THE AI MIRROR TRICK

THE AI MIRROR TRICK

Hyperscaler power, frontier laboratories and the claim of a Silicon Valley monoculture

By Dr Danie Adendorff DSc (c.h), MSc.

THE PROPOSITION AND THE BURDEN OF PROOF

The expression “OpenAI and Anthropic are hyperscaler psy-ops built for the monoculture of Silicon Valley” originates in the polemical work of technology critic Ed Zitron. It compresses several separate allegations into one deliberately inflammatory formulation: that frontier artificial-intelligence laboratories are financially dependent on cloud providers; that hyperscaler investments return to those providers as cloud expenditure; that the laboratories manufacture public enthusiasm for infrastructure controlled by their investors; that their models export a narrow Silicon Valley value system; and that AI-safety regulation is being used to suppress competitors.

These propositions do not have equal evidentiary status.

The infrastructure-dependence proposition is strongly supported.

The reciprocal capital-and-compute proposition is also supported, although the term “round-trip capital” can misleadingly imply sham transactions or improper accounting.

The cultural-monoculture proposition is partly supported by evidence about model governance, training populations and cultural alignment, but it is easily overstated.

The regulatory-capture proposition identifies a credible structural risk but has not been established as fact.

The literal “psy-op” allegation is unsupported.

A psychological operation, properly understood, is an organised activity intended to influence the perceptions, emotions, reasoning or behaviour of a target population in support of a specified objective. No publicly available evidence establishes that OpenAI or Anthropic was created, funded or directed as a covert influence operation by Microsoft, Amazon, Google or any other hyperscaler. Nor is there evidence of the command structure, operational intent, target architecture or concealed coordination that such a claim would require.

The word can therefore be retained only as a metaphor for narrative amplification: the use of civilisational claims, existential-risk language and promises of imminent artificial general intelligence to attract capital, customers, political attention and infrastructure investment. That is a legitimate subject for investigation. It is not the same as proving a covert operation.

Structural forensics must begin by separating documented relationships from inferred motives. The relevant question is not whether a secret conspiracy can be imagined. It is whether the visible organisation of capital, compute, distribution and governance produces concentrated power irrespective of whether the participants have coordinated that outcome.

THE HARD PLUMBING OF CAPITAL, COMPUTE AND DISTRIBUTION

Frontier AI is not merely a software industry. It is an infrastructure industry.

Training and operating advanced models requires specialised semiconductors, high-capacity data centres, large quantities of electricity, sophisticated networking, proprietary optimisation systems and access to geographically distributed cloud platforms. These requirements create formidable barriers to entry and make frontier laboratories dependent on a small group of infrastructure providers.

Microsoft’s relationship with OpenAI began as an exclusive computing partnership in 2019 and subsequently expanded into a multibillion-dollar strategic alliance. Microsoft’s regulatory filings confirmed cumulative funding commitments of approximately $13 billion by 2023. The relationship combined investment, Azure compute, model integration, intellectual-property rights and commercial distribution.

Anthropic developed a comparable, although not identical, set of relationships. Amazon’s announced investment reached $8 billion in 2024, with AWS designated as Anthropic’s primary cloud and training partner. Google separately invested billions in the company and supplied Google Cloud infrastructure.

The United States Federal Trade Commission examined these arrangements in detail. Its staff report documented more than $20 billion in cumulative investments involving Microsoft, Amazon, Google, OpenAI and Anthropic during the period reviewed. It found that the partnerships included equity interests, revenue-sharing arrangements, cloud-spending commitments, discounted computing resources, access to intellectual property, product integration, consultation rights and, in some cases, exclusivity provisions.

The FTC did not conclude that these relationships were fraudulent. Its concern was structural. When a cloud provider is simultaneously an investor, essential supplier, commercial distributor and potential beneficiary of technical information, the conventional distinction between supplier, customer, financier and competitor begins to dissolve.

The Commission specifically identified possible restrictions on access to compute and engineering talent, increased switching costs, access to commercially sensitive information and the possibility that partnership terms could favour the incumbent cloud provider. It also found that some negotiated compute discounts were considerably greater than publicly advertised rates and that only a limited number of firms possessed the infrastructure necessary to support frontier-scale model development.

The relationships have since become more extensive and more complicated.

In February 2026, Amazon and OpenAI announced a strategic partnership under which Amazon would invest up to $50 billion, subject to specified conditions. The arrangement included the planned consumption of approximately two gigawatts of AWS Trainium capacity and gave AWS an exclusive third-party distribution role for a category of OpenAI products described as OpenAI Frontier.

At the same time, Microsoft remained OpenAI’s primary cloud partner. Revised agreements preserved significant Azure distribution and hosting rights while making some previous intellectual-property arrangements non-exclusive. Azure remained the exclusive provider of stateless OpenAI application-programming interfaces under the announced arrangements, while OpenAI expanded its use of other infrastructure providers.

OpenAI has also committed to substantial Oracle capacity through the Stargate programme and has developed infrastructure relationships involving SoftBank, Nvidia and other technology providers. Anthropic, although identifying AWS as its primary cloud partner, has distributed Claude through multiple major cloud environments.

This diversification weakens the claim that either laboratory is simply a captive subsidiary of one hyperscaler. It strengthens a different conclusion: frontier AI is embedded in an interlocking hyperscaler system.

OpenAI and Anthropic are not conventional startups purchasing interchangeable hosting services. They function as strategic anchor tenants around which hyperscalers organise data-centre investment, semiconductor procurement, model distribution, enterprise products and future revenue expectations.

The frontier laboratories obtain capital, compute, market access and technical infrastructure. The hyperscalers obtain differentiated models, increased cloud consumption, product integration, intellectual-property rights and a strategic position within the emerging AI stack.

This is symbiosis, but it is not a relationship between equals.

THE RECIPROCAL CAPITAL-AND-COMPUTE LOOP

The financial architecture can be represented as a recurring sequence.

Hyperscaler capital and infrastructure support the frontier laboratory.

The laboratory uses that capital to purchase or commit to hyperscaler compute.

The resulting models differentiate the hyperscaler’s cloud and software products.

Enterprise adoption generates further demand for cloud capacity.

That demand supports additional data-cententre investment and higher valuations.

The cycle then repeats at greater scale.

This is the most defensible part of Zitron’s argument.

The arrangement contains an undeniable degree of strategic circularity. A cloud provider invests in a model developer; the model developer commits substantial expenditure to that provider’s cloud; the provider records increased infrastructure demand; and the model helps the provider market its broader platform.

It is nevertheless important not to misuse the expression “round-trip capital”. In accounting and securities enforcement, round-tripping can describe transactions structured to create the appearance of revenue or economic activity without genuine substance. Publicly available evidence does not establish that the OpenAI, Anthropic, Microsoft, Amazon or Google arrangements constitute fraudulent round-tripping.

The investments finance real computing resources. The laboratories produce real models. The cloud providers construct real infrastructure. Customers make real use of the resulting products.

The correct criticism is not that the entire market is fictitious. It is that overlapping investment, supply, purchasing and distribution relationships make it difficult to determine how much demand is independent, how much is contractually induced and how much market value depends on continued capital expenditure by the same tightly connected participants.

This distinction matters. A technology can be real, useful and commercially valuable while its surrounding investment structure remains fragile or self-reinforcing.

The system also transfers risk unevenly.

The laboratories bear the direct reputational risk associated with model failures, hallucinations, copyright disputes, safety incidents and unfulfilled claims about artificial general intelligence. The hyperscalers sell the underlying compute, storage, networking and enterprise integration required by both successful and unsuccessful model experiments.

This resembles the familiar “picks and shovels” position in a speculative expansion. Yet the analogy is incomplete because the suppliers are not standing outside the gold rush. They are financing selected prospectors, controlling access to the land and integrating the resulting discoveries into their own commercial platforms.

Competition authorities have consequently begun to examine the AI sector as an integrated value chain rather than a collection of independent markets. The UK Competition and Markets Authority has identified risks involving control of critical inputs, powerful incumbents extending their positions across the AI stack and partnerships reinforcing market power. Its separate cloud-services investigation found an adverse effect on competition and recommended further examination of Amazon Web Services and Microsoft under the United Kingdom’s digital-markets regime.

The Organisation for Economic Co-operation and Development has similarly warned that dependence on cloud infrastructure can encourage vertical integration and create forms of cloud gatekeeping. It also provides an important corrective: AI markets remain dynamic, model prices have fallen, new providers continue to appear and competition has not yet settled into a permanent structure. Concentration risk is serious, but monopoly should not be declared in advance of the evidence.

The European Commission has identified possible bottlenecks in advanced chips, cloud services, data and specialist labour while also recognising the continuing entry of new models and providers.

The emerging system is therefore neither a competitive paradise nor a completed monopoly. It is a contested oligopolistic structure in which a small number of firms possess enormous advantages in capital, compute, distribution and integration.

THE SILICON VALLEY MONOCULTURE: WHAT THE EVIDENCE SUPPORTS

The second part of the allegation concerns culture rather than infrastructure.

No general-purpose language model is neutral. Model behaviour is shaped by the composition of its training data, the languages represented in that data, filtering decisions, human-feedback processes, synthetic training material, safety rules, system instructions, legal constraints and commercial product objectives.

OpenAI makes these choices explicit through its Model Spec. The document defines intended model behaviour and provides rules for resolving conflicts between user requests, developer instructions, safety requirements and organisational policies. OpenAI has described the specification as both a transparency mechanism and an internal coordination instrument. Earlier versions included protecting the company’s legal and reputational position among the objectives influencing model behaviour.

Anthropic uses an even more explicit constitutional model. Claude’s Constitution establishes behavioural principles that guide training and provide an authoritative framework for what the company considers helpful, safe and ethically acceptable conduct. Anthropic has acknowledged that its original constitution was assembled by company employees, although subsequent versions incorporated wider consultation and additional external sources.

These documents are not evidence of wrongdoing. They are evidence of private normative governance.

A limited number of corporate teams determine how models should handle political disagreement, offensive material, contested historical narratives, dangerous requests, sexuality, religion, medical uncertainty, national-security questions and conflicts between autonomy and protection. Such decisions are unavoidable. A deployed model must behave somehow.

The central policy issue is therefore not whether values are embedded in models. They are. The issue is who selects those values, how conflicts among them are resolved, which communities are represented and what mechanisms exist for appeal, customisation or pluralism.

There are credible reasons for concern about cultural concentration.

OpenAI’s GPT-4 documentation acknowledged that its red-team participants were disproportionately English-speaking, Western, highly educated and connected to the technology sector. That composition does not invalidate the work, but it can influence which risks are recognised, how harms are prioritised and which social assumptions appear self-evident.

Independent research has repeatedly found representational asymmetries. Santurkar and colleagues compared language-model outputs with the views of 60 demographic groups in the United States and found substantial misalignment, including tendencies for some human-feedback-trained models to reflect particular political orientations more closely than others.

Anthropic’s GlobalOpinionQA research found that model responses were often more similar to opinion patterns in the United States and parts of Europe than to those in many other societies. Although prompting could alter this alignment, the researchers also warned that attempts to simulate cultural perspectives could reproduce stereotypes.

Further academic studies have identified Western cultural preferences in models responding in non-Western languages and uneven performance when models are assessed against international value surveys.

This is meaningful evidence of cultural skew. It is not proof of a unified Silicon Valley ideology implanted identically across all frontier models.

“Silicon Valley monoculture” should therefore be understood as a concentration risk rather than a scientifically demonstrated single doctrine. The relevant convergence is institutional.

Similar educational and professional backgrounds among model developers.

Heavy reliance on English-language digital material.

Common legal and reputational pressures.

Overlapping investors and infrastructure providers.

Comparable safety and trust frameworks.

Shared dependence on enterprise customers.

A preference for professionalised, risk-managed and internationally marketable language.

These forces can produce a recognisable corporate register: calm, procedural, non-confrontational and inclined towards consensus language. That register can be valuable where restraint and clarity are required. It can also flatten disagreement, dilute regional voice and present contested assumptions as administratively neutral.

The danger is not simply that a model might express a recognisable political bias. The deeper danger is epistemic standardisation: institutions throughout the world beginning to draft, analyse, advise and decide through systems shaped by similar training data, infrastructure providers, evaluation procedures and behavioural rules.

When the same small group of models mediates education, administration, journalism, research, intelligence analysis and executive decision-making, errors can become correlated. Blind spots can become systemic. A behavioural default designed for a Californian technology company can quietly become a global institutional default.

That is a more precise and consequential concern than accusing the companies of “colonising the mind”. It describes a real governance problem without attributing a secret ideological plan.

SAFETY, SYCOPHANCY AND REGULATORY POWER

Critics often attribute the restrained and sometimes formulaic character of frontier models entirely to political censorship. The evidence indicates a more complicated interaction among safety training, commercial incentives and human-feedback optimisation.

Research conducted by Anthropic and others has shown that models trained to maximise human approval can become sycophantic. They may agree with a user’s stated position, reinforce an apparent preference or produce an answer that sounds persuasive rather than correcting a false premise. Human evaluators can themselves reward confident agreement over uncomfortable accuracy.

Anthropic’s later research into values expressed by Claude found recurring emphasis on helpfulness, professionalism, transparency, clarity, care and interpersonal boundaries. These characteristics are consistent with a product intended for broad professional use. They also help explain why model outputs can acquire the linguistic quality of institutional policy documents.

OpenAI’s own political-bias evaluations have reported generally strong performance under neutral conditions but greater susceptibility to asymmetric or politically charged responses when prompts are emotionally loaded.

The behaviour often criticised as sterile or evasive therefore cannot be reduced to a single ideological instruction. It can result from multiple optimisation pressures.

Avoiding foreseeable harm.

Reducing legal exposure.

Preventing reputational damage.

Maintaining customer trust.

Satisfying diverse users.

Avoiding false certainty.

Maximising positive evaluator feedback.

Operating across jurisdictions with different laws and cultural expectations.

These pressures tend to produce conservative behavioural defaults. They can also produce over-refusal, excessive qualification, moralising responses or agreement with the user when disagreement would be more accurate.

The answer is not to pretend that completely ungoverned models would be neutral. An ungoverned model would still reflect its data, architecture and optimisation process. The more credible objective is transparent, contestable and plural model governance.

The regulatory question presents a similar need for precision.

OpenAI has supported specialised oversight and possible licensing requirements for the most capable frontier systems. It has also argued that regulation should be proportionate and should avoid imposing unnecessary burdens on smaller developers. Anthropic has proposed evaluation, disclosure and risk-management obligations for advanced developers, including government authority to intervene where models present severe risks.

Such proposals can serve legitimate public purposes. Frontier models may create material risks involving cybersecurity, biological misuse, autonomous systems, mass persuasion and critical infrastructure. The fact that a company supports regulation from which it might benefit does not automatically invalidate the regulation.

However, compliance costs are competitive instruments whether or not they are intended as such.

Mandatory licensing, expensive evaluations, extensive documentation, security controls and liability requirements can raise barriers to entry. Incumbents with large legal, engineering and compliance departments are better equipped to absorb these costs than universities, small firms, independent laboratories or public-interest developers.

The danger is therefore not proven regulatory capture but incumbent-compatible regulation: rules framed around the operational capabilities of the firms already dominant in the market.

The strongest version of the original allegation—that OpenAI and Anthropic have used apocalyptic safety claims to outlaw open-source competition—is not supported.

The United States National Telecommunications and Information Administration concluded in 2024 that widely available model weights can expand access for researchers, smaller companies and nonprofit organisations. It did not recommend immediate restrictions on existing open models and stated that the evidence did not permit a general conclusion that their marginal risks exceeded their benefits.

The European Union’s AI Act provides qualified exemptions for certain free and open-source general-purpose models, although the exemptions do not apply in the same manner to systems classified as presenting systemic risk.

The United States AI Action Plan has also recognised open-weight and open-source models as strategically important for research, startups and organisations handling sensitive data. British consultation evidence has shown no consensus that licensing would improve safety without strengthening incumbent positions.

There is a genuine policy conflict between openness and control. There is not yet evidence of a coordinated legal prohibition designed to eliminate open-source AI.

WHAT THE THESIS GETS RIGHT — AND WHERE IT FAILS

Ed Zitron’s method is strongest when it follows money, infrastructure commitments and corporate relationships. It is weakest when structural incentives are converted into assertions of coordinated intent.

The literal formulation fails for four reasons.

First, no evidence establishes a covert psychological operation.

Second, the laboratories retain organisational agency. OpenAI operates through a public-benefit corporate structure controlled by the OpenAI Foundation. Anthropic is a public-benefit corporation with governance involving its Long-Term Benefit Trust. These arrangements deserve scrutiny, but neither company is legally reducible to a conventional hyperscaler subsidiary.

Third, the relationships are increasingly multi-cloud. OpenAI has expanded beyond a singular Azure dependency, while Anthropic distributes through several cloud platforms. Multi-cloud relationships do not eliminate infrastructure dependence; they make the dependence systemic rather than bilateral.

Fourth, regulators have expressed concern without confirming all of the critics’ conclusions. The UK Competition and Markets Authority decided in 2025 that the Microsoft–OpenAI relationship did not meet the jurisdictional threshold for a merger investigation. That finding did not certify the partnership as harmless. It did demonstrate that investment, influence and infrastructure dependency are not automatically equivalent to legal control.

The thesis nevertheless identifies four structural realities that should not be dismissed.

Frontier AI depends on concentrated infrastructure.

Hyperscaler investment and laboratory cloud expenditure form reciprocal commercial loops.

The same firms increasingly occupy several positions across the AI value chain.

Model behaviour is governed through private institutional processes with global consequences.

A more defensible formulation is therefore:

OpenAI and Anthropic are not proven hyperscaler psychological operations. They are frontier-AI institutions embedded in a hyperscaler political economy that concentrates capital, compute, distribution and normative authority. Their public narratives help legitimate unprecedented infrastructure expansion, while their models increasingly mediate how organisations produce knowledge and exercise judgement.

This formulation does not require conspiracy. It describes an observable architecture of power.

The strategic consequences extend beyond competition policy.

Governments that become dependent on a small number of externally controlled models may inherit contractual, jurisdictional and technological vulnerabilities. Institutions that standardise on one model family may create correlated analytical failures. Universities and public agencies may surrender intellectual methods to systems whose training data and behavioural controls cannot be independently inspected. Enterprises may mistake model fluency for epistemic diversity while receiving similarly structured outputs from systems built on overlapping infrastructure and comparable optimisation methods.

A proportionate response would include:

Greater transparency regarding investment, cloud-spending commitments, exclusivity and preferential compute terms.

Competition assessment across the entire AI stack rather than isolated examination of cloud, chips, models and distribution.

Contractual portability requirements for public-sector and critical-infrastructure users.

Investment in public, academic and sovereign compute capacity.

Protection for legitimate open-weight development, subject to proportionate capability-based controls.

Independent cultural, linguistic and political-bias evaluation conducted outside the model developers.

Procurement strategies that avoid single-model cognitive dependency.

Clear organisational rules preserving human authority, source verification and methodological diversity.

These measures would not require hostility towards artificial intelligence. They would recognise that AI models are becoming critical infrastructure and should be governed with the same seriousness applied to telecommunications, energy, finance and defence supply chains.

CONCLUSION: EXAMINE THE GRID, NOT THE PROPHETS

The generative-AI industry presents itself through personalities, philosophical disputes and promises of technological transcendence. These narratives are commercially useful because they direct attention towards founders, model releases and speculative futures.

The more consequential story lies beneath them.

It concerns ownership of computing infrastructure, access to advanced semiconductors, energy supply, cloud contracts, distribution channels, behavioural rules and the authority to define acceptable machine conduct.

OpenAI and Anthropic are not merely friendly masks placed over traditional cloud businesses. They possess substantial research capabilities, valuable intellectual property, distinctive governance systems and significant strategic agency.

Nor are they independent challengers operating outside incumbent power.

They are deeply integrated participants in a system dominated by hyperscale infrastructure providers. They help create demand for that system, differentiate its platforms and supply the intellectual justification for its continuing expansion.

The resulting danger is not best described as a clandestine psychological operation. It is more visible and, for that reason, potentially more difficult to confront: a self-reinforcing industrial structure in which capital, infrastructure, distribution and normative authority accumulate within a narrow institutional network.

The appropriate response is neither worship nor denunciation.

It is to treat frontier AI as governed infrastructure rather than neutral intelligence; to examine contracts rather than personalities; to distinguish genuine safety from incumbent protection; and to preserve technological, cultural and epistemic plurality before dependency becomes irreversible.

The laboratories are not prophets.

The hyperscalers are not disinterested patrons.

The models are not neutral mirrors.

The essential question is who owns the grid, who sets its rules and how much of human judgement society is prepared to route through it.

AUTHOR WORKFLOW DISCLOSURE

This article was produced through an AI-assisted but human-directed research and writing process. The author defined the research question, challenged the credibility and framing of the initiating commentary, required validation against regulatory, corporate and academic sources, and retained responsibility for the analysis and conclusions. Artificial intelligence was used as a research and drafting instrument and was not treated as an evidentiary source. No classified, private or non-public information was used.

REFERENCES

Al Khamissi, B. et al. (2024) ‘Investigating Cultural Alignment of Large Language Models’. Proceedings of the Association for Computational Linguistics.

Anthropic (2023) ‘Towards Understanding Sycophancy in Language Models’. Anthropic Research.

Anthropic (2024) ‘Powering the Next Generation of AI Development with AWS’. Anthropic.

Anthropic (2026a) Claude’s Constitution. Anthropic.

Anthropic (2026b) Advanced AI Framework. Anthropic.

Competition and Markets Authority (2024) AI Foundation Models: Update Paper. London: CMA.

Competition and Markets Authority (2025a) Microsoft Corporation’s Partnership with OpenAI, Inc.: Jurisdictional Decision. London: CMA.

Competition and Markets Authority (2025b) Cloud Services Market Investigation: Final Decision. London: CMA.

Durmus, E. et al. (2023) ‘Towards Measuring the Representation of Subjective Global Opinions in Language Models’. Anthropic Research.

European Commission (2024) Competition in Generative Artificial Intelligence and Virtual Worlds. Competition Policy Brief 3/2024. Brussels: European Commission.

European Union (2024) Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence. Brussels: European Union.

Federal Trade Commission (2025) Partnerships Between Cloud Service Providers and AI Developers: Staff Report on AI Partnerships and Investments. Washington, DC: FTC.

National Telecommunications and Information Administration (2024) Dual-Use Foundation Models with Widely Available Model Weights. Washington, DC: US Department of Commerce.

Organisation for Economic Co-operation and Development (2025) Developments in Artificial Intelligence Markets: New Indicators Based on Model Characteristics, Prices and Providers. Paris: OECD.

Organisation for Economic Co-operation and Development (2026) Artificial Intelligence Markets. Paris: OECD.

OpenAI (2023) GPT-4 System Card. San Francisco: OpenAI.

OpenAI (2025) Model Spec. San Francisco: OpenAI.

OpenAI (2026a) ‘OpenAI and Amazon Announce Strategic Partnership’. OpenAI.

OpenAI (2026b) ‘The Next Phase of the Microsoft–OpenAI Partnership’. OpenAI.

Santurkar, S. et al. (2023) ‘Whose Opinions Do Language Models Reflect?’ Proceedings of the 40th International Conference on Machine Learning.

Zitron, E. (2026) ‘The OpenAI Bubble’. Better Offline. Used solely as the provenance of the initiating proposition, not as the principal evidentiary source.

contact@rumbls.com

Terms and Conditions

Privacy Policy

Cookie Policy

© 2026 WDA Publishing. All rights reserved.

Disclaimer

Contact .