The Scale of the Prize
Anthropic and OpenAI are on track to hit a combined $200 billion revenue run rate by the end of 2026. To put that in context: the collective Fortune 500 generates roughly $2 trillion in profit per year. These two companies alone are approaching 10% of that entire pool — while actual AI adoption across the real economy sits below 5%, concentrated almost entirely in software engineering and tech-adjacent niches. The bottleneck George identifies is not capability but capital sourcing. If enterprises are to absorb software costs at this scale, the money has to come from somewhere — which means either aggressive workforce restructuring or a rapid shift toward cheaper open-source and on-device models arriving far sooner than the market currently expects.
The Skeuomorphic Trap
Enterprise adoption is stuck in what George describes as a skeuomorphic phase — using new tools to accelerate old workflows rather than redesigning the workflows themselves. Drawing on Chris Dixon's observation that new technologies typically spend three to four years mimicking old mediums before going native, George argues most corporations are still in the documentation phase: manually converting institutional knowledge into Markdown files to maximise LLM context capture. The AI-native startups, by contrast, are bypassing keyboards entirely — researchers coordinating swarms of autonomous agents via voice. The gap between these two operating modes is widening faster than legacy enterprises can close it, and the engineers with the skills to close it are being pulled toward customer-facing product features, leaving internal automation to slower teams.
The Compressed Power Law
The numbers George and Clark cite are striking in their acceleration. From 2020 to 2024, a top 1% venture exit required reaching $10 billion. By February 2026 that floor had moved to $20 billion. By June 2026, driven by closed mega-deals, it had surged to $32 billion — with Wiz defining the baseline. The sum of every venture-backed IPO over the last six years combined is just over $1 trillion; that entire multi-year cohort will likely be smaller than the top three upcoming AI mega-IPOs alone. Clark notes that this scale creates organisational whiplash: startups are hitting big-company problems — complex infrastructure agreements, international regulatory navigation, systemic pricing overhauls — while still operating with adolescent team structures. The implication for venture firms is structural: founders now need scaled platforms over boutique seed investors.
Defensibility Has a Half-Life
George points to the Forbes AI 50 list as a proxy for how quickly moats are eroding: 40% of the definitive AI startups dropped completely off the list year-over-year. At the same time, the cost per token is falling more than 10x annually — yet total token consumption by frontier applications is outpacing that deflationary curve entirely. The long-term distribution of economic rent, George argues, depends on a single unknowable variable: whether five or more frontier labs maintain parity and compete down token prices, or whether one or two pull away and keep pricing inelastic. In the first scenario, wealth shifts to the application layer. In the second, the software layers built on top face a structural financial squeeze. Staying directly on the token path — capturing value as data is processed rather than wrapping it in static software — is the only moat with a realistic shelf life.
Why This Is Not a Bubble
The dot-com crash was driven by speculative oversupply: unutilised fibre-optic cables, conceptual websites, no revenue. The current AI cycle is bound by the opposite constraint. Hyperscale data centre capacity is entirely sold out and unavailable for new deployment until late 2028 or early 2029. The US infrastructure pipeline is running a full year behind schedule. Global technology infrastructure is on track to absorb $5 trillion in capital expenditure over the next four to five years — and to justify that, the ecosystem must yield $1–2 trillion in returning revenue, a target that becomes entirely plausible if the leading model companies clear a $200 billion run rate out of the gate. The primary bottleneck to expansion, George notes, is not software architecture but local civil resistance to data centre construction — communities pushing back on water consumption and grid strain. The operators who are winning are bundling infrastructure pitches with ecological preserves, school district fibre, and local tax guarantees.