Author: a16z

Translation: Felix, PANews

Recently, a16z explored the core trends of the current technology and business cycle from multiple dimensions, including market investment, AI applications, startup ecosystems, and the retail industry. The article argues that, driven by the AI wave, capital markets are gradually shifting from a past preference for asset-light, consumer-focused internet companies toward physical industries like hardware and robotics. At the same time, AI is reshaping corporate organization, entrepreneurial barriers, and productivity growth logic. Below is the full content.

Cycles: Same Yet Different

If you compare this cycle to the previous one, you'll find that in some ways they are identical, while in others they are complete opposites.

The similarity lies in the fact that technology has been a consistent winner across both the post-financial crisis era (2010-2020) and the post-pandemic era (2020–present). However, the landscape of other industries has undergone dramatic shifts: the winners of the last cycle have become the losers, and vice versa.

• Healthcare, consumer goods, and media all posted double-digit returns after the financial crisis, but now they hover around 3% to 6%.

• Meanwhile, returns in energy, raw materials, construction, and finance have jumped from low single digits to mid-to-high double digits.

The once-lagging industries have become leaders, and the former leaders have fallen behind.

Technology is the exception—it has been a cyclical winner throughout—but there are nuances. Hardware has been the standout performer in this cycle (it did fairly well last time too), but software has followed the broader reversal trend.

Stepping back, there is a very clear pattern previously mentioned: the market has shifted attention from asset-light, consumer-oriented sectors to the heavy-asset "physical" economy, largely driven by AI infrastructure development.

This is a rotation from bits (virtual) to atoms (physical).

"Heavy-asset" companies have turned the tide after lagging behind "asset-light" companies for more than a decade.

Of course, if this cycle parallels the previous one, the overall trend is that all this heavy-asset infrastructure eventually extends to the software/application layer. In the post-financial crisis era, chip manufacturers (and cloud service providers) dominated early on but eventually gave way to the apps, marketplaces, and enterprise software thriving on cloud platforms powered by phones, computers, and servers (driven by chips). In other words, the shift to the virtual layer was temporary and cyclical, rather than a more lasting structural change.

This could certainly happen again this time: in fact, it would be quite disappointing if the buildup of AI infrastructure never extended to the asset-light layer (though they may ultimately evolve together). Even so, at the public market level, some signs suggest that the "virtual revolution" might have its own staying power. Strictly speaking, this isn't just about AI infrastructure.

The premium on "real-world" technology is becoming evident in private markets, not only in AI infrastructure but also in robotics:

Measured by the market cap of the top 100 private companies (by category), robotics (and physical AI) didn’t even appear on the list in 2016, but a decade later, it has surpassed fintech and payments to become the second-largest category.

If you look at venture capital flows, you’ll also see a surge in interest in robotics:

According to PitchBook data, Q1 saw record highs in both investment amount and deal count in robotics and physical AI, with approximately $16 billion invested across nearly 500 transactions.

For context, the robotics investment boom is about 2x higher in volume and 4.5x higher in value compared to the 2021–2025 period.

The key takeaway is that the shift toward the physical economy (at least in private markets) is not just about chips and inference: hardware as a standalone product is on the rise.

This is not hard to understand. Superior software has immense potential, but robotics pushes technology into a range of real-world "tasks" that software alone cannot reach. AI, in some ways, unlocks the software that drives hardware, expanding the scope of demand in unprecedented ways. This is quite similar to how electricity eventually enabled machines to perform work that humans could hardly have imagined.

Currently, the most compelling new frontier in robotics is defense. Of course, the steady growth in global defense budgets is also a driving factor. If everything goes as planned, the shift toward asset-intensive industries could be deeper, broader, and more persistent than any modern tech cycle before.

Is AI the Enabler, or Is AI Being Enabled?

In the early days of the large language model (LLM) wave, management consulting firms were seen as potential winners in the AI space, at least in the short term. The logic was simple: companies want to use AI, so they will hire consulting firms to figure out how. Accenture, in particular, was seen as well-positioned because it could provide not just advice and roadmaps but also end-to-end services, known as "managed services."

For whatever reason, the market’s optimism around Accenture seems to have faded:

Accenture’s free cash flow multiple peaked at 30x in early 2025 but has since fallen to about 6x, roughly one-third of its long-term average.

Why the market lost confidence in Accenture so quickly is open to interpretation. But one thing is becoming increasingly clear: in the broader realm of "adopting AI," the implications go far beyond simply adopting AI. Not all AI adoption creates value in the same way, and truly (or more effectively) adopting AI, according to some recent research, requires nuanced strategies in the development and ideation stages.

In a study involving 515 high-growth startups, researchers focused on what it truly means to be "native AI." Specifically, they wanted to know how to transition from "AI improving tasks" to "AI improving companies," and the results were striking.

It turns out that the key lies in what the researchers call the "mapping" problem.

When the companies in the study were told how others restructured production around AI (the "experimental group"), they embarked on a very different exploration process. Instead of simply replicating existing workflows, they started upstream and integrated AI into business outcomes, creating entirely different processes.

The researchers used product development as an example:

In this case, AI didn’t replicate existing steps in the process; instead, it redesigned the process around its own capabilities, even though the goal was to achieve the same basic business outcome.

Of course, this is just one example, but the overall productivity impact of AI on the "experimental group" was immense. The experimental group:

• Saw an approximately 44% increase in AI use cases:

• Achieved about 2x revenue for the top 5% (50% increase for the top 10%):

• Reduced capital consumption by approximately 40% (with even larger gaps at the extremes).

In summary, when high-growth startups truly set out to "adopt AI," they discovered more use cases, generated more revenue, and consumed less capital than those that didn’t.

This is a fairly striking result, one that can both ease some concerns about the "AI ROI problem" and explain why AI’s ROI hasn’t been fully realized at the enterprise level, at least not to the extent some expect.

The researchers suggest this means that (a) AI-driven productivity gains at the enterprise level are indeed transformative, but (b) the real breakthrough lies in the exploration phase—that is, "figuring out where and how to deploy AI is the key bottleneck to realizing gains," which goes beyond simply "adopting AI."

In this sense, the existence of an "exploration bottleneck" means that AI’s trajectory mirrors that of previous technology-driven productivity leaps.

For example, when electrification first began to spread, many manufacturers simply replaced steam engines with large electric motors while keeping the existing overhead shafts and belt drives in place. The factory remained largely unchanged, just "with electric motors this time." It wasn’t until manufacturers realized they could install small motors on each machine (and almost completely abandon the shaft and belt system) that the true benefits emerged: factories were ultimately redesigned around embedded electrical systems (rather than the other way around). Of course, what followed was a landmark in the history of productivity leaps.

AI Startups Streamlining Operations

Regarding AI, startups, and academic research, the same group of researchers found another insight: AI startups are indeed streamlining operations—at least according to a study of Y Combinator startup data over the past four years.

The researchers analyzed data from YC batches W20–F24 (with initial funding rounds between 2020 and 2024) and correlated it with Revelio data on headcount, function, and seniority. They wanted to see if AI startups differ from non-AI startups in terms of hiring and/or organizational structure.

Their findings:

• AI startups start smaller and operate leaner:

• The distribution of startups with fewer employees is heavily skewed toward AI startups:

• AI startups tend to have flatter hierarchies, with AI startups making up the largest share of companies with few or no layers:

The implications are clear, though there may be more variables in the details. But you get the point: if you believe AI will enable companies to do more with less, this survey provides further evidence.

Additionally, Stripe Economics has weighed in again on the rise of the AI-empowered "solopreneur."

(Note: Recently, Ernie Tedeschi from Stripe Economics, based on Stripe’s own data, suggested that all types of founders seem to have grown in Q1, but "non-AI solopreneurs" saw the most significant growth, followed by "AI solopreneurs." As shown below:)

Despite various caveats about how Stripe identifies “solopreneurs” in its data, the findings provide further evidence that AI is indeed driving more entrepreneurial activity and business formation, with solopreneurs achieving considerable success.

Take a look at the proportion of solopreneurs by revenue thresholds:

Not only is the proportion of solopreneurs earning over $100,000 annually on the rise, but the share of those earning over $5 million and $10 million has also grown significantly in 2023 and 2024.

Stripe Economics notes:

• We found a significant increase in the number of solopreneurs earning over $100,000 annually in our index, but the growth was even larger at higher revenue thresholds, with the pace accelerating since 2023. In 2025, the number of solopreneurs making over $1 million annually is more than double the 2023 figure, while those earning over $5 million and $10 million are nearly triple the 2023 levels.

• Perhaps more interestingly, the proportion of solopreneurs exceeding these thresholds has also doubled over the past two years. This suggests that the surge in business registrations isn’t reflective of low-quality experimentation by a lucky few, but rather that the new cohort of solopreneurs may be of higher quality than in the past.

Of course, given numerous uncertainties—such as how solopreneurs are identified (in this case, via Stripe’s solopreneur-specific tools) and the fact that employee counts may change over time (of which Stripe may be unaware)—the data indicates that the era of AI-driven small businesses is still evolving.

Technology Makes Grocery Stores More Efficient, But Also Less Efficient

One interesting aspect of grocery stores is that, unlike the broader retail trade category, their productivity hasn’t significantly improved over the past 30 years:

Or, more precisely, since 1990, retail productivity growth has remained fairly steady, while grocery productivity first declined, then recovered, plateaued, and despite a recent dip, has begun to rebound—though it still lags far behind the surge in retail productivity.

This is interesting because it tells the story of technology (and its relationship with productivity) on one hand, and how productivity is measured on the other—broadly defined as output divided by labor hours (an imperfect metric at best).

For grocery stores (and retail), apart from the cash register, the greatest invention has been the electronic scanner. They first appeared in the 1970s, but by the 1990s, they were ubiquitous. Scanners primarily served two purposes: (1) vastly expanding the range of inventory, and (2) facilitating increasingly granular data collection for retailers and grocers to understand customer purchasing intent and required stock levels.

In the 1990s, both grocery stores and retailers began expanding dramatically in scale, benefiting from technology-driven economies of scale. This was good for consumers but more or less spelled the end for family-run mom-and-pop shops.

However, from that point onward, the fortunes of retailers and grocers diverged. Retailers massively expanded their inventory without adding many new employees, instead focusing on ready-made, pre-packaged goods that required far fewer people to manage and monitor stock. Grocers, on the other hand, decided to expand beyond groceries into specialty services like florists, bakeries, deli counters, and so on.

Of course, as the share of specialty services grew, so did the demand for specialized labor. As shown in the chart above, despite improvements in grocery productivity—such as a massive expansion in the variety of goods and services and lower prices—their "productivity" in terms of output per labor hour did not improve. This is also why retail "productivity" far outpaces grocery "productivity," even though wages in both sectors have grown at roughly the same rate.

It wasn’t until grocery stores began adopting lessons from broader retail and department stores that their productivity started to rise again:

Around 2000, the share of non-food household products began to grow significantly: higher-margin pre-packaged foods, snacks, and general merchandise nearly quintupled within a decade. Simultaneously, supermarkets outsourced more tasks like stocking and display to suppliers, akin to charging "slotting fees" for shelf space. This was a clever "productivity-boosting" strategy—though labor hours didn’t decrease, they were shifted to others.

From a "productivity enhancement" perspective, this shift increased output without adding labor hours, sparking a productivity renaissance for supermarkets.

Although labor’s share of grocery revenue steadily rose until around 2002 (while retail’s labor share declined), both shares have been steadily declining, at least until recently.

The decline in the "labor share of income" is essentially the inverse of "productivity": producing more with fewer workers leads to a drop in the labor share of income (without accounting for the growth in 401k retirement accounts from all those profits).

Interestingly, though (returning to the topic of technology and productivity), the latest wave of shopping innovation—e-commerce and home delivery—seems to coincide with a renewed divergence in grocery and retail "productivity." While e-commerce is a boon for retailers, who no longer need to lease physical stores, home delivery may mean the same or even more people wandering around grocery stores picking items. Curbside pickup may even be more labor-intensive than traditional shopping.

Whether this is causal or coincidental, the fact is that after the pandemic, grocery productivity fell again (and the labor share began to rise), while retail became leaner and more efficient. The same technology, the same productivity gains—yet the resulting "productivity" looks vastly different.

Still, the good news for grocers is that advertising on shelves always generates money (high margins).