The Doppler Trap: Why Waiting for Better Tech Backfires in Industry

The longer you wait to adopt a new technology, the better the version you’ll inherit. That sounds like wisdom. It’s actually a trap — and right now, an entire generation of industrial organisations is walking into it together.

It is 2030-something, and you are in charge of innovation at a global manufacturer. The breeze dries the sweat on your collar as you walk the park, half-present, your agents murmuring in your ear. You have some decisions to make: where to invest, what to commit to, how long to wait. The main challenge is that you messed this up before. You backed a technology pathway and six months later the market moved somewhere you couldn’t follow. The answer, as always, feels like: wait. The next best thing is always coming.

Call it the Doppler Trap

The rational, individually defensible decision to hold off on committing to a technology platform that is visibly improving — until the waiting itself becomes the competitive liability.

The term comes from physics. As a sound source moves toward you, the waves compress and the pitch rises. As it passes and recedes, the pitch drops. You’ve stood still. The world moved. In technology adoption, the longer you wait, the further the reference point moves, and the more you have to sprint just to reach where others already are. Manageable when you’re shipping software. Catastrophic when you’re rebuilding a decade-long supply chain.

This is not an argument for reckless early adoption — that way leads to its own boneyard. The option value of waiting is real. There is legitimate academic literature on it: Dixit and Pindyck’s Investment under Uncertainty (Princeton, 1994)¹ made the foundational case for treating irreversible commitments as options, and deferring them when uncertainty is high.²³ If you commit to a technology platform today and a significantly better one appears in eighteen months, you’ve burned capital, change management energy, and institutional goodwill on a slower horse.

The problem is that we are entering a period where the pace of improvement is not flattening — it is accelerating. And when everyone waits for the same reason, the competitive gap between those who move and those who don’t compounds in ways that NPV analysis cannot capture.

The Science Is Moving Faster Than You Think

The mainstream narrative is that AI is improving at language tasks. The more consequential story is what is happening at the frontier of mathematics and science — and how quickly those discoveries will cascade into industrial application.

In 2024, Google DeepMind’s AlphaProof and AlphaGeometry 2 solved four of six problems at the International Mathematical Olympiad, reaching silver-medal standard.⁴ By mid-2025, multiple systems had achieved gold-medal performance on the same benchmark.⁵⁶ Progress that researchers had expected to take a decade compressed into months.

Terence Tao, arguably the world’s most accomplished active mathematician, has been watching this closely. His assessment is neither alarmist nor dismissive. Working with Google DeepMind’s AlphaEvolve, he and collaborators obtained results across 23 open mathematical problems in a day or two that would have taken individual experts months.⁷ As he put it, AI is “very good at scouring big lists of problems for low-hanging fruit. It’s tedious and thankless and not something humans want to do.”⁸

Mathematical discovery is a bottleneck for physics, materials science, drug design, and engineering optimisation.⁹¹⁰¹¹ When that bottleneck loosens — even partially — it creates pressure across the entire downstream chain. New materials get specified. New processes become possible. New efficiency curves get drawn. The question for industry is not whether this wave is coming, but how fast it will break.¹²

Three Case Studies in Moving Fast — and Standing Still

BYD: The Factory That Learns

China’s BYD is not primarily an electric vehicle company. It is a vertically integrated manufacturing system that has learned to iterate faster than anyone thought an industrial organisation could.

BYD manufactures its own raw materials, cathode and anode compounds, battery cells, packs, battery management systems, and complete vehicles — the entire stack, under one roof. What that integration enables is speed of propagation: when BYD’s R&D team refines a cell chemistry, the change reaches production within weeks, not the months or years it takes when suppliers, component makers, and assemblers must coordinate across separate organisations. The specification and the production line are in constant conversation.

By late 2022, BYD’s total battery production capacity had exceeded 135 GWh. It delivered 1.86 million new energy vehicles that year — nearly triple the previous year’s volume. Battery pack prices across the industry fell 20% year-on-year in 2024, hitting $115 per kWh, a level that had been projected as years away. BYD’s vertically integrated cost structure sits consistently below competitors relying on external supply chains.¹³¹⁴¹⁵

The Blade Battery — BYD’s lithium iron phosphate design that eliminates conventional module structures and reduces manufacturing cost by consolidating pack assembly — did not emerge from a single R&D breakthrough. It emerged from a system designed to absorb, test, and deploy incremental improvements at speed.

The architecture of the organisation is the product.

Northvolt: A Doppler Trap in Real Time

Europe’s battery industry had a champion. Northvolt, founded in 2016 by former Tesla executives, raised over $15 billion, secured more than $50 billion in forward supply contracts from Volkswagen, BMW, and others, and set out to build the continent’s homegrown answer to CATL.¹⁶

It filed for Chapter 11 bankruptcy in November 2024, with $5.8 billion in debt and $30 million in cash.¹⁷

The proximate causes were multiple. The Skellefteå gigafactory — designed for 60 GWh of annual capacity — produced just 80 MWh of cells by the close of 2023: less than 0.2% of its projected run rate. BMW terminated a €2 billion supply agreement in June 2024 citing quality and delivery failures.¹⁸¹⁹²⁰ Internal reports pointed to mismatched production parameters, equipment that required Chinese engineers on-site to operate, and an R&D team 300 kilometres away in Stockholm because skilled talent refused to relocate to the Arctic Circle.

But the deeper lesson is architectural. Northvolt designed a monolithic system — a facility built to produce one type of battery at scale — at a moment when battery chemistry was still in rapid flux. While it was constructing its cathedral, the market moved toward lithium iron phosphate. Its competitors were iterating. Northvolt was building.

This is the Doppler Trap at industrial scale: a decision that looked rational in 2018, built on assumptions about which technology would win and how fast the market would move, that became catastrophic by 2024. The company did not fail because it waited. It failed because it committed to a fixed specification in a moving-target environment, without the architecture to absorb change. Fifteen billion dollars is a very expensive lesson in the difference between a factory and a learning system.

Varda Space: Manufacturing in a New Environment

In February 2024, a small capsule — roughly the size of a kitchen bin — streaked back through the atmosphere and landed in the Utah desert. Inside were crystals of ritonavir, an HIV medication, manufactured in orbit.²¹

Without gravity, crystals do not sediment. Materials do not deform under their own weight. Pharmaceutical compounds grow more uniform, more pure, and in crystalline structures — polymorphs — that are difficult or impossible to produce in terrestrial labs. Varda’s W-1 mission produced Form III ritonavir, a specific structure with potentially improved drug performance that cannot be reliably manufactured at Earth gravity.²²

By July 2025, Varda had completed three orbital manufacturing missions and raised $187 million to expand toward biologics — monoclonal antibody crystallisation in orbit, targeting conditions where drug purity determines whether a therapy works.²³²⁴

Varda is not a pharmaceutical company that decided to go to space. It is a proof that when you move the production environment itself — not just the process within an existing environment — you access capabilities that rewrite what is possible. That is an extreme version of the same logic that drives modular manufacturing on Earth: design the system to produce things that a fixed, optimised-for-yesterday facility cannot.

What Agile Industry Actually Requires

The standard response to technological uncertainty — “stay flexible,” “build optionality” — is too vague to act on. Nobody gets to the moon with that advice. Three concrete capabilities separate the organisations that will navigate the Doppler Trap from those that won’t.

The first is treating the specification as infrastructure, not a document. In a fast-moving technology environment, what you are building cannot be fixed at the start of a project. It has to flow — updated upstream as research produces new possibilities, downstream as production surfaces new constraints. Just-in-time supply chains were built for predictable demand with a fixed product spec. They are poorly suited to a world where the optimal process might change mid-run. The successor model is something closer to what Formula 1 teams do across a season: continuously re-optimising against a changing performance envelope, with every node in the chain capable of receiving and acting on updated instructions rapidly. The supply chain as communication network, not as pipeline.

The second is modularity in physical infrastructure. Manufacturing capacity built for a single process is a liability when the process changes. The winning bet is tools, lines, and facilities designed to be reconfigured, upgraded, or decomposed rather than replaced wholesale. This is more expensive upfront and significantly cheaper across a technology cycle. SpaceX’s reduction of a 150-part engine to two 3D-printed components is an extreme version of the same logic. The question for industrial operators is not whether to pursue modularity, but how far along that spectrum they need to be to remain competitive through the next iteration cycle.

The third is engaging with regulatory sandboxes as a competitive mechanism, not a compliance exercise. The EU AI Act, effective August 2024, mandates that every member state establish at least one national AI regulatory sandbox by August 2026 — explicitly designed to allow companies to test innovative systems in real-world conditions with conditional relief from compliance requirements written for a different risk profile.²⁵ The Financial Conduct Authority’s fintech sandbox, established in 2016, offers the template: independent evaluations have documented meaningful reductions in time-to-market for regulated products, and the process generated regulatory learning that improved the rulebook itself.²⁶ The organisations that benefit most will be those that get in early — running experiments their more cautious competitors cannot.²⁷²⁸

It is winter in the same park, a few years on. No sweating today. The calculus has changed: a generation of modelling tools now targets the inflection points — the moment when a technology’s cost curve crosses the threshold where waiting stops paying. It turns out it is cheaper to predict where technology will change, based on the overall trajectory, than to try to lead the change yourself. The hard part was never the model. It was building an organisation capable of acting on what the model said.

The Trap, Restated

The Doppler Trap is not a prediction that laggards will disappear. Many will survive. The question is what shape they will be in when the wave passes.

The companies that move now are not betting on a specific technology. They are betting on organisational architecture: on being the kind of entity that can absorb, deploy, and iterate on new capabilities faster than the field. That bet is independent of which model, which material, or which process wins. It is a bet on the rate of learning.

Terence Tao has observed that AI will make it possible to “solve thousands of problems at once and start doing statistical studies” of mathematical space — a discovery process that does not gate on patience or headcount, but on deployment infrastructure.²⁹

The bottleneck is no longer the science. It is the gap between what the science makes possible and what industry can operationalise. The organisations that close that gap first — not by waiting for the perfect solution, but by building the machinery to deploy imperfect ones rapidly — are the ones that will still be setting specifications when others are still deciding whether to move.

The sound source is already past you. The question is whether you can hear the pitch dropping.

References

Dixit, A.K. & Pindyck, R.S., Investment under Uncertainty (Princeton University Press, 1994) — the foundational treatment of irreversible commitments as real options, to be deferred while uncertainty is high. Source: Princeton University Press. ↩
Farzin, Huisman & Kort et al., “Technology Adoption with Uncertain Future Costs and Quality,” Operations Research — models the optimal timing of adoption when a technology keeps improving. Source: INFORMS / OR. ↩
“Strategic timing of adoption of new technologies under uncertainty” — on why firms rationally defer commitment while the technology is still advancing. Source: ResearchGate. ↩
Google DeepMind — AlphaProof & AlphaGeometry 2 solved four of six 2024 International Mathematical Olympiad problems, reaching silver-medal standard. Source: Google DeepMind. ↩
Harvard Gazette, “AI leaps from math dunce to whiz” (2025) — on AI systems reaching gold-medal-level performance on olympiad mathematics. Source: Harvard Gazette. ↩
Quanta Magazine, “The AI Revolution in Math Has Arrived” (2026) — on AI’s rapid advance across research mathematics. Source: Quanta Magazine. ↩
Google DeepMind — the “AI for Math” initiative (including AlphaEvolve) applying AI to open mathematical problems. Source: Google. ↩
Terence Tao (Mathstodon, 2025) — describes AI as “very good at scouring big lists of problems for low-hanging fruit … tedious and thankless and not something humans want to do.” Source: Mathstodon. ↩
“AI for Mathematics: Progress, Challenges, and Prospects” (arXiv, 2026) — a survey of AI’s expanding role in mathematical discovery. Source: arXiv. ↩
“Artificial Intelligence and the Structure of Mathematics” (arXiv, 2026). Source: arXiv. ↩
Henkel, “The Mathematician’s Assistant” (arXiv, 2025) — on AI as a working aide to research mathematicians. Source: arXiv. ↩
Renaissance Philanthropy — its AI for Math Fund announced $18M in grants (2025) to accelerate mathematical discovery. Source: Renaissance Philanthropy. ↩
Automotive Manufacturing Solutions — how BYD’s vertical integration (raw materials through cells, packs and finished vehicles) let it surpass Tesla on production and battery technology. Source: Automotive Manufacturing Solutions. ↩
PatSnap — Tesla vs BYD battery R&D: the 4680 cell versus BYD’s Blade (LFP) strategy. Source: PatSnap. ↩
Carbon Credits (2025) — China controls ~69% of the global EV battery market as CATL and BYD surge; industry pack prices fell ~20% year-on-year toward ~$115/kWh. Source: Carbon Credits. ↩
Bruegel — “Northvolt’s struggles: a cautionary tale for the EU Clean Industrial Deal,” on the €15bn-backed champion’s collapse. Source: Bruegel. ↩
Rho Motion — “What can we learn from Northvolt’s story,” covering its November 2024 Chapter 11 filing (~$5.8bn debt, ~$30m cash). Source: Rho Motion. ↩
CEEW — key lessons from Northvolt on scaling battery manufacturing: mismatched production parameters, equipment dependence and a distant R&D team. Source: CEEW. ↩
Transport & Environment — “Northvolt collapse shows stamina and leadership are lacking for Europe” to realise its battery ambitions. Source: Transport & Environment. ↩
CleanTechnica (Dec 2024) — lessons for Europe and North America from the Northvolt collapse. Source: CleanTechnica. ↩
Space.com — Varda manufactured crystals of ritonavir (an HIV medication) in Earth orbit, recovered in the Utah desert in February 2024. Source: Space.com. ↩
TechCrunch (Nov 2025) — Varda says it has proven orbital manufacturing works, producing a Form III ritonavir polymorph difficult to make at Earth gravity. Source: TechCrunch. ↩
CNBC (Jul 2025) — Varda raised $187 million to expand orbital drug manufacturing. Source: CNBC. ↩
BioProcess International — Varda secures $187M for orbital pharmaceutical production, targeting monoclonal-antibody crystallisation. Source: BioProcess International. ↩
EU AI Act, Article 57 — requires every member state to establish at least one national AI regulatory sandbox by August 2026. Source: EU AI Act. ↩
FCA — its regulatory sandbox (since 2016); independent evaluations document reduced time-to-market for regulated products and regulatory learning that improved the rulebook. Source: FCA. ↩
IAPP — how different jurisdictions approach AI regulatory sandboxes. Source: IAPP. ↩
Future of Privacy Forum — regulatory sandboxes as a tool for balancing AI innovation and oversight. Source: Future of Privacy Forum. ↩
Terence Tao (Mathstodon, 2025) — observes AI will let researchers “solve thousands of problems at once and start doing statistical studies” of mathematical space. Source: Mathstodon. ↩