Vintage Breakeven Analysis: Methods & Applications

• Vintage-breakeven analysis is a framework that conditions investment viability on asset age, commissioning cohorts, discount rates, and technical heterogeneity.
• It is applied across fields—from capital investment and multi-year energy planning to fusion and AI—highlighting that simple aggregation can misstate critical thresholds.
• Researchers use analytical tools such as PDEs, NPV formulations, and annuity factors to identify vintage-specific metrics guiding optimal investment and retirement decisions.

Vintage-Breakeven Analysis denotes a family of analytical procedures in which break-even, viability, or steady-state conditions are indexed by a “vintage,” meaning either the age of a capital good, its commissioning cohort, or its purchase-year cohort. In the cited literature, the same organizing idea appears in age-structured optimal investment, discounted project appraisal, multi-year energy-system planning, fusion-device gain analysis, and AI inference infrastructure economics. The unifying feature is that break-even is not treated as a single scalar threshold detached from time structure; instead, it is conditioned on age, cohort, discounting, technical heterogeneity, and the transition from full cost to marginal cost as assets depreciate or become sunk (Faggian et al., 2019, Tarzia, 2016, Wang et al., 1 May 2025, Matsuoka, 8 Jul 2026).

1. Conceptual scope and analytical objects

The literature uses “vintage” in several technically distinct but structurally related ways. In vintage-capital theory, the relevant index is age $s \in [0,\bar s]$, so that capital goods are differentiated by remaining productive life and by the age profile of productivity. In multi-year energy models, the vintage is the commissioning year $y$, which determines lifetime, availability, fixed O&M, efficiency, and salvage. In AI infrastructure economics, the vintage is the purchase-year cohort of accelerator capacity, such as “2026 vintage” or “2027 vintage,” and the central issue is whether that cohort can clear its full cost under the prevailing pricing regime. In fusion, “breakeven” is defined at the device level rather than the asset-cohort level, but the same logic of vintage-sensitive cost and performance appears when hardware classes, shielding, heating systems, or dynamic circuit conditions determine whether a system can reach $Q \ge 1$ (Collaboration et al., 26 May 2026, Vita, 25 Jun 2026).

The criterion called “break-even” also varies across domains. In simple corporate-finance models it is the quantity $Q_f(r)$ such that net present value is zero. In age-structured capital models it is the age-specific equality between discounted future marginal revenue and marginal investment cost, expressed through the shadow price $\zeta^*(s)$. In energy systems it is the vintage-specific price, utilization, or LCOE threshold that sets $\text{NPV}_y=0$. In AI inference economics it is the premium-share or utilization level needed for a capacity vintage to cover annualized capex and operating cost. In fusion studies it may denote scientific breakeven, engineering gain, or reaction-count thresholds, depending on the paper’s accounting convention (Tarzia, 2016, Faggian et al., 2019, Wang et al., 1 May 2025, Matsuoka, 8 Jul 2026).

A recurring methodological point is that aggregation can obscure breakeven structure. The energy-systems literature states this explicitly: a simple formulation that aggregates capacity regardless of commissioning year cannot capture changes in capital cost, efficiency, availability, lifetime, policy, or fuel price across vintages. A plausible implication is that any break-even metric computed on an aggregated state may misstate the age- or cohort-specific margin that actually governs investment or retirement decisions (Wang et al., 1 May 2025).

2. Age-structured vintage capital and age-specific cutoffs

In optimal investment with vintage capital, the state variable is the stock of capital goods of age $s$ at time $\tau$,

$K(\tau,s),$

with control variables $u_0(\tau)$ for investment in new capital goods and $y$0 for investment in non-frontier vintages. Capital accumulation is described by the age-structured PDE with boundary control

$y$1

where $y$2 is depreciation or obsolescence. Output is

$y$3

and the objective is discounted profit maximization with concave revenues and convex investment costs (Faggian et al., 2019).

The abstract formulation uses the modified translation semigroup generated by

$y$4

with a boundary control operator

$y$5

Under the assumptions $y$6, concavity of $y$7, convexity and lower semicontinuity of costs, Lipschitz properties of the convex conjugates, and the semigroup bound with $y$8, the paper derives both Dynamic Programming regularity and the Pontryagin Maximum Principle in infinite dimension with boundary control. The maximum condition takes the form

$y$9

and, in vintage variables with $Q \ge 1$0, the optimal controls become

$Q \ge 1$1

The co-state admits the integral representation

$Q \ge 1$2

This identifies $Q \ge 1$3 as the shadow price of age-$Q \ge 1$4 capital (Faggian et al., 2019).

The steady state is an equilibrium distribution $Q \ge 1$5 satisfying stationarity in calendar time. The paper gives two equivalent characterizations: a stationary MP system and a stationary closed-loop equation. In vintage variables, defining the discounted marginal product of an age-$Q \ge 1$6 unit by

$Q \ge 1$7

equilibrium reduces to a fixed point $Q \ge 1$8, and then further to a scalar equation. If

$Q \ge 1$9

then $Q_f(r)$0 is an equilibrium if and only if $Q_f(r)$1 where $Q_f(r)$2 solves

$Q_f(r)$3

Under concavity of $Q_f(r)$4 and convexity of $Q_f(r)$5, existence and uniqueness of $Q_f(r)$6 follow from monotonicity of $Q_f(r)$7, which is strictly increasing and has a unique zero (Faggian et al., 2019).

The breakeven rule is age-specific. In general convex costs,

$Q_f(r)$8

In the linear–quadratic case,

$Q_f(r)$9

At the frontier,

$\zeta^*(s)$0

When positivity constraints or linear cost components matter, investment in age $\zeta^*(s)$1 is positive only if

$\zeta^*(s)$2

Because $\zeta^*(s)$3 is decreasing in $\zeta^*(s)$4, there exists a threshold $\zeta^*(s)$5 separating younger vintages with positive investment from older vintages with zero investment. The paper explicitly interprets this as a vintage-specific “NPV equals marginal cost” rule. Under purely quadratic costs, the equilibrium stock can become hump-shaped, increasing in age up to a turning point and then decreasing as shorter remaining lifetime dominates (Faggian et al., 2019).

3. Discount-rate break-even in amortized investment projects

A separate strand of vintage-breakeven analysis studies a simple $\zeta^*(s)$6-year project with constant annual output, negligible inflation, straight-line amortization $\zeta^*(s)$7, no salvage value, and tax on EBIT. The annual after-tax cash flow is

$\zeta^*(s)$8

and the annuity factor is

$\zeta^*(s)$9

with $\text{NPV}_y=0$0. Net present value is affine in annual quantity sold:

$\text{NPV}_y=0$1

Equivalently,

$\text{NPV}_y=0$2

This affine structure is what makes the financial break-even quantity analytically tractable (Tarzia, 2016).

Setting $\text{NPV}_y=0$3 yields the closed form

$\text{NPV}_y=0$4

The accounting break-even quantity is instead defined by earnings before taxes equal to zero:

$\text{NPV}_y=0$5

Using $\text{NPV}_y=0$6 as $\text{NPV}_y=0$7 and $\text{NPV}_y=0$8, the paper proves

$\text{NPV}_y=0$9

Thus, when discounting is negligible, financial and accounting break-even coincide. The rate of convergence near zero discount rate is linear:

$s$0

The paper also proves that $s$1 is strictly increasing and convex in $s$2 (Tarzia, 2016).

As $s$3, $s$4, so the break-even curve has an asymptotic straight line

$s$5

with

$s$6

This asymptotic structure is central to the paper’s interpretation: large discount rates transform the financial break-even point into an almost linear capital-recovery requirement (Tarzia, 2016).

The sensitivity results are exact. Higher price $s$7 lowers $s$8, higher variable cost $s$9 raises it, and fixed cost $\tau$0 has the weaker additive effect

$\tau$1

The sign of $\tau$2 is parameter-dependent because the depreciation shield and the after-tax margin move in opposite directions. For project life $\tau$3, longer life typically lowers $\tau$4 at low $\tau$5, but the sign can be parameter-dependent at higher $\tau$6. In the numerical illustration with

$\tau$7

the paper reports $\tau$8 units and $\tau$9 units, with $K(\tau,s),$0 and $K(\tau,s),$1 (Tarzia, 2016).

4. Commissioning-year vintages in multi-year energy-system models

In multi-year energy investment modelling, the core distinction is between the simple formulation, which aggregates all capacity regardless of commissioning year, and the vintage formulation, which indexes capacity and production by both commissioning year $K(\tau,s),$2 and operational year $K(\tau,s),$3. The paper adds a compact formulation that keeps investments indexed by commissioning year but aggregates production by operational year. This preserves year-specific characteristics while reducing dimensionality and is implemented in TulipaEnergyModel.jl (Wang et al., 1 May 2025).

For a single technology, the classic vintage formulation uses $K(\tau,s),$4 for capacity built in year $K(\tau,s),$5, $K(\tau,s),$6 for available capacity, and $K(\tau,s),$7 for production. Capacity survival is encoded by the alive indicator

$K(\tau,s),$8

equal to $K(\tau,s),$9 if vintage $u_0(\tau)$0 is operational in year $u_0(\tau)$1. A common salvage approximation is

$u_0(\tau)$2

The compact formulation replaces the full production index by the convolution-like aggregation

$u_0(\tau)$3

with an operational-year production bound

$u_0(\tau)$4

The reduction is from roughly $u_0(\tau)$5 production variables per technology to $u_0(\tau)$6 (Wang et al., 1 May 2025).

Vintage-breakeven analysis enters through per-vintage discounted project appraisal. For vintage $u_0(\tau)$7,

$u_0(\tau)$8

If

$u_0(\tau)$9

is discounted energy output and

$y$00

is discounted operating cost, the breakeven price is

$y$01

and the vintage-specific LCOE is

$y$02

Breakeven therefore depends explicitly on the commissioning cohort, not merely on aggregate plant output (Wang et al., 1 May 2025).

The compact formulation requires post-processing to recover approximate vintage-level production. The paper proposes attribution weights

$y$03

or, when availability heterogeneity matters,

$y$04

These attributed flows are then inserted into $y$05, $y$06, and $y$07 (Wang et al., 1 May 2025).

The numerical example in the paper shows how strongly breakeven can vary by vintage. For a 2030 vintage with 100 MW, 3-year life, and €100,000,000 CapEx, the reported breakeven price is

$y$08

For a 2031 vintage with 100 MW, 4-year life, €90,000,000 CapEx, and pro-rata salvage, the reported breakeven price is

$y$09

At $y$10, the minimum capacity factor for the 2031 vintage is approximately $y$11. The paper’s broader point is that learning, availability, salvage, and price trajectories make breakeven inherently vintage-specific (Wang et al., 1 May 2025).

5. Fusion breakeven: scientific gain, engineering gain, and saturation limits

In fusion, the most basic distinction is between scientific breakeven and engineering or wall-plug gain. For the compact negative-triangularity tokamak CENTAUR, scientific gain is defined as

$y$12

with scientific breakeven at $y$13. Using modeled fusion power $y$14 and coupled ion cyclotron heating power $y$15, the paper reports

$y$16

The same paper notes that $y$17 is not explicitly reported and estimates a plausible range of $y$18–$y$19 under wall-plug assumptions for ICRH and auxiliary loads. This is a direct clarification that scientific breakeven does not imply wall-plug breakeven (Collaboration et al., 26 May 2026).

CENTAUR’s breakeven-class operating point combines $y$20, $y$21, $y$22, $y$23, $y$24, $y$25, and a 10 s DT pulse. The reported average plasma parameters are

$y$26

giving a Lawson triple product

$y$27

BALOO calculations place the pedestal in the first stability regime, the geometry allows a radiated fraction of $y$28 between the separatrix and plasma-facing components, and UEDGE gives a peak perpendicular heat flux of approximately $y$29, below the steady-state tungsten limit of approximately $y$30. The magnet system uses REBCO TF coils, an hourglass-shaped central solenoid, and six PF coils; a 12 cm $y$31 shield keeps HTS heating below the 33 K quench limit during 10 s, 40 MW DT pulses (Collaboration et al., 26 May 2026).

A different fusion use of vintage-breakeven reasoning appears in the plasma focus literature, where breakeven is constrained by dynamic circuit behavior. The Plasma Focus paper defines

$y$32

and for DT each reaction releases $y$33. For $y$34, breakeven requires

$y$35

DT reactions. The dynamic-resistance term is tied to the rundown inductance,

$y$36

with effective resistance generated by sheath motion. The paper argues that spontaneous filamentation increases the dynamic resistance, diverts energy into quasi force-free magnetic structure, and produces neutron-yield saturation above approximately $y$37 (Vita, 25 Jun 2026).

The proposed remedy is a radial magnetic field of approximately $y$38 in the rundown region to suppress thermal instability, current filamentation, rarefaction-shock corrugation, and magnetic Rayleigh–Taylor instability. Under filamentation suppression, the paper reports that the drive parameter is multiplied by a factor $y$39 at least and argues that

$y$40

Using a best recorded pre-suppression value $y$41 at $y$42 and scaling to $y$43, the paper obtains

$y$44

This yields the claim that breakeven is attainable in a 224 kV, 10 MJ Plasma Focus working with DT if rundown-phase filamentation is suppressed (Vita, 25 Jun 2026).

Taken together, these studies show that “breakeven” in fusion is not a single invariant threshold. It depends on what is counted in the denominator, whether the problem is transport-limited or circuit-limited, and whether the constraining mechanism is power exhaust, shielding, or dynamic resistance. The distinction is definitional in the tokamak case and dynamical in the plasma-focus case (Collaboration et al., 26 May 2026, Vita, 25 Jun 2026).

6. Capacity-vintage solvency in AI inference infrastructure

In AI infrastructure economics, a vintage is explicitly defined as the purchase-year cohort of capacity. The paper denominates inference economics in dollars per petabyte of bandwidth delivered for bandwidth-bound decode. With accelerator price $y$45, system multiplier $y$46, depreciation life $y$47, electrical draw $y$48, PUE, electricity price $y$49, operations allocation $y$50, theoretical memory bandwidth $y$51, and realized memory-bandwidth utilization $y$52, the straight-line hourly cost per PB is

$y$53

An annuity version writes

$y$54

which separates capital, energy, memory premium, and merchant margin. This framework is then mapped into token economics through effective bytes per token,

$y$55

and

$y$56

The paper assumes near-Shannon-limit KV-cache compression and lightweight local runtimes, so $y$57 is itself moving over time (Matsuoka, 8 Jul 2026).

The central claim is that the entrant disadvantage is structural because a depreciation conveyor continuously turns prior hardware into sunk fleets that set the incumbent floor. In 2026, new-build GB300-class capacity is reported near $y$580.174/\text{PB}$y$59\$y$60, a $y$61 gap. In 2027, Rubin narrows the gap to about $y$62. By 2029–2030, the gap re-widens: amortized B200 gives a floor near $y$630.022/\text{PB}$y$64\$y$65, and if HBM shortage persists they sit higher, for example at $y$660.086/\text{PB}$y$674\times$ (Matsuoka, 8 Jul 2026).

Vintage-breakeven is then computed under two pricing regimes. Under sticky premium pricing, premium tokens retain a fixed absolute price

$y$68

while mass prices are anchored at

$y$69

Under coupled pricing,

$y$70

If a vintage has full cost $y$71 and premium share $y$72, break-even requires

$y$73

For the 2026 vintage under sticky pricing, with $y$740.174/\text{PB}$y$75p_{\text{mass}}\approx \$y$76,

$y$77

For the 2029 vintage under coupled pricing with normalized HBM, using $y$780.072/\text{PB}$y$79p_{\text{mass}}\approx \$y$80, and $y$810.2002/\text{PB}$y$82$y$83$y$84c_{\text{full}}\approx \$y$85, the requirement rises to about $y$86 (Matsuoka, 8 Jul 2026).

The resulting pattern is U-shaped and regime-dependent. The paper states that 2026 and 2028–2029 capacity are each fatally exposed to one pricing regime, while only the 2027 vintage is robust. Under sticky pricing, the 2026 vintage is worst, needing premium share of about $y$87. Under coupled pricing, late vintages break only if premium share reaches approximately $y$88–$y$89 in the shortage branch, above plausible $y$90–$y$91. The 2027 vintage, by contrast, requires only $y$92–$y$93 premium share across regimes (Matsuoka, 8 Jul 2026).

Solvency is also expressed through required utilization:

$y$94

and through a standard vintage NPV expression,

$y$95

The paper argues that the announced buildout is solvent only in a corridor requiring roughly $y$96 annual token-demand growth for four years, net of bytes-per-token efficiency gains. Its scenario analysis assigns revised probabilities of $y$97 to Rotating Landlord Oligopoly, $y$98 to Commoditization Crash, $y$99 to Jevons Absorption, $Q \ge 1$00 to System-Layer Re-differentiation, and $Q \ge 1$01 to Geopolitical Bifurcation. A custom-silicon entrant removes merchant margin but not memory premium, with a central outcome distribution of $Q \ge 1$02 success, $Q \ge 1$03 mediocre, and $Q \ge 1$04 loss (Matsuoka, 8 Jul 2026).

The AI case shows a particularly explicit form of vintage-breakeven analysis: break-even is not merely whether a datacenter is profitable, but whether a specific purchase-year cohort can recover full cost before a new depreciated floor undercuts it. This suggests a direct conceptual parallel to commissioning-year energy vintages and age-specific capital cutoffs, though the state variable is now an economic cohort rather than a physical age distribution (Matsuoka, 8 Jul 2026).