Agriculture is thus one of the sectors most in need of development and investment. This research maps the problems AgTech 2050 companies are trying to solve, and asks which of them are actually suited to early-stage capital.

The work is divided into four different problems, each with a different appropriability for early-stage funding: a biological input (nitrogen), project finance infrastructure (phosphorus), a market for volatile carbon credits (one slice of soil) & a recurring data layer or royalty-bearing trait (soil measurement & seed genetics).

Problem 1: Nitrogen

Despite being abundant in the atmosphere, about half the world’s food depends on synthetic nitrogen, fixed through the Haber-Bosch process and applied as an ammonia-based fertilizer. It is the biggest single input cost in row-crop farming, about 17-22% of total corn production cost at roughly $226 per acre in 2022 and $187 in 2023, and the biggest chemical emitter, on the order of 1 to 2 percent of global CO2. Because it is made from natural gas, its price tracks energy markets.

The current research towards solving this borrows from legumes, which have rhizobia (a type of bacteria) in their root nodules. This allows them to fix their own nitrogen unlike crops like corn and wheat.

Some of the bigger players in this space are Pivot Bio, coming in as a space leader with revenue surpassing $100 million in 2023 (on more than 5 million acres). Their proprietary microbes coat the seed and colonize corn roots to fix nitrogen from the air, allowing for a cut in synthetic fertilizer requirements. They’ve raised over $600M at a $2 billion valuation from DCVC, Temasek & Breakthrough Energy Ventures. Second place is Kula Bio, which supercharges its microbes (fixing nitrogen is a very energy-heavy process, otherwise providing no reason for the plant to prefer it) so it promotes in-situ nitrogen fixing. They’ve raised roughly $87 million, with a $50 million Series A led by Lowercarbon Capital. Switch Bioworks uses genetic switches so its microbes establish on the roots before turning on nitrogen production, the step where most get outcompeted. They’ve raised about $23M from Acre Venture Partners, Anthos Capital & Emerson Collective.

At the earliest stage, NetZeroNitrogen, a UK company co-led at pre-seed by Revent & Zero Carbon Capital, is building a seed-applied endophyte (a microbe delivered on the seed that lets the plant fix its own nitrogen) without a green premium. This cost-parity claim is still pre-seed and unproven at commercial scale, which is an honest caveat on an otherwise solid thesis.

The main constraint here is that farmers already set nitrogen usage for profit, not yield. So to sell your product you have to give them better margins and consistent performance, which is why even the leader sits at 5 million of the roughly 90 million acres of US corn. The 50-70% input-level margins and possible exit options at input multiples to Bayer, Corteva, or Novonesis make this more of a specialist bet.

Problem 2: Phosphorus

This is an unusually interesting case for research-driven startups, and in my view a poor one for early-stage capital. Every crop & all life needs it, and unlike nitrogen it’s not all around us - it is mined as phosphate rock, with no substitute. Supply is finite & concentrated: Morocco holds ~70% of world reserves, about 50 billion metric tons. Much of it stays fixed in the soil, but the surplus and erosion losses that do reach waterways are environmentally consequential, driving algal blooms that choke waterways..

What’s being built here mainly revolves around a single theme: recovery. Phosphorus can be captured from wastewater as struvite, a crystalline fertilizer. Ostara, a leader in this space, sells Pearl, a process that pulls struvite from municipal wastewater and sells it as Crystal Green. Their last round was a $70 million Series C, led by Carica Sustainable Investments, Forage Capital, the Cibus Funds, Grosvenor Food and AgTech. The catch is that while this process recovers ~90% of the phosphorus from the side-stream, overall, it only recovers 10-15% of that plant’s total inflow. So it cannot completely replace the mined rock. At the earliest stage, NPHarvest from Finland, does a similar recovery process for nitrogen & phosphorus on a $2.5 million funding from EIT Food & the Finnish Ministry of the Environment.

This effort is better suited to project finance. The demand, at least for now, only exists when needed by regulation as otherwise the recovered product is a commodity that costs more than mined rock & it scales one capital-heavy plant at a time into municipal buyers.

Problem 3: Soil Measurement

There’s a lot of development in this space because the current soil measurement process barely works for the decisions growers are now being asked to make. Soil is measured sparsely & slowly. A common commercial sampling plan today uses one composite sample per ~ 5 - 20 acres, sent to a lab at ~ $7 - $10 per sample, & repeated only every few years. The problem is that a 1,000-acre operation ends up running on 50 to 200 data points, that are also stale.

In reality, soil composition, yield response, fertilizer rate & carbon stock all change at a much finer resolution. So the current picture is a blur & that’s what's opened up so much early-stage scope.

This space resolves into three business models with three different return profiles.

1. The Bio-Product: The economics of this one are very similar to nitrogen bio-products.

2. The Carbon Marketplace: Essentially, this layer pays farmers to store carbon in soil, verifies it, and sells the credits to companies chasing net-zero, acting as a trade platform. While the space has seen a lot of early stage investment, it has proven to be an unreliable bet as of today. The overall voluntary carbon market fell ~56% by volume in 2023 & another 29% by value to $535M in 2024 as confidence in credit quality weakened, especially around soil carbon’s additionality and permanence.

3. The Data Layer: Companies in this space measure the data in 3 ways: satellite models, in-field sensors & DNA sequencing. These are essentially software companies in terms of economics but calibration, sampling, lab work, field service & model refresh can reduce margins.

The unit economics of the last layer improve as sensing cost falls. Pairing satellite imagery with sparse ground calibration has brought per-acre measurement cost down sharply, though the widely cited figures come from vendors & should be read with caution. There is a cost-versus-accuracy trade-off between the three methods. Satellite gives you lowest cost per acre but the coarsest accuracy, in-field spectroscopy gives higher accuracy but hardware COGS & DNA sequencing has highest biological resolution but lab COGS.

In my view, a win here would strike a balance, pairing cheap satellite coverage with real soil samples & a library that gets cheaper to run per acre as it grows. AI could have a real effect on this field, letting a handful of real samples stand in as a good estimate for the whole field BUT the real moat here would be really strong data.

The market here is splitting into two kinds of companies: full-stack MRV platforms that measure, report & verify carbon end to end (Indigo, Agreena, Regrow, Boomitra), & specialist enablers that just sell the measurement (Perennial, EarthOptics, Downforce).

The companies I find most interesting here sit in biological products and the data layer. Stenon, who raised an $18 million Series B round led by Pymwymic, sells a real-time handheld nitrogen sensor, so the farmer pays because it saves them money that season. Seqana reads soil carbon straight from satellite imagery, & sells those readings to food companies like Danone & Bayer for their supply-chain reporting. They’ve raised €3.2M led by Pymwymic as well. Biome Makers sequences the DNA of the microbes living in the soil to tell a farmer how healthy their ground is, sold as a subscription, on a $15M Series B led by Prosus Ventures. There's a bunch of earlier, seed level names as well but I'll come to them in the recommendations.

Indigo is my cautionary case. It bundled biology, a carbon marketplace and data, then repriced sharply - raised around $1.4B, hit a $3.5B valuation, then repriced to $200M, a 94% drop. My conclusion is that the measurement layer is easier to underwrite on its own than the whole bundle.

Investing here will come down to three checks:

1. The buyers. Who pays & does it hold up without carbon? Agronomic ROI does, a farmer pays because the data saves him money. Scope-3 reporting can be a second buyer, but only where a contract or EU rules force it, so treat it as conditional.

2. The margin. Make sure there aren’t hidden hardware or lab costs sitting behind the “software” story.

3. The moat. Satellite access itself is not defensible. What is harder to copy is a proprietary bank of real soil samples, and the model built on top of it. The way you test what’s real is whether the company can beat the cheap lab test in a blinded, multi-season trial. If it can’t, it’s a nice map, not a product.

By margins & metrics, soil measurement is the cleanest software-esque fit in ag. But my conviction sits one problem over, in seed genetics, where the math & the timeline make more sense for early-stage money.

Problem 4: Seed Genetics

As I stated before, I believe in this the most. Mainly because of what it can actually do about the food crisis & because the shifts coming over the next few decades are exactly what will make this category crucial.

The same climate change driving the food crisis is what raises the value of a better seed. Drought already wipes out roughly 20% of global crop yield & drought events are up about 29% since 2000. By 2050, an estimated 62 countries could see peak production losses above 10% & 24 above 20%.

And as the climate shifts, growing zones move with it, so the seed itself has to change: drought tolerance, heat tolerance, varieties that hold up in ground where today’s seed fails. That’s a genetics problem before it’s anything else.

To me, a winner here is simple : transgene-free & high-value editing. Transgene-free in certain jurisdictions would allow for more efficient commercialization & high value because properties like drought & yield are polygenic, controlled by mulitple genes so a single, simple edit will not suffice. The real ones pair multiplex gene editing with AI, to narrow scope of genes to edit.

While researching, I noticed a pattern - a lot of development in this space is happening in the UK & EU. Not implying a direct correlation, but there are 2 things that have happened during the same time. Firstly, the EU enforced Regulation 2026/1388 in July 2026 (applicable from 2028), creating a lighter-touch category for gene-edited plants judged equivalent to conventionally bred ones. Eligibility is on defined criteria, not simply the absence of foreign DNA & other edited plants stay under GMO-style rules. It is not an immediate exemption, but a future pathway that shows real possibility of progress in this market.

Similarly, in 2023 (enforced Nov 2025 in England only) the UK passed the Genetic Technology (Precision Breeding) Act, which created an easier authorization route for precision-bred organisms, even lighter than the EU’s. And the UK already has a massive public plant-science base at Rothamsted & the John Innes Centre, which run the field trials for many of these companies.

On the other hand, the US went the other way: the SECURE-rule vacatur put the path to market back into uncertainty.

Still, I’d look at this whole picture in a forward-looking, long-view way.

Coming to the economics, to me, they are the best in agriculture. Value is captured as a royalty on genetics, so a proven trait licensed across tens of millions of acres becomes recurring revenue at near-pure margin, which is how Monsanto’s traits paid out for two decades, in a $70B-$80B seed market. Inari is the cleanest version, editing elite lines & licensing the genetics instead of selling to farmers, with over $575M raised toward a $720M total at a $1.65B valuation. Avalo uses whole-genome AI to find the gene networks behind complex traits & funds early work through partnership, an $11M Series A with Coca-Cola co-developing low-water sugarcane & Pairwise, Cibus & Benson Hill fill out the field.

The catch is the structure. In row crops, most startups depend on the three majors, Bayer, Corteva, and Syngenta, for germplasm and distribution, so they license to them, partner with them, or get acquired by them. Revenue is often close to a decade out, through breeding, trials, clearance, and adoption. Neither is the ideal situation but I take the trade because the ceiling & the importance are worth it, but only for patient capital & an asset-light model that can survive the wait.

To push back on a part of my own argument, I must discuss Tropic Biosciences, who raised a $105M Series C in March 2026, co-led by Forbion & Corteva Catalyst. They already sell two commercial gene-edited banana varieties, the first in more than 75 years, reaching market through growers, exporters & food-supply partners rather than the row-crop majors, with a TR4-resistant Cavendish due in 2027. Clearly revenue might not sit a decade out & high-value specialty crops can be a route around Bayer, Corteva & Syngenta.

Another exception worth noting is Ohalo, David Friedberg's ~$100M-funded venture, Their boosted breeding passes a parent plant's full genome to its offspring to produce uniform true seed for crops like potatoes and strawberries, so the value comes from a new propagation model rather than a licensed trait.

Two things can be true at once. My personal bet would still stay in the transgene-free & high value multiplex lane BUT the investable category here can be so much wider.

My early stage bets would be:

After researching & reading about all four avenues, my clear winners would be soil, a safer more software-like investment & seed, a bigger gamble with more real-world pay offs.

In seed genetics:

1. Rainbow Crops · Ghent, VIB spinout · €9.7M seed (June 2026)

   • Backers: LIFTT (lead), Corteva Catalyst, VIB, PINC, Maia Ventures

   • What: the Trait Foundry platform pairs AI with multiplex editing of 50-100 genes at once, precision breeding & automated phenotyping, targeting complex traits in corn, sorghum & rice, with a Gates Foundation-backed heat & drought project

   • Why I found them interesting: They’re using AI for multiplex editing & turning complex, polygenic genetics into deployable trait combinations. Also, quite early stage.

2. Biographica · London · £7M / $9.5M seed (Jan 2026, founded 2022)

   • Backers: Faber VC (lead), SuperSeed, Cardumen Capital, The Helm, EQT Foundation, Sie Ventures; earlier backers Chalfen Ventures, Entrepreneurs First & Nucleus Capital

   • What: an AI platform that finds high-value gene-editing targets around 12x faster than conventional methods (same AI-first approach as Rainbow Crops, one layer up, it picks the targets rather than doing the edits)

   • Monetization: This is distinctive. They’re asset-light & earning through milestone contracts with seed majors like BASF’s Nunhems & Cibus

   • Why I found them interesting: They’re a capital efficient platform at the target-discovery layer.

   • Diligence I’d look into: are the royalties contractual on every successful program or only possible; does it keep rights by crop, geography, trait, or improvement; what happens if BASF or Cibus takes a target in-house after the first project; is the revenue a repeatable platform royalty stream or one-off research revenue.

3. Ohmic Biosciences · CU Boulder spinout · $500K seed (2023)

   • Backers: YC, Climate Capital

   • What: founders P.J. Steiner & Matthew Bedewitz engineer disease-resistance genes built to stay ahead of pathogen evolution. Their first product is a soybean resistant to the cyst nematode, a pest that costs US growers over $1.5B a year

   • Why I found them interesting: the earliest of the set, light on the science & pointed at a durable trait. My one caution is that they’ve only raised the $500K 2023 seed & nothing since (at least based on public data), so for a three-year-old YC company I’d want to know why they haven’t raised again

Moving to Soil Measurement:

1. BioSensor Solutions · CU Boulder spinout · grant-funded, raising a ~$500K pre-seed (founded 2023)

   • Backers: NSF, USDA & ARPA-E grants; Colorado-Wyoming Climate Resilience Engine accelerator

   • What: founders Carl Kalin & David Beitz are building a biodegradable, in-field printed sensor that reads soil health in real time, right at the source

   • Why I found them interesting: the earliest & most stealth of the set, still pre-institutional & it goes straight at the sparse-sampling problem.

2. Elaniti · UK · €1.5M total (2025)

   • Backers: an Innovate UK grant, then a seed from biotope by VIB, Planetary Impact Ventures & PINC

   • What: They’re building DNA-level soil microbiome intelligence, reading the soil’s microbes to predict how a field will perform

   • Why I found them interesting: an early, capital-light way into the microbiome data layer

3. TerraBlaster · $4M pre-seed (Aug 2025), now raising a $10-15M seed

   • Backers: Khosla Ventures (lead), Trailhead Capital, OCP Group & The Reservoir

   • What: real-time in-field NPK by laser spectroscopy, reading nutrients live as the tractor moves, founded by Jorge Heraud (who built & sold Blue River Technology to John Deere)

   • Why I found them interesting: strong founder signal & a real product, though the pre-seed window is already closing as they move to a scale round.

I also found some earlier-stage leads, however they would need more diligence ( I sourced data from grants, trials etc).

1. Neocrop Technologies (Chile, ~$350K plus state grants): its Neotrait Engine combines editing, gene discovery & speed breeding to cut breeding time by ~60%, & it already has non-GMO clearance in Chile & Argentina with seed-company partners. Already a deployable-crop-system idea, through a compressed breeding cycle & a Latin American route to market.

2. InEdita Bio (Brazil, ~$1-1.5M seed from Vesper Bio & Ecoa Capital): non-transgenic editing for disease resistance & water/nutrient efficiency, founded by Paulo Arruda, who earlier co-founded Alellyx (later bought by Monsanto). Credible team on the non-transgenic thesis, though rights & financing still need work.

3. Biofonic (London, ~$500K from General Inception plus Innovate UK grants): below-ground acoustic sensors & ML for soil health, in a commercial trial with Riverford. The earliest of these leads, still with no provable results.

Thing to note: Corteva's venture arm, Corteva Catalyst, already backs Rainbow, Tropic & Resurrect Bio (a $10.3M Series A for reactivating plants' own disease-resistance genes). So the incumbents are already hedging.

This is a working thesis, built on public data. I'd welcome input from founders & investors active in this space, any new information, feedback, or notes I could use, especially where I've got something wrong, or an early-stage company I've missed. So please reach out.

Sources: Listed in the comments

I write about things I’m curious about and can’t find good answers to, usually somewhere across tech, finance, and culture. Subscribe if that sounds like your kind of read :)

Looking backward, the path to AI feels almost continuous.

People spent years, sometimes decades, building clockwork tigers and mechanical boys that could dip a pen and write out a line, machines that did nothing useful and that nobody needed. And yet they all seem to stem from the same human curiosity, the pull to understand a living thing so deeply you could reproduce it mechanically.

So instead of looking at AI from where we stand now, I went further back.

The first room is Neoclassicism, in the late 1700s, at the height of the Enlightenment, when reason felt like the highest thing a mind could do. David paints the Oath of the Horatii, three brothers swearing to die for Rome, every figure and shadow placed to carry you toward a single conclusion, like a proof. The painting is built like an argument. And the same Enlightenment was already treating the mind itself as a machine: Hobbes called reasoning a kind of arithmetic, Leibniz built a real calculating machine and dreamed of settling disputes by computation, and in 1747 La Mettrie published a book called Machine Man. You do not try to build a mind until you believe a mind runs on rules, and this is where people started to believe it.

A few decades on, the whole mood flips. If Neoclassicism trusted reason, the Romantics distrusted it on purpose. From the 1790s into the mid-1800s they went after everything the rational mind leaves out, the huge feelings, the storms, the sublime, the single unrepeatable person. You would expect them, of all people, to recoil from the idea of a manufactured mind, and yet the first real story about building one is theirs. Mary Shelley wrote Frankenstein at eighteen, and for all its dread the book runs on a want it never fully disowns, the pull to assemble a living, thinking thing out of parts and watch it open its eyes. What has kept it alive for two centuries is that Shelley refused to make that want simple. Victor Frankenstein is not undone by building the creature. He is undone by what he does next, walking away from it, refusing to raise or answer for the thing he made. The horror lives in the abandonment. She had worked out, impossibly early, that the real risk in making a mind sits with its maker, in whether they stay to take responsibility for it, and she held the awe and the responsibility in the same hand. It is close to the exact feeling of building these systems now, the wonder and the weight arriving together.

By the 1920s the Surrealists were chasing the same built mind from the other direction. Where the Neoclassicists trusted the reasoning surface of thought, the Surrealists were after what sat beneath it. In 1924 André Breton published a manifesto arguing that the truest art came from below conscious control, from the part of the mind that keeps working while you are asleep or not paying attention, and Freud had just made that underground respectable. Their method for reaching it was automatism: you would draw or write without stopping to think, letting the hand move faster than judgment, and see what surfaced. André Masson covered pages in loose, unplanned lines and only afterward found the image hiding inside them. What they were reaching for was a maker that did not need a conscious maker behind it, a hand that produced on its own. I do not want to overstate the resemblance, but it is hard not to hear an echo of that in a generative model, which also makes its images from a place no person is directly steering.

Where the Surrealists tried to automate the hand, the conceptual artists of the 1960s went after the idea itself. Their claim was that the art is the idea and the physical object barely matters. Sol LeWitt would write a few plain sentences describing a wall drawing, hand them to other people, and let them execute it, and different hands produced slightly different walls, which was fine, because to him the thinking was the real work and the making could go to anyone. In 1967 he put it in a sentence that reads very differently now than it did then: the idea becomes a machine that makes the art. He meant it as a metaphor, a way of saying that a good enough plan almost runs itself. Building AI is that metaphor made literal. You supply the idea, in words, and hand it to an actual machine that does the making. Which leaves one question hanging. If the idea is the art, and the machine can now make it, what happens when the machine starts to have the ideas?

That question is where the newest room begins, the one we’re in. It has no paintings in it, and yet the work itself is the art. Watch how a model actually gets made. Someone describes what they want it to become, chooses what it learns from, sets the conditions, and lets it train, then reads the results and starts again, and again, a long grind of trial and error to coax something closer to a working mind. There is real beauty in that, in the math underneath it, in a training run that finally holds, in each paper that pushes the next generation a little smarter, in the people who give years of their lives to it. This is LeWitt’s idea taken to its limit: the conception is the work, and the thing being drawn is a mind. Every new model is an attempt to make the human mind again out of numbers, and I find that as beautiful as anything hanging in the older rooms. Think back to the writing automaton, the mechanical boy built to dip a pen and copy out a line that someone else had written for him. For three hundred years that was the ceiling, a machine that could perform a thought but never have one. Then, in 2025, a research system called ASI-ARCH was handed a job that had always belonged to human scientists, designing new AI architectures, and left to do it on its own. It ran more than seventeen hundred experiments, proposed its own ideas, wrote them into code, tested them, and kept the ones that worked, finishing with a hundred and six new designs, the best of them beating anything humans had built and working in ways no one had thought to try. The automaton has started composing the sentence. The oldest want on the whole walk, to build something as alive as we are, is nearer now than it has ever been.

I think that is why AI has never frightened me much. Walking back through all of it felt like watching a ballet, each movement taking the stage in turn and giving way to the next, one after another after another, until the newest one steps out, and it is this. It carries the same infinite wonder that has moved people for as long as there have been people, the search that never really stops. Set beside the art and the culture that came before us, it feels a good deal less daunting. The want underneath it is old and deeply human. What we choose to build with it from here is entirely ours.

I write about things I’m curious about and can’t find good answers to, usually somewhere across tech, finance, and culture. Subscribe if that sounds like your kind of read :)

I’ve written before about the pull toward offline. This time I want to break down the process of un-frying your brain. The science behind the slow, uncomfortable rebuild of your attention and cognitive function.

Because “just put your phone down” doesn’t address the problem. The problem isn’t the phone. It’s what your neural circuitry learned to expect while you were holding it.

Before we get to the rebuild, you need to understand what you’re recovering from. Not screen time. Not “too much phone.” Something more specific.

Social media is built on variable ratio reinforcement. Same reward mechanism as a slot machine. You pull the lever (open the app), and sometimes you get something great (a funny post, a like on your photo, a DM from someone you care about), and sometimes you get nothing. The inconsistency is the point. If you got a reward every time, you’d get bored. If you never got one, you’d stop. But the randomness? That keeps you pulling.

B.F. Skinner figured this out in the 1950s with pigeons. Give a pigeon a pellet every time it pecks a lever and it pecks at a steady, moderate rate. Give it a pellet at random intervals and it pecks compulsively, endlessly, even when the pellets stop coming.

This isn’t a hypothesis. The engineers who designed these feeds studied behavioral psychology. The infinite scroll, the pull-to-refresh gesture, the unpredictable mix of content. All of it is tuned to exploit the same dopamine pathway that makes gambling addictive. The difference is that a casino closes. Your phone doesn’t.

When you scroll for hours a day, your brain adjusts. It has to. The constant stream of micro-rewards (a like here, a funny clip there, a notification every few minutes) floods your dopamine system with more stimulation than it was designed to handle. So your brain does what any smart system would do when it’s getting too much signal: it turns down the volume.

This is called receptor downregulation. Your neurons literally pull dopamine receptors off their surfaces. Fewer receptors means each hit of dopamine registers less. So you need more stimulation to feel the same thing. Longer scroll sessions. More tabs open. Faster content. The 30-second video that used to be satisfying now feels slow, so you skip to the next one, and the next one, and the next one.

It’s not just your reward system that gets hit. It’s your prefrontal cortex, the region responsible for sustained attention, impulse control, planning, and deciding what’s worth your focus.

The scrolling pattern itself, the rapid context-switching from post to post, trains your prefrontal cortex to expect constant novelty. So when you sit down to read something long or work on something that requires sustained focus, your brain is fighting against its own recent conditioning.

And here’s the thing: you don’t even retain any of it. Try to recall the last 30 videos you scrolled through. You can’t. Your brain processed them just enough to decide whether to keep scrolling, but never deeply enough to form a real memory. You traded your ability to focus for content you won’t remember consuming.

This is why people say they “can’t read books anymore.” It has nothing to do with them and everything to do with how scrolling changed the way their brain allocates attention.

Rebuilding your cognitive function isn’t as simple as deleting the apps or setting a screen time limit. The process is nonlinear, and it’s significantly closer to actual substance withdrawal than most people realize. Your receptor density needs weeks to rebuild, and the new receptors need even longer to calibrate.

What matters more than any list of recommended activities is understanding the science behind what needs to be rebuilt. Once you know what’s actually broken, you can pick the right activities yourself instead of blindly following advice. That process of choosing deliberately is itself a form of retraining your cognizance.

Sustained attention is the exercise. Your prefrontal cortex rebuilds through use, the same way a muscle does. But the key word is sustained. It’s not about doing something hard. It’s about doing something continuously, without switching, for longer than your brain wants to.

Reading works because it demands single-threaded cognition. One narrative, one argument, no branching paths, no algorithmic interruptions. Your prefrontal cortex has to do the work of maintaining focus, and that work is what rebuilds the neural pathways that scrolling eroded.

Playing an instrument layers in delayed gratification. The reward structure is inverted: frustration first, competence later. Your dopamine system has to learn to stay engaged through the difficulty instead of seeking an immediate exit. I started watching silent films for a similar reason. No soundtrack engineering your emotions, no rapid cuts. Your brain has to do the work of reading the frame, catching the subtlety, sitting in the quiet. It’s a form of perceptual retraining, teaching your attention to find detail in low-stimulation environments instead of needing everything loud and fast.

One of the more important things you can do to retrain your brain is to sit with boredom. When you’re bored and you don’t reach for stimulation, your brain activates what neuroscientists call the default mode network (DMN). This is not your brain shutting off. It’s your brain turning inward.

The DMN is responsible for self-referential processing. That means it’s the network that figures out what you actually think, what you value, what you like, who you are when nobody’s performing for you. It’s where your brain connects unrelated ideas, makes sense of past experiences, and generates the kind of loose, wandering thought that leads to creative breakthroughs. Newton under the apple tree, Archimedes in the bath, Darwin on long aimless walks.

It’s also the network that builds your personality. Your sense of self isn’t something you’re born with fully formed. It’s constructed through reflection, through your brain quietly processing experiences when you’re not focused on anything external.

When you never give your brain that idle time, because every empty moment gets filled with someone else’s content, you lose access to the process that tells you what you actually care about. This is why a lot of us “terminally online” folks feel like we don’t know what we enjoy anymore. It’s not that we lost our interests. It’s that we stopped giving our brain the downtime it needs to identify them.

There’s a distinction in neuroscience between two types of dopamine release that most people never hear about. Scrolling produces what’s called “phasic” dopamine: short, sharp spikes followed by rapid drops. This is the pattern that drives compulsive behavior because your brain keeps chasing the next spike.

What you want instead is “tonic” dopamine, a steady, moderate baseline. Activities that produce tonic dopamine tend to be things that are mildly rewarding over an extended period: cooking a meal, having a conversation, building something with your hands, going for a long walk with no headphones. The reward is distributed across the whole experience instead of concentrated in a moment.

The difference sounds subtle but it’s everything. Phasic dopamine trains your brain to seek intensity. Tonic dopamine trains it to sustain engagement. You’re not just choosing different activities. You’re choosing a different reward architecture for your neurons.

I don’t believe that the point of retraining your brain is to become a monk with no access to social culture. It’s more about understanding exactly how the urge to scroll works at a neurological level, so that the control over the narrative shifts back to you. When you know it’s your receptor system chasing a hit it was trained to expect, not actual boredom, not actual need, you have the freedom & awareness to make different choices.

Your brain built itself around whatever you repeatedly gave it. It can rebuild around whatever you give it next. The process is slow, it’s uncomfortable, and it won’t look like progress for a while. But the wiring is not permanent. That’s the whole point of neuroplasticity.

I write about things I’m curious about and can’t find good answers to, usually somewhere across tech, finance, and culture. Subscribe if that sounds like your kind of read :)

I work at a large company where our AI token usage blows past the internal limit every quarter. Leadership encourages it most of the time. These tools are becoming a part of how people actually work, and high adoption tells investors and clients the firm is positioned well. It’s good for the business and good for the narrative. The cost is an afterthought, for the most part.

But that can’t last forever. I wanted to understand the pricing economics behind AI inference, and the question I kept coming back to was simple: who actually pays when AI gets cheap?

The answer I landed on is that AI economics is heading toward a split regime. Intelligence becomes cheaper, but intelligence infrastructure becomes larger. The cost of a single query collapses while total system spending rises. One layer of AI turns into cheap, invisible infrastructure. Another layer stays expensive, scarce, and worth paying for. Those two layers don’t converge. They split further apart over time.

A few terms I’ll be precise about throughout: cost is what it takes to serve a query (hardware, energy, operations). Price is what providers charge for it. Spend is what customers and the ecosystem pay in aggregate. These move in related but different directions, and that distinction ends up mattering a lot.

The economics of tokens

Before I dig in, here are some basics. Every AI interaction is measured in tokens (roughly three-quarters of a word). You consume input tokens (your prompt) and output tokens (the model’s response). Output tokens cost more because they require more computation.

Here’s what the major labs charge developers today:

_{Sources: OpenAI pricing, Anthropic pricing, Google AI pricing, DeepSeek pricing. DeepSeek’s V4 Pro, released in May 2026, is a 1.6 trillion parameter open-weight model that undercuts every closed-source frontier lab on price while delivering competitive performance.}

The range within a single provider is worth noting. OpenAI’s Nano tier is 50x cheaper than its own flagship model. According to a16z’s LLMflation analysis, the average cost per million tokens has been falling roughly 10x per year. Both OpenAI and Anthropic now offer prompt caching (up to 90% off) and batch processing (50% off) that can stack, bringing effective per-call costs to roughly 25% of list price for optimized workloads.

Consumer subscriptions like ChatGPT Plus at $20/month are flat-fee, meaning light users subsidize heavy ones. Sam Altman confirmed in early 2025 that OpenAI is losing money on every ChatGPT Pro subscriber ($200/mo) because of how heavily those users push the models.

The cost deflation case

The optimistic answer.

According to a16z’s “LLMflation” analysis, GPT-3 level performance cost $60 per million tokens in November 2021. Today, equivalent performance from an open-source model costs about $0.06 per million tokens. That’s a 1,000x reduction in under four years. Three things drive this: better hardware, smarter model architectures, and competitive pressure from open-source models that force the closed labs to keep cutting prices.

If this pace continues, the subsidy problem resolves on its own. Costs fall fast enough that the labs reach profitability through efficiency. That’s already happening at the bottom of the model stack: GPT-5 Nano costs nearly nothing to serve.

But the deflation is uneven. It’s fastest at the low end and slowest at the frontier. The thousand-fold cost reduction mostly applies to last generation’s capabilities. Today’s best reasoning models still cost 80x more than their open-source equivalents. And that gap is widening, not closing.

This is the split regime. AI isn’t one market moving toward one equilibrium price. It’s two markets diverging: a commodity layer where costs collapse and a frontier layer where performance remains scarce and expensive.

And here's the part that trips people up: falling unit costs don't mean falling total spending. This is a version of Jevons paradox : when a resource becomes cheaper, demand often expands faster than efficiency gains reduce total consumption. Each new capability tier creates entirely new categories of work rather than simply replacing old ones. Unit costs fall. Total system spending rises.

Where the money comes from (and where it goes)

The major AI labs are not profitable. But it’s worth being precise about what that means.

OpenAI lost $5 billion on $3.7 billion in revenue in 2024, spending roughly $1.35 for every $1 earned. The trajectory hasn’t improved. Q1 2026 actuals showed a $6.95 billion operating loss on $5.7 billion in revenue, and internal projections peg the full-year GAAP net loss around $25-26 billion. OpenAI doesn’t expect to turn cash-flow positive until 2030. Anthropic lost $5.3 billion in 2024, though revenue scaled rapidly to a $30 billion+ run rate by mid-2026. Google has 750 million Gemini users but generates only about 13 cents per user per month in subscription revenue.

An important distinction: company-level losses don’t automatically mean inference is being sold below marginal cost. A large portion of these losses comes from frontier model training, R&D, and infrastructure expansion.

The funding structure keeps this going: VC and PE capital flows in, the labs use it to price aggressively, low prices drive adoption and developer lock-in, adoption growth justifies higher valuations, and higher valuations attract more capital. OpenAI has raised over $40 billion. Anthropic has raised over $15 billion. Meanwhile, a significant chunk of the revenue the labs earn comes from the same companies investing in them. Microsoft is both OpenAI’s largest investor and its largest customer through Azure OpenAI. Amazon invested $4 billion in Anthropic and resells their models through AWS Bedrock. Capital is effectively funding its own demand.

AWS ran a similar playbook from 2006-2025, growing from $3 billion to $115 billion in annual revenue on the back of aggressive early pricing before raising prices for the first time in 20 years in January 2026. The key difference: cloud lock-in was infrastructure-deep, and cloud costs only fell 10-15% per year. AI lock-in is lighter, open source provides an alternative cloud never had, and costs are falling closer to 10x per year. That’s the biggest reason the cloud arc might not repeat.

The infrastructure squeeze

Microsoft, Google, Amazon, and Meta have collectively announced roughly $725 billion in AI-related capital expenditure for 2026, up from $200 billion combined in 2024. Goldman Sachs projects $920 billion to $1.4 trillion in 2027 alone. These companies are funding it from revenue, not speculative debt, so the risk profile is different from the 1990s telecom buildout. But the capital still needs to generate returns.

The mechanism is straightforward: compute cost per query is falling while energy cost per query is rising. 63% of frontier-model lifecycle energy goes to inference, not training, which means energy costs scale directly with usage. AWS’s first-ever price increase in January 2026 was triggered by exactly this pressure.

Meanwhile, open models have narrowed the performance gap enough to create a credible pricing ceiling. If the labs raise prices significantly, developers have alternatives. But open source doesn’t eliminate pricing power. Users don’t buy tokens. They buy useful work. There’s also security. A senior executive I spoke to put it simply: open source is great, but for large institutions the question isn’t just “can it do the task?” It’s “can we trust it with our data?” That’s pricing power that has nothing to do with model performance.

The historical pattern in technology is that falling input costs don’t automatically flow to customers. The surplus gets captured somewhere in the stack by whoever holds the strongest bottleneck. In AI, that bottleneck may shift between chips, cloud providers, frontier models, and application layers. The question for investors isn’t just whether inference gets cheaper. It’s who captures the efficiency gains when it does.

What this means for startups

The conversation around AI startup risk often focuses on wrappers. That risk is well understood and most investors have already priced it in.

The harder question is about companies doing real work on top of inference. Vertical AI products with proprietary workflows, fine-tuned models, domain-specific data pipelines. These are legitimate businesses. But they share a structural trait: inference is a significant, variable cost that they don’t control.

AI startups are not just exposed to customer demand risk or competitive risk. They’re exposed to upstream infrastructure pricing risk. Their margins are partially determined by decisions made outside their own business.

This isn’t theoretical. In the first week of June 2026, Cursor, GitHub Copilot, and Devin Desktop all moved from flat unlimited pricing to metered usage within 72 hours of each other. Agent-mode sessions burn far more compute than a simple chat completion, and at scale, the vendors couldn’t absorb the cost anymore. That’s the upstream pricing risk in real time.

Traditional SaaS companies spend less than 5% of revenue on cloud infrastructure. AI startups spend significantly more. Raw API spend might be 4-9% of revenue, but the full inference stack (retrieval, self-critique loops, classification calls, re-ranking) pushes that to 20-23% at scaling-stage AI B2B companies. The new operating range for AI-heavy SaaS appears to be settling around 60-70% gross margins, compared to the 80-90% that investors spent the last decade treating as baseline.

Companies on usage-based or outcome-based pricing are better positioned to pass cost changes through than those on flat-fee plans, which report 40% lower margins and 2.3x higher churn.

The enterprise shift

I mentioned at the top that my firm blows past its internal AI usage limits every quarter, and that leadership encourages it. That’s not unusual. High adoption signals to investors and clients that the firm is forward-looking.

But AI now accounts for 41% of all enterprise IT spending, up from 32% last year. At that share, the bill stops being a rounding error. By Q2 2026, finance departments had largely taken over AI procurement from technology leadership.

The core problem is that AI breaks the link between software adoption and predictable cost. With traditional SaaS, you add users and pay per seat. Spend is forecastable. With AI, usage varies wildly. One power user can consume 100x what another does. You can’t tell every employee to use AI more and then ask the CFO to keep the line item flat. But you also can’t restrict usage without measuring which use cases are actually worth the spend, and right now, most companies don’t have a framework for that.

Maybe the answer looks like tiered internal access, where teams that demonstrate ROI get frontier models and everyone else gets the commodity layer. Maybe it looks like usage-based chargebacks to business units. This is where the “who pays” question gets personal. It’s every company trying to figure out how to budget for something they’re simultaneously telling everyone to use more of.

What I ended up with

The mistake is asking whether AI becomes cheap or expensive. It becomes both.

Cheap intelligence becomes infrastructure. Simple classification, summarization, routing, anything a small model handles well. This layer races toward zero. Expensive intelligence becomes a premium tier. Frontier reasoning, long-horizon agents, high-stakes analysis where reliability matters more than cost. This layer stays scarce.

AI is following utility economics rather than software economics. Unit costs improve while total system investment expands. Pricing power concentrates at the frontier while the commodity layer commoditizes.

The implications depend on where you sit. If you’re building on the commodity layer, margins improve and the subsidy question fades. If you’re building on the frontier layer, you’re still exposed to upstream pricing decisions, energy costs, and whether the current rate of efficiency improvement holds. If you’re an enterprise buyer, the question isn’t what AI costs per token. It’s which use cases clear the value bar and which ones get cut when the CFO starts asking harder questions. And if you’re investing in this space, the companies that win will be the ones that know which side of the split they’re building on and price accordingly.

If you’ve done deeper work on any of this, or you’re closer to these questions than I am, I’d like to hear about it. Drop a comment or find me on X and LinkedIn.

I write about things I'm curious about and can't find good answers to, usually somewhere across tech, finance, and culture. Subscribe if that sounds like your kind of read :)

The brand discovery funnel is changing

Consumer discovery is shifting from user-driven navigation to model-mediated recommendation. This doesn’t eliminate search but it changes who controls the decision.

In the old model, Google ranked options, users evaluated them, and brands converted on their own pages. In the new model, AI recommends directly, and the browse step compresses or disappears entirely in high-intent categories. The model picks, the user validates briefly, and the sale either happens immediately or not at all. In some cases, the user never leaves the AI environment at all. Perplexity already has built-in purchasing. ChatGPT surfaces product recommendations with buy links. Claude has a Chrome extension that can complete purchases on the user’s behalf. The funnel isn’t just compressing. Parts of it are moving inside the AI platform entirely.

This breaks two systems that have powered consumer commerce for twenty years.

First, the distribution layer. When I initially started researching for this piece, I assumed AI would completely overwrite the current SEO system. Here’s where I was wrong. The underlying signals that made Google rank you (backlinks, structured data, content depth, domain authority) still shape whether a model pulls you into an answer. If you already have strong SEO, you’re probably already showing up in AI outputs more than a brand that doesn’t. In that sense, AEO (Answer Engine Optimization, the emerging practice of optimizing for AI-generated answers) and SEO are not different sports. They share most of the same raw inputs.

What changes is the mechanic. In the old world, you optimized to appear in the top ten blue links and the user clicked through. In the new world, you’re trying to be the sentence the model quotes, the brand it names, the source it cites. Keyword density matters less. Conversational phrasing, comparative language (”X is better than Y for Z”), and explicit claims matter more. A Reddit thread can outrank your own domain in a ChatGPT response because large language models over-index on forum and Wikipedia-style content. And the unit of optimization shifts from “the page” to “the claim inside the page that a model can extract cleanly.” The infrastructure you built for ten blue links doesn’t map cleanly to that.

The open question is whether that gap between old playbook and new mechanic is wide enough to support a new category of tools, or whether existing SEO incumbents just absorb it as a feature. I’ll come back to this.

Second, the conversion layer. Static landing pages were designed for browse-mode visitors clicking from a search engine results page (SERP). AI-routed traffic arrives pre-filtered, high-intent, and expectation-heavy. A user who just got told “brand X is the best for flat feet” expects confirmation, not a homepage tour. A generic Shopify page can’t serve that visitor. The conversion surface needs to match the context the AI sent.

The dot-com parallel

The last time discovery shifted this hard was in the early 2000s, when Google Search became the dominant layer. An ecosystem of companies rose up to help brands get found: Moz in 2004, SEMrush in 2008, Ahrefs in 2011. SEO became a $100 billion category.

The structural reason this happened is that Google never built those tools itself. They built the search engine and gave brands basic analytics, but the actual work of optimization was left to third parties. Google had no incentive to make gaming the system easy. The result was an entire industry filling the gap between “the platform exists” and “brands know how to win on it.”

That industry produced real outcomes. SEMrush was acquired by Adobe in 2026 for $1.9 billion. Ahrefs generates over $100M in annual recurring revenue, still bootstrapped. Moz still operates today.

The same dynamic is playing out now. OpenAI, Anthropic, Google (Gemini, AI Overviews), and Perplexity are building the new discovery layer. None of them are building tools to help brands optimize for it. The gap is wide open again.

The opportunity in numbers

Sources: Adobe Analytics · Search Engine Land · eMarketer · Search Engine Land · Gartner · Conductor

The three layers, and who’s building in them

Layer 1: AI Visibility (the monitoring tier)

Think of this as Google Analytics for AI outputs. Profound ($1B valuation, Sequoia-backed) leads the category. Peec AI just raised a $21M Series A. Otterly, LLMrefs, Bluefish ($24M, NEA and Salesforce Ventures), and Scrunch AI ($19M raised) are all here. The question they answer is simple: when someone asks ChatGPT about your category, does your brand come up?

Of the roughly $287M in total venture funding I tracked across all three layers, about $227M (nearly 80%) has gone to monitoring alone. Profound accounts for $155M across four rounds. The monitoring tier is pretty mature in both funding and market positioning. The question now is whether Profound’s lead is catchable or whether this layer is already consolidating around a winner.

Layer 2: AEO Execution (the intervention tier)

Monitoring tools tell you what’s happening. They don’t fix it. The execution layer is about actually changing what the model says through content generation, structured data optimization, and re-tracking the output.

AthenaHQ is the closest thing to a pure play. Founded in 2024 by ex-Google Search and ex-ServiceNow engineers, YC-backed, $2.2M seed. They’ve built a catalog of 3 million AI responses mapped to the sites they cite, and their Action Center recommends the exact content changes needed to show up more often. Frase is coming from the SEO side with full programmatic access for AI agents to execute content changes directly. Cognizo is an unfunded two-person team in NYC doing the full stack at $89 a month.

The tension in this layer is whether AEO is a real category or just a feature of existing SEO tools. Right now there’s still 60-70% overlap between top Google results and ChatGPT citations. In two years, that could be 30%. Or it could be 80%. Nobody knows yet. This is the layer I’d say I’m watching closely.

Layer 3: AI-Native Conversion (the arrival tier)

When AI sends a buyer to your site, they arrive with more intent, more specificity, and less patience than a Google search visitor. A static Shopify page can’t serve them.

Spangle AI ($21M raised, $100M valuation) has clear first-mover advantage, building landing pages that adapt in real time per visitor. REVOLVE and Steve Madden report 50% conversion lifts. Remark ($27M, Stripe invested) takes a different angle with 60,000 human experts training AI shopping personas and hits 28% conversion rates against an industry average of 1%.

But this layer faces a similar “category vs feature” question. Established tools like Unbounce, Instapage, and Leadpages have been doing conversion optimization for years and are now adding AI features. Emergent launched in 2026 as a full-stack AI-native conversion system. About $58M in funding has gone into this layer, roughly 20% of total capital in the space. Spangle has a real lead, but conversion is inherently fragmented across brands, categories, and price points, which leaves room for vertical-specific wedges.

What I concluded from my search

I came into this expecting a clean thesis. SEO gets replaced by AEO, brands need new optimization tools, back the picks and shovels. Straightforward. What I found was that I’d understated how deeply AI is reshaping commerce. It’s not just one part of the funnel that’s moving. Discovery, distribution, conversion, even the purchase itself are all being restructured at the same time, and in ways that interact with each other. Some layers are already consolidating. Others might not exist as standalone categories at all.

I don’t have a clean answer on who or what wins. But I know more about the shape of the problem than I did a week ago, and I think the map is useful even without a final verdict. This is the first in what I expect will be a series as I explore the effects of AI on industries.

If you’re a founder building in any of these layers, I’d genuinely love to hear what you’re seeing on the ground. And if you’re a VC already looking at this space, I’d love to exchange notes on where you think the category lines are forming. You can reach me at rgavaskar21@gmail.com or connect with me on LinkedIn.

I write about the things that make me curious, the shifts happening around us, and what it all means for how we live and spend. If that sounds like your kind of read, subscribe and I’ll see you next week.

So I went decided to write about it. Not to add to the noise, but to actually explain what’s happening. Because the jargon makes it almost impossible for anyone outside of finance to follow. Here’s my attempt at cutting through it.

What private credit actually is

Private credit is when a non-bank lends money directly to a business. No JPMorgan, no Goldman, no public markets. Just a private fund and a company agreeing on terms.

Who lends it: Firms like Apollo, Blackstone, Ares, and Blue Owl. They raise money from pension funds, endowments, and wealthy investors, then deploy it as loans.

Who borrows it: Mostly “middle market” companies doing $5 to $100 million in annual profit. Too big for a small business loan, too risky for public markets. Regional manufacturers, software companies, healthcare chains. Real businesses, just not ones you’d recognise by name.

Why not just go to a bank? Because after 2008 (when super risky lending nearly collapsed the entire financial system), banks got heavily regulated and quietly stopped wanting these borrowers. Private credit stepped into that gap and has been growing ever since.

The Basics

Private credit talks a lot about these things. Here are some definitions.

EBITDA is a company’s real cash before any adjustments. Lenders use it to measure how much debt a borrower can handle. If a company has $10M in EBITDA and borrows $50M, that’s 5x leverage. Higher leverage means higher risk.

Leveraged loans are loans to companies already carrying a lot of debt, paying higher interest rates to compensate lenders for the extra risk.

Covenant-lite loans have fewer protections for the lender. Normally a loan includes financial tests the borrower has to pass regularly. Covenant-lite strips most of that out. Great for borrowers, riskier for lenders.

NAV loans are loans made to private equity funds using the fund's investment portfolio as collateral. So PE funds are also borrowers here, not just companies. They borrow against their own assets to return cash to investors or make new deals without having to sell anything. Regulators are not thrilled about these.

LTV (Loan-to-Value) is the ratio of the loan amount to the value of the collateral backing it. A 20% LTV means the lender is only lending 20 cents for every dollar of collateral value, which is conservative. A 90% LTV means if the collateral drops in value even slightly, the lender is in trouble.

PIK (Payment in Kind) is when interest isn’t paid in cash but added to the loan balance instead. A $100 million loan at 10% PIK becomes $110 million after year one. The debt quietly balloons.

Carried interest is the ~20% share of profits fund managers take after returning capital to investors. It’s how they get paid when things go well.

To sum it up: more risk, less protection, quieter consequences. All trends that accelerated in the last three years.

How it grew

Banks retreated after 2008, forced to hold more cash as a buffer. The Fed kept rates at zero to stimulate the economy, which pushed investors to look for better returns elsewhere. Pension funds (who need their asset pool to grow over long time horizons to fund people’s retirements and can stomach more risk because of it) piled in chasing 8-12% yields. Insurance companies followed, since private credit’s predictable cash flows (unlike PE) fit their model of covering long-term claims. Asset managers built massive platforms to meet the demand.

The market went from $600 billion in 2015 to $3.5 trillion today. Five times the size in under a decade. Either it was genuinely filling a need, or it got way ahead of itself. Probably both.

Major Players

Apollo Global Management has a secret weapon: Athene, an insurance company they own. Because insurance money isn’t subject to the same redemption pressure as a regular fund, Apollo has more stable funding than competitors. Their CEO Marc Rowan said in March 2026: “This will be a shakeout. It was foreseeable. It was predictable.” Translation: we’re fine, others are not.

Blackstone is the largest private credit manager. In early 2026, investors tried to pull $3.8 billion from their $82 billion fund. Blackstone met every redemption request but had to raise $400 million from their own balance sheet to do it. This is tiny detail the market picked up.

Blue Owl Capital is having the roughest time, stock down two-thirds from 2025 highs. In February 2026, they permanently closed redemptions on a $1.6 billion fund. Investors who wanted their money back were told, politely, no. This is the news that thrust private credit into mainstream media.

Ares Management is a large direct lender, also down significantly.

Where did things go wrong?

Three things broke at once.

Too many players, too little discipline. When a market grows 5x in a decade, everyone wants in. Experienced or not. Standards slip when nobody wants to lose a deal. Apollo’s own CEO called it a “race to the bottom.”

AI ate the collateral. About 25% of private credit portfolios are software and tech companies, which looked bulletproof in 2022: predictable revenue, fat margins, sticky customers. Then AI disrupted their business models faster than anyone expected. When your collateral loses value, your loan does too.

The bill arrived. The U.S. private credit default rate hit 5.8% in January 2026, the highest on record. The headline number looks closer to 2% because it excludes “selective defaults,” where a company restructures debt without technically defaulting. It is still very much a default. Many borrowers also switched to PIK loans because they couldn’t afford cash interest. When investors started asking for their money back, funds that had lent it out long-term couldn’t sell loans fast enough to cover withdrawals. Liquidity crunch. Here we are.

Is this really 2008-esque?

I can see why people are panicking. In 2008, careless lending got packaged into securities nobody knew how to value, and we didn’t realise how deep the rot went until it collapsed. Today, the Federal Reserve has found that private credit growth was largely funded by bank loans, meaning stress flows back into traditional banking. The IMF has warned about “valuation uncertainty” (private company valuations are stagnant and don’t always reflect reality). The DOJ warned about “creative marks,” which is government-speak for funds lying about what their loans are worth.

But there are two key differences: scale and detection.

In 2008, mortgage debt was sold globally and went bad everywhere at once. Private credit funds are mostly holding the loans they made. The damage is more contained. And private credit at $3.5 trillion, while enormous, is not the $10 trillion of mortgage debt wired into every corner of the global financial system.

Jamie Dimon put it best after JPMorgan took a $170 million loss on Tricolor (a sub-prime auto lender): “when you see one cockroach, there are probably more.” Bloomberg’s analysis landed on “private credit is bad, but not 2008 bad.”

We caught this early. We can fix it. Unless it gets worse.

Why it matters even if you’ve never touched it

If you have a pension, your fund likely has exposure to private credit. If losses get recognised, that flows somewhere. If you work at a mid-sized company with private credit debt, lending terms are about to get harder and more expensive. And if you’re just watching the economy: private credit funds a meaningful chunk of corporate America. If it contracts sharply, that’s a headwind for everyone.

In conclusion, private credit isn’t new. It filled a real gap banks left behind and for fifteen years, it worked. Then everyone piled in, standards slipped, AI disrupted a quarter of its portfolio, and the wave of consequences is still coming in.

Some funds will be fine. Others are already in distress. We just don’t know which ones yet because private credit, by definition, doesn’t have to tell you.

That opacity makes it very tough for everyone else to follow the reality of it. When you can’t see what something is worth, you find out the hard way.

We’re finding out.

I write about the things that make me curious, the shifts happening around us, and what it all means for how we live and spend. If that sounds like your kind of read, subscribe and I'll see you next week.

I deleted Instagram off my phone about two months ago. No announcement post. I just noticed I was opening it out of habit and closing it feeling worse. So I took it off and waited to see if I’d miss it.

I didn’t.

Instead, I started gravitating offline without really trying. Books, piano, less phone. Once the autopilot scrolling stopped, my brain just... filled the time with things I actually liked doing.

I figured this was a me thing. Quarter-life recalibration or whatever. But then I saw this tweet from Jonas Willett that made me go “oh, wait.” He’d tracked over $750 million deployed in the last three months into companies betting on in-person connection. Events platforms. Social planning apps. A literal device that blocks microphones so your real-life conversations stay private.

So I went digging.

The numbers are kind of beautiful

I expected to find some soft survey data and a few trend pieces. What I found was way more concrete.

Gen Z averages 7.2 hours of screen time a day. Over half of teenagers get more than 237 notifications daily (some get over 5,000, which honestly sounds like a form of torture). A 2024 Harris Poll with Jonathan Haidt found that more than one in five Gen Zers wish smartphones had never been invented. Nearly half of us said they’d prefer a life without the internet.

These are people who grew up with the internet. They’re not nostalgic for a world they never experienced. They’re just exhausted by the one they have.

And it’s showing up in how they spend. Brick phone sales among 18 to 24 year olds spiked 148% between 2021 and 2024. Smartphone use in that age group actually dropped 12%. Flip phone sales doubled at HMD in a single year. The dumb phone market is headed toward $10.6 billion. Vinyl records have outsold CDs for five consecutive years, with Gen Z making up 27% of buyers. Vinyl production in Europe grew 50% in the first half of 2025.

People are putting money toward things that are slower on purpose. Analog on purpose. That felt worth paying attention to.

BeReal understood the problem and then fumbled the bag

We cannot talk about the “anti-online” space without at least mentioning BeReal.

I was there. I used it. Loved it. Then forgot about it.

The idea was great. Once a day, random time, two minutes to take an unfiltered photo of your actual life. No performance, no curation. Peak BeReal in August 2022 had 73.5 million monthly users. Most downloaded app in the world for a stretch.

But they never made a dollar. No ads, no subs, nothing. The thing that made BeReal special was exactly the thing that made it impossible to monetize. When they finally tried in-feed ads in 2025, downloads had already dropped 60%. Voodoo bought them for €500 million, which feels generous for zero revenue and a shrinking user base.

73 million people showed up and the business model still couldn’t keep up. That’s what happens when you promise people an attention-free experience and then try to sell their attention anyway.

The companies figuring this out now are doing something different. They’re charging for the thing itself, not trying to monetize what happens around it.

What I think is actually going on

The easy read on all this is “people are tired of screens.” That’s true but it’s also kind of shallow. I think something more specific is happening.

We lost the ability to be bored. Somewhere in the last ten years, every empty moment got filled. Waiting in line, sitting on the subway, lying in bed before sleep. There’s always something to scroll, always a notification, always another 30-second video. And our brains adapted to that. They started expecting constant input. So when the input stops, it doesn’t feel like peace. It feels like withdrawal.

What I noticed in myself when I stopped the doomscrolling wasn’t some instant calm. It was discomfort first. My brain kept reaching for the thing that wasn’t there. It took a few weeks before I started actually enjoying the quiet, before I picked up the piano again not because I was trying to be “intentional” but because I genuinely wanted to.

I think that's the real behavioral shift underneath the numbers. People aren't just choosing offline experiences. They're retraining their brains to tolerate, and then enjoy, unstimulated time. And that process is creating a whole new set of needs that didn't exist five years ago

Talking about the companies mentioned in the original tweet: Partiful (500K MAU, up 400% YoY, Google’s best app of 2024), Posh (6 million users), Luma (2 million+ users) have raised a combined $60M. I think the demand for in-person experiences has grown so fast that people needed better infrastructure to organize at scale. That’s where the capital went.

Rodeo, built by the former COO and CPO of Hinge, had 20,000 waitlist signups before launch. It just helps you follow through on plans. That’s how broken the loop got. You want to see people. You just... don’t end up doing it.

Corner (55,000 users, 450 cities) is a social map where your friends recommend places. The deeper insight is about trust erosion. Years of algorithmic recommendations and sponsored content made us skeptical of everything except people we know. (If I see an influencer recommend something, I specifically know it’s not going to be what it claims to be.)

Even the weirder stuff maps to behavior. Spectre I blocks microphones in a 2-meter radius so conversations stay private. Five years ago that sounds paranoid. Now it sounds reasonable.

I’d say even Substack is growing partly because reading a long-form newsletter feels different from scrolling. Podcasts that go deep for two hours are growing while short-form attention spans keep shrinking.

And then there’s the stuff with no tech angle at all. Mark Cuban put a seven-figure check into Burwoodland, a company that just produces live events. Emo Night Brooklyn, Gimme Gimme Disco, Broadway Rave. 1,200+ events a year, 1.5 million tickets sold. No app, no platform, no AI. Just rooms full of people having a good time.

What I think is next

The experience economy is valued at $810 billion and headed to $1.2 trillion by 2035. Experiential marketing just crossed 38% of the marketing services market, overtaking digital at 35% for the first time. Money follows behavior, and the behavior has clearly shifted. The companies actually winning here all charge for the experience itself. The event, the plan, the place. Not your attention. That's the line BeReal couldn't figure out, and the line everyone else on this list got right.

The same firms writing checks for OpenAI are backing Partiful and Burwoodland. As Jonas put it, they're not hedging. They're reading the room.

People got tired and that tiredness created a gap. Capital filled it. But what’s underneath is simpler than any investment thesis. We’re social animals. You can automate a lot of things but you can’t automate sitting across from someone you like and laughing at something stupid. You can’t replicate a packed room where everyone showed up on purpose. The more AI handles the transactional parts of life, the more the non-transactional parts become the thing worth paying for.

I think we’re watching people figure that out in real time. Not because someone told them to go outside, but because their own brains started asking for it.

As for me, I now rarely open Instagram to pass time. The piano is getting more use than it has in years. And I’m writing about things I actually find interesting, which turns out to be a pretty good way to spend time online.

Funny how that works.

This is the first piece in what I'm building here. I write about the things that make me curious, the shifts happening around us, and what it all means for how we live and spend. If that sounds like your kind of read, subscribe and I'll see you next week.