The Token Tollbooth: Where AI's Value Settles When the Waves End

The Framework: Three Phases of the Token Economy

Phase 1: The Capability Race (2023 through roughly 2026)

The opening phase of commercial AI was about existence proofs. Can a model write code, pass the bar, read a radiograph, run a browser? Value accrued to whoever had the frontier model, and buyers paid frontier prices (in the range of $30-75 per million output tokens when GPT-4-class models launched) and tolerated multi-second latencies, because the alternative to a slow, expensive answer was no answer at all.

In our view, Phase 1 is ending not because capability stopped mattering but because it stopped being scarce. Frontier-class performance on most commercial tasks is now available from at least four model families, several of them open-weight. When four suppliers can all do the job, the job stops commanding a scarcity premium, and competition moves downstream.

Phase 2: Token Industrialization (now)

The economy has begun buying intelligence the way it buys electricity: metered, by the unit, in compounding volume. The disclosed and estimated datapoints, as of July 2026:

Three structural features of this phase carry the thesis:

Volume is outrunning deflation. Token prices have collapsed by orders of magnitude and revenue is rising anyway, because agentic workloads multiply consumption per task by 100-1,000x: an agent chains dozens of calls, a reasoning model burns thinking tokens, a coding agent iterates all night. This is the classic Jevons pattern, in which each efficiency gain expands the addressable workload faster than it shrinks unit revenue.

The production function is tokens per second per dollar. Once intelligence is metered, an operator's economics reduce to a manufacturing question: how many sellable tokens does a dollar of installed capital produce per unit of time, at what utilization, at what power cost? That question rewards exactly what the hyperscalers have: scale purchasing, custom silicon, power contracts, demand aggregation that smooths utilization, and a decade of datacenter operating discipline.

Speed has started to price. Dedicated inference silicon vendors have published benchmarks showing large open models served at 2,000-3,000 tokens per second (Cerebras) and roughly 500-750 (Groq), which those companies present as 10-20x typical GPU serving baselines, with first-token latency below 150 milliseconds against 400-600 milliseconds for common GPU serving, per their published figures. These insurgents are winning latency-sensitive agentic workloads and forcing price-performance transparency onto the market. In our view the insurgents validate the axis; the incumbents own the balance sheets to industrialize it.

Phase 3: The Distribution Endgame (timing unknowable, logic already visible)

Extend the lines. If model capability keeps converging at the frontier, and especially if anything resembling AGI is achieved by any lab, the scarce thing stops being intelligence and becomes delivery: latency, throughput, reliability, cost, compliance, proximity, and trust. Every endgame scenario we can construct terminates in the same place: the winner is whoever can serve the smartest available system to billions of users and millions of enterprises fastest and most accurately, with the regulatory standing and the balance sheet to be trusted with it. Electricity is the precedent we reach for: generation commoditized, and the durable economics accrued to the grid. In that scenario, in our opinion, the hyperscalers are the grid.

Speed Is Becoming Intelligence

The deepest change hiding inside the reasoning-model era is that inference speed and answer quality have become fungible. A reasoning model's accuracy scales with the compute it spends thinking. Hold latency and price constant, and the operator with 10x token throughput can afford 10x the reasoning per query: more search, more verification, more drafts discarded. The fastest infrastructure does not just answer sooner; it answers better at the same wall-clock time. Speed converts to accuracy at a known exchange rate, and that exchange rate is set by infrastructure.

Latency budgets define markets. Real-time agentic workflows (voice, trading, operations, robotics, security) are addressable below latency thresholds and not above them. Each order-of-magnitude speed gain opens product categories that slower serving cannot enter. The speed leaders are not winning share of a fixed market; they are expanding the market boundary.

Custom silicon is a strategy, not a cost line. The hyperscalers' in-house chip programs (Google's TPU, Amazon's Trainium, Microsoft's Maia, Meta's MTIA) are, in our reading, a bet that the tokens-per-second-per-dollar curve is too strategically important to rent from a merchant supplier whose reported gross margins have run near 75%. Google appears the furthest along, serving its frontier models on its own silicon at global scale and marketing Gemini on speed, which we consider a material and underappreciated element of the Alphabet story.

The speed insurgents look to us like validators, not category killers. Cerebras and Groq have demonstrated that the speed axis matters. Converting wafer-scale or LPU silicon into planetary-scale serving, however, requires exactly the capital, power, and distribution that the hyperscalers own. The likeliest end-state, in our opinion, is absorption: insurgent architectures reaching scale inside or alongside hyperscaler fleets through partnership or acquisition.

Power is the floor under speed. Fast tokens are energy-dense tokens. The nuclear and behind-the-meter power procurement announced across 2024-2026 (Meta's agreement with Oklo for up to 1.2 GW, Microsoft's Three Mile Island restart deal, Google's agreements with Kairos Power and Commonwealth Fusion Systems, Amazon's X-energy partnership, all per company announcements) is the bottom layer of the same thesis: platforms locking in the physical inputs of token production for the 2030s.

Why We Think the Hyperscalers Win: Six Structural Moats

1. The capital moat. Capital spending near $725 billion in 2026, with street estimates above $1 trillion for 2027, is a game a handful of entities on earth can play. The number is itself the barrier: a challenger that could finance a competing token factory at that scale would be, definitionally, a hyperscaler.

2. The flywheel. Capex buys capacity; capacity produces tokens; tokens produce revenue and operating cash flow; cash flow funds the next capex round. Each hyperscaler runs this loop from an existing cash-generative core (search, advertising, commerce, enterprise software), which means the flywheel survives drawdowns that would sink a standalone challenger.

3. Model integration. Each platform owns or holds privileged access to a frontier lab: Microsoft with OpenAI, Google with DeepMind, Amazon with Anthropic, Meta with Llama and its successor efforts. In the scenario where the model layer captures the value, the hyperscalers hold large stakes in the model layer too. The thesis is hedged at the layer above it.

4. Silicon verticalization. Owning the tokens-per-second-per-dollar curve rather than renting it. The spread between merchant GPU pricing and owned-silicon serving cost is a structural margin source that widens as internal workloads scale.

5. Distribution and data gravity. Billions of consumer endpoints, millions of enterprise contracts, the compliance surface (data sovereignty, security clearances, industry certifications), and the workload gravity of existing clouds. Intelligence tends to be bought where the data already lives. In the endgame scenario, trust and regulatory standing become moats that, in our view, favor institutions of this scale.

6. The power moat. Contracted gigawatts, behind-the-meter deals, and nuclear power purchase agreements signed across 2024-2026 lock in physical capacity through the 2030s, while grid interconnection queues have been reported at multi-year waits (commonly cited at 4-7 years in US markets). We consider the energy layer the least-appreciated barrier to entry in the stack.

The Four Platforms, Briefly

Without ratings or rankings, here is how each name expresses the framework. Microsoft pairs Azure with its OpenAI relationship and one of the broadest commercial distribution surfaces in software: the enterprise expression of tokens as a utility. Alphabet runs a vertically integrated stack from TPU silicon through its own frontier models to consumer products with billions of users, and its 3.2 quadrillion tokens-per-month disclosure is, in our opinion, the best single piece of public evidence of Phase 2 scale. Amazon combines AWS as a default enterprise substrate, its Anthropic relationship, Trainium silicon, and a neutral model marketplace in Bedrock that pays off if the model layer fragments. Meta is the unconventional case: it does not sell tokens, it consumes them into advertising workloads with strong internal returns, while holding an enormous consumer distribution surface and the open-weight model ecosystem.

The Capex Question, Taken Seriously

The bear case that matters is arithmetic. Roughly $1.7 trillion of combined 2026-2027 hyperscaler capital spending, on street estimates, must eventually earn its cost of capital through token-era revenue, against depreciation schedules of 4-6 years on silicon that improves 2-3x per generation. If token monetization stalls while depreciation compounds, hyperscaler earnings compress violently, and the 2000-era dark-fiber analogy stops being a straw man. Anyone who tells you this risk is small is not doing arithmetic.

Three observations keep us constructive, each stated as our reading of the evidence rather than as fact:

The revenue is showing up. Disclosed AI-related revenue at the large clouds has been growing rapidly (company disclosures across 2025-2026 earnings cycles have shown AI segments and run-rates growing far faster than the businesses around them), Google's token volume grew roughly 7x in a year per its own disclosure, and enterprise "token shock" is, at bottom, a complaint about bills that are being paid. Demand does not look like the missing variable.

The flywheel self-corrects. These are cash-generative businesses that demonstrated in 2022-2023 that they will cut capital spending within a couple of quarters when returns disappoint. Capex is a throttle, not a ratchet, and management teams have shown the will to use it.

The alternative is worse, from each board's chair. For any single hyperscaler, underspending while a peer achieves a token-economics or capability breakout risks an existential error; overspending into soft demand is a recoverable margin error. Game theory suggests they will all spend to the edge of what cash flow permits, which is what a tollbooth owner would rationally do if it expects the road to carry much of the world's traffic.

The Bear Case, Steelmanned

Beyond the capex gap, the serious objections to this framework, stated as strongly as we can state them:

Model-layer capture. The frontier labs, not the clouds, could take the margin, leaving hyperscalers as commodity landlords. Our response is that the hyperscalers hold equity stakes in and compute contracts with those same labs, so the thesis is partially hedged; but revenue-share terms in each new lab-cloud deal are worth watching, and the hedge is imperfect.

Efficiency collapse. Algorithmic gains could cut tokens-per-task faster than demand grows; the DeepSeek episode of early 2025 previewed how such a shock trades. To date, volume growth has outrun price deflation, but that is an observation about the past, not a law.

A standalone winner at the frontier. A single lab could reach a decisive capability lead and vertically integrate its own distribution. We think distribution takes years to build even with a superior model, and the labs currently rent their compute from, and share ownership with, the platforms; but the scenario is not impossible, and lab self-build datacenter announcements are the tell.

Neocloud margin competition. CoreWeave, Nebius, Oracle, and sovereign clouds can compress serving margins at the rental layer. We believe hyperscaler value sits in the integrated stack rather than bare compute, but the pricing pressure is real and visible in GPU rental spot rates.

Antitrust and sovereignty. Concentration of frontier AI in four platforms invites structural remedies, and national markets could fragment. We regard this as the genuine tail risk to the endgame claim: it is hard to model and it is not small. US and EU platform-AI proceedings are the watch item.

The energy wall. Grid limits could cap deployment for everyone. Our reading is that this constraint currently favors incumbents who pre-purchased power, but it caps the pace of the whole buildout.

What Would Prove Us Wrong

We publish break conditions on every framework we maintain. For the Token Tollbooth, the conditions are:

1. The Jevons ratio turning. If token demand growth falls below token price deflation for four consecutive quarters, the industrialization phase is not funding itself, aggregate AI revenue shrinks while capex is still landing, and the core of this framework breaks. We track token volume growth against token price deflation as, in our view, the single most important series in AI economics.

2. Distribution without a landlord. If a frontier lab demonstrably serves planetary-scale distribution without a hyperscaler partner (its own datacenters, its own power, its own enterprise and consumer reach), the tollbooth claim fails on its face. Lab self-build announcements are the early warning.

3. Structural separation. If antitrust or sovereignty remedies force a durable separation of model, cloud, and distribution, the integrated-stack advantage this thesis rests on is dismantled by rule rather than by competition.

4. The capex throttle failing. If AI revenue growth decelerates materially while capital spending guidance keeps rising for several consecutive quarters, management teams are no longer treating capex as a throttle, and the self-correcting flywheel argument in this note stops being true.

As of this writing, none of the four is in evidence. That statement is a snapshot, not a promise, and we revisit it quarterly.

Our View, Stated as Opinion

Three Douglas Research holds a constructive long-term view of the hyperscale platform complex (Microsoft, Alphabet, Amazon, Meta Platforms) on a multi-year horizon, for the reasons in this framework. We expect the group to remain volatile: a capital cycle of this size guarantees periodic drawdowns whenever the market panics about capex, and June 2026 offered a preview. We also note that in every scenario we can construct short of AI demand outright collapsing, the entities that own the capital, the power, the silicon, stakes in the models, and the distribution take a large share of whatever value exists. That is our opinion, not a recommendation, and the break conditions above describe exactly how we would come to abandon it. Our full reasoning and our company-level work are available to members.