In the latest episode of the Cambridge Executive Business Insights: Rethinking AI podcast, Simone Lenzu, Assistant Professor of Economics and Policy at CJBS and Research Economist at the Federal Reserve Bank of New York, takes us on a tour of the macroeconomic questions beneath the surface – from the allocation of scarce resources to the very real risks and opportunities of our collective investment in AI. In this blog post, we distil the insights from that conversation, exploring what “frugal AI” really means, why history teaches caution about productivity revolutions, how AI is upending traditional macroeconomic thinking, and why the winners in this new era may not be who you expect.
Frugality in the age of frontier models: more than just shrinking chips
Few terms have come to define AI’s latest wave as rapidly as “frugal AI”. To many in engineering and business, frugality implies smaller, cheaper, more energy-efficient models and data centres – a technical and financial sleight of hand for a planet under resource strain. But Lenzu and his host challenge this narrow interpretation: true frugality in AI, they argue, is fundamentally a problem of economics, not just engineering.
Frugality, when viewed at scale, is about the creative allocation of scarce resources with the aim of unlocking widespread growth, equity, and efficiency. This means asking not only “can we train larger models using less electricity?” but also “are these resources being used in ways that provide genuine welfare gains to the broader economy?” The real test of AI’s purposefulness is not simply its power, but the extent to which it flows to where it is most needed, improving lives rather than deepening divides.
At the heart of this is a more ambitious challenge: ensuring that our vast investments in chips and compute translate into improvements in living standards, not just fleeting financial returns or concentrated gains for a handful of hyperscalers.
The productivity paradox: history’s warnings about technological revolution
The allure of AI as the “next big productivity leap” is irresistible to business leaders, investors and policymakers alike. Yet Simone Lenzu reminds us that history is littered with cautionary tales. Even the computer age, which is perhaps the closest historical analogue to today’s AI, was famously described by Nobel laureate Robert Solow as being visible “everywhere except in the productivity statistics”.
Indeed, the much-vaunted computer and IT revolution of the late 20th century took over a decade to show up in aggregate productivity figures, and those effects proved temporarily transitory before productivity growth stagnated again. This lag was due not to a lack of technological potential, but to the slow, costly, and messy process by which firms reorganised, re-skilled, and retrofitted their organisations to integrate fundamentally new technologies.
Far from being a smooth leap forward, history suggests a ‘J-curve’ – an initial period where resources flow into experimentation and adjustment, with productivity falling or stagnating before rising again, if at all. The message for AI is sobering: as transformative as it may be, timelines for mass productivity gains are likely to be uneven, sector-dependent, and, above all, slow.
AI investment booms: necessary bet or bubble in the making?
Turn on any financial news channel, and you’ll see staggering figures: tens or hundreds of billions being poured into AI research, chips, and data centre infrastructure. Is this a healthy, forward-looking investment, or a speculative gamble that risks financial fragility?
The answer, says Lenzu, is both complex and finely balanced. Big, front-loaded investment is logical for a general-purpose technology with as much transformative potential as AI. If (and it is a big ‘if’) AI delivers on its productivity promise, these investments will be vindicated in hindsight.
But there are darker clouds. The financing of AI’s infrastructure is becoming more opaque and fragile, with specialised vehicles and debt instruments (often 80% leveraged or more) underwriting construction in anticipation of future demand. These bets are, at times, backed more by expectation of future leases and productivity than by proven cash flows. Should the productivity payoff be delayed or the models veer towards greater efficiency (thus requiring fewer massive data centres), these investments could sour, triggering instability across financial markets.
Moreover, as the episode reveals, expectations appear precariously high, with market valuations for AI-exposed firms justifiable only if the ultimate returns prove spectacular. In a world where capital chases speculative narratives, a single negative report can send markets tumbling, as was witnessed when a pessimistic forecast from a minor player provoked an 800-point crash in the Dow Jones.
The hidden cost: opportunity cost and the risk of resource misallocation
Perhaps the most overlooked dimension of AI’s rise is the opportunity cost, that is, the value of what is forgone so that resources (capital, energy, attention) can be marshalled towards AI. Every data centre and bond issued to finance this boom is a dollar, an electron, a skilled worker not supporting hospitals, factories, or the next generation of entrepreneurs.
For the wider economy, these shifts are not abstract. The power requirements of hyperscale data centres, for instance, have the potential to drive up electricity costs for everyone else, pushing up hospital overheads and household bills. So too, capital redirected towards AI-related ventures can crowd out lending and investment for other productive sectors, particularly as pension funds and sovereign wealth increasingly favour technology unicorns over industrial stalwarts.
Lenzu urges us to remember: assessing AI by its direct costs is “missing the entire tab” – the full spectrum of what society is forgoing as it pursues the next technological grail.
The sluggish reality: friction and organisational inertia
If AI is so important and attracts so much capital, why is adoption not swifter? Is it regulation, a lack of funds, or something else entirely?
Surprisingly, neither red tape nor capital constraints seem to be major limiting factors at present. Rather, the main friction lies inside organisations themselves: inertia, uncertainty, and the complex process of embedding unfamiliar tools in legacy systems.
Most firms, according to emerging microdata, are currently experimenting rather than transforming. Early adoption is concentrated in customer-facing functions (marketing or service bots, for example) rather than core, productivity-enhancing tasks. Deep integration – the kind that changes how products are made, services are delivered, and value is created – requires experimentation, risk, and a willingness to tear up existing processes. This, right now, is happening only at the margins.
“Firms seem to be running small reversible trials to find out where the returns of AI will be before committing to a harder, more disruptive reorganisation,” observes Lenzu. It’s a rational response to radical uncertainty, but it means the “AI revolution” is, in practical terms, a slow drip rather than a tidal wave.
Who wins the ‘frugal AI’ race? Tech builders vs. smart adopters
A second narrative often runs through the boardroom and media alike: will the spoils of AI accrue to giants building platforms and models, or to businesses who deploy these tools in creative, economic ways?
History suggests the smart money is on the latter. Drawing on analogies from the electricity revolution, Simone notes that the greatest gains accrued not to those building generators, but to those who reorganised their processes, products, and strategies to capitalise on cheap, flexible power. Likewise, in AI, the long-term winners may well be those who integrate “good enough, cheap enough, reliable enough” tools in a way that solves genuine problems and delivers superior value to their customers.
The true deployment challenge, then, is not raw model capability but integration, impact, and the economics of return on investment. The firms who will set the pace are those who view AI not as a technological trophy, but as a resource to be tailored, repurposed, and made frugal, embedded in organisational DNA.
Don’t underestimate ‘customer capital’: AI beyond productivity
While much discussion of AI circles around cost savings and operational efficiency, another – quieter – revolution may be afoot: the use of AI to build customer capital. This concept, as explored in Simone’s emerging research, refers to the strategic deepening of relationships, loyalty, and market share through the clever use of AI-powered tools.
Firms are increasingly deploying AI to refine advertising, enhance brand-building, and conduct hyper-targeted outreach. While these investments often don’t register as productivity gains in the classic sense, they can generate robust growth and justify further investment in digital transformation. Over time, such strategies may create “umbrella effects” that encourage broader deployment of AI throughout the enterprise.
Importantly, this is not just technical progress: it’s a reimagining of how value is created, captured, and sustained in the digital age. AI can help firms to “grow by making their products and services more appealing to customers”, a type of growth often underestimated in economic analysis but increasingly central in a hyper-competitive, platform-driven economy.
Monetary policy in the AI era: A harder, faster, noisier road
It isn’t only business strategy that will be challenged by the spread of AI, but also the day-to-day work of central bankers and policymakers. Lenzu argues that AI is likely to make the job of stabilising economies, through interest rates, inflation management, and financial oversight, far more difficult.
There are three principal ways in which AI, as a general-purpose technology, could disrupt central banking:
- Faster price transmission: as AI-powered pricing algorithms become widespread, cost shocks may be transmitted to consumer prices at unprecedented speed. This “steepens the Phillips curve,” making monetary policy less forgiving and requiring more precise intervention.
- Blurring signal and noise: the changing structure of the economy, powered by AI, adds layers of opacity and volatility to key economic indicators. Central banks will struggle to distinguish between cyclical shocks and genuine, structural change, risking policy errors that either strangle or overstimulate growth.
- Financial market fragility: the expectation-driven investment boom in AI, layered with complex and risky financial instruments, increases the risk of market instability. Policymakers may find themselves torn between tackling inflation and avoiding the “popping” of AI-exposed asset bubbles.
The road ahead, in short, is less a straight line than a hairpin-laden switchback – with greater demands for data, sophistication, and finely tuned judgement.
Concentration of power: can AI democratise, or will it entrench elites?
A key concern echoes throughout the discussion: will AI prove a force for democratisation or for new concentrations of power?
Current trends are worrying. The AI ecosystem is dominated by a handful of hyperscalers and chip fabricators, with smaller players dependent on expensive proprietary infrastructure. This stands in stark contrast to previous generations of technology (such as the PC revolution) which quickly became widely accessible.
Yet beneath the surface, there are glimmers of optimism. If AI enables startups to leapfrog costly investments or helps small businesses enhance value and scale faster, the balance of power could shift. But this will require conscious effort: investing in open-source models, incentivising wider diffusion, and ensuring the benefits of AI are not hoarded by a digital aristocracy.
Measuring progress: hype vs. harvest
How should we measure the impact of AI, and avoid being fooled by hype? The answer, says Lenzu, lies in better data, not just firm-level sales or revenues, but granular statistics on real output, hours worked, and how AI is actually generating value.
Most current studies, he warns, rely on measures that conflate pricing power with genuine productivity. If AI adoption results in higher markups – through clever marketing or better customer targeting – but does not increase the actual quantity or quality of products, we risk overstating its economic impact. New datasets and a clearer distinction between operational productivity and customer capital are essential for honest assessment.
In the meantime, we must be humble: “With the data we have, we’re better equipped to measure the hype rather than measure the harvest”.
A call for smarter, more democratic AI
In the final analysis, what should leaders and teams do now to make AI deployment both more productive and more frugal? The advice is straightforward: decentralise and democratise. Move away from an over-reliance on hyperscale, proprietary models; lean into open-source and local solutions. In so doing, firms not only encourage resource efficiency but help check the concentration of power that threatens to undermine the social contract of technology.
“Decreasing reliance on these large-scale models and increasing the reliance on like leaner, smaller local models… would be my recommendation,” urges Lenzu, pointing to a future where AI not only does more with less, but is more widely accessible and ultimately more transformative.
To hear Simone Lenzu’s full insights on AI, resource allocation, macroeconomic frictions, and the fate of productivity, listen to this episode of Cambridge Executive Business Insights: Rethinking AI – available wherever you get your podcasts.
