Super Intelligence | Compute, Data and Capital: Where Economic Power Could Accumulate

Economic power can accumulate around an AI system when other people need inputs, permissions or routes to customers that they cannot readily replace. The important question is therefore larger than who owns the most chips. Who can provide usable capacity at the required time? Who may use the necessary information? Who can finance the transition? And what happens when a customer tries to leave?

This guide explains those mechanisms without assuming that a single company, country or technology must win. In the eduKate series, Super Intelligence is an editorial umbrella for understanding advanced AI and its implications. Artificial superintelligence, or ASI, refers here to a hypothetical broadly superhuman capability. Evidence about present AI markets does not establish that ASI exists or determine how a future ASI economy would be organised.

The central exercise follows a fictional document service through three computing providers, one restricted dataset, a misleading backup claim and a repair that has a real cost. Every quantity required for the calculations is supplied. The exercise is an economic reasoning model, not a description of an actual supplier, a procurement recommendation or investment advice. Its purpose is to make dependence visible enough to question and improve.

1. Economic power is a relationship, not a pile of equipment

Imagine two firms with identical computers. The first rents them to customers who can move their work elsewhere overnight. The second supplies a specialised service whose customers would need months of rebuilding, permissions work and testing to replace it. The physical equipment is similar, but the bargaining relationship is different. The second supplier may have greater room to alter terms because refusing its offer is more difficult for its customers.

That is the practical meaning of economic power used in this article: an ability to influence the terms on which other participants obtain something they need. Influence can concern price, priority, quality, contractual conditions, compatibility or the direction of development. It is not unlimited control. Customers can reduce demand, redesign tasks, organise alternatives or leave the activity altogether. Suppliers can also depend on their customers, their financiers and their own upstream inputs.

The relationship must be specified. Power over a small customer’s urgent workload is different from power over all AI development. A supplier might be essential for one language, location or security requirement while facing strong alternatives elsewhere. A temporary shortage during a deployment window can also create leverage without establishing a durable position. The relevant question is always power over whom, for what service, over which period, with what realistic alternatives.

Large revenue, high profit and fast growth are clues that may deserve investigation, but they are not interchangeable measures of this relationship. Revenue can reflect serving many customers. Profit can reflect innovation, risk, scarcity, efficient operation or stronger bargaining power. Separating those explanations requires evidence about costs, quality, entry and customer substitution. Calling every successful supplier a monopoly would replace the analysis with a label.

Similarly, this guide does not make a legal finding about any firm. Competition law uses jurisdiction-specific rules and evidence. A classroom calculation of supplier dependence is not a determination of a relevant market, dominance, unlawful conduct or an appropriate remedy. The useful first step is more modest and more transferable: identify the indispensable service and examine the practical consequence of losing access to it.

This perspective also keeps benefits in view. A capable supplier may create a service that previously did not exist. Integration can make that service cheaper and more reliable. The economic question is not whether dependence is always bad. It is whether the benefits remain available on reasonable terms, whether alternatives can develop, and whether users retain workable ways to adapt when their needs or the supplier’s behaviour change.

Back to contents · Next: 2. Trace ownership, access, use rights and influence separately

2. Trace ownership, access, use rights and influence separately

Four different relationships often hide inside the sentence “they control the AI.” An organisation can own physical equipment, hold a contract for access to equipment owned by somebody else, possess permission to use particular data, or influence how a service reaches customers. These relationships can overlap. They can also belong to different organisations, which is why a simple ownership chart can miss the mechanism that matters.

A cloud company might own servers while a developer reserves their use for a period. The developer may own application code but license a model and obtain customer records under a limited processing arrangement. A distributor may own none of those components yet control the interface through which customers discover the application. The people represented in the records may retain interests or rights that neither the developer nor the distributor can simply erase.

For hardware, ask what productive service the asset actually delivers. For access, ask how much capacity is committed, when it is available and under which conditions. For information, ask what purposes, recipients and processing arrangements are authorised. For influence, ask which decisions the participant can change and what others could do in response. These are different questions, so an answer to one does not settle the others.

An owned machine that cannot be powered or maintained is not equivalent to usable computing. A valid reservation may give a renter more dependable near-term access than an owner with installation problems. A downloaded dataset is not automatically authorised for every use. A non-exclusive licence may allow several organisations to use similar information, while an exclusive agreement can narrow that particular route without conferring ownership over every fact it contains.

Avoid treating these distinctions as word games. They change the counterfactual. If a server owner raises its price, a developer with a fixed-term contract may be protected temporarily. If a data licence expires, moving the files to another server does not itself solve the permission problem. If the application loses its distribution channel, additional computing capacity may leave customer demand unchanged. Each failure calls for a different response.

A useful dependency map therefore labels relationships, not just company names. “Owns,” “reserves,” “licenses,” “processes for,” “distributes through” and “depends on” tell different stories. Record the service, period and constraint beside each arrow. The point is to understand which participant has a credible choice at the moment of negotiation, rather than to award a single control score to the biggest-looking box.

Back to contents · Next: 3. Compute access has several dimensions that a chip count hides

3. Compute access has several dimensions that a chip count hides

Computing is a productive input, but chips are not interchangeable tokens. A particular workload may require suitable memory, interconnection, software support, scheduling, location and operating reliability. A large inventory can therefore coexist with a shortage of the capacity that a particular user needs. Conversely, a smaller specialised system can be sufficient for a bounded application without resembling the infrastructure needed to train a frontier model.

Separate training from deployment. Developing a model can require a concentrated run with particular communication and memory needs. Serving an application concerns repeated requests, response times, availability and continuing cost. A source of capacity suitable for one job may not be an economical substitute for the other. Neither the number of model parameters nor the number of processors tells the reader which service is actually being purchased.

Time matters as much as quantity. Capacity offered next year does not replace capacity required this month. An interruptible service can be useful for a flexible research task but unsuitable for a deadline-dependent application. A reservation also differs from an advertised maximum: the former should describe an enforceable service commitment, while the latter may describe what the supplier can sometimes offer under conditions that need checking.

Physical infrastructure adds another boundary. The IEA’s April 2025 report Energy and AI examines the relationship between data centres and electricity systems, including the resources needed to supply growing demand. That makes power availability a legitimate part of compute-access analysis. It does not justify inferring a particular company’s usable capacity from a national energy total or treating a forecast as an already completed installation. IEA: Energy and AI

To see the difference, suppose a provider advertises twice as many accelerator hours as its rival. If the application requires a compatible environment and half the advertised hours cannot run that environment, the headline comparison is incomplete. If the remaining capacity is available only after the customer’s deadline, it is not an immediate substitute at all. This is an illustrative logic test, not a claim about any actual hardware.

Economic leverage can emerge at whichever condition is hardest to replace. Sometimes that is a processor. Sometimes it is a powered site, a scheduling commitment, a qualified software environment or the personnel who can keep the service running. The relevant unit is a usable service under specified conditions. Chip ownership remains important, but it becomes informative only when connected to that service and the alternatives available to its users.

Back to contents · Next: 4. Scale can lower costs and still create a difficult entry threshold

4. Scale can lower costs and still create a difficult entry threshold

A service with substantial initial development costs may become cheaper per unit when more customers share those costs. Suppose an entirely fictional service costs 120,000 credits to establish and one credit for each subsequent task. At 20,000 tasks, its simple average cost is seven credits: six for the allocated initial cost and one for the task. At 120,000 tasks, it is two credits under the same simplified assumptions.

That calculation explains a possible scale advantage, not an inevitable market structure. It excludes financing, maintenance, congestion, sales costs and the possibility that the service must be redesigned as it grows. It also does not prove that a large supplier will pass savings to customers. The arithmetic shows why an entrant may need enough demand to spread a fixed cost, while leaving the actual competitive outcome unresolved.

An entrant’s difficulty can be sharper when it must commit before knowing whether customers will arrive. A mature supplier has existing revenue, operating knowledge and a base of users. A new supplier may need to finance capacity and qualification while still persuading customers to switch. That sequencing creates an entry threshold even when the underlying technology is available for purchase. The barrier is a coordinated transition, not necessarily a secret invention.

But scale is not useful at every level. A small application provider can rent computing instead of constructing a facility. A specialist can serve a task that a general model handles poorly. Shared tools can spread development effort across organisations. Users can sometimes choose a simpler system that meets their needs. These routes do not abolish upstream dependence; they can change where it sits and which kinds of entrant can compete.

There can also be disadvantages to expansion. A larger service may face more complex support, different user needs and a wider set of reliability obligations. A specialised rival might compete through a better fit rather than lower raw processing cost. The right empirical question is which costs fall with scale, which rise, and whether the resulting advantage survives changes in workload, quality and customer preferences.

This is why a large spending announcement is not enough to establish lasting economic power. Spending must produce a service customers value, and that service must remain difficult to replace. An investment can be expensive and unsuccessful. It can also be socially useful while earning an ordinary return. Analysis should follow the chain from committed resources to usable capability to customer alternatives, instead of jumping from expenditure directly to domination.

Back to contents · Next: 5. Data advantage depends on usefulness and legitimate use

5. Data advantage depends on usefulness and legitimate use

“More data” is too broad to explain a competitive advantage. A collection may contain duplicates, obsolete information, unreliable labels or records unrelated to the task. Another collection may be smaller but directly relevant, carefully checked and refreshed when circumstances change. The economic value comes from the information’s contribution to a useful service, after the costs and constraints of obtaining, preparing and using it are considered.

Distinguish training material, retrieval sources, evaluation cases and operating feedback. Training material influences model development. Retrieval sources supply information during a particular application. Evaluation cases help determine whether a system meets requirements. Feedback can reveal which outputs users actually accepted or corrected. A participant’s advantage in one category does not automatically transfer to the others. An extensive text collection may offer little help with a specialised acceptance test.

Availability and authorisation also differ. A record may be technically accessible without being appropriate for a particular transfer, training run or commercial use. Contractual terms, confidentiality, intellectual-property questions and data-protection obligations can all matter, depending on the facts and jurisdiction. This article does not infer legal permission from possession, public visibility or a vendor’s marketing description. A real organisation should verify the relevant rights and obtain qualified advice where needed.

For Singapore readers, the PDPC’s 2024 advisory guidelines address the use of personal data in AI recommendation and decision systems, including development, consumer information and the responsibilities of relevant service providers. They are an appropriate starting point for that specific topic, read with the other guidance the PDPC identifies. They are not a blanket licence to reuse personal information in every AI application. PDPC: personal data in AI systems

A data advantage becomes economically important when competitors cannot obtain an adequate substitute on workable terms. The substitute need not be an identical file. It might be a different source, an independently collected dataset, a revised workflow or a service that solves the customer’s underlying problem without those records. Whether any alternative is adequate requires testing, not a theoretical assertion that information can always be recreated.

Data can also lose value. A historical pattern may become less relevant, access may expire, or a competitor may develop another source. Synthetic examples may help with some tasks while failing to replace real-world information needed for others. The disciplined question is which uncertainty the data resolves, why that matters to the service, and how long the advantage is likely to remain difficult to reproduce lawfully and reliably.

Back to contents · Next: 6. Feedback can strengthen an incumbent without creating an automatic flywheel

6. Feedback can strengthen an incumbent without creating an automatic flywheel

A service used by many people may receive more examples of success, confusion and failure. If the organisation is permitted to use that feedback, can interpret it accurately, and can turn it into better performance, adoption may support improvement. Better performance may then attract more users. This is a possible feedback mechanism, but each step needs evidence; user volume alone does not demonstrate a self-reinforcing advantage.

Feedback can be selective. Users who leave may provide less information than users who remain. A click can mean interest, habit, confusion or an accidental choice. A correction may represent one person’s preference rather than an objective error. If the organisation measures the wrong behaviour, increasing the quantity of feedback can strengthen a misleading signal. The relevant asset is an effective learning process with legitimate inputs, not simply a larger log file.

The task also matters. Repeated interactions can improve a highly specific workflow when the success criterion is observable. They may be less useful where outcomes arrive years later or depend on external conditions. A service that generates impressive suggestions may not observe whether the suggestions ultimately worked. In that case, claims that usage automatically creates superior knowledge should be treated as hypotheses awaiting outcome evidence.

An entrant can sometimes compensate through a different design. It may work closely with a narrow user group, use expert-labelled tests, support a more transparent workflow or concentrate on an underserved need. None of these routes guarantees success. They demonstrate why incumbency should be analysed as a collection of advantages with limits, rather than as a single irreversible property acquired once enough people use a product.

There is a useful distinction between scale economies and network effects. Spreading a fixed cost over more users is a cost mechanism. A network effect concerns one user’s value changing because other participants join. A service can have one without the other. Calling every improvement from increased usage a network effect can conceal whether the actual advantage lies in costs, interaction between users, better evidence or a route to distribution.

For independent judgment, ask what observation would weaken the proposed loop. Perhaps additional users stop improving task performance. Perhaps a smaller rival matches the relevant quality. Perhaps customers can carry useful records to another service. Perhaps permission boundaries prevent the assumed reuse. A credible explanation includes these possible breaks, so that future evidence can change the conclusion rather than merely decorate a story already decided.

Back to contents · Next: 7. Complementary capital can make a good alternative hard to reach

7. Complementary capital can make a good alternative hard to reach

Computing and data rarely become an operating service without other resources. Application code, connectors, evaluation suites, staff knowledge, security arrangements, support routines and customer relationships can all be necessary. These are complementary inputs: improving one is useful only when enough of the others are available. Money can help acquire them, but purchasing power is not the same as having the finished arrangement in place.

Some complements travel easily. A well-documented workflow or portable test collection may remain useful after a supplier changes. Others are specific to an environment. A team may have written extensive code around one provider’s behaviour, or built review procedures around one tool’s outputs. Specific investment can improve current performance while increasing the work required to leave. That tension is not automatically a mistake; it is a choice whose future consequences deserve attention.

Switching cost includes more than an exit fee. It can include rebuilding, testing, staff learning, duplicate operation, data migration and lost service during transition. It can also include the time spent discovering whether a proposed replacement actually works. A quoted rival price that excludes these costs may be relevant to a new customer but misleading for a deeply integrated existing customer considering an immediate move.

The timing creates bargaining consequences. Before investing, a customer may have several credible options. After committing to a specialised setup, its short-run alternatives can narrow. A supplier may then have more influence over renewal terms. This does not prove that it will exploit the situation. Long relationships, reputation, competition for future customers and contractual protections can all restrain behaviour. The vulnerability is a mechanism to investigate, not a prediction of intent.

Financial resources change how quickly the customer can respond. An organisation with sufficient transition funding may run two environments while it tests the replacement. Another may see the same long-run savings but be unable to pay the initial cost. Its dependence is partly a financing constraint. A small grant of computing credits would not necessarily solve that problem if the missing resource is skilled integration time or permission to move data.

The general capital guide explains productive assets and complements in greater depth. Here their distinctive role is in the customer’s outside option: what the customer could actually do if it rejected the current offer. Economic power often accumulates where other participants have invested heavily in a relationship but cannot quickly redeploy that investment. Portability can reduce that dependence, while carrying costs that must be compared with the value of the freedom it provides.

Back to contents · Next: 8. Partnerships and integration can enable entry or narrow alternatives

8. Partnerships and integration can enable entry or narrow alternatives

A developer may need a partner’s computing, finance, engineering support or customer reach. Combining these resources can make an otherwise impossible project feasible. Vertical integration can also reduce coordination failures between layers. A provider that operates infrastructure and software together may improve compatibility or offer customers a simpler service. An economic analysis should recognise these possible gains before examining the restrictions that accompany them.

The FTC’s January 2025 staff report on major cloud-provider and AI-developer partnerships described investment-related rights, resource sharing and spending commitments. It identified potential implications for access to inputs, switching costs and access to sensitive business information. The report explicitly bounded its evidence to information available to staff through September 2024 and public information through January 2025. It is a dated investigation of possible mechanisms, not proof that every partnership has identical effects today. FTC: AI partnerships and investments study

Consider a fictional financing offer that provides credits usable only with the investor’s infrastructure. The offer can make development affordable, but it does not give the recipient the same choices as unrestricted cash. If the recipient later finds a better provider, unused credits may not finance the move. To assess the arrangement, compare the benefit of receiving resources with the conditions, duration and alternatives, rather than describing all funding as interchangeable.

An integrated service can also bundle useful complements. A customer may value one support relationship and a coordinated release schedule. The question is whether the customer can choose a different component when that would better meet its needs, and at what cost. A bundle that genuinely improves the service differs from a restriction that prevents an otherwise workable choice. Determining which explanation fits requires the actual terms and operating evidence.

The CMA’s April 2024 update identified risks around access to critical inputs, existing routes to customers and partnerships across the AI value chain. It also recognised that partnerships can play a pro-competitive role. That combination is important: investigate the mechanism and likely effect instead of assuming that cooperation is either automatically beneficial or automatically exclusionary. CMA: foundation-model competition concerns

The relevant comparison is not always partnership versus a perfectly independent firm with unlimited resources. It may be partnership versus no viable entrant, or this arrangement versus a less restrictive financing route. Good analysis makes that counterfactual explicit. It then asks whether efficiency gains depend on the disputed restriction, who can enter under similar conditions, and how customers’ choices change as the arrangement develops.

Back to contents · Next: 9. A route to customers can matter as much as an upstream input

9. A route to customers can matter as much as an upstream input

A technically capable model still needs users who know about it, can access it and find it useful. Distribution can therefore be an economically important complement. A default placement, an established procurement relationship, a familiar interface or integration into a widely used workflow may affect which services receive attention. Owning computing infrastructure is not the only route to influence over an AI market.

Imagine two applications with similar task performance. One is already available in the software an organisation uses every day. The other requires a separate approval process, staff training and a new support arrangement. Customers may prefer the integrated application for good reasons. The outcome cannot be explained by benchmark performance alone. Convenience, trust and the cost of changing routines belong in the comparison.

That does not mean every default prevents competition. Users may switch readily when another product is better, or use several products for different tasks. An organisation may deliberately evaluate competing services before renewing. A distributor may support outside providers because choice makes its own platform more attractive. The source of influence must be demonstrated through actual customer behaviour and available routes, rather than inferred from interface placement alone.

This layer can interact with data. A route to users may produce feedback that improves an application, subject to legitimate use. It can also create knowledge about customer needs that a new entrant must acquire separately. But if customers can export their working materials and evaluation records, or if distribution is available through several channels, the advantage may be easier to challenge. Once again, the strength of the mechanism depends on practical conditions.

The OECD’s 2025 report Competition in Artificial Intelligence Infrastructure examines competition across the infrastructure supply chain, including relationships between layers and barriers to switching. It supports studying the stack rather than treating all AI supply as one undifferentiated market. This guide uses that framing without reproducing market-share estimates as timeless facts or treating a report’s concerns as findings about every participant. OECD: AI infrastructure competition

A useful map ends at the customer, not at the model. Follow the chain from equipment and access through information, application, assurance and distribution. Ask where a capable entrant would still fail to reach a user. That point may be the commercially important bottleneck even if it receives less attention than the largest facility or the most expensive training run. It also indicates what evidence would establish a meaningful improvement in choice.

Back to contents · Next: 10. Contestability means a credible challenge, not a long supplier list

10. Contestability means a credible challenge, not a long supplier list

For this guide, contestability means that customers or entrants have credible ways to challenge an existing arrangement: switch, enter, expand, substitute, self-provide or decline the service. The word is used practically, not as a claim that all assumptions of a particular economic model hold. A market with many logos may still offer few usable alternatives for a particular task and deadline.

A substitute must clear several hurdles together. It must produce an acceptable result, have sufficient capacity, be available when needed, operate within legitimate rights and fit the customer’s transition resources. Failing any essential hurdle can make the option unusable. A cheaper rival that cannot process the permitted data is not a complete substitute. Nor is an excellent system that cannot supply the required volume.

Switching and entry also operate on different clocks. An alternative that disciplines a supplier over several years may offer no protection against a disruption tomorrow. Conversely, a customer may have a short-run backup that is too expensive for permanent use. Separate emergency continuity from sustained competition. They can reinforce each other, but buying resilience does not automatically create a new economically viable competitor.

Concentration measures need equally careful boundaries. The share of one buyer’s purchases going to each supplier is not the same as the suppliers’ shares of a legally defined market. A count of cloud accounts is not a count of independent power systems. A list of model names is not a list of independent training organisations. Use a measure only after stating what is counted and why that denominator answers the question.

The following case deliberately gives a buyer three providers while leaving a large part of its work dependent on one of them. The dependence is created by the combination of rights, qualification and capacity, not by a missing supplier name. Repairing it requires addressing all three. This is why the exercise begins with a complete input packet rather than a recommendation to “diversify” and an unexplained percentage.

Before moving on, try a simple verbal test. If the preferred provider disappeared at the start of next month, what exact work could each alternative deliver, using which information, for what cost? If the answer is only “we have another account,” the outside option has not yet been established. The case turns that gap into quantities that can be calculated, challenged and improved.

Back to contents · Next: 11. The complete fictional packet: Rowan Document Services

11. The complete fictional packet: Rowan Document Services

Rowan is a fictional business producing standardised internal document checks. Its monthly demand is 9,000 accepted jobs: 6,000 ordinary jobs and 3,000 specialist jobs. Each job consumes one unit of qualified service capacity regardless of class. Jobs are divisible for allocation purposes but must finish within the month. Demand is fixed, there is no backlog carried between months, and incomplete jobs count as unmet demand rather than silently disappearing.

All prices below are in fictional credits. They are stipulated amounts inclusive of every variable processing, storage and transfer charge in this exercise. Taxes, financing charges and other unlisted expenses are zero by assumption. The case is not based on current commercial quotations. Its abstraction makes the input-dependence mechanism visible; a real procurement analysis would need actual offers and a more detailed treatment of uncertain quality, outages and demand.

Provider Alder supplies at most 6,000 accepted jobs per month at 2.00 credits each. Provider Birch supplies at most 4,000 at 2.60 credits each. Provider Cedar advertises capacity for 3,000 at 3.20 credits each. Alder and Birch are already integrated. Cedar is not yet qualified, so its advertised capacity is initially unusable by Rowan. No supplier has a minimum-purchase commitment, and Rowan pays only for jobs actually processed unless a later paragraph explicitly introduces a fixed fee.

Ordinary jobs use Rowan’s own prepared reference set. The packet stipulates that Rowan has all required permissions to process this set with any of the three providers. Specialist jobs need a separate ReferenceCo dataset. The fictional signed licence initially permits processing that dataset only through Alder. Downloading a copy does not broaden that permission. Birch’s software could handle the format, but it may not perform specialist jobs under the initial arrangement. Cedar has neither current qualification nor that permission.

Rowan’s internal review team can accept up to 10,000 jobs monthly. Customer delivery can handle 10,000. The dataset has sufficient relevant coverage for all 3,000 specialist jobs, and the stipulated qualified services meet the same task-specific quality and timing requirements. Those assumptions prevent review, delivery or quality from becoming hidden constraints in the first calculation. They do not imply that such constraints can be ignored in real systems.

An existing ReferenceCo licence costs 1,800 credits monthly. Rowan’s review and delivery operations cost a further 12,000 credits monthly and are fixed over all scenarios in this case. These common costs are included when total operating expenditure is requested. No lost revenue, compensation or customer harm is assigned a monetary value, so an unmet-job count cannot be converted into a financial loss without adding new evidence.

For later tests, Alder and Birch share a fictional NorthLink network service; Cedar uses an independent SouthLink service. A NorthLink interruption disables both Alder and Birch for the entire test month. A failure of Alder alone leaves Birch, Cedar, both network services and ReferenceCo available. These are two different stress events. The packet supplies no event probabilities, so expected annual outage loss and a probability-weighted optimum cannot be calculated.

Back to contents · Next: 12. Solve the baseline by matching each job to an eligible route

12. Solve the baseline by matching each job to an eligible route

Start with specialist work because its permitted route is narrower. All 3,000 specialist jobs must go to Alder. That uses half of Alder’s 6,000-job capacity. The remaining 3,000 Alder places can process ordinary jobs at 2.00 credits each, cheaper than Birch’s 2.60. Send the other 3,000 ordinary jobs to Birch. Cedar receives none because it is not qualified, regardless of its advertised capacity.

The complete allocation is therefore Alder 6,000 and Birch 3,000, delivering all 9,000 jobs. Variable expenditure is 6,000 × 2.00 plus 3,000 × 2.60, which is 12,000 + 7,800 = 19,800 credits. Add the common 1,800 licence charge and 12,000 operating cost to obtain 33,600 credits for the month. Every requested job is delivered, so the all-in expenditure per accepted job is approximately 3.733 credits.

Check the constraints rather than trusting the total. Alder processes exactly 6,000 jobs, its maximum. Birch processes 3,000, below its 4,000 maximum. All specialist work uses the authorised route. Internal review and delivery each handle 9,000, below their 10,000 limits. Ordinary and specialist totals equal 6,000 and 3,000 respectively. These checks establish feasibility; the price ordering then explains why this feasible allocation minimises the stipulated variable cost.

What is the marginal alternative to one more Alder job? For the first 1,000 ordinary jobs moved from Alder, Birch has unused capacity. Beyond that, there is no qualified alternative in the initial arrangement. For specialist work there is no permitted alternative even within Birch’s spare capacity. The same supplier therefore faces different substitution conditions across Rowan’s jobs. An average unit price alone cannot show that difference.

The visible purchase split is two-thirds Alder and one-third Birch. Squaring and summing those shares gives 5/9, or approximately 0.5556. On a 10,000-point scale it is about 5,556. Here this is only a descriptive concentration index for Rowan’s purchases. It is not a market-share calculation, a legal threshold or evidence that Alder controls the wider industry. The denominator is this buyer’s 9,000 jobs.

Even within that narrow boundary, the index leaves important facts out. It does not show that specialist work has only one permitted provider. It does not show Cedar’s missing qualification or the shared NorthLink dependency. A useful numerical summary can support the explanation, but it cannot replace the underlying map. The next step tests the claimed alternatives by removing one input and recalculating what remains possible.

Back to contents · Next: 13. The failed backup claim: three accounts do not protect nine thousand jobs

13. The failed backup claim: three accounts do not protect nine thousand jobs

Rowan’s manager initially says, “We have three providers, so an Alder outage is covered.” Apply the supplied failure event instead of accepting the statement. Alder is unavailable for the entire month. Birch remains available for up to 4,000 ordinary jobs. Cedar is still unqualified. Specialist jobs cannot move because the ReferenceCo permission is limited to Alder. The maximum delivered output is therefore 4,000 ordinary jobs.

The shortfall is 5,000 jobs: 2,000 ordinary and all 3,000 specialist jobs. Variable expenditure is 4,000 × 2.60 = 10,400 credits. With common fixed costs, expenditure is 24,200. That is less cash spent than the normal 33,600, but Rowan has failed to deliver more than half its demand. Calling this a saving would ignore the missing service. The case provides no monetary valuation of those consequences.

Now try a partial repair: qualify Cedar without changing the data permission or advertised capacity. Birch and Cedar could together process 7,000 ordinary jobs in principle, but demand contains only 6,000 ordinary jobs. The least-cost allocation is Birch 4,000 and Cedar 2,000, costing 10,400 + 6,400 = 16,800 in variable expenditure. All ordinary work is delivered, while 3,000 specialist jobs remain unmet.

This second result is a useful improvement but not the promised full recovery. The extra 1,000 of nominal ordinary capacity has no value for the missing specialist work while that work cannot legally use those routes under the stipulated licence. Increasing Cedar’s machines alone would not change the constraint. The right repair must address the specific permission as well as qualification and capacity.

Consider the opposite partial repair. Suppose the licence permits Birch and Cedar, but Cedar remains unqualified. Birch can now process specialist work, yet total usable alternative capacity is still only 4,000. Reassigning its slots changes which customers receive service; it does not create the other 5,000 slots. A permission change is necessary for some work but insufficient for complete continuity.

The failure reveals why labels such as “multi-cloud,” “portable data” or “backup supplier” should be treated as claims to test. Each can describe a valuable component of an alternative without establishing a complete route. Rowan needs a joint solution: permission for the specialist input, a qualified service at the alternate providers and enough firm capacity to meet demand. The repair should be judged against that whole requirement.

Back to contents · Next: 14. Repair the outside option, including its cost and lead time

14. Repair the outside option, including its cost and lead time

The fictional repair package costs 18,000 credits upfront: 12,000 for a negotiated licence amendment and 6,000 for integration and qualification work. The amendment expressly permits the specialist dataset to be processed through Alder, Birch and Cedar for the following twelve operating months. It does not transfer ownership of the dataset, permit unrelated uses or eliminate ReferenceCo’s continuing role. Its exact scope is a supplied contractual fact of the exercise, not a legal inference.

The implementation takes one full month. Work proceeds alongside the old arrangement, and the original capacity remains available during that transition by assumption. The new route is usable only at the start of the following month after qualification has succeeded. The package also secures Cedar capacity of 5,000 jobs monthly at the same 3.20 variable price, with a fixed capacity-reservation charge of 1,000 credits per month.

An additional 200 credits monthly pays for the ongoing test workload and checks that keep the alternate route qualified. This amount includes all testing resources and does not consume the production capacities listed in the packet. The combined new recurring fixed expense is therefore 1,200. The existing 1,800 dataset charge and 12,000 operating cost remain unchanged. For the twelve-month comparison, all fixed charges remain payable even in the specified outage tests.

Once the package is live, an Alder-only outage has a complete solution. Put 3,000 specialist jobs and 1,000 ordinary jobs on Birch, filling its 4,000 capacity. Put the other 5,000 ordinary jobs on Cedar, filling its new 5,000 capacity. Output is 9,000, permissions are satisfied, and review and delivery stay within their limits. Variable expenditure is 4,000 × 2.60 + 5,000 × 3.20 = 26,400 credits.

The outage-month operating total is 26,400 + 1,800 + 12,000 + 1,200 = 41,400 credits. This is a more expensive complete service than the normal baseline, but it is a complete service. The repair has converted an impossible full exit into an available, priced alternative. Whether that benefit justifies the package cannot be answered from the outage cost alone; it depends on the consequences and likelihood of losing access, which are not supplied.

Normal production can still use Alder 6,000 and Birch 3,000 because they remain cheapest at the original prices. The new normal operating total is 34,800 credits: the old 33,600 plus 1,200 for keeping the option ready. Purchase concentration is unchanged, yet Rowan’s ability to leave Alder has materially improved. This is the central result: usable alternatives can strengthen competition even before they receive a large share of ordinary production.

Back to contents · Next: 15. A price change shows how a credible outside option alters the choice

15. A price change shows how a credible outside option alters the choice

Now consider a separate price scenario, with all providers operating. Alder raises its price from 2.00 to 3.80 credits for the next operating period. In the initial unprepared arrangement, Rowan must still use Alder for the 3,000 specialist jobs. Birch can take all 4,000 of its available places as ordinary work. The remaining 2,000 ordinary jobs must also stay with Alder because Cedar is unqualified.

The unprepared allocation becomes Alder 5,000 and Birch 4,000. Its variable expenditure is 5,000 × 3.80 + 4,000 × 2.60 = 19,000 + 10,400 = 29,400 credits. Add the original 13,800 common fixed costs for a total of 43,200. Rowan has reduced its Alder use from 6,000 to 5,000, but it cannot respond fully to the price difference. The remaining dependence includes both specialist permissions and limited qualified capacity.

With the completed repair, Rowan can instead use Birch 4,000 and Cedar 5,000, at the 41,400 total established above. This is 1,800 credits less per operating month than remaining in the unprepared arrangement at Alder’s new price. Notice that the comparison includes the repair’s 1,200 recurring cost. Comparing only variable charges would overstate the ongoing financial advantage of the repaired option.

The optimal production allocation changes in stages as Alder’s price changes. Below 2.60, Alder 6,000 and Birch 3,000 minimise variable expenditure. Between 2.60 and 3.20, Birch 4,000 and Alder 5,000 do so. Above 3.20, Birch 4,000 and Cedar 5,000 are cheapest once the repair is live. At exactly 2.60 or 3.20, more than one allocation can have the same minimum cost. Permissions and capacities remain satisfied throughout these repaired scenarios.

Those breakpoints concern allocation after the option exists. They do not establish that an unprepared organisation should pay any upfront amount to build it. If the 18,000 setup cost is still prospective and the 1,800 monthly advantage is expected to persist under the exercise’s assumptions, ten complete operating months recover that setup cost. This is a simple undiscounted break-even calculation, not a risk-adjusted investment valuation or prediction of future prices.

The one-month implementation delay also remains real. Savings start only after the repaired route becomes operational. A customer who has not prepared cannot obtain ten months of savings within the first ten calendar months after starting the project. Existing commitments, price uncertainty, changing demand or a failed qualification could alter the result. The arithmetic illustrates bargaining through a credible alternative; it does not supply the missing evidence needed for a real purchasing decision.

Back to contents · Next: 16. Follow the money and test the remaining common dependency

16. Follow the money and test the remaining common dependency

At the original prices, keeping the repaired route ready for twelve operating months adds 12 × 1,200 = 14,400 credits in recurring charges. Add the 18,000 upfront package for a total incremental commitment of 32,400 over that period. Baseline annual operating expenditure is 12 × 33,600 = 403,200. Repaired normal-operation expenditure including setup is 12 × 34,800 + 18,000 = 435,600. The difference reconciles to the same 32,400.

For comparison only, spreading setup evenly over twelve months gives 1,500 credits monthly. Add the 1,200 holding cost to obtain an equivalent 2,700 monthly increment, making the comparison total 36,300. This allocation does not change when cash is paid: Rowan still needs 18,000 upfront. A firm can understand the longer-run cost perfectly and still lack the liquidity or skilled time needed to reach the new arrangement.

The repair also has a defined protection boundary. Under an Alder-only outage, it supplies all 9,000 jobs. Under a NorthLink interruption, both Alder and Birch disappear. Cedar can supply at most 5,000. Because specialist permissions now cover Cedar, Rowan can prioritise all 3,000 specialist jobs and 2,000 ordinary jobs, leaving 4,000 ordinary jobs unmet. The repair solves one failure pattern without making the service invulnerable.

A different allocation would change which work is missed, but not Cedar’s 5,000 total limit. The packet provides no valuation or priority rule beyond the illustrative choice just made, so it cannot prove that prioritising specialist work is socially or commercially optimal. The honest result is a remaining capacity shortfall under a common-dependency event. Resolving it would require another specified resource or a revised service promise, not a stronger adjective for the existing backup.

Suppose someone proposes splitting normal work equally across all three providers and calls the result fully diversified. The purchase concentration index would fall to one-third, or approximately 3,333 on the 10,000-point scale. But 6,000 of the 9,000 jobs would still use NorthLink through Alder and Birch. At the network level, the shares would be two-thirds NorthLink and one-third SouthLink, returning the descriptive index to about 5,556.

There is another common dependency: ReferenceCo. Permitting its dataset on three providers removes a restriction on processing location, but it does not create an independent source for the specialist information. If the dataset itself becomes unavailable under a separate scenario, computing diversity does not replace it. This is why economic influence can move after a repair. The bottleneck is a feature of the complete production arrangement, not a permanent title awarded to one supplier.

Back to contents · Next: 17. What evidence would establish power outside a classroom packet?

17. What evidence would establish power outside a classroom packet?

The fictional case makes constraints certain so that the reasoning can be checked. Real conditions are less tidy. An investigation would need current contracts, tested alternatives, available capacity, migration estimates and evidence about the task customers are trying to complete. It should distinguish a supplier’s advertised offer from a customer’s successful use of that offer under comparable conditions. Unknowns should remain visible rather than being filled with confident estimates.

Begin with actual substitution. When price or quality changes, where can work move? How much moves, how quickly, and what prevents the rest from moving? A customer may stay because the service remains best, because transition is expensive, because a rival lacks capacity, or because an input cannot be used elsewhere. These explanations can produce the same observed retention rate while implying very different sources of influence.

Quality comparison is indispensable. NIST’s AI Risk Management Framework 1.0 treats validity and reliability in relation to intended use and conditions, and discusses testing and monitoring. That supports comparing alternatives on a defined workload rather than assuming nominally similar models deliver equivalent results. The framework is voluntary guidance, not a certification that a proposed substitute is suitable or a legal finding about market structure. NIST: AI Risk Management Framework 1.0

Investigate bargaining in both directions. A large buyer may negotiate capacity guarantees or favourable prices. A supplier with high fixed commitments may depend on keeping that buyer’s volume. A contract that looks restrictive in isolation might finance dedicated investment that the customer requested. The question is how the actual arrangement changes outside options for both parties, including what happens at renewal or termination.

Then look for durable rather than merely temporary effects. Can competitors expand? Are they obtaining the relevant inputs? Can new methods reduce demand for the scarce component? Have customers successfully migrated? Can smaller firms finance the qualification process? Evidence across several periods can help distinguish a transient capacity shortage from an advantage repeatedly protected by structural obstacles, though it will not eliminate all uncertainty.

Finally, preserve the counterfactual. If a service improves while its price rises, assess the combined change rather than price alone. If migration is rare, examine whether the available alternatives are worse or whether obstacles prevent their use. If entry fails, investigate why. A persuasive account identifies observations that support its mechanism and observations that would undermine it. The objective is a revisable explanation, not a prosecution or defence assembled around a preferred label.

Back to contents · Next: 18. Compare responses by the constraint they actually change

18. Compare responses by the constraint they actually change

Several responses can improve usable choice, but none is universally sufficient. Interoperability can reduce technical rebuilding if systems genuinely support compatible exchanges. Portability can help users carry relevant records or configurations, subject to legitimate rights. Shared evaluation methods can make comparisons more credible. Each response should be tested against an actual blocked route: which cost, delay or uncertainty becomes smaller, and which constraint remains?

Access programmes can address a different barrier. The US National Science Foundation describes the National Artificial Intelligence Research Resource as providing routes to resources including computing, datasets, models and related support for research and education. It is an example of widening access without requiring each participant to own a facility. It does not establish universal eligibility, unlimited capacity or the success of every supported project. NSF: National Artificial Intelligence Research Resource

Standards and shared tools can help entrants build on common foundations, but standardisation can also freeze an unsuitable design or leave important behaviour outside the specification. A data export that omits meaningful context may be technically portable and practically weak. An evaluation suite can become stale. Maintaining usable choice requires updating the bridge, not only publishing its original specification.

Public procurement can create demand for interoperable services or qualified alternatives. It can also entrench an incumbent if requirements fit only its existing design. The assessment should connect a requirement to its legitimate purpose and consider whether another design can meet that purpose. This is a policy-analysis principle, not a claim that any particular procurement rule is lawful, unlawful or suitable for every sector.

Safety and privacy introduce genuine constraints. Opening every dataset or removing every access condition is not an appropriate general response to concentration. Some restrictions protect people, confidential information or critical systems. The task is to ask whether a protective objective can be met while preserving more legitimate alternatives, and what evidence would demonstrate that the revised arrangement still provides the necessary protection.

Trade-offs should be budgeted rather than hidden. A second qualified provider may require scarce engineering effort. Public infrastructure requires resources and accountable allocation. Stronger reporting may impose costs or expose sensitive information if poorly designed. The best response depends on the failure being addressed, the people affected and the feasible alternatives. An increase in the number of suppliers is useful only insofar as it produces better choices and outcomes under those conditions.

Back to contents · Next: 19. What changes in a hypothetical superintelligence economy?

19. What changes in a hypothetical superintelligence economy?

Present AI evidence cannot settle how ASI would affect economic power. A hypothetical system might alter the cost of research, programming, design or coordination, but the direction of concentration would still depend on access, reproduction, complementary resources and institutions. A scenario becomes informative when it specifies those conditions. “Intelligence becomes abundant” is an opening assumption, not a complete account of who can use it and on what terms.

In one conditional world, advanced methods sharply reduce the computing required for many valuable tasks, and several organisations can deploy them. Customers can reproduce useful services with modest resources and move their information lawfully. Some existing compute advantages could weaken. Economic influence might shift toward scarce physical assets, trusted relationships, specialised information or services that remain difficult to substitute. This is a possible mechanism, not a forecast.

In another conditional world, the most valuable capability requires facilities, operating knowledge or inputs that few participants can obtain. Better intelligence increases demand for those complements faster than supply expands. Their holders could gain leverage even while individual tasks become cheaper. The conclusion depends on continued scarcity and limited substitution; it would weaken if alternative architectures, new capacity or workable access arrangements changed those conditions.

A third world separates the model from deployment authority. The technology may be widely available, while permission to act in high-consequence settings remains deliberately bounded. Organisations with trusted assurance processes, qualified staff and legitimate operating authority might remain important. That should not be described as proof that every restriction is an economic barrier to remove. Some requirements serve protective purposes that remain relevant even when technical capability improves.

These scenarios can coexist across tasks. A cheap writing tool, a specialised industrial application and a tightly controlled public service may face different inputs and institutions. There is no requirement that one market structure govern all of them. Nor does the scenario establish how the resulting income is distributed among households. The companion distribution guide examines that separate question through wages, prices, ownership and public choices.

The useful preparation is therefore robust reasoning. Learn to identify the scarce complement, test the outside option and separate demonstrated capability from assumed future changes. Track evidence that could move the bottleneck. Avoid treating technical capability as legitimate authority or economic influence as an entitlement. A system’s ability to produce useful work does not decide who should control the terms on which others depend on it.

Back to contents · Next: 20. Independent transfer packet: Larch's specialised service

20. Independent transfer packet: Larch's specialised service

Try this problem before reading the next chapter. Larch is a separate fictional service with monthly demand of 12,000 accepted jobs: 8,000 ordinary and 4,000 specialist. All three providers are technically qualified from the start and meet identical stipulated quality requirements. Each job uses one capacity unit, all demand must be completed in the month, and internal review and delivery each have capacity for 14,000. There is no backlog, uncertainty or omitted variable charge.

Pine offers 4,000 jobs monthly at 1.50 credits each. Quartz offers 5,000 at 2.00. Reed offers 6,000 at 2.80. There are no minimum purchases. The ordinary reference set may be used with every provider, but the specialist dataset is initially authorised only for Pine. Existing fixed operating costs, including the dataset licence, total 4,000 credits monthly. All amounts are fictional, taxes are zero and no other costs apply.

A proposed package costs 10,000 credits upfront and takes one full month to complete. It extends specialist processing permission to Quartz and Reed, increases Reed’s firm capacity from 6,000 to 7,000 and adds a 600-credit monthly reservation charge. Prices and the existing 4,000 fixed cost do not change. The package’s qualification work and ongoing checks are included in those stated amounts. Capacity becomes available only after successful completion.

Pine uses an independent network. Quartz and Reed share another network. In a Pine-only failure, both other providers remain operational and the dataset remains available. In a failure of the shared network, Quartz and Reed are both unavailable while Pine remains usable. The dataset licence permits Pine to continue processing specialist work in that event. No probabilities or monetary consequences for unmet jobs are supplied.

First, find the least-cost baseline allocation and its variable and total expenditure. Second, remove Pine before the package is implemented and calculate maximum completed output, unmet jobs by class and the least variable cost of delivering that maximum output. Third, show a complete allocation after the package that delivers all jobs without Pine, with variable and total costs. Check every capacity and permission explicitly.

Fourth, calculate normal operating expenditure with the package at the original prices and the total twelve-operating-month increment including setup. Fifth, consider a separate normal-operation scenario in which Pine’s price becomes 3.40. Compare the cheapest unprepared and prepared allocations, including their respective fixed charges. Find the number of complete operating months required for the price-scenario savings to exceed the setup cost, using a simple undiscounted calculation.

Finally, test the shared-network failure after the package. Explain whether the package guarantees full continuity, whether a lower purchase-concentration index would solve that failure, and what claim about legal market power can be inferred from these buyer-level figures. A complete answer should distinguish an arithmetic result, a stipulated contractual fact and an unresolved judgment. That distinction is as important as obtaining the correct total.

Back to contents · Next: 21. Worked answers and checks for independent use

21. Worked answers and checks for independent use

The initial specialist requirement fills Pine’s 4,000 places. The cheapest ordinary allocation is Quartz 5,000 and Reed 3,000, delivering all 12,000 jobs. Variable expenditure is 4,000 × 1.50 + 5,000 × 2.00 + 3,000 × 2.80 = 6,000 + 10,000 + 8,400 = 24,400 credits. Adding fixed costs produces 28,400. Pine and Quartz are full, Reed remains below capacity, and review and delivery each process 12,000, within their 14,000 limits.

Before the package, a Pine-only failure makes all 4,000 specialist jobs infeasible because neither remaining provider is authorised for them. The 8,000 ordinary jobs can still be completed: Quartz handles 5,000 and Reed 3,000. Variable expenditure is 18,400 and total expenditure is 22,400 including fixed costs. The lower expenditure is associated with missing specialist work, so it is not evidence of delivering the original service more efficiently.

After the package, Quartz 5,000 and Reed 7,000 provide exactly 12,000 places. One valid allocation sends all 4,000 specialist jobs plus 1,000 ordinary jobs to Quartz, and the remaining 7,000 ordinary jobs to Reed. Variable expenditure is 10,000 + 19,600 = 29,600. Total expenditure is 29,600 + 4,000 + 600 = 34,200. Rights, provider capacities and internal capacities all support the complete route under the specified Pine-only failure.

At the original prices and with all providers available, the cheapest production allocation remains Pine 4,000, Quartz 5,000 and Reed 3,000. The new total is 29,000, the old 28,400 plus the 600 recurring charge. Twelve operating months add 7,200 in recurring expense, and setup adds 10,000, for a 17,200 incremental commitment. The complete baseline and prepared totals are 340,800 and 358,000 respectively, which reconcile to that difference.

At Pine’s separate price of 3.40, the unprepared service must still place all 4,000 specialist jobs there. Quartz 5,000 and Reed 3,000 serve ordinary demand. Variable expenditure becomes 13,600 + 10,000 + 8,400 = 32,000, or 36,000 including fixed costs. The completed package permits Quartz 5,000 and Reed 7,000 at a total of 34,200. The ongoing difference is therefore 1,800 monthly after including the new reservation charge.

Dividing 10,000 by 1,800 gives approximately 5.556 operating months. Five complete months save 9,000, insufficient to cover setup; six save 10,800, exceeding setup by 800. The one-month implementation delay comes before those operating savings. The calculation assumes the stated prices, demand and arrangements persist and ignores discounting as instructed. It is not a prediction that such an investment would pay off under uncertain real-world conditions.

Under the shared-network failure, only Pine remains. Its 4,000 capacity can deliver the specialist jobs if those are prioritised, leaving all 8,000 ordinary jobs unmet. The package therefore does not guarantee full continuity. Changing normal purchase shares does not create another independent network or more surviving capacity. Finally, none of these figures establishes legal market power: they describe one buyer’s stipulated dependencies and feasible alternatives, without defining a wider market or applying a legal test.

Back to contents · Next: 22. Questions that sharpen the conclusion

22. Questions that sharpen the conclusion

Does owning more hardware always create more economic influence?

No. Hardware needs usable operating conditions, customers and an appropriate service. Influence depends on what others cannot readily replace. A renter with firm access can have strong near-term capacity, while an owner may face commissioning or utilisation problems. Ownership is a meaningful relationship, but it is neither a complete capacity measure nor a sufficient account of bargaining power.

Is buying from several providers the same as having alternatives?

It can help, but Rowan shows the distinction. A provider may lack rights, qualification or capacity for the work that must move. Several providers can also share a network, dataset or other input. Conversely, an unused but maintained alternative can change negotiation even before it receives ordinary production. Ask what work the option can actually deliver after the specific disruption or price change.

Is all data an asset its holder can freely sell or reuse?

No such assumption is made here. Information can involve different rights, obligations, agreements and affected people. Technical possession does not answer every use question. The economic analysis needs a verified description of legitimate access for the intended task. Where that is uncertain, keep it uncertain and obtain appropriate advice; do not convert a convenient modelling assumption into a legal conclusion about real records.

Would open models remove the need to analyse concentration?

Access to model weights can widen some technical choices, subject to the applicable licence and operating requirements. It does not by itself supply computing, relevant information, qualified deployment, distribution or support. The question becomes which dependencies are reduced and which remain. It is possible for an application layer to become more competitive while important upstream inputs remain difficult to obtain.

Does a concentrated input necessarily make consumers worse off?

The packet does not establish that general conclusion. A strong supplier may deliver valuable innovations, scale economies or reliable integration. Concerns arise when limited alternatives allow terms, quality or innovation to deteriorate relative to a feasible comparison. Assess benefits and constraints together. The correct question concerns the arrangement and its alternatives, not whether a large firm must be good or bad by definition.

What would count as progress for the reader?

You should be able to identify a dependency without confusing it with ownership, state the permitted use of information without inventing permission, calculate a complete alternative under capacity constraints, and distinguish recurring cost from upfront transition funding. You should also be able to explain what your calculation does not establish. The independent packet tests those abilities by changing the numbers, the qualification position and the shared failure pattern.

Back to contents · Next: 23. Continue with the question you actually need to answer

23. Continue with the question you actually need to answer

If the question is why faster tools do not immediately raise economy-wide productivity, continue with Super Intelligence and productivity. That guide follows the path from task improvements to completed output and aggregate measures. This article has instead examined who can set terms around the inputs and alternatives along that path.

If the question is who receives the resulting gains, read wealth, ownership and inequality. For productive assets, depreciation, complements and financing, use how capital works. These questions connect, but a gain in productive capacity, a bargaining advantage and a household benefit are different outcomes.

For the physical service behind an interface, continue with chips, memory and the geography of computing or from data centres to intelligence factories. For oversight of strategically important infrastructure, see compute governance. Return to the Super Intelligence master guide for the broader reading map.

Back to contents · Next: Sources and scope

Sources and scope

The linked primary sources support the nearby descriptions of their published work: CMA, 11 April 2024; FTC, January 2025 with the evidence cutoff stated above; OECD, 2025; IEA, April 2025; PDPC’s 2024 advisory guidance; NSF’s NAIRR resource description; and NIST AI RMF 1.0, January 2023. Links were checked for this article on 1 October 2026. These dates matter: none of the sources is presented as a real-time census of all AI supply arrangements.

All named case businesses, providers, prices, capacities, contracts, network dependencies and calculated scenarios are fictional teaching inputs. They make no claims about actual companies or jurisdictions. The cases demonstrate accounting, allocation and dependency reasoning under stated assumptions. They provide neither investment advice nor a legal opinion, and they do not assign probabilities to the emergence of ASI or to any particular future concentration outcome.

Back to contents · Return to the Super Intelligence master guide