a16z Charts Explained: Are Prediction Markets Biased? Where AI Capital and Jobs Are Flowing

This edition of a16z's chart roundup examines four distinct topics: election forecasting, software valuation multiples, hardware infrastructure buildout, and customer service employment. Two contrasting dynamics are particularly striking: software operating scores have climbed while valuation multiples have compressed, and while broader office positions continue to expand, customer support roles have started pulling back on hiring.

When Bettors Bring Political Biases, Do Contract Prices Get Distorted?

Participants in prediction markets are not a random cross-section of the voting public. Demographic surveys cited in the literature indicate they lean younger, more male, and possess higher formal educational attainment. If traders are eager to back candidates they personally support by bidding higher prices, contract valuations could easily be pushed away from fair value. The central empirical question is whether profit-seeking arbitrageurs step in to correct those pricing distortions.

In the first round of the Brazilian presidential election, Flávio Bolsonaro secured approximately 47.0% of the vote against Luiz Inácio Lula da Silva's 45.2%. Because neither candidate crossed the 50% majority threshold, both advanced to a head-to-head runoff. The eventual president has yet to be determined by the runoff, and the ultimate winning contract has not been settled. On the Kalshi exchange, contracts tracking Bolsonaro's election victory flipped to become the favored outcome roughly three weeks before the vote, climbing further after his first-round plurality was confirmed.

Political developments also ripple into equity market expectations. a16z compared the contract movement against the EWZ ETF, which tracks Brazilian equities, noting that the fund gained about 15% after the first-round outcome became clear, having already begun drifting upward when the contract probability shifted earlier. While this demonstrates how political assessments enter broader asset pricing, the relative timing of two market prices alone does not prove which venue acquired actionable information first.

Prediction markets predominantly trade binary contracts: paying out $1.00 if the designated event materializes, and $0.00 otherwise. In market parlance, a contract trading at 60 cents is commonly interpreted as implying an approximate 60% probability of occurrence. However, this market clearing price is not identical to a candidate's popular vote share, nor is it an established certainty; trading frictions such as liquidity depth, transaction commissions, and capital opportunity costs all affect how cleanly prices translate into strict probabilities.

To test whether markets exhibit systematic bias, researchers analyze long-term settlement deviations across large catalogs of historical contracts. If a platform systematically underprices candidates bearing specific attributes, a strategy of consistently buying those contracts would yield an average positive residual when subtracting purchase price from terminal payout—measured as a settlement-price residual in percentage points. This residual reflects an accumulated average over many distinct outcomes; an individual contract can still lose completely, and the metric does not imply a risk-free investment return on principal.

A working paper released through NBER by Dartmouth economist Eric Zitzewitz compiles data spanning over a century, incorporating exchanges such as IEM, Intrade, PredictIt, Betfair, and Kalshi. The full market dataset comprises 9,642 contract observations across 216 distinct election days (the working paper has not yet completed peer review and does not include Polymarket data; observations come primarily from modern commercial platforms, with the US accounting for roughly 60%, the UK 16%, and Australia 13%).

In full-sample regressions, the primary explanatory variable codes ideology as left-wing (-1), centrist (0), and right-wing (+1), with the dependent variable defined as contract settlement value minus purchase price in percentage points (capturing pricing deviation from outcome, not return on invested capital). The empirical findings show:

1. The overall market partisan coefficient is close to zero: The right-leaning indicator registers a regression coefficient of +0.03 percentage points with a standard error of 2.66. The 95% confidence interval spans [-5.18, +5.24]. While it crosses zero and shows no evident directional bias, the interval remains wide enough to accommodate small biases of several percentage points and cannot be taken as definitive proof that the market is strictly unbiased. Because multiple contracts on the same election date are correlated, standard errors are clustered by election day (without election-day clustering, standard errors would be underestimated by a factor of roughly 2.75).

2. The non-US point estimate is negative but statistically insignificant: Within non-US election samples, the regression point estimate stands at -1.42 (standard error 1.30, 95% confidence interval [-3.97, +1.13]), which also spans zero.

3. Demographic preferences show no overall statistically significant evidence: In contracts evaluating explicitly named US candidates, the payoff differentials for female candidates (-0.95), non-white candidates (-0.85), and age per additional decade (+1.17) all fail to achieve statistical significance. (Across 18 regression coefficients in multinomial logit models controlling for longshot bias, the female coefficient was significant only once, which the author cautions against overinterpreting.)

This does not mean prediction platforms operate free of pricing distortions. A well-documented phenomenon across historical prediction markets is the "favorite-longshot bias," wherein market participants systematically overpay for low-probability longshots while underpricing heavy favorites (this evidence derives mainly from legacy commercial platforms; existing sample sizes on Kalshi remain too small to confirm whether the bias persists there).

Furthermore, across approximately 62,000 daily observations on PredictIt (where a single contract is tracked daily across its active lifespan, which typically lasts around 7 days), contracts tied to short-term polling averages overvalued left-leaning polling outcomes by an average of roughly 1.4 percentage points (p < 0.01). This test evaluated short-term polling average fluctuations, which are distinct instruments from long-term general election winner contracts; the study does not isolate which specific trading motives drove this gap. Because historical election sample sizes are inherently limited and statistical power is constrained, if general election contracts carried a comparable slight bias of 1 to 2 percentage points, existing datasets would not be able to detect it reliably. One therefore cannot claim that actual election markets have had every distortion eliminated by arbitrage. Theoretical mechanisms and historical literature indicate that when prices deviate noticeably, informed traders face strong financial incentives to take counter-positions and curb bias; however, this operates as a balancing mechanism rather than a guaranteed eradication of all minor discrepancies.

Why Didn't Software Companies Get More Expensive When Operating Scores Improved?

Software investors have long leaned on the "Rule of 40" (or Rule of X) as a benchmark for business health. Its original intent was to establish a standardized heuristic to evaluate companies across disparate lifecycle stages: early-stage firms typically exhibit rapid top-line growth accompanied by steep operating losses, whereas mature companies are expected to moderate growth while generating healthy profit margins. The heuristic directly sums "year-over-year revenue growth rate" and "net profit margin" (the charts in this series utilize net margins inclusive of stock-based compensation, or SBC), considering a combined score of 40% indicative of strong performance. For instance, 100% revenue growth with a -60% net margin yields a score of 40, as does 10% revenue growth paired with a 30% margin. Conversely, a firm expanding revenue at 150% while losing -150% in net margin scores 0; faster growth does not automatically confer a higher composite score. This heuristic is commonly employed as a quick screening rule of thumb, not a pricing theorem that dictates identical valuations for differing business profiles.

In recent years, however, publicly traded software equities have displayed a conspicuous divergence: the median Rule of X score across software companies has steadily climbed, while valuation multiples have trended downward.

According to public market data, the median Rule of X score for US-listed software companies rose from roughly 17% to near 23% over a three-year span. Concurrently, enterprise value relative to annualized revenue (EV/annualized revenue) compressed from 5–6x down to approximately 3.5x. Enterprise value (EV) represents the aggregate cost to acquire an entire business (accounting for both equity value and net debt), and the revenue multiple indicates how much enterprise value the market is willing to pay per dollar of annualized sales. When measuring the rolling eight-quarter correlation between the two metrics, the relationship collapsed from a strong positive reading near +0.95 straight into negative territory.

Why didn't operational score improvements drive multiple expansion? The primary reason is that the Rule of 40's simple addition masks fundamental structural differences in how value is created:

• Growth and profit carry different weight in valuation: The Rule of 40 adds revenue growth and profit margin directly. This is a rough, experience-based trade-off for assessing operating performance, rather than a valuation model derived from rigorous pricing theory. Using it to predict valuation multiples directly can be misleading. Profit and controlling losses still matter, and companies should not pursue unchecked expansion; however, public markets usually assign a higher premium to the potential for revenue expansion. a16z explains that recent increases in software Rule of X scores have mainly been driven by improving profit margins, while revenue growth has slowed relatively. Even when margins lift the total score, valuation multiples can remain under pressure without strong growth. This is a divergence observed in the current public-market sample, not evidence that improving the operating score itself causes valuations to fall.
• Breaking down the same score: Consider two hypothetical companies that both score 40: Company A has revenue growth of 45% and a net profit margin of -5% (45 + (-5) = 40); Company B has revenue growth of 10% and a net profit margin of 30% (10 + 30 = 40). The arithmetic scores are identical, but Company A shows the potential for rapid expansion, while Company B shows steady profit-generation capability. Valuation multiples depend on the market's combined assessment of growth potential and sustainable profitability. A single operating score cannot directly determine a company's market value.

Across four discrete Rule of X cohorts (0–<20, 20–<30, 30–<40, and 40+), each quadrant plots next-twelve-months (NTM) revenue growth on the horizontal axis against EV/NTM revenue multiple on the vertical axis (note that axis scales vary across the four individual sub-charts). While select cybersecurity and observability vendors command premium multiples, each high-scoring cohort also contains heavily discounted names. Within the 20 to 40 score range, most software multiples cluster between 5x and 10x. Among the elite 40+ cohort, a quadratic fit between forward revenue growth and forward revenue multiples yields an R² of just 0.02. This shows that even within the highest-performing operational tier, revenue growth alone shares a negligible quadratic relationship with the valuation multiple, which does not justify deriving a mechanical pricing formula from this chart.

Hardware Buildouts Are Surging, but Have Software Expectations Kept Pace?

In discussions around the commercialization of artificial intelligence, "Atoms" often refers to physical infrastructure—semiconductors, servers, electricity, and networks—while "Bits" refers to the software layers, platforms, and application services built on top.

Earnings revision breadth across the S&P 500 has pulled back from recent highs, though it remains in positive territory. This breadth metric tracks how widely analyst revisions are skewing positive across constituents, not the absolute magnitude of corporate earnings growth. Looking across sector breakdowns, larger upward adjustments are concentrated in hardware and energy.

FactSet earnings revision data demonstrates that sell-side analysts have upgraded outlooks far more aggressively for physical infrastructure:

• Hardware and energy lead positive revisions: In current-quarter earnings estimate revisions, sectors registering the most pronounced upward revisions include energy (+19.3%), semiconductors (+9.8%), technology hardware (+7.6%), and financials (+1.9%), whereas software and services saw only a marginal upward revision (+0.7%). Looking ahead to next-year earnings outlooks, semiconductors received a +27.3% upward adjustment, energy +13.3%, hardware +9.8%, and software and services just +0.4%.

These percentage figures denote the magnitude of analyst consensus revisions, not absolute capital flows or realized year-over-year profit growth rates. An upward revision of 19.3% in energy means analysts raised their profit projections for that sector by roughly one-fifth; one cannot attribute all traditional sector upgrades strictly to AI demand. Conversely, modest upward revisions in software simply indicate that expectations have stayed relatively steady, not that software businesses themselves have stopped growing.

Physical manufacturing backlogs are also expanding. In ISM survey data, the proportion of industries reporting expanding order backlogs approached 60% on a three-month moving average basis. This indicates that order accumulation is broadening beyond a handful of niches. However, because the survey covers general industrial manufacturing without isolating AI-specific initiatives, not all backlog pressure can be ascribed to AI infrastructure buildouts.

Portfolio data disclosed by the Carlyle Group illustrates the contrasting paces across supply chain tiers: an index tracking AI hardware shipments across its ecosystem shows a steep upward trajectory, whereas year-over-year spending growth on broader IT services has hovered steadily around 30%.

A common misinterpretation lies in the charting format: the left panel tracks an index level for hardware shipments, while the right panel measures the annual growth rate of IT services spending. This micro-level evidence on actual spending and shipments comes from Carlyle's portfolio-company sample, rather than a comprehensive census of global IT spending. The fact that IT service spending growth has stabilized near 30% indicates that spending volumes continue to grow at a healthy clip; it simply means the pace of expansion is not accelerating further as it is for hardware, and it should not be conflated with flatlining enterprise tech budgets.

These charts highlight divergence in timing across the technology stack: AI hardware shipments in Carlyle's sample have risen rapidly, while IT services spending in the same sample has continued to grow steadily without further acceleration. Analysts have raised earnings forecasts several times before; market consensus is dynamic rather than permanent and adjusts as subsequent business performance becomes clear. The divergence between hardware and software expectations represents a snapshot of sentiment today, not a settled conclusion.

The Decline in Customer Support Starts with Slower Hiring, Narrowing the Junior Office Gateway

Whenever the labor market impact of technological change is evaluated, customer contact centers are frequently singled out as the operational tier most exposed to generative AI. Call center work centers on handling recurring inquiries, processing standard returns, and conducting script-based phone confirmations—high-frequency, procedural administrative tasks. Because these workflows rely heavily on structured text and defined processes, they represent natural domains for large language models and automated support agents to assist with or replace some tasks.

Workforce analytics tracked by Revelio Labs reveal that after more than a decade of steady global expansion, year-over-year employment growth in call centers has dipped below zero, settling near -4%.

Geographically, contraction manifested first in high-income economies before propagating into middle-income outsourcing hubs. Across an 18-country dataset, high-income, upper-middle-income, and lower-middle-income brackets all ended with negative contributions to aggregate employment growth. While this shows headcount compression reaching lower-cost destinations, group contributions are measured in percentage points rather than individual cohort growth rates; select national markets within cohorts continue to expand.

To determine whether this weakness simply mirrors a broader softening in corporate white-collar roles, researchers benchmarked call center employment against broader domestic white-collar services across 12 middle-income countries (indexing each to 100 at the public release of ChatGPT). In the Philippines (call centers 90 vs. other white-collar 112), Mexico (94 vs. 112), and Brazil (98 vs. 114), call center contraction diverged sharply from overall white-collar job growth. China's call center index (around 107) moved broadly in line with its domestic white-collar trajectory, but it is not alone in maintaining positive momentum; India's call center index remains above 100 (around 106), and Nigeria has also sustained growth. This intra-country divergence between call centers and other office roles strengthens the plausibility of AI substitution, even though industry-specific demand adjustments cannot be entirely ruled out.

Underlying flow data shows that the shrinking headcount is not driven by a surge in departures, but by an abrupt slowdown in new recruitment. Across the 12 tracked middle-income economies, the hiring rate per 100 employees fell from 25.1 to 17.2, while the departure rate moved from 22.3 to 18.5. Because recruitment declined faster and fell below the rate of departures, incoming hires were insufficient to backfill departing staff, causing total headcount to shrink on a net basis without requiring a spike in employee turnover.

This pullback in recruitment directly impacts new entrants seeking their first corporate opportunity. Revelio's analysis underscores that customer support has historically served as a reliable on-ramp to formal office employment across lower- and middle-income economies; citing International Labour Organization research, it emphasizes that these entry-level gateways have been vital for younger workers and women. As this doorway narrows, securing a first office role may become more difficult for new job seekers. Even as the office economy and higher-paying jobs continue to grow, entry-level candidates cannot automatically step into those advanced positions.

For workers who do transition out, granular mobility tracking shows significant friction in career pathways: among observed job changers, 49.7% moved laterally into other customer service, sales, administrative, or human resource positions. Only 10.8% transitioned into higher-paying, expanding technical support, customer success, software, or data roles. This demonstrates that transitioning into higher-wage tech functions involves substantial barriers, with many workers finding their next job in categories close to their previous role (this mobility sample covers 12 middle-income countries and excludes the US; it cannot be inferred that all remaining, unclassified workers are unemployed).

At the macroeconomic level, a stable headline US unemployment rate does not prove that displaced support workers found roles with equivalent compensation; international tracking highlights noticeable labor frictions in non-US markets. While AI deployment is very likely contributing to this contraction, the exact share attributable directly to automation cannot be isolated from available aggregated data. The initial deceleration in call center growth began slightly before ChatGPT's rollout, coinciding with global interest rate increases and corporate cost-cutting initiatives; the underlying data does not track which specific enterprises replaced staff with automated agents on a company-by-company basis. While resilience in broader domestic white-collar employment highlights the distinct pressure on customer support, shifts in end-market demand for call center services may still influence these numbers.

Clarifying Which Questions Each Dataset Answers

These charts cover pricing, operational performance, earnings revisions, and labor dynamics. Markets can reprice an election outcome or re-evaluate software valuations; whether corporations actually translate that into higher revenue, and whether job seekers can secure their next role, depends on separate real-world evidence.

The key to interpreting these shifts is understanding what each metric is equipped to answer: prediction market prices measure probability expectations, not verified outcomes; earnings revisions and valuation multiples reflect shifts in analyst sentiment and market pricing of growth, not final corporate trajectories; and labor inflow, outflow, and recruitment statistics reveal how individuals actually enter or exit employment. Distinguishing among these distinct layers enables a more grounded evaluation of where structural change is genuinely taking place.