Managers rarely wake up and choose absurdity. Instead, they make a series of small decisions that are reasonable within a narrow frame. A project is delayed, so more funds are added. A competitor launches, so the budget is increased. Market expectations rise, so targets are revised. Existing contracts lock in future obligations. Internal champions protect their earlier decisions. By the time the project should be questioned, too many reputations, budgets, and strategic narratives are attached to it. Exit becomes socially and politically difficult. The project survives, not because it is strong, but because it is already large.

The AI infrastructure race shows many of these features. BloombergNEF reported in March 2026 that data-center buildout is continuing at great speed, with the capital spending of the largest data-center firms approaching very high levels and large lease agreements involving hyperscalers and specialized compute providers. At nearly the same time, the IEA reported that electricity demand from data centers rose sharply in 2025 and that AI-focused data centers grew even faster, well above the pace of overall global electricity-demand growth. These are not small experimental commitments. They are large, physical, expensive, and slow-to-reverse investments.

This matters because infrastructure changes the psychology of strategy. Software can be adjusted quickly. Large physical infrastructure cannot. Once firms sign long-term leases, secure power, commit to equipment orders, recruit specialized staff, and build client expectations around future capacity, the cost of retreat rises dramatically. The issue is not only financial. It becomes organizational and symbolic. Managers do not merely abandon a building or a server cluster; they abandon a story about the future. In public markets and competitive industries, that story has value. So continuation becomes easier to justify than reversal.

A second reason the dollar auction analogy fits the present moment is that AI competition is highly observable. Rival firms can see each other’s model announcements, product launches, partnerships, capex plans, energy procurement, and talent movement. This visibility creates signaling contests. In classic game theory, signaling affects beliefs; in modern management, it affects valuation, employee morale, supplier confidence, and customer expectations. Once signaling becomes central, spending may serve two goals at once: productive capacity and strategic theater. The danger is that the theater begins to guide the capacity decision instead of the reverse.

The IMF’s recent concern about circular financing and interdependence in AI is especially important here. When firms invest in one another, depend on one another’s infrastructure, or combine financing with commercial commitments, the system can become more difficult to evaluate clearly. Valuation, supply dependence, market narrative, and strategic necessity begin to reinforce one another. This does not mean the system is unsound. It means that transparency becomes harder and the cost of correction may become higher. In a dollar auction, escalation comes from the inability to stop without loss. In a modern financial-technology system, escalation can also come from the inability to slow down without disrupting several linked balance sheets and expectations at once.

At a deeper level, Shubik’s model helps explain the difference between opportunity cost and commitment cost. In ordinary investment analysis, firms compare a project against alternative uses of capital. In escalation settings, however, managers may focus less on future alternatives and more on past commitments. The internal question changes from “What is the best use of resources now?” to “How do we avoid admitting that previous spending may not pay off?” This shift is subtle but powerful. It changes the time orientation of strategy. Forward-looking analysis is replaced by backward-looking defense.

In the current AI race, this risk is heightened by the political and cultural language surrounding technological leadership. Many firms and governments frame AI as a decisive strategic frontier. Such language can be justified, but it also creates a context in which caution is easily mistaken for weakness. Once competition is framed as existential, ordinary cost discipline becomes harder to maintain. The result is a classic escalation environment: high uncertainty, visible rivalry, irreversible commitments, and powerful reputational stakes.

This does not mean that all AI investment is excessive. That conclusion would be as simplistic as assuming every dollar auction must end badly. Some investments will generate strong returns. Some firms will achieve real scale advantages. Some infrastructure will become essential. The point is different: when a field becomes strategically central, managers must work harder, not less, to separate genuine positive expected value from continuation bias. The more important the field, the greater the danger of overcommitting in its name.

One of the strongest lessons from the dollar auction is that escalation is often relational rather than individual. A bidder does not escalate in isolation; escalation emerges from interaction. In the same way, the AI race is not simply the product of each firm’s internal optimism. It is produced by mutual observation and mutual reaction. One firm’s aggressive expansion changes the threshold of “normal” commitment for others. A data-center lease, a chip order, a power agreement, or a public investment plan by one actor becomes a strategic signal for the rest. This is why even prudent managers can be drawn into costly patterns. They are not only making a forecast about technology; they are making a forecast about what rivals will do if they hesitate.

This relational dimension is highly relevant for management education. Students often learn strategy as if firms choose from a menu of clear options. Real strategic environments are less calm. They are interactive, emotional, path-dependent, and shaped by public interpretation. The dollar auction remains valuable precisely because it compresses these features into a simple structure. It shows that rational people can produce irrational aggregate outcomes when they react to each other under pressure. For institutions such as Swiss International University (SIU), this remains a useful lesson for teaching strategic judgment in technology, finance, and public policy.

The model also helps explain why escalation frequently moves across levels of analysis. What starts as a project-level commitment can become a division-level target, then a corporate strategy, then a sector-wide narrative. In AI, the progression is visible. Technical development became product competition. Product competition became platform competition. Platform competition became infrastructure competition. Infrastructure competition is now linked to energy, capital markets, and industrial policy. The broader the field becomes, the more difficult it is for any one actor to de-escalate without appearing strategically exposed.

Energy is a particularly important example. The IEA’s recent analysis suggests that data centers are becoming a major force in electricity systems and that supplying AI-related demand will require large additional generation over time. This pushes the AI race beyond software economics into power economics. Once infrastructure competition enters the energy system, escalation becomes more durable. Energy contracts, grid planning, and generation investments extend the time horizon of commitment. They make rapid reversal less likely and increase the number of stakeholders who become dependent on continued expansion.

Another useful insight from Shubik’s model concerns the illusion of near recovery. In a dollar auction, each additional bid feels like a possible turning point. The losing bidder tells themselves that one more move may save the situation. Organizations often do the same. Under pressure, managers reinterpret setbacks as temporary and portray additional investment as the final step needed to unlock value. In some cases this is correct. In many others it is a psychological defense against recognizing that conditions have changed. The problem is not optimism itself; it is optimism that has stopped being conditional.

This is why governance matters. Firms in escalation-prone environments need governance systems that can challenge momentum. Independent review, staged investment, explicit kill criteria, counter-scenarios, and post-investment audits are not bureaucratic obstacles. They are defenses against path dependence. In high-velocity sectors, governance is often criticized for slowing innovation. But the real purpose of governance is not to stop risk-taking. It is to protect risk-taking from becoming compulsive.

The most effective managers therefore do something difficult: they remain ambitious while preserving reversibility where possible. They recognize sunk costs but refuse to obey them. They distinguish between “we have invested a lot” and “we should continue.” These statements are not equivalent. The first is historical; the second is strategic. Good management keeps them separate.

A further benefit of the dollar auction framework is that it clarifies why escalation often survives bad news. In ordinary linear reasoning, negative signals should reduce commitment. In escalation settings, negative signals can actually increase commitment, because they raise the cost of exit. This is one of the most important and least intuitive dynamics in strategy. Weak early returns, delayed customer adoption, infrastructure bottlenecks, or financing stress do not always cool competition. Sometimes they intensify it by making withdrawal more painful. The IMF’s concern about financial amplification channels is relevant precisely because these channels can magnify stress rather than discipline behavior smoothly.

This issue also has implications for public policy. When sectors become systemically important, policymakers are not merely regulating innovation; they are also managing escalation risk. If AI supply chains, financing structures, and energy dependencies become heavily concentrated, then the costs of abrupt correction may spread across the wider economy. In such conditions, the challenge is not only how to support innovation, but how to prevent strategic races from hardening into fragile structures. Better disclosure, clearer risk separation, and stronger infrastructure planning can reduce the chance that competitive intensity becomes systemic instability.

From a tourism and service-management perspective, the same logic appears in another form. Digital transformation in hospitality, airlines, booking platforms, and customer-experience systems increasingly depends on AI tools and cloud infrastructure. When firms in these sectors feel compelled to adopt AI rapidly because rivals are doing so, they can also fall into escalation patterns. They may continue spending on automation, personalization, analytics, or platform integration beyond the point where organizational readiness supports real value creation. The technology sector provides the most visible example, but the mechanism is broader. Escalation is portable.

At the personal level, the lesson is equally powerful. The dollar auction is not only about corporations or governments. It also explains why individuals stay in failing projects, continue unproductive negotiations, prolong bad investments, or defend decisions long after evidence has turned. Human beings often value not losing face more than making a fresh decision. This is why the model has remained influential far beyond formal game theory. It captures an uncomfortable truth about judgment: once we have paid enough, stopping feels like defeat even when stopping is wise.

What, then, should managers learn from the present AI moment? First, they should recognize that strategic importance increases the danger of escalation rather than eliminating it. Second, they should evaluate projects using forward-looking criteria only. Past spending can explain the context, but it should not justify the next move. Third, they should study rivals carefully without allowing rival behavior to define their own threshold of reasonableness. Fourth, they should design governance systems that permit retreat without humiliation. Finally, they should remember that in fast-moving sectors the greatest strategic advantage may not be maximum commitment, but superior judgment about where commitment should stop.

Seen in this way, the recent surge in AI infrastructure investment is not merely a technology story. It is a management story about commitment under uncertainty. The prize is real. The opportunities are substantial. But the structure of competition matters. When heavy investment, signaling pressure, interdependence, and fear of exclusion converge, escalation becomes a serious analytical concern. Shubik’s dollar auction remains valuable because it helps us see the shape of that concern before the system reaches the point where every participant feels unable to step back.

The wider relevance of this lesson is clear. Markets, organizations, and states often assume that more commitment signals more strength. Sometimes it does. But sometimes it signals something more fragile: the inability to exit gracefully. The most mature form of strategy is therefore not endless persistence. It is disciplined persistence combined with the courage to revise. In the age of AI infrastructure competition, this may be one of the most important lessons management theory can offer.

In conclusion, Martin Shubik’s dollar auction remains one of the most effective conceptual tools for understanding how rational actors become trapped in irrationally expensive contests. The last month’s developments in AI finance, data-center expansion, and energy demand make the analogy especially timely. Recent evidence from the IMF, BloombergNEF, the IEA, and the OECD suggests that AI competition is rapidly expanding across multiple systems at once. That does not prove that the current race is unsound. It does show that the conditions for escalation are present. For managers, investors, policymakers, and scholars, the task is therefore not to reject ambition, but to govern it. The future will belong not simply to those who bid the most, but to those who know when a bid remains strategic and when it has become a trap.

Hashtags:

#GameTheory #ArtificialIntelligence #StrategicManagement #DigitalInfrastructure #EscalationOfCommitment

Sources used:

• International Monetary Fund, Global Financial Stability Report, April 2026

• BloombergNEF, AI Data Center Build Advances at Full Speed: Five Things to Know, March 24, 2026

• International Energy Agency, Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions, April 2026

• International Energy Agency, Key Questions on Energy and AI, April 2026

• International Energy Agency, Electricity 2026

• OECD, AI Compute

• OECD, The OECD Going Digital Measurement Roadmap 2026

• OECD, OECD AI Observatory Index, February 2026