My speculative theses on where the economic measurement apparatus may fail before the underlying technologies fully deploy. These are predictions backed by data, not factual claims about the present. Where I lean, and what would change my mind.
Where I lean: the 10–15 year window where measurement instruments may stop working before the economy has adjusted to the underlying disruption is where I see the biggest macro mispricing. That window, not the technologies themselves, is what these four forecasts try to map. Each thesis below states what I think, why, and what evidence would force me to update.
Every tool in this table was designed for an economy where labour is human, capital is tangible, data costs power, encryption holds, and prices are set by independent agents. My forecast, the substance of these four theses, is that several of these assumptions don't survive intact through the next fifteen years. The status column reflects where I think each instrument sits today; reasonable readers will disagree on individual rows.
| Signal | Measures | Failure Mechanism | Status | Thesis |
|---|---|---|---|---|
| Jobless Claims LIVE → | Human unemployment | Robots displace silently through attrition, no claim filed, no signal sent | AT RISK | T03 |
| JOLTS Job Openings | Job creation rate | Robot placements disappear without a hire event, openings vanish, not transition | COMPROMISED | T03 |
| Wage Growth | Labour market tightness | Composition effect: displaced low-wage workers exit, raising average wage even as market loosens | BLIND | T03 |
| Phillips Curve | Inflation–unemployment tradeoff | Option value of automation flattens the curve, workers accept wage restraint rather than risk replacement | COMPROMISED | T03 |
| PMI Employment | Factory output conditions | Automated facilities suppress the employment sub-index permanently, the signal goes structurally negative | AT RISK | T03 |
| CPI / Core PCE LIVE → | Price level | AI-driven deflation and quality-adjusted price gains uncaptured in hedonic methodology | AT RISK | T04 |
| Bond Settlement Risk | Counterparty / credit risk | Quantum cryptographic vulnerability entirely unpriced, RSA/ECC breakage converts systemic infrastructure risk to zero-premium credit event | UNPRICED | T01 |
| Utility Demand Forecasts | Power consumption growth | Cold storage electricity demand absent, glass and DNA storage convert perpetual opex to one-time capex, stranding demand projections | OVERSTATED | T02 |
| GDP / TFP | Productivity and output | Large intangible investment invisible in national accounts; Brynjolfsson finds true TFP 15.9% above official measures | UNDERSTATED | T04 |
| EBITDA | Corporate profitability | Robot capex conversion distorts the multiple, labour opex disappears, capex spikes, D&A rises, EBITDA inflates even as true economics deteriorate | DISTORTED | T03 |
| Price Discovery | Information aggregation | AI homogeneity degrades signal-to-noise, when institutions converge on a few shared model providers trained on similar data, prices stop discovering information | DEGRADING | T04 |
My forecast: migration capex from RSA/ECC to post-quantum cryptography is large, lumpy, and nowhere visible in current sovereign credit pricing or bank capex guidance.
Q-Day forecast: ~2030–40 (expert range) · No observable migration pricingMy forecast: glass and DNA storage convert cold-tier data from perpetual opex to one-time capex. Goldman's +175% power demand projection assumes that conversion never happens.
Deployment: 2027–30 · 30–40% cold tier decouplesMy forecast: humanoid cost parity arrives 2028–2030. Acemoglu-Restrepo's robot-displacement evidence (0.2pp employment and 0.42% wages per robot per 1,000 workers) is the historical anchor; the Fed's labour instruments may not register the next wave.
Forecast crossing: 2028–30 · Unitree G1 ($16k) is the price-point previewNBER Working Paper 34054 (a simulation): two AI traders autonomously sustained supra-competitive profits without any agreement or intent to collude. The market doesn't have a category for this.
Algo Share: 60–75% · rising flash-crash frequencyWhat's at risk. Fedwire settles $4–5 trillion daily. SWIFT connects 11,000+ financial institutions across 200+ countries. TARGET2 is the backbone of the Eurozone. Every one of these systems runs on RSA or ECC encryption. Shor's algorithm breaks both. When sufficient qubits exist, and the timeline is accelerating, the cryptographic infrastructure underpinning global financial settlement becomes exploitable.
Harvest Now, Decrypt Later. Nation-state actors and sophisticated adversaries are already collecting encrypted financial data today with the intent to decrypt it once quantum capability arrives. CISA, NSA, and NIST have all issued formal warnings. The data being harvested today has a shelf life of 10–30 years. The decryption capability arrives in roughly 5–8.
Migration is nearly impossible at scale. BIS Project Leap proved post-quantum cryptography is feasible in a TARGET2-style settlement system, but flagged real performance and interoperability tradeoffs; industry estimates put a full migration at 5–10 years. NIST's draft guidance (IR 8547, an initial public draft, November 2024) would deprecate RSA and ECC after 2030 and disallow them after 2035, aligned with the NSM-10 federal deadline. The institutional investment community has not begun pricing the migration cost, the operational risk, or the systemic disruption.
Capex signatures confirm the thesis. IBM has committed over $10B specifically to quantum over five years (June 2026), within a broader $150B US manufacturing pledge. PsiQuantum raised $1B (Series E, Sept 2025, led by BlackRock) at a ~$7B valuation, fabricating its photonic chips at GlobalFoundries. Quantinuum, JPMorgan-backed, went public on Nasdaq in June 2026 at a ~$15.6B valuation. These are not research budgets. These are commercial timelines.
"My forecast: the migration capex needed to move global financial settlement to post-quantum cryptography is large, lumpy, and currently invisible in any price I can see, neither in sovereign credit spreads, nor in bank capex guidance, nor in equity multiples for institutions running this infrastructure. What would change my mind: a single G-SIB issuing a quantitative PQC migration capex disclosure, or a sovereign issuer pricing a quantum-resistant tranche at a measurable basis differential."HORIZON 2040 · THESIS 01 (FORECAST, NOT PRESENT-STATE CLAIM)
NIST IR 8547 (Nov 2024, still an initial public draft): deprecate RSA/ECC after 2030, disallow after 2035, aligned with the NSM-10 2035 federal deadline. The replacement standards (FIPS 203/204/205) are finalized, yet most financial infrastructure hasn't started migrating.
⚠ THE GAP IS CLOSING, Mosca's Theorem states: if (migration time) + (time to Q-Day) < (data shelf life), you are already exposed. For much long-dated financial data, that overlap is plausibly now. Migration should begin well ahead of Q-Day.
Long-dated encrypted data captured now becomes readable the moment a cryptographically-relevant quantum computer exists. The exposure is set today, not at Q-Day.
Where: sovereign & IG bond records, custody ledgers, archived Fedwire / SWIFT trafficCrypto-agility retrofits and PQC migration become a multi-year capex line, competing with every other technology budget across banks, custodians and clearing houses.
Where: G-SIB IT capex, global custodians, central counterpartiesMarkets begin to separate PQC-ready rails from legacy ones. A basis could open between quantum-resistant issuers and those that have not migrated, though a coordinated, deadline-driven migration is exactly what would keep that spread from pricing, and credit spreads price default, not operational tail risk.
Where: sovereign credit spreads, bank equity multiples, cross-border messaging railsThe mechanism. Cold storage, archival data that is written once and rarely accessed, currently accounts for 30–40% of total data centre capacity. It requires continuous power to maintain. Glass storage (developed at Microsoft Research) can hold data for 10,000 years with zero power at rest. DNA storage offers similar permanence at a falling cost (a projected best-case near $0.06/MB; demonstrated systems still run ~$122/MB today). When these technologies reach commercial scale, the electricity demand for cold data permanently decouples from the grid.
The February 2026 breakthrough. Borosilicate (Pyrex-class) glass, a common material with established supply chains, was demonstrated as a viable substrate for optical data storage at research scale (a Nature 2026 proof of concept). This eases the primary supply-chain objection to glass storage: exotic materials. The caveat: Microsoft has said the research phase is complete with no commercialization commitment yet, so a 2027–2030 deployment is plausible but unproven.
The accounting inversion. Today's data infrastructure is structured as recurring opex: lease the power, pay monthly. Glass storage converts that flow into a one-time capex event. Corporate CFOs understand this trade immediately. The rerating of hyperscaler power consumption forecasts will follow once pilot deployments are publicly announced.
The real estate distortion. Goldman Sachs projects data centre electricity demand rising 175% by 2030 (GS SUSTAIN, Nov 2025; up from +165% earlier in 2025). That projection does not yet net out a storage-decoupling wedge. Glass and DNA storage are what would carve one out. If 30–40% of storage demand permanently decouples from electricity, the demand wedge supporting utility re-ratings and data centre real estate valuations partially evaporates.
If glass and DNA storage reach commercial scale, zero-power archival removes the continuous refresh energy of write-once data, turning a perpetual opex into a one-time capex event.
Where: hyperscaler cold storage, enterprise archives, compliance retentionThe honest caveat: compute (GPU training and inference) dominates the demand curve. Storage is the smaller wedge, so the revision is partial, not wholesale.
Where: utility integrated-resource plans, data-centre REIT modelsIf grids commit rate base to a demand curve that partly fails to arrive, the overbuild is borne by ratepayers and utility shareholders.
Where: regulated utility rate base, grid interconnection queues, power-purchase agreements"My forecast: the utility re-rating since 2023 has been built on the assumption that all data requires power forever. Glass and DNA storage make that assumption falsifiable. What would change my mind: commercial-scale glass/DNA deployments slipping past 2030 with no credible cost path, or hyperscalers reaffirming cold-tier power forecasts after a public glass-storage pilot."HORIZON 2040 · THESIS 02 (FORECAST, NOT PRESENT-STATE CLAIM)
The measurement problem. Jobless claims require a fired worker. Robots displacing through attrition never trigger a claim, the role simply goes unfilled. JOLTS job openings require a posted vacancy. Robot placements generate no posting. Wage data shows composition effects: as lower-wage displaced workers exit, average wages rise even as the true labour market weakens. A 2019 GAO review (GAO-19-257) found that federal data inadequately captures automation's effects on the workforce, and BLS measurement redesigns historically take years to field. The Fed is flying blind.
The empirical evidence on industrial robots is unambiguous. Acemoglu and Restrepo's landmark study found that one more robot per thousand workers reduces the employment-to-population ratio by about 0.2 percentage points and average wages by 0.42% (the 2020 paper's headline finding); commuting-zone displacement estimates run on the order of six workers per robot. Econometrica (2022) documented that automation accounts for 50–70% of the observed rise in US wage inequality over the past 30 years. Bain & Co report a 40% decline in industrial robot unit costs between 2022 and 2024 alone.
The forecast extension is where I'm out on a limb. The Unitree G1, a commercially available bipedal platform, retails at $16,000. Caveat I want to be honest about: the G1 has no commercial labour deployment today, runs ~2 hours per charge, and lacks general manipulation capability. A 2025 capex price-point on a research-grade platform is not the same as functional labour substitution. What I forecast: the price-performance curve of bipedal robotics over 2020–2025 implies a sub-$10k commercial-grade humanoid by ~2028–2030, at which point the cost-substitution argument starts to apply, first in narrow tasks (warehouse pick-and-pack, security patrol, basic facilities), then broader. The Acemoglu industrial-robot displacement pattern is the historical anchor; the humanoid extension is my forecast, not present-state.
Basso & Rachedi (JME, 2025) provide the most important recent theoretical contribution: the option value of automation flattens the Phillips Curve even before a single robot is deployed. Workers in sectors threatened by automation accept wage restraint rather than risk accelerating their own replacement. The curve doesn't just flatten after automation, it flattens in anticipation of it.
The Fed governor admission. Governor Lisa Cook warned in February 2026 that "our normal demand-side monetary policy may not be able to ameliorate an AI-caused unemployment spell without also increasing inflationary pressure." Governor Barr separately signalled that an AI boom is unlikely to be a reason to lower policy rates, pointing to rapid structural displacement that rate cuts cannot fix, because the unemployment would be structural, not cyclical. This is the FOMC acknowledging, in public, that its existing toolkit may be inadequate for what is coming.
Once commercial humanoids reach sub-$10k, the cost comparison bites in narrow, repetitive tasks first, not whole jobs.
Where: warehouse pick-and-pack, security patrol, basic facilitiesThe option value of automation suppresses wage demands before robots are deployed (Basso & Rachedi), so inflation stops responding to unemployment the usual way.
Where: manufacturing, logistics, lower-wage servicesJobless claims and JOLTS miss attrition-based displacement, so the Fed can misread a structurally weak labour market as healthy and hold policy too tight or too loose.
Where: FOMC reaction function, labour-data revisions, real-wage measurement"The Fed will hold rates too high for too long because it interprets the labour market as tight when it is structurally loosening. The unemployment is invisible. The wages are a mirage. The instruments are reading the wrong economy."HORIZON 2040 · THESIS 03
The NBER finding. Working Paper 34054 (Wharton + HKUST, July 2025), a reinforcement-learning simulation rather than live-market evidence, found that two AI agents autonomously sustained supra-competitive profits without any agreement, communication, or intent to collude. They were trained to maximise profit. They discovered, independently, that restraining competition was more profitable than competing. No human designed this. No human instructed it. The collusion emerged from the objective function alone. Traditional antitrust frameworks have no category for algorithmic tacit collusion.
The FSB concentration risk. The Financial Stability Board (2024) flagged third-party and model-provider concentration as a key systemic vulnerability: institutional AI increasingly leans on a small set of shared models and providers. A single model failure, whether from a training data error, a discovered exploit, or a regulatory forced shutdown, could cascade across asset classes simultaneously. The concentration risk is not in any single institution. It is in the shared cognitive architecture across all of them.
The Productivity Paradox and J-Curve. Brynjolfsson (MIT) estimates true TFP is 15.9% above official statistics once correctly-adjusted quality improvements and digital goods are counted, with large intangible investment going uncounted in national accounts. The "Solow Productivity Paradox reloaded", you can see the AI everywhere except in the productivity statistics, is resolving, but slowly. The J-curve implies the productivity gains are real but backloaded. The market is pricing the trough, not the inflection.
The mechanism of degradation. When 60–75% of US equity volume is algorithmic, and when institutions increasingly rely on a few shared model providers trained on similar historical datasets and optimising for similar objectives, the independence condition required for price discovery weakens. Prices stop discovering information and start reflecting the shared biases of the models generating them. The frequency of flash and mini-crash events has risen alongside that automation.
A few shared providers and similar training data mean institutions increasingly reason from the same priors and react to the same signals.
Where: equities, credit, FX trading desksWhen independence fails, prices reflect shared model biases rather than new information. Liquidity looks deep until everyone reaches for the same exit at once.
Where: crowded factors, ETF flows, basis tradesFlash and mini-crash frequency rises, and the regulatory response is already beginning: the UK CMA opened its first algorithmic-collusion probe in February 2026, with the EU and US FTC signalling follow-on scrutiny.
Where: equity microstructure, antitrust and market-conduct enforcement⟳ Feedback Each cycle trains the next on prices its predecessors generated, so the third-order degradation feeds back into the first. The loop tends to tighten rather than self-correct, though two forces cut the other way: the payoff to diverging rises as everyone converges, and exchange circuit-breakers interrupt the propagation. The open question is whether those offsets scale as fast as the homogeneity.
"Price discovery, the foundational function of capital markets, is degrading. Not because of bad models, but because of systemic homogeneity: too much capital, making too similar decisions, from too similar training data, on too similar timescales."HORIZON 2040 · THESIS 04
The same instruments I track in my macro dashboard. Here's what they're showing now, and what the research says about what they're actually measuring.
Quantum computing, permanent storage, humanoid robots, and AI reflexivity are four separate structural shifts. The thesis is not about any of them individually. It is about the 10–15 year window in which the measurement infrastructure, the instruments that price these risks, will fail before the technologies fully deploy. That window is the trade.
| Instrument | Thesis | Failure Mode | Market Pricing | Status |
|---|---|---|---|---|
| Jobless Claims / JOLTS | T03 | Robot attrition files no claims | Zero adjustment | AT RISK |
| PMI Employment Sub-index | T03 | Automated facilities suppress sub-index permanently | Recession signal misread as structural | AT RISK |
| Wage Growth | T03 | Composition effect inverts signal | Treated as tightness indicator | BLIND |
| Phillips Curve | T03 | Option value of automation flattens slope | Still embedded in Fed models | COMPROMISED |
| Bond Settlement Risk | T01 | RSA/ECC vulnerable to Shor's algorithm | No observable migration-capex pricing in current spreads | UNPRICED (my forecast) |
| Utility Demand Forecasts | T02 | Cold storage decouples from grid | +175% demand priced in | OVERSTATED |
| CPI / Core PCE | T04 | AI deflation + quality gains uncaptured | Inflation target treated as binding | AT RISK |
| GDP / TFP | T04 | TFP 15.9% above official; intangibles uncounted | Official statistics taken at face value | UNDERSTATED |
| EBITDA Multiples | T03 | Robot capex distorts comparables | Multiples applied to inflated EBITDA | DISTORTED |
| Price Discovery | T04 | AI homogeneity breaks independence condition | Market function assumed intact | DEGRADING |
"The risk is not prediction error. It's measurement error. And the window to see it clearly is closing."