The AI boom is sold as a great equalizer of tools, talent, and opportunity. The buildout tells a harder story. Economic freedom in this market is not only an open license or a cheaper chatbot trial. It means entrants can contest frontier capacity, customers can switch without punitive lock in, and productivity gains travel beyond firms already large enough to finance the buildout. On that definition, the boom is concentrating economic power faster than it spreads freedom. Everyday AI access is widening. Frontier power is clustering faster still.
That is not a claim that nobody else can use AI. It is a claim about where bargaining power is accumulating. Summer 2026 earnings locked the CapEx race into public guides that only a handful of balance sheets can stay in. Meta put 2026 capital spending at $130 billion to $145 billion. Alphabet raised its guide to $195 billion to $205 billion. Amazon management pointed to roughly $220 billion in cash CapEx on the call, after $54.2 billion of property and equipment purchases in the second quarter. Microsoft framed calendar 2026 CapEx near $175 billion after a lease accounting shift that changed the headline more than the underlying spend. Those are company figures with different CapEx definitions, not one official industry total. Together they describe a financing contest, not a level playing field. When only a few firms can fund the next increment of compute, the boom’s upside skews toward those already writing the checks.
Why the timing matters is that the same late July window put cloud and market concentration beside those CapEx guides. Synergy Research Group, vendor research rather than an SEC series, put worldwide cloud infrastructure spend at $143.4 billion in the second quarter and assigned the top three providers 67 percent of the market: Amazon Web Services 28 percent, Microsoft 20 percent, and Google Cloud 15 percent. NVIDIA’s Data Center franchise remains the clearest primary scale marker in the chip layer, with $41.1 billion of Data Center revenue in the quarter ended July 27, 2025. Bloomberg Intelligence has described NVIDIA near about 86 percent of AI accelerators in 2025 and toward a 70 to 75 percent long run share call, depending on the market denominator. Goldman Sachs Asset Management, citing S&P Dow Jones Indices, put the S&P 500 top ten at about 36.5 percent of index market capitalization as of March 31, 2026. Magnificent Seven weights have traded in a roughly 33 to 37 percent band by date and source. The pattern is consistent across layers: access tools can proliferate while investment, silicon leverage, and index weight still thicken in the same platform core. Spreading usage does not, by itself, disperse control.
The hard gate is frontier training. Epoch AI finds amortized frontier training costs rising at about 2.4 times a year, with billion dollar runs possible by 2027 if the trend holds. That is capital intensity, not speculation. The Federal Trade Commission’s January 2025 staff report on AI partnerships documented Microsoft OpenAI, Amazon Anthropic, and Alphabet Anthropic arrangements totaling more than $20 billion in cumulative investment and flagged concerns about compute access, switching costs, and information advantages. Staff concerns are not guilt findings. They still show how cloud equity, spend commitments, and model development can tighten around a few relationships, which is exactly where freedom as switching power gets tested. If customers and rivals cannot leave without losing compute, data advantages, or contractual leverage, then cheaper chatbots at the edge do little to reopen the frontier.
Productivity data keep the argument honest about what has and has not arrived. The Bureau of Labor Statistics revised second quarter 2026 nonfarm business productivity to a 1.4 percent annual rate, or 2.2 percent year over year. A Federal Reserve FEDS note dated July 17, 2026 called the moment a buildout phase, with limited signs yet of broad aggregate labor transformation. Micro gains show up in field experiments and firm surveys, including Atlanta Fed work finding positive labor productivity effects and smaller firms expecting modest AI linked employment growth. What is clearer than a national wage concentration story is concentration of CapEx, chips, cloud share, and equity weights. Treating AI as the proven cause of economy wide employment concentration would outrun the evidence. Spreading software access is not the same as dispersing frontier power, and the macro data still refuse to confuse the two.
A fair counterargument deserves weight, and it is real. Open weight ecosystems such as Meta’s Llama family, Alibaba’s Qwen, DeepSeek, Google Gemma, and Mistral have drawn huge secondary download counts, even though downloads are not unique users. GPU unit costs and some inference paths are falling, as Fed and Epoch charts show. Neoclouds including CoreWeave appear among Synergy’s high growth providers. Custom silicon and AMD gains erode pure merchant GPU dominance at the margin. Small firms can already buy useful inference without financing a frontier training run. Concentration at CapEx, chips, and index weights can coexist with democratizing inference and apps. Collapsing those layers into one monopoly story would overstate the case. Cedar’s point is narrower: spreading access does not automatically equal spreading freedom when the frontier financing gates keep narrowing.
That coexistence does not settle the institutional question. If the largest training runs, densest cloud shares, and heaviest index weights keep clustering in the same hyperscaler and platform complex, the boom redistributes opportunity unevenly even while it expands tools for many users. On the CapEx and cloud evidence of summer 2026, concentration is the clearer macro fact. Freedom at the application layer is real, and incomplete, until entrants can contest frontier capacity, customers can switch without punitive lock in, and gains travel beyond firms that can already fund the buildout. The boom will look more free only when power at the frontier becomes as contestable as the apps already are.
Sources: Meta Q2 2026 earnings (Jul 29, 2026 CapEx guide); Alphabet Q2 2026 IR and earnings call (Jul 22, 2026 CapEx guide); Amazon Q2 2026 earnings and call (Jul 30, 2026 PP&E and cash CapEx guidance); Microsoft FY26 Q4 earnings call (Jul 29, 2026 CY26 CapEx framing); Synergy Research Group cloud infrastructure (Q2 2026); NVIDIA Q2 FY2026 newsroom (Aug 27, 2025 Data Center revenue); Bloomberg Intelligence AI Accelerator Outlook (Jan 14, 2026); Goldman Sachs Asset Management / S&P Dow Jones Indices (top 10 share Mar 31, 2026); FTC Staff Report on AI Partnerships and Investments (Jan 17, 2025); Epoch AI training cost insights; Federal Reserve FEDS Note (Jul 17, 2026); BLS Productivity and Costs (Q2 2026, revised Sep 3, 2026); Atlanta Fed / related firm survey evidence on AI productivity.
