AI: the modern day haves and have-nots
Who actually benefits from AI tools, what the productivity research really says, who is paying the infrastructure bill, and whether the access gap is closing or widening. A data-led look at a question most coverage skips.
Picture two people opening ChatGPT on the same Tuesday morning. Same app, same icon, same blinking cursor. What happens next is where it gets interesting.
The first is on the free plan. She gets GPT-5.3 Instant capped at 10 messages every five hours, a 16,000-token context window, ads in the US, and no Deep Research, Sora, Agent Mode, or image generation. The second pays $20 a month for Plus. She gets GPT-5.5, 160 messages every three hours, double the context, Deep Research, Sora, Agent Mode, image generation. That is not a better version of the same product. It is a different product category wearing the same logo. Go to $200/month for Pro and you get a 1-million-token context window (an entire codebase, a book manuscript, years of filings) and twenty times the message volume of Plus.
The gap between free and Plus is $240 a year. In isolation that is trivial. The catch is that $240 a year is roughly the line where AI stops being a novelty and becomes a serious productivity tool, and that line sits at a meaningful income level for a very large share of the planet. This is the AI access question, and it is more tangled than the price tag suggests.
What does the productivity research actually show?
Here is the finding almost nobody leads with: when access is equal, AI narrows the gap between workers rather than widening it.
Noy and Zhang’s 2023 controlled experiment in Science randomly handed ChatGPT to 444 college-educated professionals. Completion time fell 40%, quality rose 18%, and (the part that matters) ChatGPT compressed the productivity distribution, helping weaker performers more than stronger ones. Their words: “inequality between workers decreased, as ChatGPT compressed the productivity distribution.” Brynjolfsson, Li, and Raymond’s NBER paper followed 5,179 customer-support agents through a staggered AI rollout: 14% productivity gain overall, but 34% for novices and almost nothing for the already-experienced. The mechanism: “The AI model disseminates the best practices of more able workers and helps newer workers move down the experience curve.” Inside a firm where everyone has the tool, AI behaves like an equalizer.
There is a sharp constraint, though. The BCG/HBS “Jagged Frontier” study of 758 BCG consultants found that on tasks inside AI’s frontier, GPT-4 users were 25% faster and rated 40% better, but on tasks outside it they did 19 percentage points worse than colleagues with no AI at all. Knowing when not to use the tool is itself a skill, and it does not distribute evenly. Peng et al. clocked developers 55.8% faster on coding tasks with Copilot, and GitHub’s internal analysis reported 87% felt less mental effort on repetitive work, with the now-familiar caveat that a September 2024 analysis found higher bug rates at constant throughput. The through-line is consistent: with equal access, the gains equalize performance. These were controlled experiments where researchers handed everyone the same tool. They did not measure a world where AI is universally free, and that distinction is load-bearing. The access question and the productivity question are separable, and the inequality problem lives in the gap between them.
Who has access, and who does not?
The Anthropic Economic Index tracks real usage. Computer and mathematical roles are 37.2% of Claude conversations but 3.4% of the US workforce. Usage clusters around $75,000 to $100,000 earners. Internationally, a 1% rise in GDP per capita tracks a 0.7% rise in Claude usage per capita; within the US, a 1% income rise tracks 1.8% higher adoption by state. Pew’s February 2025 survey found 91% of degree holders aware of workplace AI versus 76% without a bachelor’s, and by October 2025 degree holders were using it at nearly double the rate (28% vs. 16%).
The teen data flips the script in a way worth pausing on. Pew’s February 2026 teen report found US teens in households under $30,000 are more likely to use AI for most schoolwork than teens in households over $75,000 (20% vs. 7%). Lower-income teens use it more. They just use it on weaker, free accounts while higher-income peers use it less often on premium ones. Frequency favors one group; capability favors the other, and capability is the part that compounds. Geography stacks on top: OECD analysis puts urban AI exposure at 32% versus 21% rural (Stockholm and Prague at 45%, rural Cauca, Colombia at 13%), and the Federal Reserve’s broadband data explains why: 74% of nonmetro US households have broadband versus 85% metro, and slower when they do. Access needs a stable connection before it needs a subscription.
What do free AI tiers actually cost?
Free tiers exist; the terms tell you what the transaction really is. OpenAI’s privacy policy splits by tier. Free and Plus: conversations stored indefinitely unless deleted, used for training unless you opt out. Enterprise: not used for training, zero-retention options available. The API: not trained on by default. Free users pay with data; enterprise customers pay with money and receive privacy as a feature.
The FTC flagged exactly this asymmetry in January 2024, warning that “AI companies’ business incentive to constantly ingest additional data can be at odds with obligations to protect users’ data,” and reiterated in February 2024 that quietly amending a privacy policy to cover AI training is not transparency. Positions, not enforcement, but the direction is clear. The commercial logic is clean: free users generate training signal, training signal improves models, better models attract paying subscribers. The free tier funds the paid tier. Whether that is fair depends on what you think a conversation is worth and whether the person having it understands the deal. The EU AI Act’s Article 10 will require documented data governance for high-risk training data, with full compliance not due until June 2027. Until then: a privacy policy and an opt-out, if you read closely enough to find it.
What are the infrastructure costs behind AI?
You cannot read the access gap without the supply side, because the companies building these tools are not running a comfortable business. They are running a multibillion-dollar bet that the economics eventually work. Internal OpenAI documents reported by Fortune show a $5 billion net loss in 2024, over $13.8 million burned daily in early 2025, and a projected $74 billion operating loss in 2028, against fast-growing revenue (projected $13.1 billion for 2025, up 118%). Profitability is targeted at 2029-2030, with a cumulative cash need through 2029 estimated at $115 billion. OpenAI is private, so these are leaked projections, not audited filings.
The build-out is the Stargate Project: $500 billion over four years, $100 billion immediately for Texas data centers, partners including SoftBank, Oracle, MGX, NVIDIA, Microsoft, Arm, and Cisco, roughly 7 gigawatts of compute. NVIDIA’s H100s run $25,000-$40,000 each, B200s $30,000-$50,000, at roughly 84% gross margin. Hyperscaler capex from Epoch AI, via SEC filings, went from ~$162 billion in 2022 to ~$443 billion in 2025, with ~$602 billion projected for 2026 and roughly 75% of that aimed at AI. Meta alone guided to $125-145 billion for 2026, a number that knocked 15% off its stock the day it was announced. That is the market reacting to one company’s annual AI budget. The environmental bill runs in parallel: MIT research puts data-center cooling at roughly two liters of water per kilowatt-hour, and AI was 15-20% of global data-center electricity in 2024. For contrast, Anthropic reports a $30 billion annualized run rate (April 2026, up from $87 million in January 2024) while spending roughly four times less than OpenAI on training. The headline either way: frontier infrastructure is staggeringly expensive, and the people writing the checks are not funding access programs out of goodwill.
Who benefits most from AI’s wealth concentration?
These numbers are big enough that reciting them risks sounding like hyperbole, but they are documented, and they matter for the question of who the AI economy is actually for. OpenAI closed a $122 billion round in March 2026 at an $852 billion post-money valuation, up from $28 billion in April 2023, running at $2 billion a month in revenue. Anthropic hit a $380 billion valuation in February 2026; CEO Dario Amodei and his sister Daniela have pledged to give away 80% of their wealth, stating publicly that “the thing to worry about is a level of wealth concentration that will break society.”
The Sam Altman detail is the one that sticks. His OpenAI salary is $76,001 and he holds zero equity; his estimated $2 billion comes from early bets on Reddit, Stripe, Airbnb, Helion, and 400-plus others. The person setting AI access terms earns $76,000 from the company doing it; the wealth flows through the ecosystem around it. Meanwhile Microsoft’s Satya Nadella took $96.5 million in FY2025, Google’s Sundar Pichai $10.73 million in 2024 plus a performance package worth up to $692 million, and Mark Zuckerberg’s net worth sits near $226 billion while Meta authorizes its $125-145 billion 2026 capex. The industry concentrates enormous wealth in a few hands and makes frontier capability more widely available at the same time. Both are true. Whether one offsets the other depends entirely on which comparison you think counts.
Is the access gap actually closing?
Genuinely, yes, in places. Stanford HAI’s 2025 AI Index documents a 280-fold drop in inference cost for GPT-3.5-level performance between late 2022 and late 2024, and Epoch AI puts GPT-4-class capability at ~$37.50 per million tokens in March 2023 and ~$0.40 by mid-2025: a 99.7% decline in two and a half years. Open-weight models moved from second-tier to frontier-competitive: DeepSeek-V3 is a 671B-parameter MoE at far lower training cost, and Kimi K2.6 reaches top-10 on graduate-level science at $0.95 per million input tokens. Not browser-runnable on a free plan, but self-hostable without paying OpenAI or Anthropic. The moat narrowed.
Access programs are real too. OpenAI’s ChatGPT for Teachers launched November 2025, free to verified US K-12 teachers through June 2027 with normally-paid features and no training on teacher data (Dallas ISD, Fairfax County, Houston ISD and others participating; weekly users reported saving 5.9 hours a week). Perplexity offers students Pro at $10/month, Google added Veo 3.1 and Deep Research to a $19.99/month tier, and OpenAI introduced an $8/month Go tier with previously Plus-only features. The cost curve is real and fast. It does not fully close today’s gap, but the direction is not in doubt.
What does the evidence really show about AI inequality?
The accurate summary is more specific than “AI helps the rich and hurts everyone else,” and more interesting.
Inside organizations with uniform access, AI narrows the performance gap, with the NBER customer-support result (34% for novices, near-zero for experts) the cleanest demonstration. That finding is underreported precisely because it complicates the tidy inequality story. Across organizations and income brackets, access is not uniform: the Anthropic Index shows 1.8% more adoption per 1% of state income, and Pew’s education gap (28% vs. 16%) rhymes with it. The equalizing effect simply cannot operate on people who do not have the tool. Infrastructure gaps come first: OECD reports 93% internet adoption in high-income countries against 27% in low-income ones, and the Federal Reserve broadband gap is the domestic version of the same problem. The teen pattern is what structural inequality looks like in miniature: the less-resourced group has access, but worse access, and the capability gap is the part that matters.
What is being amplified is the productivity of people who already have the tools, the skills, and the slack to experiment; the BCG/HBS work showed a 40% boost for skilled consultants on suitable tasks, compounding an existing advantage. What is not being captured are the larger gains that NBER and Noy/Zhang suggest lower-skilled workers and low-connectivity regions would see if access were equal. The 99.7% cost decline, the open-weight ecosystem, the teacher and student programs are a real countertrend, not marketing. They just run on a slower clock than the benefit concentration does. None of this looks like deliberate exclusion; it looks like ordinary market dynamics, where early adopters pay for capabilities that eventually commoditize, playing out in a domain where the capability curve moves faster than the access curve. The data describes the current distance between those two curves. It does not tell us whether it closes before it does lasting damage to who gets to participate in a knowledge economy increasingly run by these tools. That part is still open.