The AGI race is an all‑pay auction. That’s why “over‑investment” is rational.
This unprecedented race has an AI bubble built in by design
When the prize is “winner‑takes‑all” and everyone must pay their costs whether they win or lose, you don’t get measured competition—you get value (rent) dissipation [3]. That is what contest theory calls an all‑pay auction [0]. In expectation, participants spend roughly the entire value of the prize in aggregate trying to win it [1][2]. What happens when the perceived value of the prize is nearly infinite?
For AGI—where the imagined prize is monopoly‑like profits across software, science, society, the next industrial revolution, the whole fabric of human civilization—equilibrium spending is enormous by construction. In this worldview, the seemingly excessive capital allocation is rational: if you cut spending while rivals do not, you lose the race and everything you’ve already invested. Google co‑founder Larry Page has allegedly asserted (as relayed by investor Gavin Baker): “I am willing to go bankrupt rather than lose this race” [4].
These massive investments are well known: Microsoft told investors it expects capital expenditures (capex) to exceed $30 billion in a single quarter [5], while Alphabet lifted its 2025 capex plans to ~$85 billion [5]. McKinsey maps a ~$6.7–$7 trillion global data‑center build through 2030 [6]. Nvidia’s Blackwell‑class chips are $30k–$40k apiece [7]. OpenAI’s Broadcom deal targets 10 GW of custom accelerators [9]. The Stargate alliance (OpenAI–Oracle, with SoftBank) is scoped at up to $500 billion and 10 GW [10]. And the IEA estimates data‑center electricity demand will roughly double by 2030 [8].
For scale: the Manhattan Project cost roughly $2 billion in 1940s dollars (≈ $30–$50 billion today) [11], Apollo $25.8 billion then (≈ $250–$300 billion today) [12], and the Interstate Highway System is about $634 billion in 2024 dollars [13]. AGI‑related capex now rivals or exceeds the total costs of these projects every year.
A minimal model
An all‑pay auction is as it sounds: all bidders pay regardless of who wins. Once you have sunk enough resources into the race, the marginal cost to continue playing is small relative to the prize. If you bid $0.99 for a $1.00 prize, and a rival bids $1.00, it can be rational to raise by $0.02 to win—even though your total outlay is now $1.01 to win just $1.00. In this type of auction, the best strategy is often to not enter at all.
What makes this race look particularly like an all‑pay auction is the lack of a clear definition and finish line for what AGI even is [27]. There is an exuberant belief that sufficient investment will get us there based on known scaling laws, without clarity on sufficiency and the technological roadmap. While theory suggests that group spend should approach perceived value, there lacks clarity as to what that final value looks like. In addition, real‑world experiments show that people tend to over‑bid relative to theory [15]. This mirrors the current tech race, where status, imperfect information, “learning while doing,” and fear—fear of missing out and being a loser—can cause participants to overshoot theory.
A common instance of an all-pay auction are the infamous penny auctions, exploiting ill-informed bidders, for example bidders have paid $7,264 for a $180 item [28].
An unprecedented moment in human history
There has never been this combination of scale, competition, and investment in fast‑depreciating assets—burning through GPUs, electricity, and the time of ultra‑scarce (and ultra‑expensive) AI scientists—all financed primarily by private balance sheets. Unlike rail rights‑of‑way or overbuilt fiber that retained residual value, today’s frontier spend sits on a few‑year clock—servers and GPU accelerators that turn over quickly (even as shells and interconnects endure)—while data‑center power demand is set to roughly double by 2030 [8].
The rapid depreciation dynamics combined with the sheer scale and intense private competition makes this the largest and most unprecedented all-pay-auction in the history of humanity, with enormous implications for both our future and for the economy regardless of the outcome.
What happens if AI leaders are correct? A bubble.
If all the tech billionaires chasing AGI are correct—that the prize is singular (winner‑takes‑all or most) and that AGI is attainable—the economic situation looks ominous:
Most of the value gets burned en route. In canonical all‑pay contests, aggregate spend tends to the value of the prize [1]; experiments often find over‑dissipation [15]. With a dozen serious AGI bidders, more than 90% of ultimate value will be burned by the group in expectation [0][1]. The prize can be hot, and the bonfire can burn just as hot.
Losses concentrate in ordinary savings. The losers are not just founders and a handful of VCs. Public markets, pensions, and retirement accounts are bankrolling these investments. About 53.7% of U.S. households own mutual funds [16]; pensions are a massive balance‑sheet item in the national accounts [17]. If all but one of the dozen major bidders lose, those losses propagate back into the vehicles most people own.
AGI’s value will take significant time to pay out. There’s a difference between creating value and capturing value. There’s also a difference between inventing a technology and deploying it. Given the severity of the economic fallout, there won’t be a wellspring of capital to flow to the winner, and deployment will not be instantaneous; the world will still need to adapt to the new reality.
This is why “AI bubble” worries do not require technological failure. Contest logic alone can yield bubble‑like P&L: one epic winner, many deep losers, and an index that looks over‑capitalized until cash flows land. Markets are already keying on capex as the hinge variable [21]; buybacks are giving way to record spending, and patience has limits.
How the auction might end
All‑pay auctions do end. In the real world, they stop—or morph—through a few predictable doors:
The prize becomes legible. If we converge on an operational definition of “AGI” and even a clear roadmap to attaining such a thing, firms can price the end state and reduce the fog of war. When the prize is better measured, marginal bids shrink toward measured value.
A decisive lead becomes common knowledge. Once one player’s edge is credible and durable, rivals’ expected payoff from the next dollar collapses; they quit. The cousin model here is the war of attrition: when others are expected to quit first, the rational move is to stop before you bleed out [24].
The prize fragments. If the market sustains multiple niches—foundation models vs. specialized agents; chips vs. base models vs. vertical apps—effort is spread across several “first prizes,” and aggregate waste falls. Contest theory shows prize structure matters; multiple prizes typically dissipate less value than a single jackpot [25][26].
Coalitions / Mergers form. Joint ventures, chip alliances, and long‑term supply pacts turn rivals into partial teammates on the costliest inputs (silicon, power). Think OpenAI–Broadcom’s 10 GW accelerator program [9] and the Stargate build‑out with Oracle and SoftBank [10]; these are coalition bids on the input stack rather than duplicative concrete and copper.
Budget constraints bind. If revenue trails capex for long enough—or bottlenecks bite—capital markets ration the marginal dollar. Citi projects AI infrastructure capex of ~$490 billion by 2026 [14]; investors are already grading firms on whether spend turns into cash flow [21], with grid constraints looming large [20].
Policy or physics shrink the prize. Licensing, export controls, interconnection queues, and power availability can lower the feasible return to scale. The IEA’s scenarios point to tight power fundamentals [8]; grid frictions and interconnection hurdles are mounting [20][22][23].
AGI is discovered to be unattainable. If the frontier stalls, capital reprices and the contest ends abruptly. Then nobody wins and everyone, including society, loses. That possibility is precisely why coordination now is cheaper than a disorderly unwind later.
Can we spend smarter? Five ways to reduce waste
Merge or cross‑license earlier. If the game is truly winner‑takes‑all, consolidating upstream to avoid proving the same point four times is efficient. There are precedents for pre‑competitive collaboration that pooled risk but kept rivalry alive downstream—SEMATECH in U.S. chips; Airbus in European aerospace [18][19].
Form public–private “mission” vehicles. The scale is already approaching or exceeding traditionally nation-scale projects. Manhattan, Apollo, and the Interstate Highway System were centrally coordinated because duplication would be wasteful. Today’s AI build is already larger in many scenarios; a mission vehicle can concentrate procurement, reduce duplication, and publish interfaces that split the prize into layers [11][12][13][6].
Standardize and “split the prize.” Open interfaces and clear safety baselines make it easier for multiple winners to coexist—chips here, base models there, vertical apps elsewhere—defusing the incentive to torch the commons for a single crown [25][26].
Gate capex to real demand and real constraints. Tie the next tranche of GPUs or gigawatts to observed utilization, revenues, and interconnection progress. Grid operators and the IEA are already warning about the bottlenecks; investors are as well [8][20][21].
Make “coopetition” boring. Long‑term supply contracts, shared fabs, site‑level partnerships—think Stargate’s shared footprint—move competition up the stack toward product differentiation and away from duplicating fixed assets [10].
Conclusion: It’s rational for them to risk a recession for all of us
One spectacular winner, many deep losers, and the losses that reverberates into the rest of the economy. This is the consequences of the current structure of the AGI race: an all-pay single jackpot auction that is proven to expect massive rent dissipation. This is bubble-by-design, even if the technology proves to be everything the AI leaders claim. The firms are acting in their rational self interest and as predicted by contest theory, so we cannot just ask them to change their behavior. The only way to get them to steer clear of economic catastrophe is if the game changes or we change the game.
Sources
[0] All‑pay auction (Wikipedia) – overview and intuitive framing. Wikipedia
[1] Baye, Kovenock & de Vries (1996), “The All‑Pay Auction with Complete Information,” AER – formal equilibrium results and dissipation. Digital Commons
[2] Skaperdas (1996), “Contest Success Functions,” Economic Theory – how prize structure and effort allocation work. SpringerLink
[3] Rent‑seeking (Wikipedia) – intuitive treatment of rent dissipation. Wikipedia
[4] Yahoo Finance (2024), citing investor Gavin Baker relaying Larry Page’s alleged line. Yahoo Finance
[5] Reuters (July 2025): Microsoft >$30B quarterly capex; Alphabet lifts 2025 capex to ~$85B. Reuters
[6] McKinsey (2025), “The cost of compute: a $7T race to scale data centers.” McKinsey & Company
[7] Reuters (Mar 2024): Nvidia Blackwell pricing $30k–$40k. Reuters
[8] IEA, Electricity 2025 / “Data centres and AI” – data‑center demand roughly doubling by 2030. IEA
[9] Reuters (Oct 2025): OpenAI–Broadcom to deploy 10 GW of custom accelerators. Reuters
[10] Reuters (Oct 2025): Stargate (OpenAI–Oracle, with SoftBank) up to $500B / 10 GW. Reuters
[11] National Park Service / CMU: Manhattan Project cost (~$2B then; $30–$50B today). ethos.lps.library.cmu.edu
[12] Planetary Society: Apollo (~$25.8B then; $257B 2020 dollars). Planetary Society
[13] Interstate Highway System (Wikipedia): ~$634B in 2024 dollars. Wikipedia
[14] Reuters (Sept 2025): Citi projects AI infra capex $2.8T by 2029; ~$490B by 2026. Reuters
[15] Gneezy & Smorodinsky (2006), “All‑pay auctions—an experimental study.” Rady School of Management
[16] ICI (2024): 53.7% of U.S. households own mutual funds. ICI
[17] Federal Reserve, Z.1 Financial Accounts (Table L.229): pension entitlements are a large household asset. Federal Reserve
[18] CSIS (2023): SEMATECH lessons for pre‑competitive consortia under the CHIPS Act. CSIS
[19] Airbus (Wikipedia): origins as a European industrial consortium. Wikipedia
[20] S&P Global Market Intelligence (Oct 2025): grid congestion and interconnection bottlenecks. S&P Global
[21] Reuters (Oct 27, 2025): buybacks take a back seat as record AI capex stretches investor patience. Reuters
[22] Reuters (Apr 2025): Google–PJM collaboration to use AI for faster interconnection. Reuters
[23] Utility Dive (Sep 2025): PJM stakeholder debates on data‑center demand response and curtailment. Utility Dive
[24] War of attrition (Wikipedia): “others quit first” logic. Wikipedia
[25] Barut & Kovenock (1998), “The symmetric multiple‑prize all‑pay auction.” ScienceDirect
[26] Faravelli (2012), “Single vs. multiple‑prize all‑pay auctions to finance public goods.” IDEAS/RePEc
[27] A Definition of AGI (2025) Arxiv
[28] Penny Auctions - How to sell a $180 tablet for $7,264. Curious Gnu