The Cost Curves

When does everything get cheap enough? A live prediction engine: measured price series, fitted decline rates with honest intervals, computed gate crossings — plus our public prediction ledger and our own backtested misses. Check us, don't trust us.

the curves

Each card is one tracked metric: its measured history (solid), the fitted projection with an 80% interval (shaded), and — where a price threshold activates something real — the gate. Cost curves fall; climb curves (transistors, supercomputers) rise — both are the same frontier read from opposite ends. Observations flow in two ways: a weekly fetch of authoritative series (OWID and friends, graded benchmark) and hard figures our daily AI editor extracts from the news, each one fact-checked against its source article before it may touch a fit. The editor also proposes new series to track when the news keeps quantifying a frontier we don't yet measure — the engine grows its own instrument panel. Rates and crossings are calculated from the data — nobody types a conclusion here. For the AI capability climb, model by model and lab by lab, see the intelligence ladder.

You are now standing in December 2020. Every card below is re-fitted on only the data that existed then — the cone is what this engine would have published that day, drawn before the future arrived. The amber dots are what actually happened. This is the honest answer to "you drew the line after the points landed" — here the line comes first, and you can watch reality land inside (or outside) it.

the build clock — what becomes buildable, when

The engine's standing answer to the only question that matters here: what is cheap enough to build now, and what joins it each year. Computed from every gate's fitted crossing window — the years shown are 80% windows, the bold year the central fit. Every weekly steward run exists to make this table more accurate.

the cascade — the curves feed each other

The deepest dynamic here isn't any single curve — it's the loop: AI accelerates the other curves, and several of them feed back into AI (cheap solar powers datacenters; robots build fabs and farms; AI designs the next battery). Couplings can't be honestly fitted from a handful of points, so this engine treats them three ways: measured — each card above carries a curve check (is its own rate steepening or flattening, computed from the record); declared — the edges below, each with a mechanism and a confidence grade; evidenced — the daily feed tags stories that demonstrate an edge in action, and the evidence trail accumulates against it. The news is how the not-yet-predictable earns its way into the model.

The scenario table is a scenario, not a fit. The "with cascade" column applies each declared edge's boost from the year its trigger gate opens (measured where open, the fitted prediction where not), with decline rates clamped at −60%/yr. When an edge's evidence trail grows — or a curve check flips to accelerating — that's the world voting the scenario toward the baseline.

the prediction ledger

Every time the engine's conclusion changes — a state, a predicted crossing window — a row is appended here, timestamped, never edited, never deleted. When a predicted crossing resolves in the real world, the prediction gets scored, hit or miss, on this page. That's the deal.

calibration — the engine grading its own intervals

For every feasible origin-and-horizon pair in every metric's history we re-fit the live model on only the data available at that origin, predict the later actual, and check whether it landed inside the 80% band. Honest bands contain ~80%.

the backtest — how this method actually performs

We ran the engine's own fit on data as it stood at the end of 2020 and let it "predict" the years that followed. The results — including the misses — are below.

the theory — why curves like these hold, and when they don't

This engine practices a specific, hundred-year-old body of theory. Naming it matters, because the theory also names its own failure modes — and the critics set the bar this page is built to clear.

• The feedback loop (Kurzweil's Law of Accelerating Returns). Technology builds the tools that build the next technology, so the rate of progress tracks the current state of the art — better computers design better computers. The cascade table above is this loop, made explicit and graded: AI accelerating the curves that feed back into AI is not a metaphor here, it's declared edges with evidence trails.
• No single exponential exists — only stacked S-curves. Moore's law is the fifth paradigm to carry the compute trend (electromechanical → relay → vacuum tube → transistor → IC); each paradigm is an S-curve that saturates, and the smooth exponential is the envelope of the stack. That's why the trend survived the 2005 death of Dennard scaling (clocks froze; GPUs picked up the envelope) — and why every projection here still carries an interval: paradigm succession is a historical pattern, not a law. A curve flattening on this page may be one paradigm saturating, or the whole trend ending; only the next observation tells.
• Wright's law is the engine; Moore's is the special case. Costs fall a roughly fixed percent per doubling of cumulative production — experience, not time (Farmer–Lafond 2016, 53 technologies; the best out-of-sample forecaster known). When production grows exponentially in time, Wright reads as Moore. This is literally the fit these cards run — and it's also why some curves never fall (nuclear's negative learning, Eroom's law in pharma): no learning curve, no law.
• The live question is whether the clock itself is speeding up. Kurzweil's strongest claim was never "exponential" — it was that the doubling time shrinks. That's now a measured dispute: METR's task-horizon doubling ran ~6.5 months across the record, ~4.3 since 2023, ~3 since 2024 — while the walked-back AI-2027 forecasts and Epoch's skepticism argue it won't persist. This page doesn't vote; the doubling-time-over-time gutter on the Climb is the instrument, and the intervals here widen exactly because that answer is unknown.
• The critics set the honesty bar. Modis: the early half of an S-curve is indistinguishable from an exponential, so a fit alone proves nothing about the future. Nordhaus: the macro data point away from an economic singularity. Kevin Kelly: a chart of accelerating milestones shows an explosion at "now" whenever you draw it. The answers this engine can actually give: intervals instead of dates, the calibration score below (are the bands honest?), the ledger (predictions timestamped before resolution), and the time-traveler view above (the cone drawn at end-2020, before the points landed). A forecaster who can't show you those is asking for faith either way.

method, honestly

• Fits: the Farmer–Lafond stochastic model — log cost as a random walk with drift over the trailing window, the field's canonical spec (validated by hindcasting on ~50 technologies). Shocks accumulate, so the 80% bands widen with horizon (variance ∝ τ + τ²/m), with Student-t tails, their global autocorrelation θ = 0.63, and their volatility-from-drift prior for short series. Where cumulative-deployment data exists, a Wright's-law fit runs alongside — first-difference regression through the origin (the Way et al. spec; levels regressions of trending series are spurious), recovering the published ~20–25% learning rates — plus a deployment-conditional Monte Carlo: 400 sampled paths under a growth-decaying-to-saturation scenario, shown as the "Way scenario" percentiles on those cards.
• Why annual declines slow while learning holds: Wright's law prices per doubling of cumulative deployment, not per year. As the installed base grows huge, each doubling takes longer (the cards show the doubling time, now vs the prior era) — so the same learning rate yields smaller annual declines. When the Wright-implied annual rate and the time fit disagree (they do for battery packs right now), the time fit is what predictions use; the gap is displayed, not reconciled away — it usually means material-cost floors or a regime break.
• Data grades: curated benchmark history > weekly authoritative fetch > fact-checked news extraction. Unverified figures are stored, labeled, and never fitted. A 0.5×–2× year-over-year plausibility band quarantines unit confusion (a cell price filed as a pack price).
• Calibration fuel: completed historical series (Santa Fe Performance Curve DataBase — transistor prices, the Model T, aluminum, crude oil as the non-learning counter-example) join the hindcast pool without appearing as live cards — every extra technology tightens the engine's knowledge of its own error structure.
• No invented ETAs: gates without a price series are labeled editorial, same rules as the rest of the site.
• The data is open: the raw ledgers are public JSON — observations · predictions — and the fit code ships to your browser as readable source. This page computes everything you see, live, from those inputs.