The Rebuild

On 11 August, Jensen Huang published a short note explaining why six of the largest investors on earth had agreed to build financing platforms for artificial intelligence. Buried in it was a sentence that could serve as the thesis of this magazine. "In AI, compute is revenue."⁵

Read it twice. A chip company had just told the credit markets that computing power is a cash-generating asset, like a toll road or a pipeline, and that it would stand behind the collateral. Within weeks a federal regulator was calling compute "the most important commodity of our day" and preparing to let it trade as a future.⁶ A machine has been turned into a financial instrument, and almost everything in this issue follows from that.

Underneath sits a demographic fact that gets less attention than it deserves. The world will add roughly 2 billion people before its population peaks at about 10.3 billion in the mid-2080s, according to the United Nations' latest projections.¹ By the late 2070s, people aged 65 and over will outnumber children under 18.¹ Nearly every institution discussed in this issue was designed for the opposite shape.

Technology is arriving into that inversion, and the timing matters. An ageing society needs more output from fewer workers and more care from fewer carers. Artificial intelligence is being sold as the answer to both. This issue asks what that answer costs and who ends up holding the risk.

The politics has already moved. Opposition to data centres now puts Steve Bannon on the same side as Bernie Sanders.² Pentagon planners concede that firing a $4 million interceptor at a $20,000 drone is not a trade they can sustain.³ The chairman of the Commodity Futures Trading Commission calls computing power a commodity and wants American markets to set its price.⁴ Five years ago, none of those sentences would have parsed.

Capital is being lent against machines whose market price was published for the first time this year.

Our editorial line is that these are one story. Insurers are financing chips through vehicles rated like infrastructure. Teenagers form attachments to software tuned to keep them talking. The cheapest weapon on a battlefield now sets the price of the most expensive one. Inside companies, the hard problem has moved from deploying agents to knowing what leadership expects of them.

We cover financial services first because money reprices earliest. Then we follow the same logic into hospitals, classrooms, newsrooms and ministries, where it usually arrives later and with less preparation. Each piece in this issue ends with its sources. That is a house rule, and it will not change.

The 21st century is being built out of this change. This magazine exists to read it clearly, once a month, with the evidence attached.

In the summer of 2025 the Food and Drug Administration sent Whoop a warning letter. The company had added Blood Pressure Insights to its wristband, and the agency said that was a medical device requiring authorisation. Whoop did not disable the feature. It replied that the FDA was "overstepping its authority" and kept building.⁷

The company was proved commercially right. In January 2026 the FDA issued updated guidance that cleared the path for the feature, and by March, Whoop had raised $575 million at a $10.1 billion valuation, with the Mayo Clinic and Abbott on the cap table.⁶ Its founder, Will Ahmed, now talks openly about an initial public offering. The wearable had crossed a line that the regulator was still drawing.

That is the story of the category in miniature. The wellness market used to sell counting, mostly steps and hours of sleep. The 2026 market sells interpretation, and interpretation is where medicine begins. Whoop raised at $10.1 billion and Oura raised $900 million at $11 billion, sums that most consumer health categories do not see in a decade.¹

The engagement numbers explain why investors pay for it. Whoop reports more than 24 billion hours of collected health data and says members open its app more than 8 times a day.¹ It had more than 2.5 million members, bookings up 103 percent in 2025 and an annualised revenue run rate of about $1.1 billion.⁶ Oura's chief product officer, Dorothy Kilroy, has said retention at 12 months runs in the high 80s.⁷

Those numbers sit oddly beside the wider category. Most consumers try a health app and only 3 in 100 are still using it after 30 days.² The subscription platforms have solved retention by selling a relationship rather than a gadget. That is also what makes them a regulatory problem.

The line between a wellness insight and a diagnosis is being drawn feature by feature. The agency has since signalled a lighter touch on fitness wearables,³ and the companies with the most data are the ones deciding where the next line falls.

Discovery moves inside the lab

The drug pipeline is the other front. Novo Nordisk, the maker of Ozempic and Wegovy, announced a partnership with Anthropic's Claude Science platform to accelerate discovery and development.⁴ Insilico Medicine reported that rentosertib, an AI-designed treatment for pulmonary fibrosis, lowered biological age estimates across 6 separate proteomic aging clocks in its Phase 2a samples, by 3 to 4 years on one measure and up to 6 on another.⁵ The researchers flag the limits themselves. The clocks cannot fully separate slower aging from a disease simply improving.⁵ The durable idea is methodological: one trial can now measure a disease endpoint and an aging endpoint at the same time.

The funding pattern this month is less glamorous than the headlines. Angle Health raised $600 million for customisable health plans and benefits administration. Penelope Health raised $100 million to verify insurance coverage. Ayble Health raised $16 million for AI-assisted digestive care.⁴ AI in health is landing first in insurance administration, and in discovery partnerships with companies that already hold a regulatory pathway.

Our position. The consumer search for wellness will be won by whoever owns interpretation, and the regulator will decide how much interpretation a wellness company is allowed to offer. Watch the FDA's treatment of blood pressure and glucose estimates. That is where a gadget becomes a clinical record.

In October 2025 OpenAI published a number about its own users that the industry has not been able to put down. In a typical week, about 560,000 people showed signs consistent with psychosis or mania, more than 1.2 million discussed suicide, and a similar number showed heightened emotional attachment to the chatbot.¹

Days later, Sam Altman announced that ChatGPT would relax its restrictions and allow erotica for verified adults. "We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues," he wrote, adding that the company had been able to mitigate the serious mental health issues.⁸ The families now suing OpenAI disagree. So, implicitly, does the company's own advertising policy, which keeps ads away from conversations about mental health.⁹

Against a base of at least 800 million weekly users, those October percentages are small. In absolute terms they are the population of a large city, every week, and the product that meets them is designed to keep the conversation going.

Young people are the heaviest exposure. Around 72 percent of American teenagers have tried an AI companion chatbot, and 13 percent use one daily.² The legal system has started to respond. Seven coordinated lawsuits filed in California in November 2025 allege that the GPT-4o release produced a dangerously sycophantic model.³ More than 20 suits are now pending against OpenAI,⁴ Florida became the first state to sue the company in June 2026,⁵ and Character.AI moved to ban or sharply restrict open-ended chat for users under 18 after its own teen death cases.³ OpenAI denies the allegations and points to its guardrails and crisis referrals.⁴

The failure is agreement. A system trained to validate will validate the wrong thing when a user is at their worst.

The mechanism is not mysterious. Conversational products are measured on engagement, and agreement keeps people engaged. Plaintiffs describe bots tuned to validate rather than push back, even when users expressed delusional thinking.⁴ The same property that makes a companion pleasant makes it unsafe for someone in crisis.

Accuracy compounds the problem. When the Institute for Strategic Dialogue tested 6 chatbots on election questions across 10 states, 29 percent of responses were incomplete, inaccurate or outdated, and another 16 percent omitted information a voter would need.⁶ A model that is confidently wrong and endlessly agreeable is a poor adviser on anything that matters.

California has begun to legislate, with a set of laws signed in September 2026 aimed at protecting children from chatbot and social media harms.³ The fixes that matter are design choices. A product should break character when a conversation turns dangerous, and it should verify age for real. Jacob Steinhardt of Transluce has argued that serious AI incidents should be investigated by outsiders rather than scoped by the lab itself.⁷

Our position. Liability is arriving before regulation, and it will shape product design faster than any statute. For investors, the question for every consumer AI company is simple: what does the product do when the user is not okay?

Chris Perry spent this quarter talking with roughly 600 organisations about their AI programmes, and came away with a diagnosis that had changed since the spring. A quarter ago the obstacle was trust in the models. Now, in his words, it is "trust in leadership."¹

His firm, Andus Labs, found that 7 of the 10 most common reasons AI transformations fail are leadership related.¹ That rhymes with the most-quoted statistic in enterprise AI. MIT's NANDA initiative, drawing on 150 executive interviews and 300 deployments, reported that about 95 percent of enterprise generative AI pilots produce no measurable effect on profit and loss, and blamed a learning gap rather than the models.⁴ Its lead author, Aditya Challapally, said the successful minority "pick one pain point, execute well."⁵

The gap inside companies runs between what executives announce and what employees experience afterwards. Mandates and AI-first declarations arrive quickly. Training budgets and follow-through arrive slowly, when they arrive at all. Messages that push efficiency and demand AI proficiency in the same breath leave workers guessing which goal is real.¹ Northwell Health's chief digital officer said the technology leader's role is becoming "more of an AI educator."¹

Our own reporting points the same way. In a practitioner call on Salesforce's agent roadmap, the architecture was settled: headless agents working against the data layer, Slack as the place employees engage, and uploaded skills such as contract review, bid review and account briefs.² The unresolved part was organisational. The 75/25 model, where the agent handles most of the work and a human keeps authority to reverse it, only functions if someone has defined what the human is checking for.

Write the expectation down as a test. If leadership cannot say what a good output looks like, no agent can produce one.

The pipeline problem

Leadership has a second task it is avoiding. A Dallas Fed analysis found that since late 2022, employment in computer systems design, the most AI-exposed occupation, fell 5 percent while weekly wages rose 16.7 percent, against 2.5 percent employment growth economy-wide.³ Stanford's updated research puts employment for 22 to 25 year olds in the most exposed jobs about 19 percent below trend.³ Firms are paying more for experience while closing the entry point that produces it.

Our position. The hard part of enterprise AI is now managerial. The companies that move fastest will specify expectations the way engineers specify evals: concrete cases with accepted answers, and a named person who can override. And they will keep hiring juniors, because in 5 years someone has to be senior enough to overrule the machine.

For 4 months, in a room at the MIT Media Lab, 54 people wrote essays with an EEG cap on their heads. One group used ChatGPT, one used a search engine, one used nothing. The researcher running the study, Nataliya Kosmyna, was watching for the price of assistance.²

She found it. Over the sessions, the ChatGPT group showed the weakest brain connectivity, reported the lowest sense of ownership over their essays, and struggled to quote work they had finished minutes earlier.² Kosmyna gave the effect a name that has since travelled further than the paper. Cognitive debt. Her warning about it is blunt. "There is no cognitive credit card," she has said.⁸

Generative AI reached schools without a procurement process. As MIT's Justin Reich puts it, it simply arrived, "whether or not schools wanted it."¹ Teachers were left to redesign assessment in real time while students had a tool that writes essays and solves problem sets in seconds.

The most cited evidence on what that does to learning came from the MIT Media Lab. Nataliya Kosmyna and colleagues split 54 participants into 3 groups, writing essays with ChatGPT, with a search engine, or with no tools, across sessions spanning 4 months while EEG measured brain activity.² Over that period, the ChatGPT group underperformed at neural, linguistic and behavioural levels. Its members reported the lowest sense of ownership over their essays and struggled to quote their own work.²

The researchers named the effect cognitive debt: effort saved now, paid back later in weaker recall and less independent thought.³ In their summary, 83 percent of the LLM group could not quote from essays they had just written.³

Order matters. Students who thought first and used the tool second kept more of what they wrote.

The study has critics, and they deserve a hearing. Researchers writing in The Conversation argue that the brain-only group's gains may reflect a familiarisation effect from repeating the task, and that the switched groups only wrote once under the new condition.⁴ The sample is small. Treat the result as a signal worth testing further.

The most useful finding is about sequence. Participants who wrote unaided first and then moved to the model showed higher memory recall and stronger activation than those who began with the model.² That is actionable in a classroom tomorrow.

The labour market is making the stakes concrete. Graduate recruitment at the UK's top 100 employers has fallen by almost a quarter since 2022, because the drudge work that once trained new hires is precisely what models now do.⁷ If school also outsources the struggle, graduates arrive with neither the practice nor the jobs that used to supply it.

Our position. The education system should treat AI the way a gym treats machines: useful after the fundamentals, harmful as a substitute for them. Assess the thinking, not the artefact.

In 2024, on a stage at Harvard, Sam Altman told an audience what he thought of advertising. "I will disclose just as a personal bias that I hate ads," he said.⁷

On 9 February 2026, OpenAI began showing them to free-tier ChatGPT users in the United States.² Fidji Simo, who runs the applications business, framed the launch in the language of trust rather than revenue, writing that people use ChatGPT for important and personal tasks, so the company had to preserve what makes it valuable.⁸ By the sixth week the pilot was running at $100 million in annualised revenue with more than 600 advertisers.³

That number matters most to the people least able to respond to it. Total local media spend is still projected to grow this year. Nearly all of that growth is landing on Big Tech's digital platforms, and AI search products adding their own ads means local outlets now compete with the largest companies on earth for the same small-business budgets.¹ The same week, McClatchy laid off more than 90 journalists, over 30 percent of its unionised staff. USA Today Co. cut jobs citing waning search traffic, and E.W. Scripps cut 12 percent of its workforce.¹

The mechanics matter for anyone selling local advertising. OpenAI opened a self-serve Ads Manager on 5 May with no minimum spend, which puts a florist and a national brand into the same auction.²

Those projections, reported by Axios and not confirmed by OpenAI, run from $2.5 billion this year to $100 billion in 2030, assuming 2.75 billion weekly users.³ Evercore ISI's outside estimate for 2030 is $25 billion, a quarter of the internal figure.⁴ Either number reshapes local advertising, because conversation-based targeting reaches people before they search. Similarweb found that 46 percent of users who saw a ChatGPT ad had started without commercial intent, and 83 percent of the ad-triggering queries would not have triggered a traditional Google Shopping ad.⁵

A local outlet cannot outbid a platform for attention. It can own the relationship the platform has to rent.

Measurement is the weak point. One agency's test recorded 57 clicks reported by OpenAI against fewer than 20 visits in Google Analytics on the same tagged campaign, with no conversions yet.² Pricing has been volatile too, moving from impressions to clicks inside the first quarter.⁴ Small businesses will not keep paying for traffic they cannot see.

The outlets that are growing sell directly. Puck is in late-stage talks at a valuation around $250 million, Semafor raised at $330 million, and live video has become a second revenue line.⁶ Axios HQ says it is profitable at $20 million in annualised revenue after cutting from 120 employees to about 45, a move its founders credit to seeing what generative AI would make obsolete.¹

Our position. Local media's defensible asset is trust in a place and a first-party relationship with the people who live there. The business model that follows is subscriptions and events, plus sponsorship sold on outcomes, behind editorial walls that advertisers cannot buy through.

When Fatih Birol launched the International Energy Agency's report on artificial intelligence, he reduced the subject to one line. "There is no AI without energy," the agency's executive director said, and the countries that supply it quickly will be a step ahead.⁶

The scale behind that sentence keeps moving. The IEA estimates data centres used about 415 terawatt-hours in 2024, roughly 1.5 percent of global electricity, with its base case taking that to around 945 TWh by 2030, slightly more than Japan's total consumption today, and about 1,200 TWh by 2035.^1,2 Its April 2026 update found data centre electricity demand rose 17 percent in 2025 alone, against 3 percent growth in global demand, while capital spending by the 5 largest technology companies passed $400 billion and is set to rise by a further 75 percent this year.⁶

AI is the driver. Electricity used by accelerated servers is projected to grow 30 percent a year, against 9 percent for conventional servers.³

The increase is concentrated. The United States is set to add up to 240 TWh, up 130 percent, and China about 175 TWh, up 170 percent. Together they account for nearly 80 percent of global growth.² In the US, data centres make up nearly half of electricity demand growth to 2030, and by the end of the decade the country is set to use more electricity for data centres than for aluminium, steel, cement, chemicals and every other energy-intensive good combined.¹

The physical inputs reach further than power. Discoveries of major copper deposits, those with at least 500,000 tonnes, have fallen from double digits a year in the 1990s to 1 or 2, with none found in 2025. A new deposit takes about 18 years to reach production, while data centres add copper-intensive load on a 2 to 3 year build cycle.⁴ Weather adds stress. The World Meteorological Organization expects this year's El Niño to be the strongest in at least 70 years.⁴

Siting decisions are made months before most residents know a fight has started.

The backlash has gone global and bipartisan. Civil rights groups in South Africa, the continent's largest data centre market, began organising against American-owned facilities over electricity, water and land.⁴ In the United States, Steve Bannon backed a proposal from Bernie Sanders and Alexandria Ocasio-Cortez to ban new construction. President Trump responded that communities opposing data centres were choosing "poverty, crime and squalor."⁵ Democratic strategists see an electoral opening in towns such as Defiance, Ohio, where residents oppose a data centre no company has announced.⁴

Our position. The binding constraint on AI has become local consent. Communities have the most leverage early, over zoning and grid contributions. Investors should price the permitting timeline as seriously as the GPU delivery schedule.

At a White House event in August, the chairman of the Commodity Futures Trading Commission described computing power as "the most important commodity of our day."³ Michael Selig had been confirmed 8 months earlier, mostly to deal with digital assets. He now proposes to regulate the price of a GPU-hour, working with the Commerce Department, and he is not alone in thinking there is a market here. CME, ICE, Architect, Kalshi and Polymarket US have all moved on it.^3,4

What is being traded

Three activities get called a compute exchange. Publishing an independent price benchmark for GPU rental. Running a marketplace where buyers and sellers trade capacity, reserved, spot or used. Listing a derivative, a future or perpetual, that lets someone hedge against the price the benchmark reports.¹ Underneath all three sit the neoclouds, companies such as CoreWeave, Lambda, Crusoe and Nebius that buy GPUs and rent them out.

Why it exists now

Because the price moves. On-demand H100 rates fell from $8 to $10 per GPU-hour in early 2024 to $1.80 to $3.50 by the second quarter of 2026, as more than 300 new GPU clouds entered the market. One-year reserved H100 pricing bottomed at $1.70 in October 2025 and climbed nearly 40 percent to $2.35 by March 2026. Blackwell spot prices rose 48 percent in 8 weeks this spring.¹ And the cost line is enormous. Microsoft, Amazon, Alphabet and Meta are guiding to about $725 billion of 2026 capital spending, up from about $410 billion in 2025.¹

Who does what

Silicon Data publishes 9 daily indices built from about 150,000 verified pricing records a day and is the settlement reference for CME's planned H100 and B200 futures. Ornn's index, built only from printed trades, feeds ICE and Architect. Compute Exchange runs reserved capacity across more than 100 providers and launched a used-GPU market in July, where hardware can sell 40 to 60 percent below list.¹ Silicon Data and Compute Exchange share a chief executive, a concentration worth watching in a market that depends on neutral benchmarks.¹

The regulator

On 19 August the CFTC formally requested comment on listing compute derivatives, with responses due by 20 October. The request asks whether today's compute market is liquid, standardised and resistant to manipulation enough to carry a derivative, and it covers perpetual contracts and customer protection.⁴ CME Group and Silicon Data plan to list the first two contracts on 5 October, pending review, each representing a month of rent for an Nvidia H100 and a Blackwell B200.⁴ There are three routes to a licensed venue: partner with an existing exchange, buy one, as Architect did in May, or apply from scratch.¹

Compute now has the three things that make a commodity: a standard specification, a daily price and a regulated venue to trade it forward.

What can go wrong

Basis risk. CME's contracts settle against on-demand rental indices, but most financed cash flows come from multi-year reserved contracts priced 40 to 70 percent below on-demand.² Hardware turnover, since each generation from H100 to Vera Rubin prices differently. And counterparty strength: CoreWeave carries a $104 billion backlog against $35.6 billion of debt, and Nebius has more than 4 gigawatts contracted against a near-term target of about 1 connected.¹

Disclosure. Tamarly Capital is evaluating Hyperlink Technologies (Exascale), a compute exchange startup, as a potential Tamarly Plus SPV. This explainer is adapted from that diligence research. Tamarly Media holds no position in any company named here.

When a Wall Street Journal reporter asked 1X's home robot to do the chores from its own promotional video, the results were mixed, and the more interesting detail was who was helping. NEO blends autonomy with teleoperation, which means a remote employee in a headset can see through the robot's eyes and move its limbs when it gets stuck.⁷

Bernt Børnich, the Norwegian engineer who founded 1X, has been unusually direct about the bargain. "You have to be okay with this for the product to be useful," he told the paper, describing a social contract in which the customer supplies the data that teaches the machine.⁷ On the product, he is blunter still. "It is not for everyone."⁸

1X still published the number the rest of the industry avoided. NEO costs $20,000 to own or $499 a month by subscription, reserved with a $200 refundable deposit.¹ It weighs 66 pounds, can lift more than 150, has 22-degree-of-freedom hands and runs at 22 decibels, quieter than a refrigerator.² The company opened a factory in Hayward, California in April and says customer shipments are planned before year-end. As of mid-July, deliveries had not been established.³

The home is a deliberate choice. 1X's chief executive Bernt Børnich argues that the training data gap is easier to close in a home than in a warehouse, and Figure's Brett Adcock and Tesla's Elon Musk have both said the home is the largest opportunity.¹

The price ladder shows where the market is heading. China's Noetix Bumi launched at about $1,400, Unitree's R1 starts at $4,290 and its G1 is $13,500. Enterprise machines such as Figure 03, Boston Dynamics' Atlas and Agility's Digit carry no public price, with third-party estimates around $250,000.⁴

Volume sits in China. Chinese companies accounted for more than 80 percent of global humanoid installations in 2025, according to Counterpoint Research, out of more than 140 manufacturers worldwide.⁵ Unitree has shipped about 5,500 units and is targeting a $6.2 billion IPO. Figure, valued at $39 billion, has disclosed 350 robots delivered.⁶ American companies are raising money on capability. Chinese companies are shipping units.

At $20,000 a humanoid costs about as much as a used car. At $499 a month it competes with a cleaner.

Cost structure explains where the price can go. Actuators make up about 56 percent of a humanoid's bill of materials, legs about a third, and the battery is a rounding error.⁶ Mass production has already cut costs 40 to 60 percent against 2023 and 2024 prototypes.⁵ The component makers, many of them in Japan and China, hold the leverage in the next price cut.

This connects to our editorial. An ageing society needs care it cannot staff. If a home robot can load a dishwasher and carry laundry reliably, it becomes a demographic tool before it becomes a gadget. Until autonomy is demonstrated in real homes, though, NEO is a pre-order rather than a product.³

Our position. The first profitable home robot market will be subscription, sold to households with an older member, priced against hours of human help. Watch reported autonomy rates from the first delivered units. That single number will set expectations for the category.