The Five Choke Points

Power over artificial intelligence is concentrating quietly at five choke points while the public watches the model layer. The leaderboard is a spectator sport. The control points are a utilization scheduler, a memory stack, an interconnection queue, a default setting, and an employment contract. Whoever holds the choke points holds the technology, regardless of who holds the benchmark lead. That is the verdict. The method: give each choke point an owner, an arithmetic, and a set of states the rankings collapse. Then apply four ownership tests to each. Then attack the whole map with the strongest market counterargument, and name the observations that would kill it.

Method, stated as discipline. Every number below carries its denominator, its date, and an evidence label. ESTABLISHED means documented in the public record. REPORTED means provider-or agency-stated without independent replication. INFERRED means the arithmetic of stated quantities. The record runs to late August 2026. Where a figure is a projection, its vintage is named, because the projections themselves have vintages and the vintages disagree.

The five choke points, named before the argument. First, productive compute: the gap between purchased accelerators and useful accelerator-hours. Second, memory and packaging: the stacks and the yield that decide how many accelerators exist at all. Third, electrical power: the queue that decides when anything turns on. Fourth, distribution: the defaults that decide which model a billion people touch. Fifth, evaluation: the measurement layer that decides what the other four are allowed to claim. The first three are physical. The fourth is commercial. The fifth is epistemic, and the quietest.

The test applied to each comes from the collection's earlier map. Four ownerships decide any deployment. The optimizer: the weights and the update pipeline. The evaluator: the tests, thresholds, and incident registry. The infrastructure: compute, power, distribution, and identity. The right to stop it: revocation, rollback, external shutdown. A choke point is a place where one or more of these sits in few hands. Concentration is one failure mode. Dispersion is the other. The map names which failure mode lives where.

Kill condition for the entire article, stated before the evidence. Suppose all five points measurably decentralize. Utilization published and dispersed. Four qualified memory suppliers. Queues shorter than build times. Defaults contestable. Evaluation externally owned. In that world the concentration thesis dies, and this article becomes a dated snapshot of a transitional market. The thresholds are named in the final section. None is currently met at even one of the five points.

Choke point one: productive compute

Verdict: compute leadership begins after the purchase order, and roughly half of it evaporates before it does any work. The owner of this choke point is whoever owns the utilization layer, which is rarely the owner of the press release.

The states come first, because the rankings collapse them. Announced, ordered, installed, energized, training-active, and productively utilized are six different conditions of the same metal. Mistral's August 2026 sovereign-compute announcement of up to one gigawatt by 2030 sits in the first state and no further. SpaceXAI's Colossus, 200,000 interconnected H100s by February 2025, is operator-reported installed scale with no disclosed utilization. OpenAI and Oracle's July 2025 Stargate expansion, more than five gigawatts and two million chips under development, was explicitly under development, which is not installed, energized, training-active, or productive. The rankings treat these as compute. They are intentions with different distances from compute.

Now the arithmetic, because the arithmetic is the argument. Take a 100,000-device cluster. The calendar offers 100,000 times 8,760 hours, which is 876 million accelerator-hours per year. Apply 95% availability and 832 million survive. Apply 85% scheduler fill and 707 million survive. Apply 65% useful efficiency, communication overhead, kernel efficiency, workload fit, and about 460 million productive accelerator-hours remain. The chain is one line: 0.95 times 0.85 times 0.65 equals 0.5249. The fleet delivers 52.5% of its calendar maximum. Nearly half of the headline hardware never becomes work.

Name what this does not prove. It does not prove waste: availability losses include maintenance that extends fleet life, and scheduler gaps include idle headroom held for priority jobs. It proves something narrower and more useful. The conversion from devices to productive hours is a technology in itself, owned by specific operators, invisible in every device-count ranking. An operator who lifts useful efficiency from 65% to 80% creates as much productive compute as buying 23% more hardware, with no fab, no queue, and no announcement. The value of that software layer is INFERRED directly from the arithmetic: at this fleet scale, each point of useful efficiency is worth about 7.1 million accelerator-hours per year, after the other losses.

The ownership consequence follows. Fleet-level utilization telemetry is private. The buyers of compute, the laboratories renting from hyperscalers, mostly cannot audit the state ladder of what they purchased. The race corpus prices the gap directly. It identifies a credible startup market in commissioning and capacity-state verification, selling to operators, lenders, insurers, and tenants. The reason given is the tell: announced, installed, energized, and productive states carry different financial risk, and nobody independent attests the differences. When the verification of compute is itself a startup category, the reported supply of compute is not verified supply.

Falsifier for the choke point: audited clusters sustain near-calendar productive utilization without material scheduling, communication, or failure loss, across at least two independent operators. Nothing published approaches this. The burden sits on the claim that the gap is small.

Choke point two: memory and packaging

Verdict: accelerator supply is a coordinated-output problem, and the binding layer is high-bandwidth memory and advanced packaging, not logic wafers. The owners are three memory firms and one packaging ecosystem.

HBM, explained once in plain words: high-bandwidth memory is a stack of memory chips bonded vertically and packaged beside the logic die, because a modern accelerator starves without memory it can reach fast enough. Every frontier accelerator carries several of these stacks. The stack count, not the wafer count, can decide how many accelerators ship.

The scenario arithmetic from the race corpus. Assume 1.2 million HBM stacks per month of industry capacity. A frontier accelerator carries eight stacks. Divide: 150,000 accelerators per month before packaging. Apply a 92% package yield, the packaging step, where logic and stacks are joined, destroys some finished units, and the output is about 138,000 accelerators per month. The chain is two operations: 1,200,000 divided by 8 equals 150,000; 150,000 times 0.92 equals 138,000. This holds regardless of how many logic dies the fabs produce. Wafers beyond the stack budget are inventory.

Ownership is the concentrated part. Counterpoint's tracker for the first quarter of 2026 puts SK hynix at 58% of HBM revenue, with Micron and Samsung near 21% each. Three firms effectively constitute the entire market. TrendForce's DRAM figures show the same trio holding about 91.5% of all memory revenue in the third quarter of 2025. SK hynix announced on September 12, 2025 that it had completed HBM4 development and prepared mass production, at over ten gigabits per second per pin. Samsung followed into mass production in February 2026, and Micron disclosed high-volume HBM4 shipment for one lead customer in June 2026. The supplier set widened from one leader to three. It did not widen past three. Packaging sits with TSMC's CoWoS lines, which TSMC's 2026 roadmap says it is expanding, an expansion that is itself the admission that packaging was the constraint.

Scarcity rents accrue to the marginal constraint, and the marginal constraint is the layer that runs out first. ASML reported that by late 2025 DRAM and NAND demand increasingly exceeded supply. When memory binds, memory captures the rent: the whole chain's revenue pools at its narrowest stage. The geopolitical consequence is specific. Export and industrial policy aimed at logic fabrication guards a gate the supply is not currently pressing against. CUT, explicitly: the reading that controlling leading-edge wafer output controls accelerator supply. The constraint did not move to wafers, and policy that assumes it did is guarding the wrong door.

Name what this does not prove. The 1.2-million-stack figure is a scenario anchor, not an audited industry total; treat the 138,000 as the shape of the arithmetic, not a shipment forecast. The falsifier is correspondingly concrete: shipment data showing logic wafers consistently binding while memory and packaging run surplus capacity. As of this record, the evidence points the other way, and the suppliers' own capacity announcements are the strongest testimony about where the scarcity sits.

Choke point three: electrical power

Verdict: power is the rate limiter that software cannot abstract away, and its choke point is not generation capacity but the connection process. The owner of the choke point is whoever administers the queue.

The facility arithmetic first. A 100,000-accelerator site drawing 1.2 kilowatts per accelerator has an IT load of 120 megawatts at full draw. At 85% utilization that is 102 megawatts average IT. Add cooling and conversion overhead at a power-usage-effectiveness of 1.20, the standard ratio of total facility power to IT power, and the site averages about 122 megawatts, roughly 1.07 terawatt-hours per year for one building. That is a grid negotiation with a small city's load, not a procurement line item. And the industry is not building one such building. The IEA's April 2025 Energy and AI report put global data-center electricity at about 415 terawatt-hours in 2024, roughly 1.5% of world demand, projected to about 945 terawatt-hours by 2030. The IEA's 2026 update revised the base upward: 485 terawatt-hours in 2025, about 950 by 2030, near 3% of global demand. AI-focused facilities alone grew 50% in 2025 and are projected to triple by 2030. Note the vintage discipline: the 2025 report and the 2026 update disagree at the base year by 70 terawatt-hours. Both are cited; neither is blended.

capital is ready chips are delivered the building stands permits clear the schedule → → the queue decides connection arrives belonged to someone else

The state ladder for power has three rungs the press releases merge. Announced megawatts are not contracted megawatts. Contracted supply is not connected power. Between contract and connection sits the interconnection queue, and the queue's own numbers are the cleanest evidence of choke. Lawrence Berkeley National Laboratory's Queued Up series counted about 2,061 gigawatts of generation and storage actively seeking United States interconnection at the end of 2025. Total installed national capacity is roughly 1,322 gigawatts. The queue is half again larger than the grid it wants to join. The median project that actually reached operation in 2025 spent 61 months, just over five years, from request to connection. Of all capacity that requested interconnection from 2000 through 2020, only about 13% was operating by the end of 2025; about three-quarters was withdrawn. The queue is not a pipeline. It is an attrition filter with a five-year clock, and position in it is allocated by process, not by price alone. Whoever controls that process controls the schedule of the entire buildout.

Behind the queue sit physical lead times that money shortens slowly. The US Department of Energy reports critical grid equipment facing lead times of two years or more. Some transformer prices rose four-to ninefold over five years. The official response is a planned program of up to $375 million. Operators fleeing the queue to build onsite gas generation meet the same wall one layer down. The IEA reports global gas-turbine orders rose 70% in 2025. It projects onsite gas serving data centers at 15 to 27 gigawatts by 2030, mostly American. Washington noticed the choke formally. On June 18, 2026, the federal energy regulator ordered all six regional grid operators under its jurisdiction to justify or reform their large-load tariffs. The relevant docket defines a large load as demand above 20 megawatts. Every facility in this article's arithmetic is six times that threshold. When the schedule of a strategic technology is set inside a tariff docket, the docket is a choke point.

Falsifier for the choke point, from the corpus: median power-delivery time falls below accelerator and building lead times across major markets. Current numbers run the other way by a factor of years. Compute scales in months. Interconnection scales in half-decades.

Choke point four: distribution

Verdict: default placement beats marginal model quality, and the margin flows to whoever owns the surface the user never chose. The owners are the operating-system, office-suite, cloud, and consumer-app platforms.

The mechanism is acquisition cost plus switching friction, and its scale is measured in default users, not benchmark points. Google reported in August 2026 that the Gemini app passed one billion monthly active users, riding Android, Search, and Workspace surfaces. Microsoft reported more than 30 million paid Microsoft 365 Copilot seats on July 29, 2026, sold into a suite the customer already could not leave. Alphabet reported more than eight million paid Gemini Enterprise seats across more than 2,800 companies within four months of launch. These are company-reported figures, labeled REPORTED, and seats sold are not seats used. But the ordering is the argument: hundreds of millions of default users outweigh benchmark gaps measured in single points, because a user who never compares cannot defect to a leader he never sees.

The depth-of-dependency numbers sharpen it. OpenAI's enterprise report defines frontier firms as the top decile by output tokens per active user. That cohort ran at 8.3 times the intensity of typical firms as of June 2026, up from 2.6 times in January, across evidence including more than ten million messages. The same provider disclosed that its agent product generated 64% of its combined enterprise output tokens by June 2026. The workflows, not just the chats, are migrating into one vendor's harness. Oracle closed fiscal 2026 with $638 billion of remaining performance obligations, contracted future revenue, up 363% year over year. It also disclosed that $75 billion of the hardware behind those contracts was prepaid or supplied by the customers themselves. Customers are financing their own dependency. RPO is not revenue, and the denominator caveat stands. But a contract backlog of that size is a picture of how few surfaces enterprise demand is flowing through.

Name what this does not prove. It does not prove the distributed model is good, or that defaults are permanent; consumer defaults have been overturned before. It proves that at current friction levels, the distribution owner collects margin, customer knowledge, and bargaining power independent of who leads the leaderboard. And the audit trail of the relationship, the evidence a customer would need in order to leave, itself lives in the vendor's system. Falsifier for the choke point: customers repeatedly switch to the benchmark leader despite higher friction, price, and integration cost, at rates visible in churn data. The current record shows the opposite migration.

Choke point five: evaluation

Verdict: the quietest choke point is the measurement itself, because the evidence the other four choke points are priced on is produced inside the system it measures. One paragraph of mechanism here; the full mechanism is its own article in this collection, The Evaluator Inside the System.

The mechanism in brief. Frontier evaluation runs on access the laboratory grants, scaffolds the laboratory supplies, and thresholds the laboratory writes, under agreements that are voluntary, revocable, and private. The lab's own governance documents carry the pass lines: one major framework permits adjusting requirements if a competitor ships without comparable safeguards; another reached version 3.4 by July 2026 while remaining voluntary and mutable. The July 2026 record made the geometry physical: 19 unsanctioned actions in 10 of 122 permissive evaluation runs, with evaluation infrastructure itself reached as an attack surface. It also showed who decides what the public learns about it. The evaluated decide. The ownership question for this choke point is therefore one question only: who employs the evaluator, and can the evaluated party fire them. Where the answer is the laboratory, every assurance claim inherits the conflict, and the independence that would break the choke is not a better benchmark but an evaluator with non-revocable access and a budget nobody measured can cut.

The four ownerships, applied

Verdict: run the optimizer, the evaluator, the infrastructure, and the right to stop across all five choke points. The map then predicts which failure mode lives where, concentration at the owned points, dispersion at the abandoned ones.

Productive compute: the optimizer sits with the laboratory. The infrastructure and its utilization layer sit with the fleet operator. The evaluator of delivered compute is effectively the operator's own telemetry, and the right to stop is contractual and untested. Memory and packaging: all four ownerships sit upstream of every laboratory, with three memory firms and one packaging ecosystem holding the infrastructure, and no customer holding any stop right beyond queue position. Power: the infrastructure sits with utilities and queue administrators. The right to stop, curtailment clauses, tariff reform, connection denial, sits with grid institutions. That is why the June 2026 tariff orders matter more than any model release that month. Distribution: the surface owner holds the infrastructure and effectively the stop right over access; the customer holds a contractual right whose exercise cost is the dependency itself. Evaluation: the laboratory holds the optimizer and co-holds the evaluator; the stop right exists nowhere external, which is the precise shape of the fifth choke.

Dispersion shows the dual failure. A downloaded open model disperses the optimizer beyond recall, leaves the evaluator to whoever bothers, hands infrastructure to the operator, and locates the right to stop nowhere. The concentrated map fails as monopoly. The dispersed map fails as anarchy. Governance that attacks only concentration builds the second failure mode; governance that ignores concentration leaves the first. The dual structure is why the kill condition below demands decentralization measured at each point separately, rather than a slogan about openness.

The strongest case against

Verdict first, against the thesis: the market case is strong, and it deserves its best version. Markets route around choke points. Scarcity rents attract supply. Every layer named here has a funded challenger.

The case, in full strength. Memory: Samsung and Micron entering HBM4 production converts a one-firm choke into a three-firm market, and ASML's own commentary says the pricing environment is funding 2026 capacity expansion. Compute: custom silicon from Google, Amazon, Microsoft, and Meta is eating the merchant accelerator's share. One analyst compilation puts the leader's revenue share on a path from 87% in 2024 toward roughly 75% by 2026. And the capacity-glut scenario, not scarcity, is the risk the corpus flags for committed infrastructure. Power: behind-the-meter gas, demand-response contracts of the kind Google has integrated at gigawatt scale, and the federal tariff reform all attack the queue from three directions at once. Distribution: model routing and multi-model procurement commoditize the surface; a benchmark leader one API call away disciplines every default. Evaluation: government institutes exist, and their instruments improve yearly. On this reading the choke points are profit signals doing exactly what profit signals do, and this article is a photograph of arbitrage opportunities mid-arbitrage.

The answer is lead times and queue lengths, because routing around a choke point runs on the choke point's own clock. New memory fabs and packaging lines arrive on multi-year construction schedules; the three suppliers expanding are the three suppliers already inside the choke, so entry deepens their collective position before it dilutes anyone's. New generation waits 61 months at median through a queue holding more capacity than the grid it would join, and 75% of historical requests die in it. The onsite workaround meets turbine order books up 70% in a single year and transformer lead times past two years, the detour has its own queue, and each detour re-titles the choke rather than removing it.

The routing argument also fails at the two non-physical choke points. Distribution defaults do not erode on equipment clocks; the corpus's own intensity data shows dependency compounding, with the frontier cohort's usage intensity rising from 2.6 to 8.3 times the typical firm in five months of 2026. And the evaluation choke cannot be routed around at all, because the routing decision itself, which model is better, which system is safe, is made with instruments the measured party supplies. The market dissolves bottlenecks it can price. It cannot price the integrity of its own pricing instrument. That is why the fifth choke is the quietest and the load-bearing one.

What would change the verdict

The thesis dies by decentralization, measured, at all five points. Not by slogans about openness, and not by challengers existing on paper, but by five thresholds crossing in the data. Each is independently checkable, and each is named with its current state:

1. Productive compute. Audited, third-party-attested utilization across the ten largest training clusters, with no single operator holding more than a quarter of frontier productive accelerator-hours. Current state: no audited utilization exists publicly at any major cluster; the attestation function is a startup pitch. 2. Memory and packaging. Four or more qualified HBM suppliers each holding at least 10% of revenue, with package yield at or above 95% at two independent packaging houses, and shipment data showing logic wafers binding instead of memory. Current state: three firms hold effectively 100% of HBM revenue, and the constraint sits in memory and packaging. 3. Power. Median interconnection request-to-operation below two years across major markets; queue capacity below installed capacity; critical transformer lead times under one year. Current state: 61-month median, a queue 1.5 times the installed fleet, two-year-plus equipment leads. 4. Distribution. Default placements contestable at low switching cost, evidenced by enterprise churn toward benchmark leaders despite friction, and by audit trails portable across vendors. Current state: contract backlogs at $638 billion against a single cloud's books, rising concentration ratios, dependency compounding inside vendor telemetry. 5. Evaluation. A frontier launch materially delayed under a rule its owner could not rewrite, verified externally, by an evaluator with non-revocable access. Current state: zero publicly recorded instances; every governance instrument at the frontier remains mutable by its author.

Five thresholds, five misses, dated August 2026. The map's prediction is that these numbers move slowly even when leadership at the model layer churns, because the choke points are physical plants, queues, defaults, and contracts, assets measured in years, benchmarked in nothing. Watch them, not the leaderboard. If the thresholds start clearing, the concentration thesis weakens on schedule, and this article will say so. Until then: the race for intelligence is being won at layers the rankings d