Frontier AI Is Being Enclosed and the Labs and Washington Are Both Holding the Fence
AI Datum Point
SubscribeSign in
Frontier AI Is Being Enclosed and the Labs and Washington Are Both Holding the Fence<br>Two actors, opposite directions, identical outcome: the people locked out are the same in both cases
AI Datum Point<br>Jun 25, 2026
Share
Frontier AI Is Being Enclosed and the Labs and Washington Are Both Holding the Fence
Four companies absorbed 65% of all global venture capital in Q1 2026. OpenAI, Anthropic, xAI, and Waymo collectively raised $188 billion — 80% of total global AI funding that quarter. The Anduril defence round ($5 billion) and Cognition agentic coding round ($1 billion) followed the same pattern: mega-capital flowing to companies with proven enterprise revenue traction. Every tracker covering the period confirms the same structural condition: serious money is concentrating in a small group of established players, and the funding environment for everyone else has bifurcated into seed-stage or nothing.<br>Anthropic built that trajectory on a specific customer base. The company grew from $87 million ARR in early 2024 to $47 billion by May 2026 — the fastest B2B revenue ramp in recorded history — on enterprise API contracts. Eight of the Fortune 10 are Claude customers. Over 1,000 companies spend more than $1 million annually. Anthropic monetises at roughly $211 per monthly user versus OpenAI's $25 — an 8x difference that SaaStr called "the enterprise playbook working exactly as designed." The individual researcher, the public institution, the hospital in a resource-constrained economy: they don't appear in that record at all.<br>Token pricing is the gate. Claude Opus 4.7 costs $5 per million input tokens and $25 per million output tokens. GitHub Copilot switched from flat-rate subscriptions to token-based billing on June 1; developers reported costs jumping from $29 to $750 per month as agentic workflows consumed credits at scale. TechCrunch called it the end of "the golden age" of GitHub Copilot. Average SMB spend on Anthropic runs to $24,000 a year. Enterprise spend runs to $85,000. Those numbers describe customers with procurement departments and IT budgets. They don't describe a teacher in Karachi, a clinician in Nairobi, or a first-generation student who doesn't have access to an institution willing to sign a volume contract.<br>Claude Pro runs at a flat $20 per month globally, with no purchasing power parity adjustment. OpenAI introduced a locally priced ChatGPT Go tier at the equivalent of $5 per month for Vietnam and fifteen other Asian markets. Anthropic has not. One analysis tracking AI access across Asia described Claude Pro as "the conspicuous outlier" — available, technically, but priced at a level that in Vietnam represents five to seven percent of average urban monthly income.<br>The Trump administration arrives at the same destination from the opposite direction. The Commerce Department's export directive pulled Fable 5 and Mythos 5 from every non-American user. Helen Toner of Georgetown described the practical effect: barring foreign nationals from frontier models is "essentially equivalent to preventing any company affected from doing any further AI R&D work." State authority achieves through security controls what enterprise pricing achieves through the market. The people excluded are identical in both cases.<br>Open-source models from Meta, Alibaba, DeepSeek, and MiniMax have closed the performance gap with proprietary APIs faster than most predicted. Epoch AI's research shows frontier inference prices falling by 9x to 900x per year depending on the benchmark. The MMLU performance gap between open and proprietary models narrowed from 17.5 percentage points to 0.3 in a single year. Anyone with a consumer GPU can now run models approaching frontier capability at near-zero marginal cost. If that trajectory continues, the enclosure argument weakens.<br>Gartner's response is the one that lands: "Chief Product Officers should not confuse the deflation of commodity tokens with the democratisation of frontier reasoning." Cheaper tokens don't equal cheaper outcomes when agentic workflows multiply consumption faster than unit costs fall. Microsoft cancelled Claude Code licences after its engineers burned through the entire 2026 AI budget in four months. Uber's CTO said the same. The models are cheaper per token. The capability that matters — extended agentic reasoning, long-horizon tasks, frontier cybersecurity — remains gated behind pricing structures that run at enterprise scale.<br>And open-source access still requires hardware. Running a 70B parameter model locally needs a GPU cluster costing thousands of dollars. Self-hosting frontier-capable open weights at production scale requires infrastructure that a university in Lagos or a clinic in Dhaka cannot provision. The accessibility argument for open-source holds for developers in high-income economies with existing infrastructure. For the...