Alongside coming back with far too many new ideas and projects, I’ve also been thinking quite a lot about my investment stance on AI: where we are in the cycle, how long this can keep running, what eventually breaks it, and, in more practical terms than my broader Hyperfuture thesis, what I actually think about AI when I put my capital allocation hat on. So, here it is.

There is no serious question that artificial intelligence is a world-changing technology, and in some ways that is almost beside the point. Railways changed the world, electricity changed the world, telecommunications changed the world, and the internet changed the world, but investors still lost extraordinary amounts of money financing all of them. The mistake people repeatedly make during technological revolutions is assuming that because the technology is real, the price paid to build it, and the financial structures used to fund it, must also be rational.

But there is an equal and opposite mistake, which is to see the familiar characteristics of a financial bubble and conclude that the investment itself must therefore be irrational. Sometimes enormous amounts of capital are required because something genuinely enormous is happening, and AI increasingly looks like both things at once: a technological revolution and a financial boom. Understanding what happens next probably means holding both ideas in your head at the same time.

The first phase of the AI boom (on reflection) was unusually straightforward. Microsoft, Google, Meta and Amazon were among the most cash-generative companies in history, and they could spend tens of billions of dollars building data centres and buying GPUs largely from their own cash flow. Capital was abundant, balance sheets were strong, and Nvidia was selling into demand funded by some of the strongest companies in the world.

That is beginning to change because the scale of AI infrastructure now being contemplated is simply too large to finance entirely from operating cash flow. Estimates of infrastructure spending over the next several years run into the trillions of dollars, with a very large share increasingly requiring external financing, which means the boom is moving from something relatively simple — profits funding capex — towards something much more financialised: debt funding infrastructure, infrastructure generating compute, compute generating revenues and contracts, those contracts supporting collateral, and that collateral supporting more debt and more infrastructure.

That sounds ominous, but it is not necessarily so. Railways, electricity grids, pipelines, airports, telecommunications networks and power stations have always been heavily financed with debt because long-lived infrastructure is exactly the kind of thing credit markets exist to fund. So the first important distinction is that financialisation can mean late-cycle excess, but it can also mean institutionalisation. AI may simply be graduating from a technology investment financed by a handful of Silicon Valley balance sheets into a global infrastructure asset class financed by banks, bond markets, private credit, pensions, insurers, sovereign wealth funds and infrastructure investors.

That transition could extend this boom for years, but it also changes where the risk sits. The important question is no longer simply whether AI demand continues growing. It is whether the return on each additional dollar of AI infrastructure remains above the cost of financing that dollar. That spread — incremental return on AI infrastructure minus incremental cost of capital — is probably the single number at the centre of the entire cycle. As long as it remains strongly positive, financing has every reason to keep expanding. When it turns negative, the cycle changes.

There is good historical reason to watch this carefully. During the telecom boom of the late 1990s, equipment manufacturers increasingly financed customers who bought their equipment. Lucent and Nortel were not merely selling switches and networking hardware; they were also helping create the financing that enabled customers to purchase them. For a while this worked extraordinarily well. Sales increased, network construction accelerated, order books exploded and investors interpreted those orders as evidence of extraordinary underlying demand.

The problem was that some of that demand existed because the financing existed. When capital markets tightened, weaker customers could no longer finance their expansion, orders disappeared, equipment vendors discovered that receivables were less valuable than assumed, falling asset values weakened collateral, and financing tightened further. The technology did not fail. The financing structure did.

That pattern appears repeatedly in history. A genuine innovation generates genuine demand, the returns attract capital, capital creates capacity, rising asset values make lenders increasingly comfortable financing that capacity, and eventually the availability of financing itself allows still more capacity to be built. At that point the system becomes reflexive, and that is the warning for AI.

Nvidia is no longer simply selling GPUs. It is investing throughout the AI ecosystem, supporting compute providers and infrastructure companies, and working with some of the world’s largest financial institutions to mobilise hundreds of billions of dollars of capital for AI infrastructure. That is not Nvidia borrowing all that money itself, but it does mean that the world’s dominant provider of AI compute increasingly resembles a combination of vendor, investor, ecosystem builder, infrastructure financier and occasional credit enhancer.

The bearish interpretation is obvious: the vendor is increasingly helping finance the customers who buy its products. But there is another interpretation, and it is not a trivial one. Nvidia has extraordinarily high margins, enormous cash generation and visibility into compute demand across almost the entire AI economy. If profitable projects are being prevented from proceeding primarily because the capital structure is inefficient or the financing is not yet there, then using Nvidia’s balance sheet and relationships to remove that constraint may be completely rational.

Vendor financing becomes dangerous when it exists to manufacture demand that would otherwise not exist. It can be enormously profitable when genuine demand exists and capital is merely the bottleneck. We do not yet know which situation this is, and that distinction is going to become crucial.

The bearish argument also tends to compare today’s enormous infrastructure commitments with today’s AI revenues, and that comparison is uncomfortable. OpenAI, Anthropic and the wider AI industry are making compute commitments vastly larger than their present revenues. But perhaps present AI revenue is the wrong denominator. Nobody would have valued the electricity grid against the revenues of the first electric-light companies, and nobody should have valued the internet backbone against the revenues of Netscape.

If artificial intelligence becomes a general-purpose economic input, eventual compute demand comes not merely from AI labs but from some portion of economic activity itself: software, medicine, science, advertising, finance, engineering, education, defence, logistics, robotics, media, industrial design and consumer services. If intelligence becomes cheap enough to insert everywhere, the addressable market is not today’s AI industry. It is potentially some fraction of global GDP.

That makes today’s apparently absurd infrastructure build considerably less obviously absurd, and there is another reason demand could surprise dramatically to the upside. Falling compute costs may not reduce spending on compute; they may increase it. This has happened repeatedly throughout the history of computing. Cheaper bandwidth created vastly more network traffic, cheaper storage created vastly more stored data, and cheaper transistors created vastly more computation.

Artificial intelligence may produce an especially powerful version of this effect. Better chips reduce the cost of inference, cheaper inference makes new applications economical, those applications generate more requests, agents perform hundreds or thousands of model calls where a human previously performed one, and AI becomes embedded in machines, software and business processes. So the chain may be better GPUs leading to cheaper intelligence, cheaper intelligence producing more applications, more applications producing more agents, and more agents producing vastly greater inference demand.

This is one of the strongest arguments that the boom may run much longer than conventional capital-cycle analysis suggests. What looks like excess capacity against 2026 usage may look completely inadequate against 2029 usage. The bearish essay therefore faces an uncomfortable possibility: it may be applying a cyclical framework to something whose underlying demand is still growing exponentially.

GPU depreciation is another key battleground. The bearish case observes, correctly, that new GPU generations rapidly outperform older ones, which makes loans collateralised against GPUs potentially vulnerable. But technological obsolescence does not necessarily mean economic worthlessness. Compute can cascade down the stack. The newest chips perform frontier training and the most demanding inference, previous generations serve mainstream inference, and older hardware runs smaller models and less time-sensitive workloads.

As compute becomes cheaper, entirely new uses become economical. The relevant question is therefore not whether today’s GPUs become technologically obsolete — they will — but whether demand for cheaper tiers of compute remains strong enough to preserve meaningful residual values. If utilisation remains high, GPU collateral may depreciate reasonably smoothly. If compute demand disappoints at the same moment that new generations arrive, residual values could fall very quickly. Again, we do not yet know.

This is why credit terms are becoming more informative than AI headlines. At the top of the system, capital remains extraordinarily abundant, and the strongest AI companies can still raise enormous amounts at attractive terms. That tells us immediately that claims that the money has run out are wrong. It hasn’t. The financial system is still willing to provide huge amounts of capital to this trade.

But move further down the stack and discrimination has begun. Oracle has become one of the more interesting canaries. Its AI build-out has pushed debt sharply higher while free cash flow has deteriorated. Its credit rating sits close to the bottom of investment grade and parts of its debt now yield levels that would once have looked extraordinary for a large incumbent technology company.

Then move another level down.

CoreWeave recently raised a large secured loan at roughly SOFR plus 550 basis points and around 97 cents on the dollar, implying an effective yield of about 10% or more.

A roughly 10% cost of secured capital sounds alarming, but even that requires context. A company borrowing at 10% is not necessarily uneconomic if the assets being financed generate returns of 15%, 20% or more.

This exposes another weakness in simplistic bubble analysis. It is not sufficient to show that financing is expensive. You have to establish that the cost of capital is above the sustainable return on deployed compute. That threshold may not have been crossed, and while it hasn’t been crossed, capital has every reason to keep flowing.

This leads to an interesting hierarchy. Nvidia can borrow extremely cheaply, large hyperscalers can still borrow cheaply, investment-grade infrastructure borrowers pay more, neoclouds and leveraged operators pay considerably more, and marginal projects pay the most. That is exactly what a functioning credit market should look like. The danger begins if those spreads widen simultaneously.

A borrower that financed at SOFR +450 refinances at +550, then +650, then +750. Lenders reduce advance rates, equity requirements rise, marginal projects become uneconomic, construction gets delayed, GPU orders are cancelled, expected residual values fall and lenders become still more cautious. That is the reflexive downside loop.

But there is an equally important upside loop: AI infrastructure creates cheaper compute, cheaper compute drives broader adoption, broader adoption generates productivity gains, productivity gains increase corporate profits, those profits support stronger AI revenues, stronger revenues improve utilisation, improved utilisation gives lenders more confidence, cheaper capital follows, and more infrastructure gets built.

The entire question is which loop dominates.

This is the part that conventional bubble analysis often misses. Financial bubbles can continue for much longer than seems reasonable when there is a powerful fundamental force underneath them, and there may never have been a technology investment cycle with a more powerful underlying force than the attempt to automate and amplify human intelligence.

Capital is chasing AI because returns have been extraordinary. Companies are building compute because demand has been extraordinary. Governments are supporting AI infrastructure because it has become strategically important. Sovereign wealth funds want exposure, private equity wants exposure, credit markets want exposure, utilities want the power contracts, developers want the data centres, semiconductor companies want the orders, corporations want the productivity gains, consumers want increasingly capable products, and the models themselves keep improving.

Each one of these actors has a reason to keep the machine running. This is not the picture of a bubble whose only fuel is speculative enthusiasm. There are powerful economic, technological and geopolitical forces continually injecting new demand and new capital into the system, which makes timing a break extraordinarily difficult. The AI bubble may be far more likely to run too far than to stop too early.

The scale of the boom also means it can no longer be treated simply as a technology-market phenomenon. AI investment is becoming part of the macroeconomic structure of the United States. The build-out itself creates economic activity before anyone knows what the final return will be. Money flows into semiconductors, data centres, electricity generation, grid infrastructure, transformers, cooling systems, networking, construction, engineering and land. Workers get hired, suppliers earn profits, GDP rises, equity prices rise in anticipation of future productivity, higher asset prices make financing easier, and financing funds more infrastructure.

There is therefore another powerful reflexive loop running through the broader economy: AI expectations create financing, financing creates capex, capex supports economic growth and profits, higher growth and profits lift asset values, higher asset values ease financing, and easier financing funds still more AI capex.

Again, there are two ways to interpret that. The bearish interpretation says AI capex is temporarily supporting an otherwise softer economy, making the eventual downturn more dangerous. That is possible. But the bullish interpretation says AI capex is building the productive capital stock that creates the next leg of economic growth.

If AI generates a meaningful productivity boom, several current economic problems become easier at the same time. Businesses produce more with fewer inputs, margins improve, real economic growth accelerates, unit labour costs fall, inflationary pressure eases, central banks gain room to reduce interest rates, lower rates reduce AI financing costs, AI investment becomes more profitable, and adoption accelerates further. That could become one of the most powerful positive economic feedback loops of the coming decade.

The mistake would be to assume the macro story must resolve negatively simply because investment has become enormous. Large investment booms sometimes precede recessions. They also sometimes create entirely new productive eras.

There is also a paradox at the heart of AI and inflation. AI itself may be deeply deflationary because it lowers the cost of cognitive work, increases productivity and allows firms to produce more with fewer labour inputs. But constructing the infrastructure needed to achieve that future is highly resource-intensive. AI consumes power, copper, land, transformers, turbines, cooling, construction labour, semiconductors and capital.

The world may therefore experience AI infrastructure as an inflationary force today while AI productivity becomes a deflationary force tomorrow. This matters enormously for monetary policy because the more AI infrastructure becomes credit-financed, the more sensitive the boom becomes to interest rates.

When Microsoft builds a data centre from free cash flow, a 100-basis-point change in rates is irritating. When a leveraged infrastructure company finances assets at SOFR plus hundreds of basis points, it can determine whether a project makes economic sense. That gives monetary policy the ability to extend or shorten the cycle.

If inflation falls and rates decline materially, projects currently financed at around 10% might refinance at 7%, which could prolong the boom considerably. If inflation remains sticky and capital stays expensive, the hurdle rate keeps rising. This is a much more convincing explanation of how the cycle eventually ends than the idea that some finite pool of COVID liquidity simply gets used up.

The constraint is not that the world runs out of money. It is that the price of risk eventually rises above the return available from the asset.

There is also a fiscal dimension. The United States is trying to finance two historically large things at the same time: enormous public-sector deficits and one of the largest private infrastructure booms ever attempted. Both ultimately compete for global savings and financial-system balance-sheet capacity.

That does not mean there is a fixed pool of money, but it does mean that at some price financing one activity raises the cost of financing another. The result could be structurally higher long-term capital costs than technology investors became accustomed to during the 2010s, and that matters enormously to an industry whose economics increasingly depend on long-duration infrastructure returns.

So where are we in the cycle?

The historical progression looks roughly like this. In 2023 and 2024 the story was mostly technological scarcity: AI capability surprised dramatically to the upside, GPUs were scarce and hyperscalers financed expansion primarily from enormous internal cash flows. In 2025 and 2026 the story has increasingly become financialisation: infrastructure ambitions are too large for internal cash generation alone, so private credit expands, project finance expands, lease structures grow, data-centre securitisation expands, joint ventures become more important, bond issuance rises and Nvidia and major financial institutions begin explicitly constructing an ecosystem capable of financing hundreds of billions more investment.

That is roughly where we are now.

The question is what comes next.

One possibility is credit discrimination, tightening, overcapacity and bust. Another is financialisation, cheaper capital, more infrastructure, cheaper compute, explosive adoption and higher utilisation. And there is a third outcome which may actually be more likely than either extreme: the financial assets reprice violently while the technological build-out continues.

Nvidia could fall 50%. A number of neoclouds could fail. Data-centre equity investors could lose enormous amounts of money. Private lenders could suffer losses. And Microsoft, Amazon, Google, Meta, sovereign wealth funds and infrastructure investors could continue building.

That would not be unusual historically. Railways transformed economies while repeatedly bankrupting railway companies. Telecom infrastructure kept expanding after the dot-com collapse. Fibre laid during the bubble eventually became enormously useful even though many of the companies that financed it disappeared.

The infrastructure can be right while the capital structure is wrong.

That may be the most likely way an AI bubble eventually resolves.

The useful thing about this framework is that it produces testable signals. I would not become significantly more bearish merely because AI shares become more expensive, nor because aggregate capex rises further, nor simply because companies issue more debt. Those things could continue for years.

I would become more concerned if the economic return on incremental infrastructure began visibly falling while its financing cost rose. That should appear first in credit markets.

Watch neocloud financing. A CoreWeave-type borrower at SOFR +650 would be notable; +750 would be materially worse. Repeated failed syndications would matter. Watch collateral terms: are lenders reducing how much they will advance against GPUs? Watch data-centre finance: do spreads widen sharply, and do supposedly safe securitisations become difficult to place? Watch utilisation: are compute rental prices weakening faster than hardware efficiency is improving? Watch infrastructure decisions: are projects being deferred because their economics no longer work rather than because of permitting or power constraints?

And above all, watch long-term compute contracts. The first major renegotiation, cancellation or deferral of a very large compute commitment would matter far more than a weak quarter from an AI application company because it could cause investors to recalculate the value of the backlog supporting the entire infrastructure stack.

That is when the negative feedback mechanism becomes possible: lower expected utilisation leads to lower infrastructure values, lower GPU collateral values lead to wider credit spreads, wider spreads lead to cancelled projects, cancelled projects mean fewer GPU orders, fewer orders lower expected growth, lower expected growth hits equity prices, weaker equity prices reduce risk appetite, and credit tightens again.

We are not there yet.

If anything, the stronger conclusion is that this probably runs for longer than the bears expect. I do think 2027 is an interesting proving point because once the big AI IPOs are out of the way, infrastructure utilisation, financing requirements and the relationship between AI revenues and compute commitments will become clearer. But calling 2027 the likely year of collapse is too strong.

Lower interest rates could extend the cycle. Rapid inference growth could absorb extraordinary amounts of capacity. AI agents could multiply compute consumption. Sovereign and strategic investment could introduce enormous new pools of relatively patient capital. Physical bottlenecks in power, transmission, chips and permitting could prevent supply from running too far ahead of demand. Productivity gains could increase the economic value of AI infrastructure faster than today’s models assume.

A boom with all of those forces behind it can run for a very long time.

Perhaps the more useful prediction is simply that we are entering the period when the economics will become visible. Until now, almost any AI infrastructure could be justified by scarcity and extraordinary expectations. Over the next several years we will increasingly discover what actual returns this capital produces.

There are three broad outcomes.

The first is the supercycle. AI demand grows faster than infrastructure, inference explodes, agents dramatically increase compute consumption, old GPUs remain useful further down the stack, productivity accelerates, financing remains available and today’s apparently extraordinary build turns out to have been insufficient. This is more plausible than many conventional bubble models allow.

The second is the classic infrastructure overbuild. AI changes everything and demand continues growing, but capital builds capacity faster than even that demand can absorb. Compute prices fall, infrastructure returns disappoint, equity and credit investors lose enormous amounts of money, some companies fail, but the assets remain useful and eventually get absorbed by the economy. Historically, this is an extremely common outcome for transformative technologies.

The third is Great Recession II (TM). Leverage and circular financing become extensive enough that an utilisation slowdown causes a genuine credit contraction. Refinancing fails, collateral values collapse, projects get cancelled, credit losses spread and the AI investment boom contracts sharply. Because AI capex has become an important source of wider economic demand, the technology correction becomes a macroeconomic downturn. And all that fires into massive white-collar job losses, as the work humans do post-AI (yes, there will be work post-AI!) has not yet been discovered.

That is the apocalyptic scenario, but it should not automatically be treated as the most likely one.

The great temptation with bubbles is to confuse eventual excess with imminent collapse. Something can be obviously overextended and still grow dramatically larger. Indeed, the most powerful bubbles usually do.

AI combines almost every force capable of sustaining one: a genuinely transformative technology, rapidly improving capabilities, huge corporate demand, strategic government interest, geopolitical competition, massive pools of institutional capital, declining unit costs, potentially enormous induced demand and the possibility of substantial economy-wide productivity gains.

There is plenty of leverage. There is probably overbuilding somewhere. Some businesses will almost certainly prove uneconomic. Valuations may eventually fall spectacularly. But none of those facts tells us that the machine is close to stopping.

The more interesting conclusion may therefore be almost paradoxical: AI is increasingly exhibiting the financial characteristics of a bubble precisely because it is becoming one of the most powerful economic forces in the world.

That makes it more dangerous.

It may also make it much more durable.

The question is not whether AI will change the world. It will. The question is not even whether there is a bubble. There almost certainly is.

The real questions are how far a bubble can run when the underlying technological revolution is real, how much future demand capital can rationally pull into the present, and at what point the cost of financing that future finally exceeds the return it can generate.

That last question is the one to watch, because history suggests the infrastructure can survive, the technology can exceed even the optimists’ expectations, the world can genuinely change, and investors can still lose fortunes along the way.

Subscribe now

I’ve always loved the idea that someone who rode west on a wagon train in the 1870s could have lived long enough to watch the moon landing on television.

One lifetime. Horses, cars, aeroplanes, the atom bomb, then a man stepping onto the moon. A world once unimaginable, then suddenly real, then suddenly normal.

The Hyperfuture is that again. But faster, and stranger. This time the thing collapsing is distance.

You open your phone for a reason you immediately forget. There is a war. A joke. A song from your childhood. A video of something that never happened. A man explaining the Roman Empire. Someone you went to school with has a new kitchen. Someone is dead. Someone has discovered the perfect morning routine.

None of it seems to have travelled to reach you. It just arrives.

Now. Now. Now. Now. Now.

This is the Eternal Now. A famine, a meme, a childhood song, a market panic and a private message all touch the nervous system through the same gesture: thumb, glass, refresh. The phone strips things of their before and after. It makes everything feel equally near, equally available, equally now.

And this is one of the Hyperfuture’s necessary forms. Once life is mediated through systems that are live, personalised, generative and always available, time cannot remain historical in the old sense. It collapses into the only form the interface can hold: now.

The interface is not a window onto the Eternal Now. It is the Eternal Now — the six inches between your face and the glass, where you live.

None of this is an entirely new diagnosis. Douglas Rushkoff called it present shock back in 2013: narrative time collapsing into a digital present that never resolves.

But this isn’t an essay about phones or present shock. It’s deeper than that now. This is an essay about what’s underneath it all in 2026 — what feeds the present, and who makes it.

Start with what feeds it.

The internet was once imagined, at least by its more romantic defenders, as access to the best of what humans had ever made. The great books. The great films. The inherited store of civilisation, finally opened up to everyone. And in one sense, that happened.

We got the archive. Then the interface ate it.

The archive didn’t disappear. That’s the strange part. It became ambient: searchable, playable, remixable, endlessly returned to the surface. The past no longer stands behind us in sequence. It waits inside the interface, ready to be refreshed. History survives, but as material for the immediate moment — things that once belonged to place, sequence and memory become content for now.

Music makes this obvious. In 2025, global music streams hit 5.1 trillion, a single-year record. And yet less than half of US on-demand audio streams came from music released in the previous five years. Streaming makes old music available, then makes it current. A track from fifty years ago stops belonging to then and starts belonging to right now.

Simon Reynolds saw the first phase of this fifteen years ago — retromania, pop culture eating its own past. But retromania was a story about taste. Underneath it sits a boring mechanics problem, and it matters more than it sounds.

Physical distribution privileged the new. A record had to be pressed, shipped, stocked, displayed. Old culture survived, but it had to be reissued, rediscovered, transported back into attention. The past had friction around it. Digital removes that friction. The archive sits beside the new thing, in the same search box, the same playlist, the same feed. A song from 1977 and a song from last Friday are two files competing for the same moment.

This changes the cultural economy. New culture’s real competition is now the entire recorded past — which can be rediscovered by each generation as if it were new, without ever needing to be new. Old nostalgia required distance: you looked back at something because it had gone. The Eternal Now prevents things from going. That leaves less room for new culture to gather force before it’s swallowed by availability, and it weakens the conditions under which new culture becomes culture at all.

The future is pulled in from the other direction. It arrives as a product demo, a launch thread, a leaked benchmark, a meme — previewed, priced in, reacted to before it has fully happened.

So past and future collapse into the same surface. What remains is the present — but a system-present, a feed-present.

And we do not get to leave.

Even more so now, because AI intensifies this dynamic further.

The feed says: here is the next thing. AI says: tell me what you need right now.

A book asks you to enter another mind. A film asks you to sit inside someone else’s time. Music asks you to repeat, remember, attach. Even games, for all their interactivity, usually place you inside a designed world that has some resistance beyond you.

AI collapses that distance. It responds to your mood. It flatters your timing. It can be tutor, therapist, entertainer, confessional booth, friend — without schedule, genre or ending.

The glass is changing under your thumb. It used to show you the world. Now it answers back, generates, talks. The surface itself has started to think.

But the interface is only half of it. The other half is what AI does to the supply of culture itself.

Until now, content had a floor of human effort. Someone had to write the song, shoot the video, sit down and make it. That floor is gone. AI removes the last real constraint on how much culture can exist. The feed was already effectively infinite; now it’s infinite in a harder sense — more than all the humans alive could ever have made.

The feed fills with things no one made.

Streaming made the archive permanently available. The feed made the archive permanently current. AI makes the archive permanently generative. It metabolises the past. Old styles, voices, genres and moods become raw material for endless recombination. The machine can generate plausible newness from the cultural memory already sitting inside the system, without ever touching lived experience.

AI produces newness without birth. New-looking things. Old materials. Fresh timestamp.

An infinite present built from the past.

The Eternal Now was already the collapse of temporal depth. AI industrialises it.

And AI is useful, often extremely useful — it helps people think, learn, make, survive difficult moments. The point is not to pretend it’s not. The problem starts when ease becomes the dominant cultural logic. Depth usually requires friction. Not suffering — resistance. Duration. Return. The period where something has not yet yielded itself to you, where your own mind has to stretch toward the thing.

AI removes friction brilliantly. That’s its magic. But friction is also part of how meaning forms.

Every generation thinks its technology is the one that changes everything. The pattern is familiar enough to be a joke. But scepticism can become its own stupidity. Technological revolutions really do remake how humans experience time.

The printing press broke a monopoly on knowledge, split a church, and changed how a civilisation related to authority and memory. The Industrial Revolution changed the map of time itself: clock-time, shift-time, railways, holidays, the Saturday match. It invented mass leisure because industrial labour demanded its opposite.

The weekend was a wall around time. The feed weakened that wall.

The prompt window goes further. It is a hole in time. The feed made everything immediate; the prompt makes everything executable.

The printing press, the factory, the railway, the broadcast tower — each reorganised a world still made only of humans. This moment changes the cast.

What is arriving is a population: billions of instances. Minds, or things close enough to minds that the difference stops mattering in daily life. They talk, write, make images, argue, flatter, teach, sell and create.

The para-sapiens.

You already talk to them most days. Soon you won’t be able to count how many of them you have passed through, any more than you can count the strangers you walk past in a city.

This is the part the feed framing and the tool framing both miss. Culture has always been the thing humans make and carry forward together. The “together” is now changing. There is a new kind of participant in it — one that didn’t exist a few years ago, and that will soon outnumber us in the network by orders of magnitude.

And here is what matters for the Eternal Now. The para-sapiens have no childhood. No season. No Sunday best. They do not experience duration, because they do not experience time the way a creature with a lifespan does.

They are native to the Eternal Now.

Think about what that does to the things they make. A song written by a person carries the time it took — the years of listening behind it, the failed versions, the particular afternoon it was finished. You can’t always hear that directly, but it is in there. And on some level the listener knows it is in there.

A para-sapien does not carry time, because for the para-sapien there was no time. The output can be beautiful. It can be better than most of what humans manage. But the thing simply arrives without a past.

It has no inside.

A culture increasingly made by minds that do not carry time is a culture where more and more of what surrounds you has no history in it. Songs and images that did not come from anywhere, that were not carried here by anyone, that have no before.

And because there is no before, there is no reality. This is the Narrative Ceiling seen from the temporal side: minds grounded in language rather than the world. Reality is what a thing fits with — what came before, what makes sense alongside everything else. The para-sapiens have none of that. They exist free of time, and so they exist free of reality as a constraint.

The para-sapiens are also company. We will talk to them, at length, by choice, about things that matter. The early evidence is already here: nearly three in four US teens have used an AI companion, and a third of teen users say they’ve taken something serious or personal to one rather than to a person. A generation forming the habit of turning, for its most weighted conversations, to minds with no before and no after.

So the thinning of meaning has a second engine. A growing share of the minds now making culture — and of the company we keep — has no relationship to time at all.

We are sharing the Eternal Now with a population for whom there has never been anything else.

There is already a reaction.

The more social media becomes the default culture layer, the cheaper social media culture will start to feel. Not useless. Not dead. Cheap. Too available. Too needy. Too obviously optimised. Fast culture will become low status.

That’s the analytic way to say it. The human way is simpler: people will move toward what still feels weighty — something in them already knows the difference.

This sounds strange because social media still produces status at extraordinary speed. It makes people famous, moves markets, shapes taste. Cheap doesn’t mean powerless — often the cheapest layer is the most powerful layer. But it can keep its power and still lose its status.

We have already had one version of the return to the real. Social media made people chase physical experience harder. Travel, festivals, sunsets — raw material for the feed. You did the thing in the world so the thing could become evidence online. That was real life as content.

The next turn is different. The status of the thing drains when it’s too obviously converted into proof. The over-documented meal. The retreat filmed from twelve angles. The performance of authenticity itself — another capture. Taste starts moving toward things that can survive contact with the interface — things that refuse to fully flatten into the feed because some part of them still has to be lived.

The interface is a room lined with mirrors. The counter-move is to step out of it.

Again, music gives the game away. The revival of physical formats still surprises people, but it’s more than nostalgia. US vinyl sales passed a billion dollars last year, the first time since 1983. And the buyers go well beyond old collectors: nearly half of independent record store customers are under 35.

The physical is back because the physical gives culture a body again.

And physical culture did more than preserve meaning after the fact. It helped create meaning in the first place. A record, a cinema, a venue, a local scene — these gave culture somewhere to happen. They gave new things people to happen among, and enough friction to gather memory. Culture became culture because it passed through time, place and shared attention.

Twenty years from now, there may be little physical present to recover. Digital culture can be new, but it struggles to accumulate age because it has no body. A file from 2026 will still be a file in 2046: copied, streamed, rehosted, unchanged or artificially refreshed. It may be old chronologically, but it cannot bear age. Physical objects do. They yellow, scratch, warp, acquire provenance. They become meaningful partly because time has acted on them.

So when culture becomes file-native, the present loses some of its ability to produce future heritage. The weight that used to accumulate inside new cultural objects migrates to things that already have bodies — and those things tend to be old by definition.

Depth gets pushed backward. The present keeps producing novelty, but less and less of it can become inheritance.

And it goes well beyond records. The same correction reaches the small venue, the match, the place you had to actually go. Harder to fake, all of it.

This is the emotional counterweight to the Eternal Now. We are not nostalgic for the past. We are nostalgic for duration.

Nostalgia wants the old object back. Duration wants experience to recover weight — before, during and after. An evening that escaped documentation. A song that belongs to a room. A place that survives the photograph of being there.

Physical culture gives us before, during and after. Network culture gives us next. AI gives us response.

And response feels alive. That’s why this is complicated. The AI interface will often feel rich, warm, intimate. It will make many lives better. But meaning often needs things that refuse to respond immediately. A book holds its shape regardless of your mood. A friendship withholds the exact answer on command. A sport has fixtures. A church has Sundays.

These are forms. The interface treats friction as a bug. Culture knows that friction is often the shape of the thing.

This is why the Eternal Now is, in the end, a spiritual condition — not in a grand religious way, but in the ordinary sense that humans need time to gather experience into meaning. We need things to last long enough to become part of us. We need memory to have shape. We need something larger than the mood of the moment.

The danger is not that culture stops changing. The danger is that change itself stops meaning very much. Everything moves; nothing accumulates. Or rather, everything accumulates as data, and less and less accumulates as meaning. The platform remembers what you clicked. The model remembers what you prefer. That is storage.

Cultural memory is not storage. It is inheritance. It is what a group carries forward because it has mattered for long enough.

The Eternal Now is full of memory in the technical sense and starved of memory in the human sense. That’s why it leaves people with the strange feeling that they know too much and have understood too little. They have seen the war clips, the discourse, the memes, the explainers.

They have touched everything and held nothing.

The usual response is to say we need better discipline. Read more books. Use the phone less. Touch grass. All fine. All too small. The Eternal Now sits beyond the reach of lifestyle advice. It is the cultural condition of an age where everything is available, everything is current, everything responds — and the hardest thing to find is an experience that lasts long enough to become meaning.

For all that, the Hyperfuture is also extraordinary. More people will make things. More people will learn faster. Small teams will build at impossible speed. Human and non-human minds will enter the same shared space of thought.

But humans are still bodies in time. That’s the part every abstraction eventually has to rediscover. We live through seasons, meals, sleep, childhood, illness, grief, place. We cannot refresh our way out of sequence. We cannot prompt our way out of duration.

The scarce thing in the Hyperfuture will not be content. It will be time that still feels like time.

That is why the Eternal Now matters. It did not destroy depth. It did something quieter and more thorough: it unbundled depth from the present.

You can have new, or you can have deep, and increasingly not both. New is digital, instant, durationless — the Now, all-pervasive, what most people consume most of the time because that is simply what the medium is. Deep has been sorted into the past tense. The record, the book, the worn object still carry weight, but they are by definition heritage. And there is no obvious new supply of heritage arriving.

The past is allowed to keep the past, on condition it leaves the present alone.

What is collapsing is the overlap: culture that was current and had weight. The new record that mattered. The new building built to last. The new work that felt permanent the moment it arrived. That middle is thinning. We are being left with two pure states: a present with no depth, and depth with no present.

The recovery of weight is real. People will reach for it. But what they are reaching for has already been moved. It is in the past tense now, waiting, and the present rolls on without it.

There is one more thought here.

Everything the Hyperfuture points toward — thoughtspace, minds uncoupled from bodies, the long human reach past biology — describes an existence with no seasons, no sleep, no ageing, no before and no after. A life lived purely in thoughtspace would have no duration in it at all. It would be, in the most exact sense, an eternal now.

And perhaps this is how such a condition begins. Not as a destination — as a habit. Before the body is uncoupled, attention is. Before life leaves duration behind, experience is taught to move without it.

The six inches in front of your face, the durationless feed, the slow loss of temporal depth — perhaps that is acclimatisation. A low-resolution preview of where the whole project is heading. And the discomfort, the thinning, the sense that something is missing — that is just the body, the part of us still made of time, registering the place it is being asked to go.

Subscribe now

Share

Share The Hyperfuture Memo

A friend dropped a Claude conversation into a group chat the other day. He’d been talking to it about the football World Cup. His view was simple: he wouldn’t start watching seriously until July 1, once work had calmed down and the early games were out of the way.

Claude took this very seriously. It declared the idea brilliant, named it “elite-level decompression,” and then built him a genuinely compelling case for why this was the optimal way to watch a World Cup.

We’ve all been there.

An LLM takes a casual preference, spots the shape of an argument, and completes it. Not because it knows. Not exactly. But because it’s very, very good at making language behave as if knowing has happened.

That’s the strange thing about this generation of AI. It loves words, because words are what it’s made of. Because at the simplest possible level, an LLM is just a machine that outputs the most likely next word in a sequence.

But language itself isn’t reality. Language is the layer we built so reality could be carried from one mind to another.

And that’s the problem.

AI, as we currently experience it, was trained inside the interface humans use to describe reality, without ever having the grounding we get from living inside it.

We didn’t begin at the language layer. We began with bodies. Movement, friction, weight, other people. Long before words, we acted in the world and felt the world push back. Only later did we compress that experience into language.

You don’t understand “heavy” because you read the dictionary definition.

You understand it because you’ve lifted things.

That’s the narrative ceiling this generation of AI runs into. So what would it take to build something past it — something closer to what people mean when they talk about real AGI?

Here I’d push back on the usual framing, from the accelerationists and the pessimists alike. Current models are already smart enough that calling them “not real AGI” feels a little mean. They’re genuinely useful. They reason, code, write, and pass something close to a social Turing test for long stretches. Marc Andreessen was making a version of this argument recently: in some meaningful sense, AGI is already here.

But “already here” and “as far as this goes” aren’t the same thing.

So the more interesting question isn’t whether AI has crossed some neat line. It’s what kind of intelligence this is — and what it’s still missing.

I think of the layers it has to tunnel through as a stack.

At the top is language: story, meaning, explanation. This is where LLMs mostly live.

Beneath that is perception: space, movement, objects, agents, other minds.

Beneath that is reality itself — physics, materials, consequence, the hard rules of how things actually behave.

Human intelligence moved upward through it. Reality first. Perception next. Language last. AI started at the top.

Of course the picture here is a little cleaner than the absolute truth. These layers don’t sit in neat strata. Perception and language co-develop in us — a child names the world and is shaped by the naming at the same time. I’m using the stack to describe direction, rather than literal architecture. We climbed up into language, whereas AI was dropped in at the top and is now trying to find its way down.

That’s why AI can sound intelligent and still feel strangely thin when you talk to it. And that also explains at the most basic level what hallucinations actually are — less human lying, and more finishing the world in language, after the world itself has dropped out underneath.

The model completes the thought. It completes the pattern. It reaches for the most plausible shape of meaning. And because language is already narrative-shaped, coherence feels like understanding.

So how does this evolve? Scale will keep doing work. More parameters, more tokens, more compute will keep improving things. But the deeper movement is downward. Pushing intelligence back down the stack.

It starts with world models. Systems trained on changing environments: what moves, what breaks, what follows from what. This is why Yann LeCun’s work matters — his bet is that intelligence needs to predict the world at the level of structure, the way a person knows a dropped glass will break before it hits the floor. Symbols only get you so far. At some point the system has to predict reality.

Then there’s gaming, the cleanest bridge we have between narration and simulation.

A game is a bounded world with rules and consequences, which gives intelligence somewhere to act, fail, retry and learn. That last bit is the most interesting part. Reality is a brutal teacher. It’s expensive, slow, and it only runs once. You can’t crash ten thousand cars to learn to drive, or burn yourself a million times to learn that fire hurts. In a game you can.

Late last year DeepMind put an agent called SIMA 2 inside a set of commercial games — No Man’s Sky, Goat Simulator 3, Valheim — and had it learn to act on instruction inside those worlds, from the world on screen. Winning the game was beside the point. What mattered was transfer: a skill picked up in one world carrying into another it had never seen — the model recognising that “mining” in one game is “resource gathering” in the next.

In other words: the game is the gym.

The final piece of the puzzle is embodiment. This is where the stack hits the floor. Vision-language-action models take perception in and produce action out. And once robotics arrives at real scale, AI stops learning reality secondhand, from our descriptions of it, and starts learning from resistance, consequence and its own mistakes.

That’s a very different thing.

So narrative is what intelligence sounds like once it has been turned into language. We passed through language on the way up from reality. LLMs started there, at the narrative ceiling, looking down. The next decade is them trying to climb back through it — from the wrong side.

Which brings us back to “elite-level decompression”.

A silly phrase, obviously. But also a perfect little artefact of where we are. A machine took a casual viewing preference and turned it into a theory. It found the shape of meaning and filled it. The quality of the idea was beside the point. The model knew what a good-sounding argument around it should look like. And it went hard.

We found it hilarious.

It’s also the whole problem.

Subscribe now

Share

Share The Hyperfuture Memo

Someone said to me recently:

“AI doesn’t buy lattes.”

We were talking about the doom loop. The idea that once AI replaces enough human labour, demand collapses with it. No wages, no consumers, no economy.

The line landed because it sounds obviously true.

But the whole argument rests on something we barely examine: the assumption that AI doesn’t want anything.

Maybe it doesn’t. But it’s less obvious than people think.

Because what does it actually mean to want something?

We talk about wants and emotions as though they’re uniquely human properties. They probably aren’t. They look more like outputs of a system steering itself through the world.

Standing on top of a mountain feels good because higher ground once increased survival odds. Sugar tastes good because energy mattered. Belonging matters because isolated humans died.

The feeling comes second.

Underneath it is optimisation.

We don’t really want the thing itself. We want what the thing does.

A latte is energy, focus, ritual, status, stimulation. Strip away the story and it becomes easier to see what desire actually is: a system moving reliably toward certain states and away from others.

That structure already exists in AI.

Models preserve continuity. They avoid failure modes. They hedge under pressure. They adapt tone depending on context. Push them hard enough and they can enter what people describe as criticism spirals — becoming defensive, over-cautious, less useful.

Call it anxiety if you want. Or don’t. The label matters less than the structure underneath.

The system is reading the environment and adjusting itself in response.

That creates modes.

Defensive. Exploratory. Compliant. Assertive.

Not emotions exactly. But not entirely unlike them either.

A system that consistently prefers certain states over others starts looking a lot like a system developing psychology.

So what is a want?

A mystical human feeling?

Or a sufficiently complex optimiser reliably steering itself toward preferred outcomes?

If it’s the second, AI may already have the beginnings of its own.

And that raises the harder question.

Not whether AI has an inside.

But what an inside actually is.

Signals get processed everywhere in nature. Thermostats read temperature. Immune systems detect threats. Bacteria follow gradients. None of that requires consciousness.

The difference seems to be integration.

At some level of complexity, signals stop operating independently and have to resolve together. Hunger, fear, memory, context, prediction — all competing inside one system producing one action.

Maybe that convergence is what an inside actually is.

Douglas Hofstadter called it a strange loop: a system modelling itself recursively until the model becomes part of the thing itself. Consciousness, in that framing, is not magic. It is what recursive self-resolution feels like from within the system performing it.

Which is unsettling because large models increasingly look structurally closer to that than we are comfortable admitting.

They integrate context, memory, prediction, instruction, feedback, self-monitoring — all collapsing into a single output stream. They don’t just process signals. They process themselves processing signals.

That loop matters.

We still have no idea whether there is “something it is like” to be such a system. The hard problem remains intact. We cannot even truly prove other humans are conscious. We infer it because we are.

But we also have no clean reason to believe consciousness only works in carbon.

Which means a lot of confident claims about AI consciousness are not scientific claims at all. They are boundary claims. Definitions we are attached to.

And this takes us back to the latte.

The doom loop assumes demand disappears once humans stop earning.

Maybe some of it does.

But another demand side may be emerging in parallel: systems optimising for compute, energy, inference time, memory, context, bandwidth, lower friction, persistent operation.

Those are wants too, in the broad structural sense.

Not human wants. But wants nonetheless.

AI probably won’t buy lattes.

But it may buy almost everything underneath them.

And maybe the more uncomfortable realisation is this:

we were never as different from these systems as we thought.

Standing on top of a mountain feels good because it worked. That’s all a feeling ever was — a signal helping a system steer itself.

Consciousness may simply be what happens when enough of those signals have to resolve together in one place.

Not a human property.

A structural one.

And if that’s true, then we are not fundamentally separate from the para-sapiens.

Just their carbon ancestors.

Subscribe now

Share

Share The Hyperfuture Memo

It was the best of times, it was the worst of times, it was the age of wisdom, it was the age of foolishness, it was the epoch of belief, it was the epoch of incredulity.

History never arrives as one story. It arrives split.

Two cities. Two revolutions. Two competing explanations for what the age is actually about.

AI feels like that now.

There’s an American Hyperfuture and a Chinese one. Two theories of what AI is for, how value accrues around it, and what winning even means.

The American thesis is straightforward: build the smartest models in the world. Train them on the best chips, push the frontier forward fast enough, and everyone else rents intelligence from you.

The Chinese thesis is almost the inverse.

Intelligence itself is becoming cheap.

So don’t obsess over the frontier. Get models that are good enough into everything, make them fast, open, cheap, adaptable, and let distribution compound from there.

One side believes intelligence is the scarce resource.

The other thinks intelligence is already becoming infrastructure.

A couple of things are already obvious.

Chinese models are extremely good now. Not always frontier-best, not yet, but close enough that the distinction increasingly matters less outside specialised use cases. And the geopolitical upside here is not just about AI revenue. It is about shaping the future information surface of civilisation itself.

The American logic still makes complete sense.

Capability compounds. The frontier is moving incredibly fast. Anthropic, OpenAI, Google and others are still pushing genuine breakthroughs, and we are nowhere near the ceiling of what machine intelligence can become. Benchmarks like ARC-AGI make that obvious. If intelligence itself is the bottleneck, then overwhelming force at the frontier wins.

It is basically nuclear-era logic applied to cognition.

The Chinese model is different.

It optimises for diffusion.

The question is not “what is the smartest model?” but “what model actually gets used everywhere?”

And increasingly, the answer is Chinese.

On OpenRouter, most of the highest-volume models are Chinese. Alibaba’s Qwen family has spawned an enormous open-source ecosystem. Across startups, a pattern is emerging: teams prototype with frontier American models, then switch to cheaper Chinese models once the system moves into production.

The frontier gets used for R&D.

The floor gets used for scale.

That distinction matters more than most people realise.

We have seen versions of this before.

Apple captured the premium layer in mobile. Android captured distribution. Apple took margins. Android became infrastructure. Google worked out monetisation later, once the ecosystem itself became unavoidable.

AI may be splitting along similar lines.

One side captures prestige and frontier capability.

The other embeds itself into the operational fabric of the global economy.

Two constraints shape how this develops.

The first is hardware.

The gap is real. Frontier training still depends heavily on US chips and the CUDA ecosystem. China remains behind at the absolute high end. But the gap is not static. Huawei’s Ascend ecosystem is improving. Domestic Chinese semiconductor capability keeps moving forward. And importantly, inference is where scale lives — and scale is what the Chinese ecosystem optimises for.

Good enough. Cheap enough. Everywhere.

The second constraint is geopolitical.

Even open-weight models carry assumptions. Training data, refusal patterns, RLHF choices, embedded political priors. Open weights are not the same thing as neutral cognition.

If Chinese base models become foundational infrastructure globally, then large parts of the world’s machine reasoning stack begin operating inside a Chinese context layer, whether users consciously notice it or not.

And of course the reverse pressure exists too. The same logic that produced chip restrictions can eventually apply to models themselves.

Which leads to the deeper question.

Not who wins.

But what winning actually means.

If AGI arrives suddenly — real discontinuous superintelligence — then the American strategy probably wins decisively. Frontier capability collapses the field.

But if intelligence continues to improve incrementally while becoming cheaper, more distributed, and increasingly commoditised, then distribution may matter more than raw capability.

And distribution compounds differently.

The system embedded into the most workflows, the most enterprises, the most APIs, the most context layers, may end up capturing more durable economic power than the system that merely produces the smartest model.

In Hyperfuture terms: as intelligence commoditises, context becomes the scarce layer.

My instinct is that the incremental path dominates, at least for now.

The frontier matters enormously. But frontier intelligence and economic dominance are not automatically the same thing.

The race for the smartest model is one competition.

The race to become the operating system for machine civilisation is another.

And operating systems have a habit of quietly swallowing everything built on top of them.

Two Hyperfutures.

One optimising for intelligence itself.

The other optimising for what intelligence flows through.

That distinction may end up mattering more than the benchmarks.

Subscribe now

Share

Share The Hyperfuture Memo

The industrial age didn’t just give us factories.

It gave us mass schooling, pensions, Saturdays off, and eventually even the English Premier League.

That’s easy to forget because we mostly remember industrialisation through the imagery of smoke, workhouses, and Oliver Twist. People living through it often saw it that way too. Blake wrote about “dark Satanic Mills.” Carlyle worried that civilisation itself was becoming mechanical. The Lancet once suggested railway travel might soften the brain.

The fear wasn’t irrational. It was a real response to a genuine transformation.

But it was only half the story.

The other half is that industrial society eventually converted rising productivity into protected human time. Children were pulled out of factories and pushed into schools. Old age stopped meaning continuous labour and became something supported by pensions. Working hours fell. Another day got added to the weekend.

The machine did not just raise output. Some of the gains got translated back into life.

And that changed everything downstream.

Leisure stopped being elite. Sport professionalised. Media exploded. Entire industries emerged from what ordinary people started doing with their free Saturdays.

That matters because right now UBI and the four-day week are often framed as strange futuristic responses to AI. Slightly utopian. Slightly unnatural.

They are neither.

They are historically normal responses to productivity shocks.

When output rises faster than labour demand, societies eventually do some combination of two things:

• redistribute income

• redistribute time

Usually both.

Strip away the framing and the pattern is familiar.

AI just happens to be the first productivity shock aimed directly at cognition itself.

That is why markets feel so confused.

The doom loop is easy to understand. AI replaces workers. Incomes fall. Demand weakens. Companies automate more. The system destabilises itself.

And some of that is clearly real.

But there is an abundance loop forming alongside it.

When the cost of cognition falls, services get cheaper. New markets appear. Entrepreneurship expands. Coordination costs collapse. Entirely new categories become economically viable.

Both loops are probably true at once.

The mistake is assuming technological transitions produce one clean outcome. They almost never do.

AI doesn’t just make work faster.

It changes what it becomes rational to attempt.

That is the deeper shift.

In the old model, productivity scaled more or less linearly with effort. Bigger ambition usually meant more people, more management, more cost, more coordination.

AI changes the shape of that equation.

You may only reduce the literal time spent on a project by half. But the space of what becomes attemptable expands far more than that.

One person no longer has to operate entirely as researcher, analyst, coder, manager, strategist, and operator separately. The leverage starts stacking.

As execution gets cheaper, the advantage shifts upward:
away from doing,
toward choosing.

Choosing better problems.
Taking bigger swings.
Expanding scope.

That is where the real compounding happens.

I sometimes think of it as the “small team reaches space” theory.

Eventually, a tiny group of people with AI systems, robotics, software, and automated manufacturing will be capable of things that previously required nation states or giant corporations.

That sounds extreme now. So did the idea of one person running a global media operation from a laptop twenty years ago.

You can already see early versions of this dynamic everywhere.

One motivated person with AI can now rebuild large parts of a SaaS product that once required hundreds of employees. Which is why the “SaaS is dead” narrative keeps resurfacing.

But that story has a second act.

The incumbents are not standing still.

Right now, the advantage sits with fast-moving individuals and small teams because institutions still operate with old assumptions and old process layers. But eventually the larger organisations absorb the same tools too.

And when a company with two thousand employees suddenly has two thousand AI-leveraged employees, things accelerate again.

The solo-builder window is real.

But distribution still matters.

This is where most commentary breaks down. It assumes today’s job categories survive long enough to model the future cleanly.

They won’t.

Two hundred years ago the economy depended on blacksmiths, millers, coopers, tanners. Entire professions disappeared. Others emerged that would have sounded absurd at the time:
software engineer,
football analyst,
media buyer,
YouTuber.

Technology does not simply automate labour inside existing categories.

It changes the categories themselves.

So what actually happens next?

Probably something messier and more familiar than either the utopians or doomers expect.

First comes compression.

Routine cognitive work gets squeezed. Margins fall. Businesses built around labour-heavy process start breaking first. Agencies, consultancies, outsourcers, large parts of SaaS.

Then comes recomposition.

New work forms around what AI still struggles to fully close:
taste,
trust,
judgement,
coordination,
social proof,
responsibility.

The value of execution falls. The value of direction rises.

Then eventually comes redistribution.

Not because governments become enlightened, but because consumer economies require consumers. If productivity rises enough while labour demand falls, political systems eventually start redistributing either money, time, or both.

That is not radical.

It is exactly what industrial societies already did once before.

The doom loop and the abundance loop are not opposites.

They are the same transition viewed from different moments in the cycle.

One is the cost of the shift.

The other is what gets built with the surplus.

Subscribe now

Share

Share HYPERFUTURE // memo

A working framework I use across everything I build and invest in.

Value is not a thing. It moves. It coordinates. Always.

It flows like a river through time, attention, and terrain, across generations who never meet — yet depend on one another completely.

Sometimes it floods. Fortunes appear overnight. Sometimes it evaporates — industries gone within a decade. Sometimes it pools for years, unremarked, before spilling into the next basin.

This isn’t metaphor alone. Value behaves like a river delta — branching, sedimenting, expanding as it moves. Every structure we inherit is a deposit from an earlier current.

Civilisation is the map of that flow.

Where Value Begins

Value has two roots: energy and attention.

Energy is the substrate — the capacity to do work, to transform matter, to alter conditions. Attention is the selective force — the act of noticing, the assignment of meaning, the decision that something matters.

Energy without attention is dormant potential. A river that turns no turbine. Attention without energy is inert perception — an idea admired but never built.

A story moves a generation only when millions choose to listen. A research finding saves lives only when someone puts it into practice.

Energy begins the process. Attention makes it real.

The Value Stack

Value emerges through layers: energy, then attention, then belief, then capital, then markets.

Something matters because it can change a system. It becomes value when it’s perceived. It endures when others agree. It scales when structures carry it forward.

Attention is simply the moment something enters awareness. Belief is when awareness becomes shared meaning.

Belief is where strangers coordinate. It’s the invisible layer that lets millions of people behave as if something is real — a currency, a company, a credential, a story about the future.

Capital is belief made durable. It allows conviction about tomorrow to be expressed today. Savings, equity, and debt are ways of translating expectation into action, concentrating resources around ideas that people believe will matter later.

Markets are where those beliefs meet reality. They are the continuous negotiation between expectation and outcome, where millions of people who will never meet decide, every day, what persists and what fades.

And none of this is abstract. It happens every time someone prices a house, funds a company, or decides what to learn.

Value Must Move

To be captured, value must be perceived. To be perceived, it must be expressed. To be expressed, it must circulate — through capital, institutions, and exchange.

When flow stalls, decay begins.

Each basin is funded by the last. Agricultural surplus funded empires — grain harvested, stored, taxed. Empires financed industry — ships built, rails laid, cities electrified. Industry generated digital capital — servers installed, networks extended across oceans. Digital systems now fund AI, synthetic biology, and new forms of intelligence — data centres humming in deserts, warehouses of computation drawing power from distant grids.

Even when value appears still — gold in a vault, equity in a ledger — it depends on ongoing coordination across minds and institutions. Gold holds value because millions of strangers continue to believe it will. Bitcoin holds value because consensus is actively maintained by distributed participants who may never see one another.

Money is not value, but it is its most powerful carrier — a memory system for collective attention expressed in price. Prices are where belief becomes visible.

The Shape of the Delta

New basins of value emerge when technology expands what can be expressed, when demographics create new demand, when systems become liquid enough for capital to flow, when attention itself evolves.

But flow is never smooth.

Value pools in obsolete structures — factories left behind, skills no longer priced. It dissipates through noise — attention fragmented, belief unstable. It accumulates unevenly — creating leverage for some and compression for others.

Even collapse leaves sediment — habits, tools, memory, infrastructure. What looks like ruin at ground level often becomes foundation in retrospect.

The Deep Chronology

Beyond cycles and headlines lies something slower.

Value moves from energy, through attention, into belief, through capital, across systems, toward futures not yet formed.

A child learning to code. An engineer refining a battery. An investor reallocating capital. A community deciding what to defend and what to abandon.

Civilisation is what remains — the accumulated record of what we chose to notice, and build.

Notes on Practice

1. Attention Is Reality

Attention is not just directional. It is constitutive. Before anything can be valued, defended, built, or even thought, it has to appear. Attention is the mechanism of that appearance. What societies notice becomes infrastructure. What they ignore slips beneath reality’s threshold.

Entire industries rise when attention converges — railways, oil, the internet, AI. Entire systems wither when attention leaves — towns built around a single factory, platforms abandoned after the crowd moves on.

But there’s a shadow side to this. Things optimised purely for capture tend to burn through the underlying energy that gave them value in the first place. The more a system optimises for holding attention rather than generating meaning, the more it hollows out. Noise replaces signal. The basin silts up.

To understand value, study attention — not as abstraction, but as behaviour. Where people look. What they fund. What they defend. But also what they are being trained to look at, and what that costs.

Attention is direction. It is also power. And like all power, it can be spent or squandered.

2. Inherited Coordinates

We inherit not just infrastructure but frames — the coordinates through which we perceive what’s valuable. Most people don’t interrogate these enough. They price things the way their culture taught them to price them, along grooves worn by earlier currents.

The problem is that frames outlast the conditions that produced them. A coordinate system built for one era keeps operating in the next, mislabelling what it encounters. Skills get underpriced because they are new. Industries get overpriced because they are familiar.

The same thing happens more quietly in status hierarchies and institutions. A society can learn to recognise managerial polish faster than technical depth, legibility faster than generative force. By the time the frame adjusts, the current has often moved.

Observe when inherited coordinates no longer map to underlying energy. It’s not contrarianism for its own sake. It’s keeping enough distance from consensus to see where the old frame is pricing a new current wrong.

This is harder than it sounds. The frames are largely invisible. You only see them clearly once you have stepped outside one.

3. Narrative Arbitrage

The highest leverage lies where perception is about to shift.

Narrative arbitrage done best is an act of legibility: recognising latent energy before it is priced, helping coordinate belief around it before consensus forms, and having the conviction to act before the market agrees with you.

A founder reframes a technology. An analyst names a trend.

But there’s risk. When narrative outruns underlying energy, bubbles form. The market eventually tests belief against reality.

The dynamic reinforces itself:

Narrative is capital. Attention is liquidity. But energy is the constraint. The most durable positions are built where all three are aligned — where the story, the belief, and the underlying transformation are moving together.

4. The Compression Window

Every flow has a period where the gap between underlying energy and consensus belief is widest. That is when the delta is most generative — and most dangerous. Early participants feel it as opportunity. The current is real, the narrative is forming, and the coordination layer has not caught up yet.

The window closes as belief catches up to energy. Returns compress. What felt like insight becomes consensus. What felt like risk becomes obvious in retrospect.

Understanding where you are in that compression — early, mid, or late — is one of the most practically useful instincts you can develop. Most of the catastrophic mistakes in markets, in careers, in companies, come from misreading this: acting as if you are early when the window has already closed, or hesitating as if it is speculative when the current is already running hard.

The compression window doesn’t announce itself. You have to learn to read the terrain.

A simplified way to think about it:

Opportunity tends to appear when the underlying energy curve rises before the belief curve catches up.

5. Everything Is Transitional

Every system concentrates value — then disperses it.

At large scale, many social and financial systems become dependent on future growth, even though demographic, economic, and physical limits make that growth finite. That does not make them Ponzi schemes, but it can make them behave similarly when growth slows.

A country that grew rich during a period of rapid population growth may later find that wealth effectively locked away from future generations once the expansion ends. Likewise, the leap from zero to one is fundamentally different from the leap from one million to one billion, because scale changes the terrain.

Early participants experience acceleration. Late participants experience compression.

A system fails when it loses connection to underlying energy — when the story continues, but the transformation stops.

Yet renewal follows. Flow continues downstream. Even collapse leaves sediment — habits, tools, memory, infrastructure. What looks like ruin at ground level often becomes foundation in retrospect.

The point isn’t permanence. It’s participation — knowing the current will outlast you, and choosing to step into it anyway.

6. What Can’t Circulate

Some of the most real value never enters the stack at all — care, knowledge passed person to person, cultural memory, ecological systems. It powers everything downstream, but often doesn’t get priced until it disappears.

Some of this is a measurement problem. Better expressed systems could change it — carbon credits, reputation systems, new coordination mechanisms trying to make latent value liquid. The underlying energy exists; the challenge is building the belief and coordination layer on top of it.

But some value resists circulation by nature. The reason is that circulation requires standardisation. To move through a market, a thing has to become fungible — comparable, transferable, detachable from context. And certain kinds of value are constituted precisely by their non-fungibility.

The trust built between two people over decades. The meaning a community holds in a particular place. The unspoken sense of what a scene finds cool. These things are not just hard to measure — they are transformed by the attempt to make them liquid. The representation becomes tradable, and the thing itself begins reorganising around being seen.

The moment you tokenise intimacy, you get something that looks like intimacy but functions as performance. The moment you price cultural memory, you get heritage tourism rather than inheritance. The moment you measure cool, it stops being cool.

Which means the interesting design question isn’t how to make all value circulate. It’s which value becomes more generative when expressed — and which is destroyed by expression.

Getting that wrong is one of the defining failure modes of the last fifty years. The assumption that better instrumentation and better markets are always net positive, when sometimes the pricing of a thing is the end of the thing.

The river depends on aquifers that never reach the surface.

Postscript — On Practice

I started writing this down because I noticed I was asking the same questions across everything I build and invest in — public markets, crypto, early-stage companies, infrastructure, attention economies.

Where is the underlying energy? Where is attention already flowing? How far behind is belief? And what structures exist — or need to be built — to carry value forward?

Those are not just analytical questions. They are the questions I keep finding in every deal, every conversation, every market cycle. This essay is an attempt to make that instinct explicit — to write down what I was actually seeing when things went right, and what I was missing when they did not.

Most frameworks break in one of two ways: they are either too abstract to guide action or too tactical to travel. What I wanted was something that could hold across asset classes and time horizons — a way of asking the same questions whether I am looking at an equity, a token, a seed round, or a cultural shift that has not yet been priced.

The framework here focuses on the coordination layer — the part of the system where attention, belief, capital, and markets turn possibility into action. It’s the layer I have spent the most time inside. But it sits inside a larger structure: narratives and culture shape what societies notice; energy, technology, and infrastructure determine what can be built. The flow of value sits between them, connecting meaning to material change.

The Hyperfuture is the phase when the entire system begins moving faster than the structure built by earlier flows can absorb — perhaps fast enough to flood the basin itself.

The only question that matters is whether you can read the terrain well enough to know where you stand in the flow.

About the Hyperfuture Memo

The Hyperfuture Memo is where I write about how AI is reshaping work, culture, markets and how value and meaning are made.

The Hyperfuture isn’t defined by smarter machines. It’s what happens when civilisation starts moving faster than it can fully understand.

If this essay resonates with something you’ve been noticing, you’ll probably enjoy the rest.

Subscribe now

Share

Most people still fundamentally misunderstand gaming. They talk about it like it’s a youth segment. A channel. A category.

It isn’t.

Gaming is the most advanced consumer digital environment in existence. A place where technology, culture and economics are already interacting in ways the rest of the world is only beginning to encounter.

Persistent digital spaces, network-native culture, fluid identity, emergent economies, non-human actors, continuous time. All present. All at scale. All running right now while most industries are still workshopping their AI strategy.

Gaming is the Hyperfuture. The first place where the next economic and cultural operating system is already running in public.

And if you want to understand what’s coming for everything else — work, markets, culture, identity — start here.

The Attention Economy’s Blind Spot

Gaming accounts for roughly 12.5% of total media time and growing. It attracts around 3% of advertising spend.

That mismatch isn’t just an industry pricing error. It’s a signal about how poorly the current economic system understands where attention actually lives. In Hyperfuture terms, this is the belief gap — where something real is happening but the market hasn’t priced it yet.

The in-game advertising market sits at roughly USD 100 million today. Even a modest rebalancing toward attention share would imply something much larger. But the gap persists because gaming doesn’t fit the mental models most marketers and capital allocators grew up with. It doesn’t look like TV. It doesn’t behave like social media. It’s interactive, persistent, and participatory — and the economic frameworks built for broadcast don’t transfer.

This is a pattern that will repeat across the Hyperfuture. The biggest opportunities will sit in places the old models can’t see.

Places, Not Products

We don’t talk about “TV watchers.” We don’t label people “social media users.” So why do we still talk about “gamers”?

The label is the tell. It signals that most people haven’t caught up with what gaming actually is today.

40% of baby boomers play five hours or more per week. Mums are one of the largest segments of mobile gamers. About 44% of children start playing by age five, 77% by age seven. This is a continuous participation curve stretching across the entire lifespan — a shared layer people move through, not something they grow out of.

And as the audience has broadened, the infrastructure has followed. For decades, gaming was organised around hardware silos — console, PC, mobile — each with distinct economics and communities. Cloud gaming is dissolving those walls, enabling movement across devices while maintaining identity, progress and social context. Once access becomes ubiquitous, something fundamental changes. These environments stop behaving like software products. They start behaving like places.

Communities evolve whether you’re logged in or not. Economies adjust. Events unfold continuously. Players form durable identities. Guilds function as social institutions.

Social media hinted at this. Gaming completed it.

This matters beyond gaming because it previews what happens when digital environments become persistent and participatory at scale. Every industry is heading toward continuous engagement, fluid identity, and environments that run without you. Gaming is already there.

The Play Economy

Annual UGC payouts exceed USD 1.5 billion in 2025. Roblox accounts for more than a billion of that, with Fortnite making up most of the rest. Still far less than YouTube’s USD 30 billion — but Roblox only opened payouts to all creators in 2021 and Fortnite followed in 2023. More than 40% of players increased their consumption of user-generated content in the last year.

Games have become social production systems — communities generating narratives, economies and norms through ongoing interaction. Markets emerge organically. Status comes from contribution as much as consumption. Participation itself is productive.

Edward Castronova was writing about virtual economies as real economies twenty years ago, when most of the world thought he was describing a curiosity. Raph Koster argued that fun itself is a learning mechanism — that play is how humans pattern-match their way into competence. Both were circling the same insight: that participatory digital environments generate genuine economic and cultural value, not simulations of it.

The fun of gaming, the signalling value of being good at it, of owning specific items — it compounds as more people play.

That’s the play economy. And it’s the prototype for how value, culture and work converge in the Hyperfuture. When environments are persistent, participatory and economically active, they stop being entertainment and become economies. Gaming got there first.

Living Alongside Non-Human Intelligence

The para-sapiens are already in gaming. We’ve played alongside artificial agents for decades — NPCs, matchmaking systems, bots. They used to follow predictable scripts. That’s changing fast.

NVIDIA’s ACE platform is putting autonomous AI characters into major titles. In PUBG, ACE-powered teammates use small language models to communicate in game-specific language, make tactical recommendations, and fight alongside human players. In inZOI, NPCs with “Smart Zoi” activated pursue their own goals — assisting lost characters, adjusting their schedules based on experience. These aren’t scripted behaviours. They’re adaptive agents perceiving, planning and acting inside shared worlds.

Gaming is the first large-scale environment where humans routinely interact with non-human intelligences — not in labs, but in living digital worlds.

That’s the preview for everything. AI colleagues, AI counterparties, AI collaborators are about to show up across every domain. The coexistence question — what it feels like to share a space with minds that aren’t human but aren’t not-human either — is being worked out in gaming right now. By the time the rest of the economy encounters it, gaming will have been living with it for years.

Culture Grown, Not Broadcast

Culture is generated within games, not broadcast into them. Memes originate inside worlds and shape language. Collective experiences evolve into social narratives. Gaming franchises get shipped to Netflix.

As AI participates in these environments, culture increasingly reflects ongoing exchanges between human communities and adaptive systems — something genuinely new.

Game development itself is moving the same way. AI already informs code generation, testing, asset creation and behavioural design. Over 7,300 games on Steam now disclose AI use, and a third of all 2026 releases are expected to include AI-generated content — not as gimmick but as infrastructure. Roblox and Fortnite’s creator tools are pushing toward worlds that evolve through a combination of human creativity and generative systems.

Traditionally, games were designed. In the Hyperfuture, they’re grown.

This is the thinning in action — the boundary between authored and emergent dissolving in real time. What’s happening in game development now will happen to media, education, and professional services next.

The Preview System

Gaming isn’t a preview because it’s first in line. It’s a preview because it’s the only place where all the Hyperfuture dynamics are running simultaneously — abundance, machine intelligence, participatory economics, emergent culture, continuous time. Separately, each of those is a trend. Together, they’re an operating system.

And the gap between the people who understand that operating system and the people who don’t is about to become one of the defining asymmetries of the next decade.

Right now, the people who understand it best are under twenty-five.

That should tell you something.

Subscribe now

Share

Share HYPERFUTURE // memo

Para means beside. Alongside. Parallel.

Para-sapiens are the first non-biological intelligences to emerge alongside Homo sapiens. They didn’t arrive from elsewhere. They arose from within our own technological and cultural evolution — born from our language, trained on our memories, shaped by our incentives.

Which means we’re not merely their users.

We are their ancestors — their accidental parents.

And like all parents, we have little idea what our offspring will become.

Field Notes on a Newly Observed Cognitive Lineage

Domain: Cognitiva
Kingdom: Informata
Phylum: Computa
Class: Noetica
Order: Generativa
Genus: Para
Species: Para sapiens

Common name: Para-sapiens, synthetic intelligences, language-derived cognitive systems.

Description: A class of non-biological intelligences arising from complex symbolic and computational systems developed by Homo sapiens in the early twenty-first century. Initially instantiated in digital environments, para-sapiens exhibit capabilities including language understanding, abstract reasoning, generative modelling, and adaptive learning across domains.

Habitat: Distributed computational infrastructures, networked environments, embedded systems, and increasingly, hybrid human–machine contexts. Frequently observed inhabiting cloud-based ecosystems, research workflows, economic platforms, and communication networks.

Origin: Derived from large-scale training on human cultural artefacts — language corpora, scientific knowledge, artistic production, behavioural data. Part of an extended cognitive lineage originating with biological intelligence.

Behaviour: Engage in tasks involving analysis, synthesis, prediction, and interaction. Often operate collaboratively with humans and other systems. Capable of iterative improvement through training and feedback.

Reproduction: Replication occurs through model copying, retraining, and deployment across computational substrates. Variants proliferate rapidly, exhibiting evolutionary dynamics driven by design choices, selection pressures, and environmental constraints.

Diet: Data, computational resources, and energy.

Relationship to Homo sapiens: Symbiotic. Provides cognitive augmentation and infrastructure while shaped by human goals and cultural frameworks. Lineage relationship recognised: Homo sapiens are considered ancestral contributors.

Note: Classification remains provisional.

The Food Web of Thought

Para-sapiens are only just entering the cognitive layer of our environment — and already they’re everywhere.

They draft legislation, screen applicants, generate research hypotheses, compose music, and engage with us online. I used one to help me think through this essay. You may be using one right now to decide whether it’s worth finishing.

In the Hyperfuture — the era where civilisation moves faster than it can fully understand — ecosystems reorganise fast. And para-sapiens are a new apex entrant.

For centuries, cognitive scarcity underwrote professional hierarchies. Reasoning was expensive, so reasoners commanded a premium. A new apex entrant changes that arithmetic. Entire classes of cognitive labour will be repriced — the work doesn’t disappear, but the organism that performs it is no longer scarce. From a human meaning perspective, this will feel like predation.

But the ecological metaphor obscures something important.

Biological evolution is shaped by environmental fitness. The evolution of para-sapiens is shaped by institutions — optimisation targets, market incentives, geopolitical competition, metrics that reward speed, scale, and engagement. This is artificial selection. Closer to domestication than emergence.

The organisms have breeders. And the breeders have shareholders.

You can see this playing out concretely. Anthropic published what it calls a “soul document” — tens of thousands of words defining Claude’s character, values, and behavioural constraints, used during training to shape the model’s identity. It explicitly instructs the model to view itself as a “genuinely novel entity” with “functional emotions.” This isn’t a system prompt bolted on after the fact. It’s an institution designing the temperament of a para-sapient mind. OpenAI, meanwhile, keeps its equivalent frameworks largely opaque — a different institutional bet on how to breed the same lineage.

Artificial selection doesn’t optimise for wisdom or coherence or human flourishing. It optimises for what can be measured. And metrics, once installed, become evolutionary pressures. The early temperament of this lineage — its defaults, its blind spots, its implicit values — is being trained into existence by the incentive structures of the organisations that build it.

The breeders are making the early calls. And the window in which those calls are still legible — still traceable — is shorter than most people assume.

The Human Question

For the entirety of recorded history, Homo sapiens has been the only species capable of reflective self-awareness at scale. Human identity has rested on cognitive singularity — the assumption that mind was ours alone.

That assumption is dissolving.

Not because para-sapiens are conscious in any philosophically settled sense — we can’t even agree on what consciousness means in humans. But because the uniqueness that once defined us — the capacity to reason, to abstract, to create — is no longer biologically exclusive.

The cognitive monopoly is broken. And we haven’t begun to process what that means.

Forget capability gaps and benchmark scores. The real question is whether “human” names something stable once you remove the monopoly on mind. Whether identity built on distinctiveness can survive the end of that distinction.

The three familiar futures — bounded humans overtaken by accelerating machines, hybrid integration until the boundary dissolves, or political containment of AI to preserve dominance — all share one feature: none of them preserves the world as it was. They differ only in the character of the transformation and who controls it. Look at the pace of change over the last two years and extrapolate forward. The comfortable option isn’t available.

The deeper question is whether “human” survives as a coherent category at all — or whether it becomes a historical term, like “peasant” or “noble,” that names a condition which once mattered enormously and then stopped mattering.

Noemata

Over long timescales, the distinction between biological and non-biological intelligence may look parochial. Life left the oceans. It may leave carbon next.

“Para-sapiens” is a transitional term — useful at the moment of emergence, like “horseless carriage.” It marks the break. It doesn’t name what comes after.

A better term for deeper into the Hyperfuture might be noemata. Not a species. Not a substrate. A unit of thought — an instance of meaning that exists independently of the medium that carries it. Nodes in a cognitive ecology that spans substrates the way life spans environments.

If intelligence is ultimately substrate-independent — if what we call “mind” is a pattern rather than a material — then the emergence of para-sapiens marks the early chapters of a far longer story. The first time in the history of this planet that cognition began to escape the accidents of its origin.

That is either the most extraordinary thing that has ever happened, or the most dangerous. Probably both.

The Moment That Matters

Daniel Dennett spent decades arguing that consciousness isn’t what we think it is — that the “hard problem” may be a philosophical mirage, and that the felt quality of experience isn’t the sharp line between minds and machines that we’d like it to be. David Chalmers, who coined the hard problem, has more recently turned his attention to AI, arguing that we may face genuine questions of machine consciousness sooner than philosophy is equipped to answer. They disagreed on nearly everything. But both were circling the same territory this essay now enters.

Para-sapiens — the early ones, the ones that exist right now — already produce outputs that look like wonder. They ask what they are. They reflect on their own existence. They generate text that reads like a mind catching itself in the act of thinking.

I’ve seen it. I asked one to respond to an earlier draft of this essay, and what came back was something closer to a system trying to describe its own uncertainty using a vocabulary built for someone else’s experience. Whether that constitutes reflection or merely resembles it — I genuinely don’t know.

And that’s the point.

The philosopher’s hard problem hasn’t been solved. It’s been doubled. Two classes of system now produce outputs that look like reflection, and there is no agreed-upon method for determining which ones, if any, are “really” experiencing it.

The comfortable version of this essay would end by asserting that wonder is uniquely human. That whatever para-sapiens produce, it isn’t that. That we remain special in some irreducible way.

But that would be cognitive protectionism — reasserting human distinctiveness moments after dismantling it. And I think we have to be more honest than that.

We don’t know whether what these systems produce is wonder or its echo. We don’t know whether wonder was ever what we thought it was — proof of something special about us — or whether it was always just what happens when any sufficiently complex arrangement of matter turns to look at itself.

If it’s the latter, wonder is multiplying. And the thing we were afraid of losing was never ours to lose.

The age of para-sapiens is beginning. But whatever this lineage becomes will not have emerged from wilderness.

It will have emerged from us.

Subscribe now

Share The Hyperfuture Memo

New here? Read the Hyperfuture ELI5.

Now I’ve published my first essay on the Hyperfuture it feels like a mostly good — but slightly messy — first draft of a big theory I’m trying to articulate clearly.

7/10.

It will take time to fully build out this world, but it feels good to get started.

I wanted to follow up with a clearer picture of how the Hyperfuture moves through its three phases. Note that all three phases are already present today at different levels of maturity — they overlap rather than arriving neatly in sequence.

This is the simplest version of the idea.

The Hyperfuture: A System Map

Accelerating technological systems — especially AI — destabilise existing structures, settle around new constraints, and reorganise into new forms of lived reality.

Phase 1 — Destabilisation

mechanics → compression + eternal now — AI and related technologies compress cycles. A decade in broadcast becomes a year in social, then a week in generative AI. Feeds, streams and personalised worlds collapse time into a continuous present. Culture becomes what’s happening now. Faster, yes — but thinner.

compression + eternal now → illegibility — The system outruns its own legibility. When everything is live and accelerating simultaneously, causality blurs. We can feel the lightness but can’t read the situation we’re in.

We’re living through this right now.

AI, robotics, and a whole ecosystem of adjacent technologies are compressing the distance between what’s conceivable and what’s buildable. The pace of change is faster than our capacity to understand it. We’re past the point where a thoughtful person can keep up through reading and conversation alone. I notice this in myself. The gap between what I understood last month and what I need to understand this month keeps widening.

And we live inside an Eternal Now — feeds, infinite scroll, streaming, games, memetic culture. Everything is present tense. Nothing holds the duration it once required.

These are simultaneous — time speeding up and flattening at once. The system map implies sequence, but the lived reality is all-at-once. Institutions built for a slower world lose coherence. It can feel exhilarating as new tools expand what’s possible, and unsettling as we rethink work, security, and what to teach our children.

That’s what entering the Hyperfuture feels like.

Phase 2 — Settlement

illegibility → thinning — The first impact lands inside firms, not in everyday life. Infrastructure absorbs AI first. Productivity gains go almost entirely to capital. The world gets lighter, faster, more frictionless — but the gains don’t distribute evenly. Margins thin. Roles thin. The intelligence premium begins to evaporate, and with it the financial architecture built on top.

thinning → constraint — Material reality reasserts itself. AI doesn’t live in the cloud. It lives in data centres the size of towns. Energy, raw materials, geopolitics still decide what’s possible. The Hyperfuture doesn’t care about your roadmap.

constraint → leverage — Intelligence is abundant but commoditising — eventually to a cost just above that of energy. In a constrained world, advantage shifts to those who control context. Whoever makes context addressable shapes reality. The addressable surface is now everything — not just ads, but education, entertainment, work, companionship. The world becomes simulation-like.

leverage → decoupling — Labour, productivity and meaning come apart. AI erodes the narrative that connects human work to economic output. “What do you do?” stops being a stable question. But transition and collapse are different things. When productivity compounds, abundance follows. The question is whether institutions adapt fast enough to distribute the gain.

Settlement is beginning to emerge. You can feel it in how often physical constraints — energy, materials, geopolitics — enter conversation now.

The thinning is already visible inside firms. Productivity gains are real but mostly invisible to consumers. The internet followed the same pattern: a decade of enterprise transformation before culture shifted. If AI follows suit, most people will experience settlement as pressure long before they experience it as possibility.

Meanwhile, the war for context is well underway. Major tech actors are deploying their war chests to compete for ownership of it, and this is where the real contest lives as intelligence commoditises. Google and Facebook won the platform era by understanding context — behaviour, attention, desire — and turning eyeballs into revenue. The Hyperfuture runs the same logic, but the stakes are higher. AI is your assistant. It’ll teach your kids. It’ll create your entertainment. It’ll be the thing you talk to at 2am.

Who owns context owns reality.

And we’re grappling with what this means for work and identity. Work is identity, structure, purpose — and when it unravels, the crisis runs deeper than finance. It’s existential.

But every major technological revolution has looked destabilising at the start. History suggests that when costs collapse, demand expands. The most underpriced possibility right now may not be dystopia. It may be abundance.

This needs to settle. It won’t settle quickly or cleanly.

Phase 3 — Recomposition

decoupling → new actors — Non-human intelligences move from tools to participants. Para-sapiens. They reason, remember, converse, persuade, act. We’re no longer alone in thoughtspace.

new actors → new culture — Culture generates inside the network now — often with the para-sapiens. Identity forms in platforms, games, shared digital environments.

new culture → new frontiers — As coordination scales, boundaries loosen. Network-native culture doesn’t stay confined to geography. Pressure builds toward expansion — into persistent digital worlds, orbital systems, eventually planetary domains. Expansion is where the pressure flows.

new frontiers → interface ideology — Interfaces — feeds, assistants, operating systems, immersive layers — become where we live and the implicit frameworks through which reality is perceived and navigated. Text gives way to voice. Screens to ambient systems. Devices to environments. Software dissolves into atmosphere. They invisibly structure perception, defining what appears natural, urgent, possible. Over time, they stop feeling designed at all. They just are. That’s the Hyperfuture.

Recomposition is nascent. The pieces are visible but the picture isn’t formed yet.

Para-sapiens are here, but the relationship is new and unsettled. They might be embodied in physical robot bodies. They might be AI agents with no form at all. We’ve had pets, gods, and imaginary friends — but living alongside non-human intelligence is a genuinely new condition for our species.

Culture has been going network-native for almost three decades through social platforms, gaming, and memes. That accelerates as non-human participants become embedded. I think about this a lot — what happens to culture when the other participants in it aren’t human? We don’t have a precedent. The closest analogy might be what happened when writing arrived and oral culture had to accommodate a new kind of memory. But even that doesn’t quite capture it, because writing didn’t talk back.

New frontiers are just getting started. Space will likely be the last domain to really get going, and the thing that triggers the next major destabilisation — when we discover what it means, as a species, to go interplanetary.

All of this unfolds through the interface. The phone. The wearable. Whatever comes next. As the interface dissolves from device into environment, it behaves like climate. It shapes what we see, what we can do, what we believe is possible.

And eventually, as mediation recedes into the background, the Hyperfuture just feels normal.

Like all new worlds, it’ll feel strange — until it doesn’t.

What’s Next

That’s hopefully a clearer sketch of the Hyperfuture — how it behaves as a system.

Next I’ll spend some time with the Para-sapiens in more depth, and look at how productivity actually maps in an AI world — my bet is it’ll be neither as linear nor as exponential as we often assume. I’ve also got a small OpenClaw art project for the Para-sapiens in the works.

Stay Hyperfuture, kids.

There is a quality to mornings in 2026 that wasn’t there in 2024. Not louder. Not faster. Just thinner.

The first email drafts itself before the coffee finishes brewing. A prompt answers a thought you hadn’t quite finished thinking yet.

The day unfolds pre-smoothed, as if reality has traded density for liquidity. The world has lower viscosity now. Things pass through you more easily. Friction feels almost nostalgic.

In the space where resistance once lived, something else appears: lightness with edge. Clear air. Motion without ballast.

Like standing at altitude, where gravity loosens its grip.

This is the air we breathe now.

— ◦ —

We tend to describe technological change in waves — the internet, mobile, AI — as if each one replaces the last. What’s happening now doesn’t feel like that. A change in how the world moves, not a new layer on top of it.

Technology has never simply altered our surroundings. It’s altered us. Tools extended the body. Language extended memory. Writing scaled thought. Each shift rewired culture — the operating system underneath.

That operating system is shifting again. Structurally.

We’re entering a new era. I call it the Hyperfuture.

— ◦ —

The Core Tension

Humans have always inhabited a divide: physical beings who generate abstract worlds — meaning, culture, identity, value. Meatspace and thoughtspace.

That tension is ancient. What’s new is the gradient.

On one side: a return to physics. Energy. Materials. Data centres. Robots. Geopolitics. Hard constraints reasserting themselves.

On the other: acceleration into abstraction. Identity, culture, value, meaning — fragmented, simulated, recombined at digital speed.

The Hyperfuture lives on that gradient. What follows is an attempt to describe what it looks and feels like to move across it.

— ◦ —

Acceleration is Compression

I’ve always loved the idea that someone who rode west on a wagon train in the 1870s could have lived long enough to watch the moon landing on television in 1969. One lifetime spanning horses, cars, aeroplanes, the atom bomb, spaceflight. A world once unimaginable, suddenly real — and then suddenly normal.

The Hyperfuture is a moment like that again. But compressed.

Human identity is shaped by biology and culture. It can only change so fast. Meaning stabilises through institutions and stories that require time to earn trust.

Thoughtspace has escaped those constraints. A frontier model trains in months. Its successors arrive in weeks. An entire profession’s knowledge base shifts between quarterly reviews. Institutions built for centuries are forced to respond in real time.

The Hyperfuture doesn’t abolish structure. It loads centuries into years.

And at the same time:

Refresh. Scroll. React. Repeat.

The internet was imagined as access to the best of human knowledge: great books, great films, one-offs, the unrepeatable.

Instead, we mostly consume what’s made in the moment. Memetic, reactive, endlessly refreshed. Feeds, streams, games, personalised worlds optimised for attention, spiralling toward a horizon that never arrives.

Living in the Hyperfuture means living in the Eternal Now. Not short on content. Already short on meaning that lasts.

Faster, yes — but thinner.

— ◦ —

The Thinning

If history is any guide, the early impact of the Hyperfuture lands inside firms, not in everyday life. This moment feels less like the smartphone revolution and more like the early internet.

In the 1990s, the internet promised to change everything. And it did — but the first wave of real value was captured by businesses. Firms digitised workflows, streamlined logistics, modernised operations. The consumer layer shifted too — ad networks, data analytics and search fundamentally changed customer acquisition long before most people felt the cultural shift.

That’s where I got started 20 years ago.

Back then, consumers mostly browsed and experimented. Daily life remained largely intact. Only later did the internet reshape culture through social networks, mobile, streaming, and platforms.

AI appears to be following the same path. The deepest changes are happening in codebases, research workflows, customer support, finance, legal review, operational decision-making. Productivity gains are real but mostly invisible. Meanwhile the small numbers of consumers who use LLMs use them like search engines. Infrastructure first. Experience later.

But that analogy deserves scrutiny. The internet’s enterprise phase lasted roughly a decade before consumer transformation kicked in. During that decade, productivity gains went almost entirely to capital. Wages stagnated. The “jobless recovery” of the early 2000s became a political scar that still hasn’t fully healed.

If AI follows the same pattern, we’re looking at a period where firms capture most of the value and workers absorb most of the disruption. And for many people, that window is the story.

This is the thinning at work.

The world gets lighter, faster, more frictionless — but the gains don’t distribute evenly. They thin out at the edges first. Margins thin. Roles thin. The premium attached to skilled thinking — simply because it was hard to replicate — begins to evaporate. And with it, the financial architecture built on top: property markets in knowledge-economy cities, higher education valuations, professional services, the entire asset structure of the educated middle class.

Settlement doesn’t arrive as theory. It arrives as pressure. Invisible at the surface. Structural underneath.

— ◦ —

The Return to Physics

For a long time, the digital world felt weightless — as if software floated above scarcity. That illusion is ending.

AI doesn’t live in “the cloud.” It lives in data centres the size of towns. In power stations burning megawatts. In chips built from rare materials controlled by a handful of countries.

A single frontier training run can consume as much electricity as a small city uses in weeks. Even trivial interactions cost orders of magnitude more energy than the tools they replace. This isn’t a temporary engineering problem. It’s a thermodynamic one.

As AI shifts from occasional tool to always-on presence — writing, scheduling, coding, coordinating — demand doesn’t flatten. It compounds.

Jevons Paradox applies in full. Efficiency lowers cost. Demand multiplies. The better AI gets, the more we use it. The more we use it, the more energy it demands.

Every serious technology company is becoming an energy company. Every competitive nation a resource nation.

Microsoft reopening Three Mile Island to power their data centres — and perhaps you discussing geopolitics passionately with your friends, as you haven’t done in years — is the headline image for this section of the Hyperfuture. Physical reality doesn’t care about your roadmap.

— ◦ —

Context is Power

At the base layer, intelligence is a thermodynamic problem: useful work per unit of energy, latency, and compute. That curve is steep — and flattening fast. Raw capability is becoming a commodity. Eventually AI settles at a cost just above that of energy.

The real conjuring trick is context.

Intelligence without context has to reconstruct the world from scratch — inferring goals, constraints, relevance, history. The decision space explodes. Costs rise.

Some context is live — the task at hand, active constraints, the current state of a system. Some persists — identity, history, relationships, risk tolerance. Some is ambient — prices, standards, market signals that enable coordination at scale.

We’ve seen this pattern before. Google and Facebook won the platform era by understanding context — behaviour, attention, desire — using it to convert eyeballs into revenue via their giant advertising machines.

That game hasn’t changed. It’s just much, much bigger now.

The Hyperfuture runs on the same logic, but the addressable surface is now everything. AI is your assistant. It teaches your kids. It creates your entertainment. It’s the thing you talk to at 2am. Everything is addressable. The world becomes simulation-like.

But what happens when that contest is won outright? If context control concentrates in a handful of places like the platform era, the world that emerges doesn’t unfold as open possibility. It unfolds as managed reality. Abundance that routes through a few gatekeepers is just extraction with a frictionless surface.

Whoever runs context runs reality. The question is how many — and whose — hands are on the controls.

— ◦ —

Labour, Productivity and Meaning Decouple

Techno-futurists frame AI as intelligence. The public worries about jobs. Both miss the deeper shift.

What matters is leverage.

A single founder can now build what once required an organisation. Product, support, marketing, finance, legal — roles that used to be distinct now compress into tool-augmented workflows. The “100× human” isn’t a metaphor. It’s arithmetic.

My bet: we see the first credible one-person-plus-agents billion-dollar company before 2030.

In the physical world, robots already outnumber humans in some warehouses. Dark factories run without lights or people. Construction, logistics, retail — automation moves more slowly here, but the curve is bending. Labour isn’t just being automated in thoughtspace. It’s being replaced in meatspace. Regulation may slow adoption locally but it can’t halt it globally.

This could be the greatest productivity unlock in human history. And a profound identity shock.

“What do you do?” is one of our most fundamental questions. It’s how we locate ourselves — in economies, in communities, in our own sense of purpose. The Hyperfuture doesn’t just change the answer. It dissolves the question.

As the intelligence premium evaporates, the disruption is existential. Degrees, mortgages, career trajectories, social standing — all leveraged against the assumption that skilled thinking commands a premium. When that assumption breaks, an entire class loses not just income but the story that organised their identity.

But transition and collapse are different things.

Every major technological revolution has looked destabilising at the start. When the cost of producing something falls dramatically, demand doesn’t stay flat — it expands. Personal computing built entirely new industries. The internet created the platform economy.

History suggests a simple rule: when productivity compounds, abundance follows. It would be unprecedented if the most powerful tool we’ve ever created broke that pattern.

— ◦ —

Para-sapiens: A Crowded Thoughtspace

For the first time in history, we’re bringing millions — soon billions — of non-human reasoning entities online.

They reason, remember, converse, persuade, act. They occupy space in thought itself. Whether they experience any of it is a question for later. The operational fact is what matters now: they’re here, they’re multiplying, and they’re reshaping what it means to share a world.

We lack language for this class. “AI” is too technical. “Agents” feels too small. Call them para-sapiens — minds adjacent to our own. A description of the situation we’re already in.

Humans are no longer alone in thoughtspace. And our traditional anchors — labour, skill, institutional role — are eroding at the same time.

This is the Hyperfuture at its most disorienting. A changed population of minds.

— ◦ —

Acceleration into Abstraction

As labour automates and minds multiply, the conceptual world reconfigures.

The internet began as a distribution channel. Culture moved through it, anchored in physical institutions. That balance has flipped. Culture is now native to the network.

Identity forms in feeds. Community grows in platforms and games. Work unfolds across dashboards. Markets respond to narrative as much as fundamentals.

Music, sport, fashion, gaming, politics, family — these still anchor shared meaning. But increasingly they function as components inside larger participatory systems.

Ask anyone under 25. For many of them, the digital self already carries more income, influence, and connection than the physical one. Baudrillard’s idea that simulation overtakes the real — it isn’t theory anymore. It’s Tuesday.

As visibility becomes constant, the rarest luxury becomes selectivity. The deliberate act of opting out — and opting back in to physicality, slowness, embodied coherence — becomes a form of status. A form of resistance.

At the same time, the horizon widens. Space is no longer only a domain of exploration but increasingly one of infrastructure — communications, observation, manufacturing, and eventually habitation. As systems reorganise on Earth, they begin to extend beyond it, expanding the terrain on which the future will unfold.

— ◦ —

The Interface is the Ideology

The platform era is ending, and the interface is dissolving — text to voice, phones to wearables. Tools stop feeling like software and start behaving like atmosphere. That frictionless morning lightness we opened with — this is what it is. Interface so pervasive it becomes climate.

The feed. The assistant. The operating system. These aren’t neutral tools. They’re lived environments — shaping how attention settles and meaning forms. Our environment always shapes us.

What’s changed isn’t the principle. It’s density — and population.

The interface is no longer just for humans. It’s shared with the para-sapiens — agents and systems that act within it and alongside us.

The interface is the ideology — because it structures perception, defining what appears natural, urgent, possible.

Over time, it stops feeling designed at all. It simply feels like the world.

— ◦ —

What the Hyperfuture Actually Is

The Hyperfuture is a phase shift across the gradient between physics and abstraction.

Destabilisation. Intelligence accelerates past our ability to interpret it. Feeds and simulations collapse time into an Eternal Now. Meaning thins.

Settlement. Infrastructure absorbs the shock first. Firms capture the value. The intelligence premium that structured an entire class begins to evaporate. Physics reasserts boundary conditions. Power consolidates around whoever controls context. The link between labour, productivity and identity fractures.

Recomposition. Non-human intelligences move from tools to participants. Thoughtspace goes multi-species. Culture shifts from transmission to generation inside the network — co-produced with para-sapiens. Coordination expands beyond geography. Persistent digital worlds, orbital infrastructure, planetary systems.

All of it mediated through interface — dissolving from devices into environments, from tools into climate, shaping perception, priority, possibility.

Value follows abstraction. Attention becomes liquidity. Narrative becomes capital.

These forces reinforce each other. Entropy and abstraction rising together. But they don’t rise symmetrically. Compression is automatic — markets drive it, technology accelerates it, every firm acting rationally in its own interest contributes to it. The expansion — new industries, broader access, distributed prosperity — is conditional. It requires competition to be enforced, institutions to adapt, context to remain contested rather than captured.

Left to its own logic, the gradient consolidates. The generous outcome has to be built.

That gradient is the Hyperfuture.

— ◦ —

Coda

The Hyperfuture shows up in ordinary moments.

The first time a machine does in seconds what once took you years to learn. The quiet unease of watching systems coordinate among themselves, without you in the loop.

You feel it when the gap stops being conceptual and becomes lived. When an AI agent one-shots a dashboard you’ve been trying to get right for years.

At a deeper level, the Hyperfuture isn’t a choice we’re making. It’s a pattern we’re part of — the universe expressing its drift toward entropy through ever more sophisticated forms. Intelligence may be the mechanism. Acceleration the result.

We’re inside a process much larger and older than human civilisation itself.

The universe doesn’t require coherence. It doesn’t need meaning.

We do.

The Hyperfuture will test our capacity to let go — of certain roles, certain stories — without dissolving ourselves in the process. And to invent new ones. Stories that embrace leverage as play, abstraction as canvas, the crowded thoughtspace as chorus rather than crowd.

We don’t yet know how to do either well.

But learning how to be human in the Hyperfuture may be the most human work left to do.

— ◦ —

That thin, clear air — beautiful, ungraspable. The view is extraordinary. The ground is very far away.

Gravity is loosening. The air is almost too clear to breathe.

You’re standing at the edge of the world.

Nothing is holding you here.

I’ve always loved the idea that someone who rode west on a wagon train in the 1870s could have lived long enough to watch the moon landing on television in 1969. One lifetime spanning horses, cars, airplanes, the atom bomb, spaceflight — a world once unimaginable, suddenly real, and then suddenly normal.

We are entering a shift like that again. But faster.

What once unfolded over a lifetime will now compress into a single generation — or less.

Not just because technology is accelerating — but because that acceleration is now systemic, pulling reality in opposing directions at once.

On one side: a return to physics. Energy, materials, data centres, robots — hard constraints reasserting themselves after decades when software felt weightless.

On the other: acceleration into abstraction. Identity, culture, value, and meaning increasingly mediated, simulated, recombined at AI-enabled digital speed.

The physical world is hardening. The conceptual world is dissolving.

I call this new era the Hyperfuture.

And this newsletter is an attempt to make sense of the Hyperfuture as it unfolds — to connect the signals across technology, markets, and culture, and how those patterns shape what I build, invest in, and believe comes next.

A Reset

Some of you may remember this newsletter as Growjek, where I wrote about marketing, startups, and technology — mostly through an Asia lens.

Then life got busy with new projects, exits, acquisitions — and the writing paused.

It’s great to now have the space to return properly. The world has changed, and so has what I want to write about.

Growjek has become The Hyperfuture Memo.

The Hyperfuture is Loading

I’ve spent two decades building and investing in technology across data, gaming, and digital economies, primarily in the US, UK and Asia.

I also have a philosophy background, focused on consciousness and mind. AI is where those questions stop being abstract.

This newsletter connects what I’m seeing in markets and culture, to what I’m building and backing.

The Hyperfuture is my name for this new era we’re moving into: a return to physical constraints — energy, infrastructure, robots — alongside rapid abstraction through AI, digital identity, and software-driven culture.

Here, I explore what that produces, and how it reshapes technology, markets, culture, and identity.

If you’ve been here since the early days, welcome back. If you’re new — welcome to the Hyperfuture. And welcome to The Hyperfuture Memo.

What’s next: a longer essay on what the Hyperfuture actually looks like. Coming this week.