a16z: Software Developer Jobs Rise 14.6% After Claude Code Launch, AI Enhances Human Value

Author: a16z

Compiled by Deep潮 TechFlow

Shenchao Summary: While the entire tech industry cut jobs by 7%, software development roles increased by 14.6%—because AI tools now enable beginners to write production-grade code. This isn’t a story of AI taking jobs; it’s proof that tools are making people more valuable. From a surge in internships to teenagers leading wage growth, the labor market sends a clear signal: those who know how to use AI are becoming highly sought after.

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

The struggles of publicly traded software companies continue, but as we previously mentioned, the current situation is not an indiscriminate "massacre."

The sell-off in software stocks is not a reflection of current or near-term performance, but rather a market-wide skepticism about whether these companies can sustain their performance over the long term:

The valuation multiple for next December’s free cash flow is at or below 2014 levels. In other words, investors are assigning the lowest premium in over a decade to software companies’ cash-generating ability.

The key point is that even though evidence that "SaaS is dead" is still weak, investors have a responsibility to form views about the future, and the future, in their view, looks somewhat bleak.

However, not all software companies have a bleak outlook. Differentiation and discernment are increasingly becoming the rules of the game. You can see this from several different perspectives.

First, examine the difference in performance between the median, top quartile, and bottom quartile of the IGV index:

Over the past 30 trading days, the top quartile and median IGV stocks outperformed the overall ETF—this was due to the bottom quartile, which includes some of the largest companies, dragging down the overall performance.

Compared to the same period last year, the situation is very different:

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

Returning to the 30-day view, it is clear that fundamental performance has not been a strict driver of recent stock price movements:

From the perspective of revenue growth, there is essentially no correlation between growth and recent performance (although it should be noted that much of this can be explained by changes in growth rather than the level of growth). When segmented by performance quartiles, the same pattern emerges: a large number of software companies cluster in the 10–20% revenue growth range, yet performance is nearly vertically distributed, with some of the fastest-growing companies—also among the largest in the ETF—located in the bottom quartile of performance.

More important than growth is the perceived long-term durability. Below is the same chart broken down by industry:

Although it is difficult to clearly categorize every company, top- and second-quartile performers are typically dominated by (1) cybersecurity/observability; and (2) vertical SaaS. In contrast, the bottom two quartiles consist of various combinations of horizontal SaaS, cloud/infrastructure, and "other" categories, including marketplaces, advertising technology, point solutions, and companies like MicroStrategy that are difficult to classify.

The point is clear. Regardless of right or wrong, the market has already perceived a distinct difference between software companies—those that appear to have some AI defense and/or tailwind, and those that don’t.

For cybersecurity and observability, AI is expected to increase buyer urgency (and existing vendors benefit from a trust premium that is harder for AI newcomers to disrupt).

For vertical SaaS, specialized knowledge around workflows, data, and customer relationships is considered an entry barrier for AI challengers—something that horizontal platforms cannot claim.

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

Years of sticky, accumulating ARR, a vast feature set, and widespread adoption are merely yesterday’s news. A few quarters of stable and/or improving growth may turn the tide, but the market currently isn’t convinced.

You may have recently heard about the shift in perception from token-maxxing to token-optimizing—and it’s a big deal.

Although the cost of training state-of-the-art models continues to rise, the marginal benefits of these advancements may not be sufficient to justify the additional expense for customers. Customers may opt to downgrade to more affordable, less capable models—including open-weight models—wherever possible, rather than maximizing the use of the latest and most powerful tokens.

For some, this shift raises questions about the sustainability of frontier model development—if customers aren’t willing to support the cost of the latest and most powerful models, how can labs sustain themselves? Fair point, though this objection is somewhat ironic, since the rapid obsolescence of non-frontier models was previously seen as a threat to the sustainability of frontier model development... and now that non-frontier models are becoming less obsolete, it’s clearly still a problem. Well.

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

Expanding the demand side is exactly what you want to see.

Let’s begin with another innovative precedent that started centralized and expensive, then became decentralized and cheap: computers. Compared to the last technological leap driven by PCs, the cost of intelligence is falling even faster in AI:

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

This is an extraordinary trajectory; similar to computers, falling costs appear to be helping drive AI demand upward and to the right—there is still significant room for growth.

On the consumer side, according to PNC Bank data, the paid penetration rate remains small but is indeed growing:

The household share and average monthly spending on paid AI are both continuing to rise—monthly spending has increased more sharply, rising by approximately 25% since the beginning of the year.

They are clearly not a perfect comparison, but as a point of reference, in 1997, only about 45% of adults aged 35–54 reported owning a PC—decades after computers were commercially introduced and primarily adopted by businesses. Today, approximately 90% of households own a computer, and nearly 97% if smartphones are included—the key takeaway is that mass market adoption takes time and expands with utility and cost.

Other data also shows that AI demand rises with cost efficiency.

According to YipitData’s analysis of OpenRouter data (which measures only a subset of total token consumption), frontier token usage continues to grow, while open-weight tokens continue to rapidly gain market share:

The share of "Asian suppliers" (a proxy for open-weight alternatives) in total tokens has grown to approximately 60%, triple the level at the beginning of the year. While OpenRouter’s sample may be at least partially biased toward open-weight users, this does align with the price-differentiation narrative and also with the Jevons-style story.

Similarly, according to OpenRouter, although both tokens per user and spending per user have been growing rapidly, the former has increased much faster than the latter since the beginning of the year:

Just a reminder: OpenRouter can only see what it can see, and this pattern aligns closely with Jevons’ paradox (demand grows in tandem with the rapid cost-efficiency gains of AI).

At least for now, it’s clear that as intelligence becomes cheaper, it pushes the demand curve further up and to the right. Cheaper tokens from sub-frontier models have indeed gained market share, but the net effect is an exponential increase in overall spending. This is hardly a pessimistic story.

Pessimists may still claim that open-weight models are eroding the frontier, but alternative scenarios in which efficiency gains have no clear impact on demand and/or are driving down total spending are actually closer to a doomsday scenario. In contrast, Jevons is precisely what bulls hope for.

Regarding AI demand, although AI is said to be neither useful enough to generate meaningful ROI nor useless enough to eliminate all jobs, we are pleased to report that there is currently little evidence that AI is actually causing any job displacement.

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

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

Internships are not equivalent to jobs, but they at least signal some level of demand for young people; the 2026 cycle far exceeds those of '25 and '24 (though it falls short of '23).

Even better, wage growth among teenagers and young adults also appears to be rising:

According to ADP data, wage growth for young workers (ages 16–24) has rebounded from its low point in 2025. Wage increases are almost certainly a sign of demand—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:

Based on an analysis of Revelio’s job data and Ramp’s spending data, "high-intensity AI adoption" corresponds to approximately a 6-percentage-point increase in entry-level employee counts two years after adoption. In contrast, "low-intensity AI adoption" corresponds to a roughly 0.5-percentage-point decline.

Chart: Companies with high AI adoption saw an average increase of 1.15 percentage points in the proportion of entry-level employees after two years, while low-adoption companies saw a decrease of 0.52 percentage points, indicating that AI is creating rather than displacing entry-level roles. Source: a16z

There are several ways to interpret this—from “AI is creating entry-level hiring” to “AI adopters are growth-oriented companies, so naturally they’re hiring”—to “Ramp’s data may not represent the broader economy, making it hard to draw any conclusions.” These all make sense, but one interpretation that is almost certainly incorrect is “AI is eliminating entry-level jobs.”

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

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

According to data from Indeed’s Hiring Lab, since May 2025, job openings with higher AI exposure have shown stronger recovery. This trend is particularly evident for software engineers, with job postings increasing by approximately 15%, while overall postings have declined by about 7% (since the release of Claude Code).

From a broader perspective, it’s still too early to draw conclusions, but the data does not support those who claim that “AI is making humans obsolete.” If anything, the opposite is true—exposure to AI appears to be positively correlated with job growth.

Chart: Since May 2025, job postings in roles with higher AI exposure have rebounded more strongly, with software development positions rising nearly 15%, outpacing the broader market. Source: a16z

Fairly speaking, it is still unclear how meaningful the category of "AI exposure" is, as there is little consensus on what or who is exposed to AI—and to what extent:

The higher the average "AI exposure" for a given occupation, the greater the divergence in views on the level of exposure.

Chart: The higher the average AI exposure score for a profession, the greater the disagreement in academia regarding its exposure, indicating that this concept still lacks consensus. Source: a16z

As expected, rational people may have differing views when predicting the future impact of new technologies.

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

Earlier this week, New York Governor Hochul announced a one-year pause on data center development. Among other things, the stated goal of the pause is "to protect electricity consumers... due to the threat that data center development poses to rising utility bills." Without intending to diminish her position, the governor provided further context for her reasoning on the Odd Lots podcast—worth a listen.

Nevertheless, this is still a confusing statement.

Although data centers are indeed increasing electricity demand, evidence suggests that they are actually helping to reduce user costs:

Based on research from Lawrence Berkeley National Laboratory and The Brattle Group, at the state level, increases in electricity prices are negatively correlated with demand—consume more, pay less—strange but true.

Indeed, over the past six years, states with some of the fastest load growth (most data center development), such as Texas and Virginia, have seen little to no price increases. In contrast, states with the largest price increases (few new data centers), such as California and New York, have experienced slower load growth—California’s electricity demand decreased by approximately 5%, while its prices rose more than anywhere else in the nation, about 33% higher than the second-highest state.

Chart: From 2019 to 2025, electricity load growth across U.S. states shows a negative correlation with electricity price changes; California and New York experienced declining demand but the highest price increases. Source: a16z

Moreover, the inverse relationship between energy demand and energy prices also appears to hold in Europe:

According to EIA data, the EU countries with the largest price increases, excluding Ireland, also experienced the largest declines in electricity demand.

Chart: From 2019 to 2024, electricity demand and prices in major economies in Europe and the U.S. showed a similar negative correlation, while demand in Texas and Virginia increased with relatively stable prices. Source: a16z

Similarly, consume more, pay less. This goes against intuition, right?

We all know that, all else being equal, higher demand should push prices up, not down (as Governor Hochul has claimed). But when it comes to the power grid, the opposite is true. The reason demand and costs have an inverse relationship is that power grids often benefit from economies of scale: the greater the electricity demand, the more the fixed infrastructure costs are spread out, resulting in lower overall prices for users.

In other words, Governor Hochul’s data center moratorium may not “protect electricity consumers” but could instead have the opposite effect: electricity prices could rise far higher than they would if data centers helped share the grid’s fixed costs. This relationship may not hold indefinitely, but it reflects the current situation.

Setting aside energy costs, broader economic impacts must also be considered. The reality is that without infrastructure investment related to AI, there is virtually no investment growth:

Tech-related investments were the only growing category, so the pause order struck precisely at the most critical point.

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

The governor certainly has her reasons, but a policy that could both (a) raise electricity prices (as part of an effort to "protect electricity consumers") and (b) strip New York State of its most important investment tailwind is difficult to understand.