STRATEGIC DISEQUILIBRIUM SERIES — SUPPLEMENTAL NOTE
Why the Abundance Thesis Fails as a Bayesian Game
A Multi-Agent Scenario Analysis of the Musk Extrapolation, 2026–2036
Farid Novin
Executive Summary
On July 23, 2026, Elon Musk told The Economist that artificial intelligence will exceed the sum of human intelligence within roughly five years, and that the resulting deployment of humanoid robots will usher in what he called an “age of amazing abundance” by 2036, sufficiently vast that traditional money loses much of its function. This note treats that forecast not as a single prediction to be accepted or rejected, but as a hypothesis embedded in a system of strategic actors — AI developers, resource-holding states, grid operators, capital markets, and affected communities — each pursuing distinct payoffs under uncertainty. Recasting the abundance thesis this way, as a problem in Bayesian inference layered over a non-cooperative game, produces a sharper and more falsifiable critique than a simple appeal to physical limits. The conclusion is unchanged from the intuitive version — Musk's timeline is very unlikely to hold as stated — but the mechanism is now explicit: the binding constraint is not computation, it is coordination among self-interested players over a small set of contested physical bottlenecks.
I. Reframing the Question: A Decision Problem, Not a Forecast
This vision relies on the assumption that the recent surge in digital intelligence can be seamlessly extrapolated into physical reality. However, applying short-run technological trends to predict long-term global outcomes fundamentally ignores the mathematics of complex systems, resource finitude, and negative feedback loops.
Here is why that extrapolation breaks down.
I.i. The Extrapolation Fallacy: Exponential vs. Logistic Growth
The core flaw in Musk's prediction is treating early-stage technological growth as a permanent exponential curve rather than the beginning of a logistic S-curve. In the short run, feeding more data and compute into AI models yields rapid, seemingly unbounded improvements. But in physical systems, exponential growth inevitably encounters a "carrying capacity"—a hard limit where the cost of the next marginal improvement skyrockets.
Key insight: What looks like an infinite trajectory early on is often just the steep middle section of an S-curve before resource exhaustion forces a plateau.
I.ii. The Rugged Landscape of Multivariable Optimization
Predicting global economic outcomes is not a single-variable problem where we simply maximize compute. It is a problem of complex multivariable optimization.
When a system optimizes for one variable (like rapid AI scaling), it uses gradient ascent to climb the nearest peak until the derivative approaches zero, ∇f(x) = 0. However, this is almost always a local maximum. To reach Musk’s hypothetical "global maximum" of limitless abundance, society would have to traverse deep valleys—discontinuities representing severe economic restructuring, resource wars, or systemic failures. AI cannot simply compute its way across physical and societal chasms.
I.iii. Limitless Abundance in a Finite Environment
Musk suggests that millions of humanoid robots powered by superintelligence will create a "quasi-infinite economy." This directly violates fundamental laws of thermodynamics and material science.
Even if digital intelligence approaches infinity, the physical "end-effectors" (robots) are bounded by the physical world. While Musk correctly identifies electricity and chips as current bottlenecks, he glosses over the sheer physical footprint required to scale them:
Materials: Building millions of robots requires finite rare-earth metals, copper, lithium, and steel. Mining these is heavily constrained by geology and time.
Thermodynamics: Expanding computation at this scale generates immense heat, requiring vast amounts of cooling water and infrastructure that disrupt local ecosystems.
Energy: Data centers and physical robots require massive, continuous baseload power that renewable grids currently struggle to supply without relying on fossil fuels.
An algorithm can be perfectly optimized, but it cannot mine a ton of copper using zero energy.
I.iv. Negative Externalities as a Systemic Brake
Finally, the distribution of outcomes in an expanding industrial system always includes negative externalities—such as climate warming, grid instability, and ecological degradation. These are not just unfortunate side effects; mathematically, they are negative feedback loops built into the optimization function. If the drive toward physical abundance requires exponentially more energy, the resulting environmental degradation creates severe friction. Rising temperatures impact crop yields, extreme weather disrupts the very supply chains building the AI infrastructure, and resource scarcity drives geopolitical conflict. This feedback loop acts as a powerful, unavoidable brake. It ensures that before "limitless abundance" can be achieved, the system's own waste heat and friction will force a deceleration.
The above critique of the abundance thesis rested on a sound intuition — that exponential curves in digital systems do not translate automatically into exponential curves in physical output. But intuition alone under-specifies the problem. The more disciplined approach is to treat the future as a small set of competing hypotheses, assign each a prior probability grounded in the historical behavior of comparable general-purpose technologies, and then update those priors against the specific physical and strategic evidence available in mid-2026. Three hypotheses span the plausible outcome space.
The first, Hₐ (the abundance hypothesis), holds that compounding digital intelligence converts smoothly into physical abundance on Musk's stated timeline, with money becoming largely irrelevant by 2036. The second, Hₘ (the managed-plateau hypothesis), holds that digital capability continues to compound while physical deployment saturates against materials, energy, and delivery constraints, producing large but bounded gains rather than a phase change. The third, Hᶜ (the friction hypothesis), holds that the scramble for the same scarce inputs — critical minerals, grid capacity, water, and political consent — actively degrades the growth path through resource nationalism, regulatory backlash, and conflict, producing volatility rather than a smooth curve of any kind.
Every general-purpose technology in the historical record — steam power, electrification, aviation, semiconductor scaling itself — has followed a logistic trajectory: a steep early phase that looks exponential from inside the technology, followed by saturation as it collides with complementary physical infrastructure it does not control. That base rate assigns the heaviest prior weight, before any 2026-specific evidence is considered, to the managed-plateau hypothesis rather than the abundance hypothesis. The task of this note is to show that the specific evidence available today updates that prior further in the same direction, and to explain why through the language of strategic games rather than physics alone.
II. The Players and Their Payoffs
Musk's model implicitly treats materials, energy, and political stability as freely available inputs that scale automatically with capital and compute. A game-theoretic reading rejects that assumption and instead identifies four classes of strategic actor, each with a distinct objective function.
Frontier AI and robotics developers
These firms face a classic race dynamic. Because market share and investor confidence accrue to whoever scales fastest, the individually rational strategy for any single developer is to keep building regardless of the aggregate strain that collective scaling places on shared inputs. This is structurally a common-pool-resource problem: the atmosphere for compute-driven emissions, the regional power grid for electricity, and the refined rare-earth supply chain for actuators and magnets are all shared, depletable pools that no single firm has an incentive to conserve unilaterally. Tesla's own Optimus program illustrates the point — Musk has stated publicly that Optimus production is constrained by rare-earth magnet supply even as the company targets a long-run capacity of roughly one million units a year from a single facility
Resource-holding states
China currently controls approximately 91 percent of refined rare-earth production, the material base for the permanent-magnet motors every humanoid robot design depends on
Grid operators, utilities, and regulators
Global data-center electricity demand is on a trajectory from roughly 415 terawatt-hours in 2024 toward something approaching 945 terawatt-hours by 2030 on International Energy Agency figures
Host communities and downstream consumers
Local communities bear the water, land, and grid-stability costs of hyperscale buildout, and are increasingly organized, litigious, and politically salient veto players rather than passive bystanders. Every additional gigawatt-scale site is now also a local political negotiation, not merely an engineering project.
III. The Rugged Landscape, Restated as a Coordination Problem
The intuition that global economic optimization resembles a rugged landscape of local peaks rather than a single smooth hill remains correct, but it is better understood as a consequence of the multi-player structure above than as an abstract property of mathematics. Each AI developer climbs the one hill it can see and control — model scale, training compute, algorithmic efficiency — because that is the variable inside its own objective function. Materials, grid capacity, water rights, and geopolitical stability are variables inside other players' objective functions, and those players are not committed to Musk's abundance outcome as their own goal. Reaching anything resembling a global optimum of “limitless abundance” therefore requires simultaneous, voluntary alignment of incentives across developers, resource-holding states, regulators, and host communities — a coordination problem with no natural mechanism to resolve it, rather than a computation problem that more model capability can simply solve. Superintelligence can search a landscape faster; it cannot compel other rational, self-interested players to walk toward the same peak.
IV. Physical Bottlenecks as Evidence That Updates the Priors
Three bodies of 2026 evidence are directly relevant to updating the priors set out in Section I, and each points the same direction.
Materials
Industry analysis of the 2026–2027 period describes it as a genuine transition toward production-readiness for humanoid robotics, but one still critically dependent on rare-earth permanent magnets, specialized alloys, and engineering plastics whose supply chains are concentrated and slow to expand
Energy delivery
The evidence here is unusually clean because it is not a forecast of scarcity but a description of an already-binding constraint: grid operators such as ERCOT have revised their own 2030 data-center demand estimates upward from 29 gigawatts to 77 gigawatts within a single planning cycle
Thermodynamics and water
Even accounting for liquid-cooling designs that cut direct water intensity by an estimated 70 to 90 percent relative to air cooling, the underlying electricity demand — and the heat it generates — is unchanged
Taken together, this evidence does not merely qualify the abundance hypothesis — it actively transfers probability mass toward the managed-plateau and friction hypotheses, because the specific mechanisms Musk's forecast depends on (near-costless scaling of robots and their power supply) are the mechanisms most directly contradicted by the data.
V. Negative Externalities as an Endogenous Term in the Payoff Matrix
The first part of the essay treated environmental and grid externalities as an unavoidable brake on growth. The game-theoretic reading sharpens this: externalities are not an exogenous force acting on the system from outside, but the endogenous response of other strategic players — regulators tightening permitting, communities contesting water allocations, states rationing exports — to the costs the scaling race imposes on them. Because this is a repeated game rather than a single move, each round of visible strain (a halted auto plant, a stressed regional grid, a contested water permit) changes the strategies available to those affected players in the next round, typically toward more defensive, less cooperative postures. The brake the first part of the essay described mathematically as a feedback loop is, in strategic terms, the predictable reaction of rational third parties who were never party to the optimization objective in the first place and have no reason to subsidize it indefinitely.
VI. A Bayesian Scenario Tree Through 2036
Weighing the historical base rate for general-purpose technologies against the 2026 evidence on materials, energy delivery, and strategic resource control yields three scenarios with distinct posterior weights, described qualitatively below rather than as point estimates, in keeping with the genuine uncertainty involved.
Scenario one: abundance by 2036 (Hₐ)
This requires materials expansion, grid delivery, and geopolitical cooperation among rival resource-holding states to resolve simultaneously and within a decade — a conjunction of favorable outcomes across independent, only loosely correlated constraints. Musk's own framing implicitly concedes the fragility of this path by conditioning it on the absence of “a massive global thermonuclear war or something like that”
Scenario two: managed logistic plateau (Hₘ)
Digital intelligence continues compounding on the software margin — reasoning, coordination, design, scientific discovery — while physical deployment saturates against the materials and energy ceilings documented above, producing large, genuinely transformative productivity gains that fall well short of Musk's post-scarcity threshold. This is the scenario most consistent with the historical base rate for general-purpose technologies and with the specific 2026 evidence on grid and materials constraints, and it carries the heaviest posterior weight of the three.
Scenario three: friction-driven deceleration (Hᶜ)
Resource nationalism, grid backlash, and strategic competition over the same chokepoint inputs — chips, magnets, power — produce volatility, partial deceleration, or a redirection of AI-era gains toward military and strategic applications rather than consumer abundance. This is not a speculative tail case; it is already visible in the 2025 rare-earth licensing regime and the resulting production halt at a major automaker
A separate line of expert criticism reinforces the plateau and friction scenarios from the demand side rather than the supply side: American Institute for Economic Research analyst Peter Earle has argued that abundance in mass-produced goods does not eliminate scarcity but relocates it — toward irreducibly scarce inputs such as time, land, positional and status goods, and peak-hour capacity even in an energy-abundant system. This suggests that even in the most favorable branch of Scenario one, money would not become obsolete so much as it would be repriced around a narrower set of goods that remain genuinely scarce, a qualification Musk's account of the Treasury simply issuing checks does not address.
VII. Strategic Equilibrium and Its Policy Implications
Because the underlying structure among AI and robotics developers is a race with a dominant strategy of continued unilateral scaling, voluntary coordination to avoid straining shared physical inputs is not the likely equilibrium. The brake on the system, when it arrives, is far more likely to be imposed externally — by grid interconnection queues, by mineral-exporting states exercising leverage, by regulators and communities slowing permitting — than to emerge from self-restraint among the firms racing to build. This has a direct implication for policy audiences: the binding constraint on the AI-driven abundance trajectory is best treated as a resource-security and infrastructure-delivery problem — diversifying rare-earth supply chains, accelerating grid interconnection, and securing water and siting agreements — rather than as a problem that further gains in model capability will resolve on their own.
Conclusion
Musk's forecast is not wrong because exponential curves are impossible; it is unlikely to hold on the stated timeline because it models a single-agent optimization problem in a world that is actually a multi-agent game over a small number of genuinely scarce, unevenly distributed physical inputs. The 2026 evidence on rare-earth concentration, grid interconnection queues, and water-constrained cooling does not merely qualify the abundance thesis — it shifts the balance of probability toward a bounded, contested, and considerably slower path than the one described in the July 2026 interview. Whether the outcome by 2036 resembles a managed plateau or a friction-driven deceleration will depend less on how fast frontier models improve than on how the strategic contest over magnets, megawatts, and political consent is resolved between now and then.
Sources are drawn from primary interview transcripts, industry and market-research reporting, and peer-reviewed supply-chain analysis current as of July 2026; see footnotes.
No comments:
Post a Comment