Author: a16z

Translation: Shen Chao TechFlow

Shen Chao Introduction: While the entire tech industry has laid off 7% of its workforce, software development positions have actually grown by 14.6%—the reason is that AI tools enable newcomers to write production-level code as well. This is not a story of "AI stealing jobs," but rather evidence that tools make people more valuable. From internship positions soaring to lead the wage growth for teenagers, the signals in the labor market are very clear: people who can use AI are becoming hot commodities.

Image: Software development job postings have rebounded to 114.6 since the release of Claude Code, while the overall market has dropped to 93. Source: a16z

Selective Sell-off of Software Stocks

The challenges facing publicly traded software companies continue, but as we previously mentioned, the current situation is not an indiscriminate "bloodbath."

The sell-off of software stocks is not a reflection of current or recent performance, but rather a collective market skepticism regarding these companies' ability to sustain performance in the long term:

The valuation multiples for free cash flow for the next December period are at or below the levels of 2014. In other words, the premium investors assign to the cash generation capability of software companies is the lowest it has been in over a decade.

The emphasis is that even though evidence for "SaaS is dead" is still thin, it is the investors' responsibility to have an outlook on the future, which appears a bit dim to them.

However, not all software companies have a bleak outlook. Differentiation and discernment are becoming increasingly crucial. You can see this from several different angles.

First, look at the performance gaps between the median, top, and bottom quartiles of the IGV index:

In the past 30 trading days, the top quartile and median IGV stocks have outperformed the overall ETF—it's the bottom quartile (including some of the largest companies) that has dragged down the overall performance.

In contrast to the annual review, the situation is quite different:

Over the past year, the median has more closely tracked the overall ETF performance—until around the turn of the year, the performance quartiles began to diverge (now there is about a 50 percentage point gap between the top and bottom).

Going back to the 30-day view, it is evident that fundamental performance is not a strict driving factor for recent stock price performance:

From the perspective of revenue growth, there is virtually no correlation between growth and recent performance (although to be fair, many can be explained by changes in growth rather than growth levels). Segmenting by performance quartiles, a similar pattern emerges: a large number of software companies cluster around a revenue growth range of 10-20%, but their performance is almost vertically distributed, with some of the fastest-growing companies (also the largest in the ETF) located in the bottom quartile.

More important than growth is the perceived long-term durability. Here is the same chart segmented by industry:

While it is difficult to clearly categorize each company, the top and second quartile performers are often dominated by (1) cybersecurity/observability; and (2) vertical SaaS. In contrast, the bottom two quartiles comprise a mix of horizontal SaaS, cloud/infrastructure, and various "other" categories, which include market platforms, ad tech, and point solutions (along with companies like MicroStrategy that are difficult to categorize).

The key point is clear. Whether right or wrong, the market has recognized clear differences between software companies—those that seem to possess some AI defensiveness and/or tailwinds (and those that do not).

For cybersecurity and observability, AI is expected to heighten buyer urgency (and existing vendors have a trust premium that is harder for new AI entrants to disrupt).

For vertical SaaS, specialization information around workflows, data, and customer relationships is seen as a barrier to entry against AI challengers (while horizontal platforms cannot say the same).

But for everyone else, the message is clear: software itself is not a moat.

Years of sticky and steadily accruing ARR, vast feature sets, and broad adoption are just—old news. Several quarters of stable and/or improving growth might turn things around, but the market is not buying it right now.

AI Demand Growth > AI Spending Growth

Recently, you may have heard about the perceived shift from token-maxing to token-optimizing, which is significant.

The gist is this: While the training costs of cutting-edge models are rising, the marginal returns of these frontier improvements may not be enough to persuade customers to absorb the additional costs. Customers may not maximize the use of the latest and strongest tokens, but instead downgrade to cheaper, less capable models (including open-weight models) where possible.

For some, they believe this shift calls into question the sustainability of cutting-edge model development—if customers will not support the costs of the latest and greatest models, then what happens to the labs? Fair enough, though this opposition is a bit ironic since the rapid obsolescence of non-cutting-edge models was also thought to threaten the sustainability of cutting-edge model development... and yet, now that non-cutting-edge models are becoming less obsolete, this is evidently still a bad thing. Well.

Without delving into the merits of that debate, we simply observe that the costs of AI are indeed declining rapidly, which appears to have a positive impact on AI demand. In other words, the Jevons dynamic continues to hold: the cheaper AI becomes, the more people/companies want to use it (for more things).

The expanding demand side is exactly what you want to see.

Let’s start with another precursor of innovation that also began centralized and expensive and then became decentralized and cheap: computers. Compared to the last technology leap driven by PCs, the costs of intelligence in AI are decreasing more rapidly:

It took nearly two decades for PCs to achieve affordability gains that AI has achieved in about three years.

This is an extraordinary trajectory, and like computers, the cost decline seems to be helping to drive AI demand upward (with plenty of room for further growth).

On the consumer side, according to data from PNC Bank, the penetration rate of paid AI is still quite small, but it is indeed growing:

The share of households with paid AI and their average monthly spending are both steadily climbing—the growth in monthly spending is steeper, increasing by about 25% since the beginning of the year.

They are clearly not perfect comparisons, but as some perspective, only about 45% of adults aged 35-54 reported owning PCs in 1997 (many years after computing was commercially released, primarily on the enterprise side). Currently, the share of households with computers is around 90%, and nearly 97% if smartphones are included—highlighting that mass market adoption takes time and evolves with utility and cost.

Other data also shows AI demand rising with cost efficiency.

According to YipitData’s analysis of OpenRouter data (measuring only a subset of total token consumption), the use of frontier tokens continues to grow, while open-weight tokens are rapidly gaining share:

"Asian suppliers" represent a share of around 60% of total tokens (acting as proxies for open-weight alternatives), tripling since the start of the year. OpenRouter's sample may at least somewhat lean towards open-weight users, but this does align with the price differentiation story and also fits the Jevons narrative.

Similarly, according to OpenRouter, while the number of tokens per user and user spending are both growing rapidly, the former is growing much faster than the latter since the beginning of the year:

Again, it's worth noting that OpenRouter can only see what it can see, and this pattern aligns very consistently with the Jevons thesis (demand rises in sync with rapid cost efficiencies in AI).

At least for now, it is apparent that as intelligence becomes cheaper, it will further push the demand curve upwards. Cheaper tokens from less cutting-edge models are indeed gaining share, but the net effect is driving overall spending to rise exponentially. It is hard to call this a pessimistic narrative.

Pessimists may still claim that open-weight models are cannibalizing cutting-edge ones, but a scenario where efficiency gains have no clear impact on demand and/or are driving total spending down would more closely resemble a doomsday scenario. In contrast, Jevons is exactly what the bulls hope for.

Tailwinds for Entry-Level Positions

Speaking of AI demand, while it is said that AI is not useful enough to generate meaningful ROI, yet it is too useful to eliminate all jobs, we are pleased to report that there is currently little evidence suggesting that AI is actually creating any job-killing effects.

In fact, even a soft spot in the labor market (entry-level hiring) has recently gained some momentum, and if anything, AI seems to be helping rather than hurting.

First, according to Revelio's data, tracking for summer internships in 2026 is far higher than in previous years:

Internships do not equate to jobs, but they signal some level of demand for young people, with the 2026 cycle far exceeding '25 and '24 (though not matching '23).

Better yet, wage growth for teenagers and young adults also seems to be rising:

According to ADP, wages for young workers (ages 16-24) have rebounded from a low in 2025. Wage growth is a clear signal of demand, and if wages are rising, it is reasonable to infer that entry-level job prospects are also improving.

As for the impact of AI (if any), the situation is even better. AI appears to be significantly accelerating entry-level hiring:

According to an analysis of Revelio's job data and Ramp's spending data, "high-intensity AI adoption" corresponds with an approximately 6 percentage point increase in the number of entry-level employees two years later. In contrast, "low-intensity" AI adoption corresponds with about a 0.5 percentage point decline.

Image: The share of entry-level employees in companies with high-intensity AI adoption increases by an average of 1.15 percentage points two years later, while those in low-intensity adoption companies actually decrease by 0.52 percentage points, showing that AI is "creating" rather than "stealing" entry-level jobs. Source: a16z

There are a few ways to interpret this, from "AI is creating entry-level hiring," to "AI adopters are growth companies, so of course they are hiring," to "Ramp's data may not represent the broader economy, making it hard to draw any conclusions." All these make sense, but one interpretation that is almost certainly not valid is that "AI is stifling entry-level jobs."

Maybe one day it will (though there is plenty of reason to believe it won't), but not today.

Overall, while the decline in "AI exposed" positions in the post-zero interest rate era has been widely discussed, there has been relatively little conversation about the fact that "AI exposed" positions are leading the recovery:

According to data from Indeed Hiring Lab, since May 2025, the higher the degree of AI exposure, the greater the recovery in job vacancies. This is particularly evident for software engineers, with job postings increasing by about 15% while the overall market has declined by around 7% (since the release of Claude Code).

On a larger scale, it is too early to draw conclusions, but for those who tend to proclaim that "AI is making humans obsolete," the data does not support them at all. If anything, the opposite appears to be true—AI exposure seems to be positively correlated with job growth.

Image: Since May 2025, job postings for roles with higher degrees of AI exposure have rebounded more significantly, with software development positions increasing by nearly 15%, leading the overall market. Source: a16z

Fairly speaking, it is still unclear how meaningful the category of "AI exposure" is, as there is almost no consensus on what or who is considered exposed to AI (and to what extent):

It turns out that the higher the average "AI exposure" level for any given occupation, the greater the divergence on the degree of exposure.

Image: The higher the average AI exposure score for an occupation, the greater the academic divergence on its degree of exposure, indicating that this concept still lacks consensus. Source: a16z

Unsurprisingly, when predicting the future impact of new technologies, rational people may hold differing views.

There are indeed some exceptions. Everyone seems to agree that proofreaders, insurance underwriters, statisticians, and interestingly, economists, are highly exposed to AI. If AI eventually replaces human economists, will it acknowledge this crime, or will there be no economists left to tell the story?

Data Centers Make Energy Cheaper

Earlier this week, New York Governor Hochul announced a one-year moratorium on data center development. Among other things, the stated goal of the moratorium is to "protect electricity users... because data center development poses a threat to raising electricity rates." No intention to diminish her, the Governor provided more elaboration on her reasoning in an Odd Lots podcast, which you can listen to.

That said, this remains a perplexing statement.

While data centers do increase electricity demand, evidence suggests that data centers are actually helping reduce user costs:

Based on research from the Berkeley National Laboratory and The Brattle Group, on a state level, electricity price increases are negatively correlated with demand increases. Consume more, pay less—strange but true.

Indeed, in the past six years, some of the states with the fastest load growth (most data center development) such as Texas and Virginia have seen minimal price increases. Conversely, states with the highest price increases (where new data centers are scarce), such as California and New York, have seen declines in load growth—California's electricity demand decreased by around 5%, while its price increase is the highest in the nation, exceeding the second highest state by approximately 33%.

Image: The growth in electricity loads by state in the U.S. from 2019 to 2025 is negatively correlated with changes in electricity prices, with California and New York seeing declines in demand but the highest price increases. Source: a16z

Furthermore, the inverse relationship between energy demand and energy prices seems to hold in Europe as well:

According to EIA data, with the exception of Ireland, the EU countries with the highest price increases also saw the largest decreases in electricity demand.

Image: Changes in electricity demand and price among major economies in Europe and America from 2019 to 2024 are similarly negatively correlated, with Texas and Virginia experiencing growth in demand but stable prices. Source: a16z

Again, consume more, pay less. This is counterintuitive, right?

We all know that all else being equal, higher demand should push prices up, not the opposite (as claimed by Governor Hochul). But in terms of the power grid, the situation is quite the opposite. The reason there is an inverse relationship between demand and cost is that the power grid often benefits from economies of scale: the more electricity demanded, the more the fixed infrastructure costs are spread out, leading to lower overall prices for users.

In other words, Governor Hochul's data center moratorium may not only fail to "protect electricity users," but may actually have the opposite effect: electricity price increases could far exceed scenarios where data centers help to share the fixed costs of the grid. Perhaps this relationship will not continue, but that is the situation so far.

Setting energy costs aside, there are broader economic implications to consider. The reality is that without AI-related infrastructure investment, there is almost no growth in any kind of investment:

Technology-related investment is the only category of investment that is growing, so the moratorium hits the most critical area.

Image: Since 2022, technology-related investments (software, R&D, IT equipment, and data centers) have been the only positive contributors to private fixed investment growth. Source: a16z

The Governor certainly has her reasons, but a policy that might both (a) increase electricity prices (as part of "protecting electricity users" efforts) and (b) deprive New York of its most crucial investment tailwinds is indeed difficult to understand.