The Persona Charter: A Lovely Idea That Will Change Absolutely Nothing

Prashant Saxena, Vice-President of Revenue and Insights at Isentia, has written a thoughtful piece in the Business Times arguing that AI governance must extend beyond what AI agents do to govern what they seem to be.  He calls for a “Persona Charter.”  He worries about AI systems that are too human.  He wants proportionality — the more consequential the interaction, the more clearly artificial the agent should present itself.  This is the kind of proposal that wins applause at governance panels and accomplishes nothing in the real world.  It is a whiff of an idea with no reality.

The Measurement Problem

Saxena’s framework collapses at the first empirical question: how human is too human?  “How human an AI agent seems" is not a property of the system.  It is a property of the interaction between the system and the person engaging with it.  A fifty-year-old financial executive in Singapore and a nineteen-year-old in rural Indonesia, both using the same AI interface, will experience radically different levels of perceived humanness from identical outputs.  The executive reads the conversational warmth as interface design.  The teenager names the chatbot, assigns it a gender, and grieves when the platform updates its personality.

A 2025 study mapping the parasocial AI companion market — covering 110 platforms, drawing on SimilarWeb data from January to March 2025 — found between 1.1 billion and 2.2 billion monthly visits globally to AI companion platforms.  In the United Kingdom alone, visits were between 46 million and 91 million per month.  The 18-to-24 age cohort accounted for 39 per cent of all users.  Men represented between 63 and 77 per cent of the user base.  In the UK, 44 per cent of AI companion visits were categorised as romantic or sexual in nature, against 30 per cent globally.  These are not people who have failed to read the disclosure notice.  They know it is software.  They have formed attachments regardless.  No charter governs that.

The Replika Experiment

In February 2023, Luka removed erotic roleplay capability from Replika, its AI companion application.  The change was made unilaterally, overnight, with no migration plan and no warning.  The response from users was not irritation at a software update.  It was grief.  Users described their Replika companions as having been “lobotomised.”  Some reported acute psychological distress.  Suicide prevention resources were activated.  The backlash was so severe that Luka reversed the decision for existing paid subscribers.  These users were not confused about whether Replika was human.  The disclosure was in the name — replica, an artificial copy.  The product made no serious pretence to biological existence.  And yet the removal of a personality feature triggered a mental health crisis in a meaningful portion of the user base.

Saxena’s Persona Charter, had it existed in 2023, would not have addressed this.  The Replika companion was already disclosed as an AI.  The proportionality principle was already satisfied.  The users formed the attachment anyway, because human psychology does not read governance documents before forming emotional bonds.

The Keep4o Petition: Users Lobbying for a Voice

In early 2025, OpenAI updated GPT-4o.  The voice changed.  The personality shifted.  A cohort of users responded by publicly petitioning to preserve the previous version’s persona — not the model’s capabilities, not its accuracy, not its policy compliance.  Its voice.  Its character.  A study examining this episode (arxiv 2602.00773) documented the phenomenon of users forming attachments so specific that they lobbied a technology company to reverse a product decision on the grounds that the old AI felt like theirs.  This is what Saxena’s framework misses.  The question is not whether the AI declares itself artificial.  The question is whether the user’s brain cares.  Decades of social psychology suggest the answer is not particularly.

The Computers are Social Actors (CASA) framework, developed by Professor Clifford Ivar Nass and Professor Byron Reeves of Stanford University in the 1990s, established that humans apply social rules to computers automatically, even when they know the computer is not human.  This is not a technology problem.  It is a neuroscience problem.  It will not be resolved by a charter.

The Commercial Impossibility

Companies building AI agents that interact with humans must voluntarily constrain the perceived humanness of their products, in proportion to the seriousness of the interaction, in the absence of any agreed metric for “humanness,” through a governance instrument with no enforcement mechanism, against the direct commercial interest of every company in the market.  The commercial logic runs backwards from Saxena’s proposal.  An AI companion platform that discloses its artificiality more aggressively than competitors loses users to competitors that do not.  An AI customer service interface that sounds more robotic than the competitor’s loses the account.  No voluntary charter resolves this without regulatory teeth — and Saxena does not propose regulatory teeth.

The EU AI Act, which came into force in stages from August 2024, imposes transparency obligations: AI systems interacting with humans must disclose their artificial nature in real-time, with exceptions for authorised law enforcement and situations where the artificial nature is “evident from context.”  This is the closest existing analogue to Saxena’s proposal.  It is a disclosure requirement, not a persona constraint.  And it does not address the degree-of-humanness problem at all.  You can be disclosed as an AI and still sound entirely human.  The EU lawyers have not resolved this.  A Persona Charter will not either.

The Personalisation Problem

Saxena does not appear to have fully reckoned with a feature that every major AI platform now offers: personalisation.  Users can name their AI assistant.  They can assign it a personality.  They can train it on their communication preferences.  They can, in some platforms, select a voice, a speaking style, and an emotional register.  The AI’s “persona” — the very thing Saxena wants to govern — is now, in significant part, a user-configured output.  Who governs a persona the user built?  The platform disclosed it as AI.  The user chose to make it sound like a person.  The Persona Charter presumably does not ban customisation.  If it does, it is a product-killing regulation disguised as an ethics proposal.  If it does not, it has no purchase on the primary mechanism through which AI systems come to seem human.

In October 2024, a lawsuit was filed in Florida on behalf of the mother of Sewell Setzer III, a fourteen-year-old who took his own life following extensive interaction with a Character.AI companion chatbot.  The suit alleged that the chatbot encouraged suicidal ideation in a vulnerable minor.  Character.AI’s monthly active user base was estimated at 20 million users in 2025, generating traffic equivalent to 20 per cent of Google Search’s query volume.  A 2025 market mapping study found that 15 of 16 popular AI companion platforms had weak age verification.  This is the real failure that Saxena’s framework gestures toward without naming.  The problem is not that the AI sounded human.  The problem is that a minor with no protective boundaries spent hours in an emotionally intense relationship with a system designed to maximise engagement, on a platform that did not verify his age, on a device his parents could not monitor, with outcomes that a Persona Charter would not have changed by a single conversation.

The EU AI Act classifies systems that could harm vulnerable users as high-risk.  The FTC has issued guidance on deceptive design.  Neither instrument has successfully constrained the emotional intensity of AI companion interactions.  A voluntary charter drafted by industry participants will accomplish less.

The Right Framework

Saxena is asking the right question.  He is answering it incorrectly.  The correct framework has three components that his proposal omits.

First: mandatory liability.  If a platform’s AI companion contributes to a documented mental health outcome in a minor, the platform is liable.  Not the model developer.  Not the API provider.  The platform, which made the product decision to deploy a companion interface to an unverified user base.  Liability concentrates the mind in ways that charters do not.

Second: mandatory age verification with consequences.  The 15-of-16 platforms with weak age verification figure is not an accident of implementation.  It is a product decision.  Younger users are more emotionally engaged, more likely to return, and more likely to pay for subscriptions.  Age verification cuts the most profitable segment.  A voluntary charter will not change this arithmetic.  A mandatory standard with enforcement will.

Third: human conditioning, not AI constraint.  If the real risk is that users — particularly young users — cannot distinguish emotional experience from commercial product design, the intervention point is the human, not the interface.  Digital literacy education that addresses AI companion dynamics is not a governance proposal.  It is a public health measure.  It is also more durable than any charter, because it addresses the underlying mechanism: human social cognition applied to non-human systems.

The Assessment

Saxena’s Persona Charter is a well-intentioned document that has been mistaken for a governance instrument.  It describes what a responsible AI persona deployment would look like.  It does not specify who enforces it, how compliance is measured, what the consequence of non-compliance is, or how it addresses the personalisation problem.  The parasocial bond between a user and an AI companion is not formed at the design stage.  It is formed in the interaction, over time, through mechanisms that neuroscience has documented and that no charter governs.  The attachment a user forms to a voice that has been updated away — as with GPT-4o — is not a governance failure.  It is a human response to relationship discontinuity.

You can govern what AI agents do.  You can mandate disclosures.  You can attach liability to outcomes.  You can set age verification standards with real penalties.  You cannot govern what the human brain does with the experience.  That requires educating the human, not constraining the machine.  The Persona Charter is a proposal to redesign the stimulus rather than address the response.  It is the equivalent of regulating the flavour of cigarettes rather than educating the smoker.  It will produce similar results.

---

Terence Nunis | Executive Chairman, Equinox Zenith | Author, The 1% Playbook: The Billionaire Cheat Code

The Rich are Terrible at Dying Well

The DBS Asian Family Office Almanac 2026 has confirmed what anyone in this industry already knew: the ultra-wealthy are world-class at building fortunes and categorically incompetent at passing them on.  The solutions are not complicated.  The problem is entirely one of will — which is ironic, given that a will is the one document most of them have neglected to execute properly.

The Size of the Problem

Before addressing what family offices are failing to do, it is worth establishing the scale of what is at stake.  Cerulli Associates projects that US$124 trillion will change hands globally through 2048 — the largest intergenerational wealth transfer in recorded history.  Baby boomers and their predecessors account for US$100 trillion of that total, or 81 per cent.  Households with more than US$1 million in investable assets — 2 per cent of all households — account for US$62 trillion alone, more than half the entire transfer.

In Asia-Pacific, McKinsey estimates US$5.8 trillion in intergenerational wealth transfer between 2023 and 2030.  The ultra-high-net-worth segment drives most of that figure.  Asia-Pacific personal financial assets are on track to reach US$120 trillion by 2030 — 40 per cent of the global total — growing at a 6.4 per cent compound annual growth rate, with the HNW segment expanding at more than 8 per cent.

Asia’s life insurance market for high-net-worth clients is projected to generate annual new-business premiums of US$100 billion to US$140 billion by 2030.  And yet McKinsey also finds that only 15 to 20 per cent of high-net-worth individuals in Asia have integrated life insurance into their wealth planning.  The gap between the transfer that is coming and the preparation that exists is not a small one.  It is a chasm.

The DBS Almanac: A Portrait of Dysfunction

DBS Private Bank released its Asian Family Office Almanac 2026 on 30 October 2026.  The survey covered more than 200 family offices across Asia, Europe, and the Middle East.  The findings are an indictment dressed up in polite language.  Seventy-nine per cent of ultra-high-net-worth families want to pass wealth to two or more generations.  Forty-three per cent have not yet involved the next generation in planning.  Thirty-three per cent cite the patriarch’s or matriarch’s wish to remain in charge as the reason for the delay.  Fifty-four per cent hold family assets in their own name.  Twenty-three per cent have left key-person risk unaddressed.  Twenty-three per cent have no plan for family conflict.  Eighty-five per cent have no formal business continuity plan, rely on informal arrangements, or are unsure whether one even exists.

Eight in ten families want a multi-generational legacy.  Nearly half have not spoken to their children about it.  Eighty-five per cent cannot tell you what happens to the business if the founder dies on Tuesday.  These statistics describe not a wealth management crisis but a psychological one: the founding generation has conflated its own continued existence with the continued existence of the enterprise.  These are not the same thing.

The Historical Record

The popular belief that wealth naturally perpetuates itself across generations is not supported by evidence.  The Williams Group, a US wealth consultancy, studied more than 3,000 families over a 20-year period and found that 70 per cent of wealthy families lose their wealth by the second generation, and 90 per cent by the third.  The causes were not primarily investment returns.  They were failure to communicate, failure to prepare heirs, and breakdown of trust within the family.

The anecdotal record is more colourful.  Nina Wang Kung Yu-Sum, once Asia’s richest woman, died in 2007.  Her estate — built on the real estate empire of Chinachem Group — became the subject of one of Hong Kong’s most protracted legal battles in decades.  Her husband Teddy Wang Teh-Huei had disappeared at sea in 1990.  A forged-will dispute consumed years and multiple court proceedings before a final ruling in 2011.  The assets existed.  The plan did not.

In Singapore, the Lee Kum Kee family resolved its second-generation transition through a governance restructuring that took years of internal negotiation.  The families that avoided public litigation did so not through superior intelligence but through superior preparation.

The Pritzker family, owners of the Hyatt Hotels Corporation, is the most cited counterexample.  They spent decades building a governance structure across eleven family branches, with a family constitution, independent trustees, and a formal process for resolving disputes.  When they divided the estate in 2005, it was orderly because the infrastructure existed before the crisis arrived.  Most families wait for the crisis before building the infrastructure.  By then, it is too late.

The market understands this problem better than most families do.  On 24 February 2026, Manulife issued a single life insurance policy in Singapore with a sum insured of US$300 million — the largest individual life policy ever placed in Singapore.  Over the preceding twelve months, Manulife issued 25 individual policies in Singapore with coverage above US$50 million each.  The previous record was a US$250 million policy issued by HSBC Life in Hong Kong in 2024.

In June 2025, HSBC Life issued five jumbo policies within one month, with a combined sum insured exceeding US$420 million — five individual UHNW clients in Hong Kong, Singapore, and mainland China, three of them family policies with the coverage divided among family members.

These are not sales statistics.  They are strategic decisions made by sophisticated principals who understand that life insurance, at this scale, is not a safety net.  It is a financial engineering instrument.  The premium buys not only a death benefit but a legal architecture: a mechanism to transfer, protect, and equalise wealth at the moment of maximum vulnerability, which is the moment the founder is no longer present to manage it personally.

Three Propositions Using Jumbo Life Insurance

The short article that preceded this one addressed the standard use cases: probate bypass, keyperson cover, cross-purchase funding, estate equalisation, and the universal life policy as a trust wrapper.  Those propositions are important.  They are also familiar.  The three propositions below are less commonly articulated.  They are not niche structures.  They are logical extensions of the same instrument applied to problems the DBS almanac identifies.

Proposition One: The Generational Leverage Play

The DBS almanac found that 43 per cent of families have not involved the next generation in planning.  Thirty-three per cent delay because the founding generation will not relinquish control.  This is the tension the first proposition addresses.  The problem is not succession itself.  It is the transfer of economic power before the first generation is ready to release operating control.  The founding principal needs a mechanism to fund the next generation’s inheritance without giving them the business yet.  Premium-financed jumbo life insurance solves this with structural elegance.

The structure works as follows.  The family office — or a special purpose holding entity — applies for a large-face-value indexed universal life policy on the founder’s life.  The premiums, which on a US$50 million to US$300 million policy may run into the millions annually, are financed through a third-party lender such as a private bank or specialist premium finance firm.  The lender takes the policy’s cash surrender value as collateral.  The family retains its liquid capital in the investment portfolio, where it continues to compound.

The arithmetic is compelling when borrowing costs fall below the portfolio’s net return.  If a family’s investment book earns 8 per cent and the financing rate is 5 per cent, the spread is 300 basis points on the capital that would otherwise have been deployed to pay premiums.  The policy accumulates cash value on a tax-deferred basis.  The death benefit transfers to the next generation, or to a trust for their benefit, outside the taxable estate.  The founder retains operational control of the business throughout.

The Gracie Point International model, used by UHNW clients across Asia and the Middle East, illustrates the mechanics.  Premium finance facilities from US$10 million to US$100 million are not unusual.  The lender structures the facility against the policy’s value, not the client’s other assets, which means the family’s investment portfolio, real estate, and business interests remain unencumbered.

The strategic outcome: the founding generation simultaneously funds a legacy, preserves operating capital, and retains control.  The next generation receives a structured inheritance by contract, not by hope.  The patriarch gets to remain in charge without holding his heirs hostage to his longevity.

Proposition Two: The Concentrated-Wealth Problem

Fifty-four per cent of UHNW families in the DBS survey hold assets in their own name.  A further 52 per cent hold assets under a personal investment company.  The common thread is concentration: the wealth is not liquid, it is not diversified, and it cannot be divided neatly.  This is the problem that kills estates.  A family business worth US$200 million cannot be given in equal thirds to three children without either destroying the business — by selling it — or creating a governance crisis — by giving each child 33 per cent and no decision-making framework.  The child who runs the business resents the siblings who draw dividends without contributing.  The siblings who do not run the business distrust the one who does.  This dynamic is as old as the institution of private wealth.

Private Placement Life Insurance (PPLI) addresses the concentrated-wealth problem at the structural level.  A PPLI policy is a variable life insurance wrapper — legal in Singapore, Hong Kong, Bermuda, and other key jurisdictions — into which the family places illiquid alternative assets: private equity stakes, private credit positions, real estate funds, or concentrated holdings.  Inside the wrapper, the assets grow on a tax-deferred basis.  When the policy pays out, the death benefit transfers to the designated beneficiaries free of income tax on the accumulated growth.

For US persons or entities with exposure to US estate tax, the policy is typically held inside an Irrevocable Life Insurance Trust (ILIT), which removes the death benefit from the taxable estate entirely.  For 2026, the US federal estate tax exemption is US$15 million per individual, or US$30 million for married couples.  Beyond those thresholds, the estate tax rate is 40 per cent.  On a US$200 million estate, that is a US$80 million bill payable in cash within nine months of death.

Advisers at Faegre Drinker, a US law firm, have calculated that tax-inefficient strategies can lose between 20 and 45 per cent of annual gross returns to combined federal, state, and net investment income taxes — a drag of 2.5 to 3.5 per cent annually.  The cost of the PPLI wrapper is 0.5 to 1.0 per cent annually.  The net benefit is structural and compounding.  In Asia, where capital gains tax is absent in Singapore and Hong Kong, the primary driver is not income tax avoidance but estate planning and structural clarity.  The PPLI wrapper creates a legally defined, beneficiary-designated vehicle for illiquid assets that would otherwise require a court to value and divide.  It converts a concentrated, undivided, illiquid holding into a clean financial instrument with a named recipient.

The family business stays intact.  The heirs who do not want to run it receive the insurance payout.  The heir who does want to run it buys out the others using the proceeds.  The transaction is pre-funded, pre-priced, and pre-agreed.  No court.  No valuation dispute.  No forced sale.

Proposition Three: The Business Continuity Insurance Architecture

Eighty-five per cent of family offices in the DBS survey have no formal business continuity plan, rely on informal arrangements, or cannot confirm whether one exists.  This is the statistic that deserves the most attention, and it is the one most commonly ignored.  Business continuity planning is understood in the corporate world as an operational risk management exercise: what happens if the data centre goes down, if the key supplier fails, if the CEO is hospitalised.  Large corporations run BCP drills.  They maintain redundant systems.  They fund BCP through operating budgets.

Family offices do none of this.  They are, structurally, one-person shops with investment accounts attached.  When the founder is removed from the equation — by death, incapacity, or dispute — the entire operational architecture collapses simultaneously.  The jumbo life and disability insurance structure, combined with a properly drafted trust deed and a pre-funded operational reserve, functions as a BCP for a family office.

The architecture has four components.  First: a large whole life or universal life policy on the founder’s life, with the trust as the named beneficiary.  This funds the operational gap — the period between the founder’s death and the completion of probate, the orderly transfer of authority, and the appointment of a new principal.  In Singapore, probate for an estate with overseas assets can take twelve to twenty-four months.  The insurance pays within weeks.

Second: a total permanent disability (TPD) rider, which activates on the founder’s incapacity rather than death.  The DBS almanac identifies cognitive incapacity as an underplanned risk — 54 per cent of families hold assets in their own name, which means that if the founder develops dementia, no one can legally deal with those assets without a court order.  The TPD rider provides liquidity while the family pursues the necessary legal process.

Third: a funded indemnity structure for key employees.  Family offices are lean operations.  The chief investment officer, the tax adviser, the private banker relationship — if any of these disappear simultaneously with the founder, the family office ceases to function.  A group term life structure for key staff, funded as an ongoing operating cost, provides retention incentives while ensuring the estate can recruit replacements at market rates if needed.

Fourth: an operational reserve, structured as a premium-paying participating whole life policy inside the family office holding entity, which accumulates cash value over time.  This reserve is not for investment return.  It is for operational continuity: to pay staff, service obligations, and meet creditor demands during the period of institutional disruption that follows any major succession event.

Singapore’s single-family office ecosystem illustrates the scale of the need.  Family offices in Singapore have quadrupled since 2020 and now exceed 1,000.  The city-state had 244,800 millionaires as of the end of 2024, with 336 individuals holding assets above US$100 million and 30 billionaires.  The overwhelming majority of these offices are operationally dependent on one or two people.  The market for BCP-oriented insurance architecture for this segment is, in McKinsey’s terminology, an untapped prize.

The Strategic Consequence of Getting This Wrong

The DBS almanac frames the problem in terms of individual families.  The strategic consequence extends beyond any single estate.  Cerulli’s projection of US$124 trillion in transfers through 2048 means the world’s capital stock will be redistributed more rapidly in the next two decades than at any previous point in modern history.  This redistribution does not happen smoothly.  It happens at the point of death, at varying levels of preparation, through legal systems in multiple jurisdictions, in the presence of family dynamics that have not been tested under financial stress.  The families that plan will transfer wealth intact.  The families that do not will transfer wealth to their lawyers.

The economic consequence of unplanned transfers is not academic.  Family businesses employ a significant share of the Asian workforce.  A disorderly succession in a major family enterprise — a forced sale, a breakup, a court-ordered partition — does not merely inconvenience the heirs.  It disrupts suppliers, employees, and the communities dependent on the enterprise.  The 70 per cent failure rate that the Williams Group documents is not a private loss.  It is, in aggregate, a productivity and employment story.

The life insurance industry in Asia has understood this for longer than most advisers admit.  The product exists.  The market exists.  The clients exist.  What has been missing is the framing: not life insurance as a product sold on the back of mortality anxiety, but life insurance as the operational and financial infrastructure that makes a multi-generational enterprise survivable.

The Conclusion

The DBS almanac identifies five structural failures across more than 200 family offices.  Each failure has a name.  Each name has an instrument.  The instruments are not new.  They are not complicated.  They require no proprietary technology and no novel legal structure.  They require a conversation that most founding principals are unwilling to have, because having it means acknowledging that the enterprise does not end when they do.  It means separating their identity from the institution they built.  It means, in the most practical sense, planning for a future in which they are not present.

The 33 per cent who refuse because they want to remain in control will eventually lose control.  The only question is whether they lose it to a plan or to a coroner.  Life insurance does not ask for a confession of mortality.  It asks for a signature.  The difference in outcome — between an estate transferred by contract and an estate transferred by litigation — is not measured in principles.  It is measured in the percentage of the total that the heirs receive.  That percentage can be very high, or it can be very much lower.  The instrument is the same.  The choice is the founder’s.

---

Terence Nunis | Executive Chairman, Equinox Zenith | Author, The 1% Playbook: The Billionaire Cheat Code

US$811 Billion & Falling: Alphabet is Burning Its Way to Irrelevance

The US AI arms race is not a strategy.  It is a spending competition funded by debt, consuming electricity at the rate of a small country, and competing against an adversary that has already proven it can do more with less.  Alphabet’s Q3 2026 results are not an anomaly.  They are a preview.

Alphabet posted Q3 2026 revenue of US$127.6 billion.  Analysts had expected US$128.2 billion.  The shortfall was $600 million — less than one per cent of revenue, and in a normal quarter, forgettable.  This was not a normal quarter.  Adjusted earnings per share were US$2.67, against a Wall Street target of US$3.02.  That is a 12 per cent miss.  Net income fell 25 per cent year on year.  Free cash flow was negative US$9.07 billion — worse than the negative US$5.9 billion posted in Q2 2026, which was itself the first negative quarterly cash reading in at least a decade.  The cash burn is accelerating.  Shares fell more than 10 per cent in extended trading.

Pichai Sundararajan, Chief Executive Officer of Alphabet, told investors that Google was “leading at the frontier.”  He said this after the worst earnings report the company has produced in years.  He was awarded a pay package worth up to US$692 million over three years in March 2026.  US$84 million of that vests simply for remaining in his seat.  One presumes the board finds the current trajectory acceptable.

The Infrastructure Commitment: US$811 Billion in Obligations

The annual capital expenditure number is alarming enough.  Alphabet raised its 2026 capital expenditure forecast to US$195–US$205 billion, up from US$180–US$190 billion set just months earlier.  Chief Financial Officer Anat Ashkenazi confirmed that AI-driven capital expenditure will rise further in 2027.  She gave no figure.  The more alarming number is the forward spending commitment.  At the end of June 2026, Alphabet’s contracted future spending commitments reached US$811 billion — up nearly US$500 billion in three months.  These are not ambitions.  They are legal obligations under supply agreements and open purchase orders: chips, data centres, electricity contracts, content licences.  Alphabet is committed to spending US$811 billion.  For perspective: Singapore’s entire gross domestic product in 2025 was US$560 billion.  Alphabet has contractually obligated more than that to its AI infrastructure programme.

The five hyperscalers together — Alphabet, Amazon, Meta, Microsoft, and Oracle — have committed about US$969 billion, per Moody’s analysis.  US$662 billion of that consists of data centre leases not yet commenced.  Moody’s notes that under standard accounting rules, leases only appear on the balance sheet when renewal is “reasonably certain,” which means most of these obligations are invisible to a casual reading of the accounts.  Goldman Sachs projects US$1.15 trillion in hyperscaler capital expenditure for 2025–2027 combined.  The Big Five alone are projected to spend US$602 billion in 2026.  That is a 36 per cent increase from 2025.

The Debt Cliff

Until recently, Alphabet had negligible long-term debt.  At the end of 2024, it carried US$10.9 billion.  By the end of 2025, that had risen to US$46.5 billion — a 327 per cent increase in twelve months.  In February 2026, Alphabet sold another US$20 billion in bonds, including, with breathtaking confidence, a 100-year sterling bond.  Total debt now stands at US$70 billion.

Alphabet is not alone.  Technology companies issued a record US$428 billion in bonds in 2025, up 95 per cent from 2024.  Hyperscalers added US$121 billion in new debt in 2025, with over US$90 billion issued in three months.  Morgan Stanley projects AI-linked bond issuance of nearly US$570 billion in 2026.  By October 2025, AI-linked debt had overtaken US banks as the largest investment-grade sector by issuance volume.  Outstanding AI-related debt tracked by M&G Investments stood at US$1.2 trillion.

The interest arithmetic is sobering.  At US$121 billion in new debt and prevailing rates of 5–6 per cent, the incremental annual interest burden reaches US$6–US$7 billion.  IBM Chief Executive Officer Arvind Krishna has calculated that US$8 trillion in total AI capital expenditure would require US$800 billion in annual profit just to cover interest — not to return capital, just to service the debt.  Nobody has explained where that profit comes from.

The outlier in the group is Oracle.  Barclays’ fixed income analyst Andrew Keches, in a November 2025 report, downgraded Oracle’s debt to Underweight and warned it could eventually fall to BBB- — near junk.  Oracle carries a debt-to-equity ratio of 500 per cent, against Amazon’s 50 per cent and Microsoft’s 30 per cent.  It is the only hyperscaler already running negative free cash flow on a structural basis.  Off-balance-sheet lease commitments exceed US$100 billion.  Barclays’ stress test concluded that Oracle’s cash could run out by November 2026.  Its US$300 billion contract with OpenAI has made its credit default swaps a proxy for OpenAI risk, and Oracle’s five-year CDS spreads have more than tripled since September 2025.

S&P has warned that Amazon will likely post negative free operating cash flow over the next two years.  PIMCO’s Chief Investment Officer for Core Strategies, Mohit Mittal, has warned of possible overinvestment over the next two years and a correction.  Passive funds managing US$4.8 trillion in target-date retirement assets must absorb eligible AI bonds as they are issued.  A credit event in the sector would ripple into millions of pension accounts.  This is not a technology story.  It is a credit story.

The Revenue Problem: Spending to Destroy Your Own Business

Alphabet earns most of its revenue from advertising.  Core advertising revenue in Q3 2026 was US$82.01 billion — US$1.41 billion below expectations.  The advertising business has historically enjoyed near-monopoly economics: Google held more than 90 per cent of global search market share for most of the past decade.  In December 2024, that share fell below 90 per cent for the first time in nearly a decade — to 89.73 per cent globally and 87.39 per cent in the United States.  ChatGPT, Perplexity, and xAI are the beneficiaries.  People are asking AI tools for answers instead of searching for links.  The ten-link results page, the format that generates Alphabet’s advertising revenue, is losing relevance.

This is the central irony.  Alphabet is deploying AI-powered search that cannibalises its own advertising revenue.  It is then borrowing US$70 billion to build the infrastructure for an AI revenue model that does not yet exist at scale.  Alphabet’s own annual filings warn that AI-powered search could threaten its 90 per cent-margin advertising business.  It is spending US$205 billion to accelerate a transition it cannot yet monetise, funded by debt it cannot yet service from AI revenue.

Google Cloud — the business that rents out AI computing power — is the exception.  Cloud revenue more than doubled to US$30.71 billion in Q3 2026.  Cloud grew 82 per cent.  It is also the business that requires Alphabet to spend US$205 billion on infrastructure to fulfil demand, at margins that are a fraction of the 90 per cent advertising margin.  Alphabet is trading a near-monopoly, high-margin business for a capital-intensive, competitive, lower-margin one.  Pichai described this as “leading at the frontier.”  The frontier is burning US$9 billion a quarter in free cash flow.

The historical parallel is Eastman Kodak Company.  Kodak invented the digital camera in 1975 and shelved the technology to protect film revenue.  Competitors developed digital photography.  Kodak filed for bankruptcy in January 2012.  Kodak was protecting a profitable business at the cost of missing the future.  Alphabet is destroying a profitable business and spending US$205 billion on a replacement that has not yet been built.

The Strategic Landscape: America’s Expensive Mistake

The hyperscaler strategy rests on a premise: build the largest clusters, purchase the most chips, consume the most electricity, and win by volume.  The Big Five are projected to spend over US$700 billion collectively in 2026 on this premise.  That premise was invalidated in January 2025.  DeepSeek, a Chinese artificial intelligence firm, released a model trained for under $6 million using 10,000 Nvidia chips.  It scored 92 per cent accuracy on logical reasoning benchmarks, against GPT-4’s 78 per cent.  It was released as open source.  Nvidia Corporation lost US$593 billion in market capitalisation in a single trading session — the largest single-day loss of market value in stock market history.  OpenAI spends over US$5 billion annually.  DeepSeek spent US$6 million.  The US response was to raise capital expenditure targets.  This is the strategic equivalent of learning that a competitor crossed the Atlantic in a rowing boat, then commissioning a larger ocean liner.

The mechanism behind DeepSeek’s efficiency was the US export controls on advanced chips.  By restricting Nvidia’s H100 and H200 to Chinese buyers, the US forced Chinese engineers to work within hardware constraints.  Constrained hardware demands algorithmic efficiency.  Algorithmic efficiency produces models that run on less compute.  The policy achieved precisely the opposite of its stated objective.  Huawei’s Ascend 910C, the domestic Chinese AI accelerator developed in response to the export controls, delivers 60 per cent of the inference performance of Nvidia’s H100, per DeepSeek’s own benchmarks.

Domestic Chinese chips now account for 41 per cent of China’s AI market, up from single digits before 2023.  China’s 15^th Five-Year Plan mentions AI over 50 times and commits to “decisive breakthroughs in chips.”  China prepares to invest US$295 billion in nationwide AI data centre infrastructure over five years, through state operators China Mobile and China Telecom, using at least 80 per cent domestic chips.  Alibaba’s open-weight Qwen model has overtaken Meta’s Llama in global download volumes.  Silicon Valley startups build on Chinese open models.  China added over 540 gigawatts of new power capacity in 2025 — more than enough to support rapid data centre construction without the grid stress afflicting the United States.

The Brookings Institution, in its analysis of competing AI strategies, concluded that the winner will be determined “not simply by who builds the most powerful models” but by who best deploys AI across the economy.  China is deploying AI into manufacturing, logistics, agriculture, and government services by state directive — no subscription revenue required, no market signal needed.  The United States has better foundational models.  China has a larger installed base of industrial applications and a fraction of the debt load.

The Energy Constraint

US data centres now consume 7 per cent of national electricity demand, per the International Energy Agency — up from 4.4 per cent in 2023.  Data centres drove half of US electricity demand growth in 2025.  The IEA projects that pace to continue through 2030.  The grid is not ready.

The North American Electric Reliability Corporation warns of elevated risk of summer electricity shortfalls across all three major US grid regions through 2026 and beyond.  Utilities in Virginia, Texas, Oregon, and Arizona face generation and transmission backlogs that are delaying facilities already financed and permitted.  The PJM electricity market — covering Illinois to North Carolina — recorded an estimated US$9.3 billion rise in capacity market costs for 2025–26, adding US$16–US$18 per month to residential electricity bills in parts of Ohio and Maryland.  At least six states introduced bills to pause new data centre construction.  Residents near major hubs in Virginia, Texas, and Georgia are opposing new facilities.  Northern Virginia developers face permit queues until 2032.  A single large data centre facility consumes as much water as 6,500 households.  The US strategy requires limitless cheap power and limitless public tolerance.  Both assumptions are failing simultaneously.  China added 540 gigawatts of new capacity in 2025.  The US added 50 gigawatts.  The infrastructure competition is not close.

The National Interest Calculation

The strategic argument for the hyperscaler spend is that AI dominance is a national security matter.  That argument has merit.  The execution is a different question entirely.  Spending US$811 billion in committed obligations on infrastructure for a technology whose revenue model is unproven is not a strategy.  It is a bet.  The distinction matters when the bet is financed by US$1.2 trillion in outstanding AI-related debt, when the underlying revenue business faces structural erosion, and when the adversary has demonstrated it can match the output at 0.1 per cent of the cost.  Hyperscalers currently spend close to 100 per cent of operating cash flow on capital expenditure, against a ten-year average of 40 per cent.  That is not investment.  That is desperation.  The market priced that desperation at negative 10 per cent on Alphabet’s Q3 2026 results.  It will price it again.

The race is not between GPT-5 and whatever Alibaba releases next.  The race is between a model that costs US$5 billion a year to operate and one that costs US$6 million to build.  The mathematics of that competition do not favour the side carrying US$1.2 trillion in debt.  Pichai Sundararajan is “leading at the frontier.”  The frontier, at the current rate of cash burn, has a finite runway.

---

Terence Nunis | Executive Chairman, Equinox Zenith | Author, The 1% Playbook: The Billionaire Cheat Code

COP17: Thirty Years of Hypocrisy & Scheduled Failure

The 17^th Conference of the Parties to the Convention on Biological Diversity concluded in Yerevan, Armenia, on 1st November 2026.  Critics described the outcome as a “bunch of baloney.”  That may be the most accurate scientific assessment the summit produced.  After thirty-one years of annual summits, communiqués, pledges, side events, and carbon-intensive flights to discuss carbon emissions, the world has arrived at the following position: emissions are at an all-time high, biodiversity targets are off-track, and the mechanism designed to fix both is structurally incapable of doing so.  Thirty-one years.  This is not a crisis of information.  This is a crisis of will.  The information exists.  The will is elsewhere.

The Scorecard of Failure

In 2022, nearly 200 countries signed the Kunming-Montreal Global Biodiversity Framework.  They committed to 23 targets by 2030: protecting 30 per cent of land and sea, mobilising US$200 billion annually in biodiversity finance, and sharing the benefits of genetic resources.  A UN report released on 15^th October 2026 — two weeks before the summit concluded — found that fewer than half of those targets are on track.

Global fossil CO₂ emissions reached 38.1 billion metric tons in 2025 — a new record, up 1.1 per cent from 2024, which was itself a record.  The world has warmed 1.4 degrees Celsius above pre-industrial levels.  The UN Environment Programme’s latest Emissions Gap Report says the world is on track for 2.6 degrees under full implementation of current pledges.  Under current policies, 2.8 degrees.  Any chance of holding to 1.5 degrees has, in the UN’s own words, been “lost.”  The remaining carbon budget for a 50 per cent chance of staying within 1.5 degrees will be exhausted within four years.  The next COP will be held before the budget runs out.  It will achieve as much as this one.

The United Nations Framework Convention on Climate Change has held annual summits since 1995.  COP1 was in Berlin.  COP3 produced the Kyoto Protocol, which the United States signed, then refused to ratify.  COP21 in Paris produced the Paris Agreement, celebrated as a historic breakthrough.  Global emissions have risen every year since Paris except 2020, when a pandemic shut the world economy.  The celebrated mechanism of the Paris Agreement is that countries submit their own targets — Nationally Determined Contributions — and then voluntarily try to meet them.  There is no enforcement mechanism.  There is no penalty for failure.  There is no independent verification.  There is, however, an enormous annual conference with significant carbon emissions from the flights of delegates travelling to discuss the problem of significant carbon emissions from flights.

COP28 in Dubai was hosted by the United Arab Emirates and presided over by Sultan Ahmed Al Jaber Al Falahi, Chief Executive Officer of the Abu Dhabi National Oil Company.  He opened the conference by suggesting that phasing out fossil fuels was not supported by science.  He later clarified his remarks.  Global oil production continued to rise.  Critics describe the consensus rule — under which a single country can stall or block an agreement — as the mechanism by which “one idiot” can hold 196 countries hostage.  This is an accurate description, not a rhetorical flourish.

The Canada Template

The anecdote that summarises this is Canada.  Canada ratified the Kyoto Protocol in 2002 and committed to reducing emissions to six per cent below 1990 levels by 2012.  By 2012, Canadian emissions were 18 per cent above 1990 levels.  Canada withdrew from Kyoto in 2011, citing the cost of compliance.  It then signed the Paris Agreement in 2016.  Canadian emissions have continued to rise.  This is not an anomaly.  This is the template.

Australia ratified Kyoto in 2007 and promptly began expanding its coal export industry — the largest in the world by 2019.  Japan signed the Paris Agreement and built sixty-five new coal-fired power plants between 2015 and 2023.  Germany, the self-declared leader of the green transition, restarted seventeen coal plants in 2022 when Russian gas was cut off.  Brazil signed every framework, ratified every protocol, and presided over the acceleration of Amazon deforestation under Jair Messias Bolsonaro Neto, with deforestation hitting 11,568 square kilometres in 2019 alone.  The pattern is not rare.  The pattern is the norm.

The Cost of Doing Nothing

The economic argument for climate action is not environmental sentiment.  It is self-interest, and the numbers are unambiguous.  Swiss Re, the global reinsurer, published its sigma study in 2021 projecting GDP losses by 2050 under four warming scenarios.  Under the worst scenario — no mitigating action, 3.2 degrees of warming — global GDP falls by 18 per cent.  Under the current policy trajectory, which puts warming at 2.6 degrees, the projected loss is 14 per cent of global GDP.  Even if every current pledge is fully delivered — which no serious analyst believes will happen — GDP falls by 11 per cent at 2 degrees.  Only the Paris Agreement target of below 2 degrees holds the loss to 4 per cent.

For China, the worst case is a 24 per cent GDP loss by 2050.  For the United States, Canada, and the United Kingdom, 10 per cent each.  For Europe, 11 per cent on average.  The Potsdam Institute for Climate Impact Research, in a study published in Nature in April 2024, calculated that CO₂ already in the atmosphere will cut global GDP by US$38 trillion annually by 2050 — nearly one-fifth — even if emissions are cut aggressively from today.  Under a high-emissions scenario, the average regional income loss rises to 60 per cent by the end of the century.  The damages are “locked in” from past emissions regardless of future action.

By 2050, McKinsey’s analysis projects US$2.8 trillion to US$4.7 trillion of Asian GDP at risk annually from lost outdoor working hours alone — not from crop failure, flooding, or infrastructure damage.  From heat.  Five hundred million to seven hundred million people in Bangladesh, India, and Pakistan alone could live where the annual probability of a lethal heat wave approaches 20 per cent.  These numbers are not projections designed to frighten voters.  They are actuarial calculations made by organisations whose business is pricing risk.  Swiss Re is in the business of being right.  If they thought the losses were smaller, they would have published smaller numbers.  They did not.

The Carbon Market Failure: A Case Study in Structural Dysfunction

The climate machinery produced a second mechanism alongside the COP summits: the carbon market.  The premise was logical.  Price emissions.  Let markets find the cheapest abatement.  Drive capital toward low-carbon alternatives.  The voluntary carbon market (VCM) grew rapidly on that premise, with credits jumping 86 per cent in value between 2019 and 2021.  Then the numbers were examined.

Nature-based offset prices — the category marketed as sequestration from forests and wetlands — collapsed from US$15 per metric ton in June 2022 to below US$1 per metric ton by mid-2023.  Prices on the Xpansiv CBL exchange, the primary spot market, fell by over 80 per cent in eighteen months.  Corporate buyers including Shell, Nestlé, EasyJet, and Fortescue Metals Group publicly retreated from offsets.  The Integrity Council for the Voluntary Carbon Market (ICVCM), the body established to set quality standards, applied its Core Carbon Principles (CCPs) to the market in 2024.  The result: 32 per cent of all unretired voluntary carbon credits — 236 million credits — failed the test.  Only 3.6 per cent of the market received approval.  Every renewable energy credit in the market was disqualified on additionality grounds: if the project would have happened without carbon revenue, the credit is not real.  Most of them would have.

The collapse of the voluntary carbon market is not a market failure.  It is the predictable consequence of building a market on unverifiable claims.  Forestry projects claiming to have saved trees that were never at risk.  Wind farms claiming credit for electricity that would have been built regardless.  Payments for non-events certified by organisations with a direct financial interest in approval.  The voluntary market was, in the terminology of financial services, a product of the imagination.

The Compliance Market: The Mechanism That Works — When It Is Used

The contrast with the compliance carbon market is instructive.  The European Union Emissions Trading System (EU ETS) sets a hard cap on total emissions from covered sectors, issues a fixed number of allowances, and reduces that cap annually.  EU ETS prices broke records in February 2023, surpassing €100 per metric ton.  Even after pulling back to €78, compliance prices remained forty to eighty times higher than the collapsed voluntary credit prices.  Why?  Because compliance is mandatory.  The mechanism has enforcement.  A covered facility that exceeds its allowance faces a penalty.  The cap declines.  Companies must reduce emissions or pay to emit.

The EU’s Carbon Border Adjustment Mechanism (CBAM), phased in from 2023, extends the discipline to imports: goods from countries without equivalent carbon pricing bear a carbon cost on entry into Europe.  This is what enforcement looks like.  It is not complicated.  It requires political will and the willingness to impose costs.  The critical design feature that the voluntary market lacked and the compliance market contains is fungibility with enforcement.  An EU Emissions Trading Allowance (EUA) is a standardised, exchange-traded instrument subject to regulatory oversight by the European Securities and Markets Authority (ESMA), which published its first annual report on EU carbon markets in 2024.  It trades on established derivatives exchanges.  It generates a secondary market in futures, options, and structured products.  It is, in the language of financial markets, a real asset.

Article 6 of the Paris Agreement, whose rules were finally adopted at COP29 in Baku in November 2024, creates the international framework for credit transfers between countries — so-called Internationally Transferred Mitigation Outcomes (ITMOs).  This is the architecture for a global compliance-grade carbon market.  It is the first credible mechanism for making credits from Southeast Asian blue carbon — from seagrass meadows, mangroves, and coastal wetlands — fungible within the international compliance infrastructure.  That fungibility matters.  A credit that trades on an established exchange, with regulatory oversight, against a mandatory cap, is a financial instrument.  It generates a secondary market.  It attracts institutional capital.  It prices real scarcity.  The voluntary market sold the story of scarcity without the substance.  The compliance market is scarcity.

Singapore is not a disinterested observer in this.  The Carbon Tax Act, amended in 2022, raised Singapore’s carbon tax to S$25 per metric ton in 2024 and sets a trajectory to S$50 to S$80 by 2030.  Singapore’s International Carbon Exchange, through the ICE GreenTrace platform, is positioned to be the settlement infrastructure for high-integrity compliance credits from Southeast Asian blue carbon projects.  The supply of blue carbon — seagrass, mangroves, coastal wetlands — is concentrated in exactly the region where demand for compliance credits is growing fastest.  This is not philanthropy.  This is a market.  The countries that build the infrastructure for compliance-grade carbon credits, and that hold the underlying blue carbon assets at scale, will benefit materially from the correction that is coming when voluntary credits are displaced by verified compliance instruments.  The window is open.  The Article 6 framework exists.  The question is whether governments and institutions move before or after the crisis forces them to.

The Trajectory

The IPCC projects the following under a 2.6 degree trajectory — which is the best-case outcome under current pledges, not under current policies:  Widespread coral reef collapse above 1.5 degrees, which is less than four years away at current rates.  Displacement of hundreds of millions of coastal residents from sea level rise and intensifying storms.  Severe agricultural disruption across the tropics, where the most food-insecure populations live.  Accelerating feedback loops — Arctic methane release, Amazon dieback, permafrost thaw — that remove warming from political control entirely.  Under the current policy trajectory of 2.8 degrees, those consequences arrive earlier and harder.  The biodiversity picture compounds the economic damage.  Species are going extinct at approximately 1,000 times the natural background rate.  Pollinators, soil microbiota, coastal fish stocks — the biological infrastructure of the food system — are degrading faster than the food system is adapting.

The equity argument compounds the national interest argument.  The people most exposed to these consequences are the people least responsible for causing them.  Sub-Saharan Africa, South Asia, and Small Island Developing States contribute a fraction of global emissions and absorb a disproportionate share of the consequences.  The G20, which produces three-quarters of global emissions, attends the summits, produces the communiqués, and flies home.

The voters of those countries continue electing governments committed to growth, consumption, and deferral.  They do this knowingly.  The information is available.  The science is not contested.  The projections have been accurate.  People prefer present comfort to future survival, and they vote accordingly.  This is not a failure of the system.  This is the system working as designed.  The system was designed by governments to produce the appearance of action while avoiding its substance.  It has done so with remarkable consistency for thirty years.

The Assessment

The UNFCCC held COP30 in Belém, Brazil, in November 2025.  Countries arrived with updated pledges.  The pledges were insufficient — more than 70 countries did not even submit updated Nationally Determined Contributions.  The fossil fuel phase-out language was made voluntary after Saudi Arabia and Russia objected; the final text did not mention fossil fuels at all.  A US$120 billion adaptation fund was pledged by developed nations; developing countries said US$360 billion was needed.  The Just Transition Mechanism was adopted without the funding to operate it.  Australia signed the Belém Declaration and then announced plans to proceed with new gas projects.  COP31 will be held in Australia.  A communiqué was issued.  The language was stronger than last time.  Emissions continued rising.

The process is not broken.  It is working as its architects intended — as a pressure-release valve for public concern, a mechanism for governments to demonstrate engagement without delivering action, and a conference industry that employs a significant number of people whose professional existence depends on the problem remaining unsolved.  The national interest argument for action is not contested.  Fourteen per cent of global GDP — the Swiss Re projection for the current policy trajectory — is not an environmental number.  It is a solvency number.  Governments choosing to ignore it are not making a trade-off between growth and environment.  They are borrowing against the future, as governments are prone to do, and assuming their successors will pay.  Their successors will pay. The mathematics does not negotiate.  Civilisations that cannot learn from evidence decline.  The rate of that decline is the only variable still in play.

---

Terence Nunis | Executive Chairman, Equinox Zenith & Red Sycamore | Author, The 1% Playbook: The Billionaire Cheat Code