This is not idle sentiment. Culture shapes technology, and technology shapes culture in return. It has always been this way. Every generation builds the future it first dared to imagine. Recent technological advancements have annealed this bond. Large language models are made from humanity’s written record. Quite literally, they embody the zeitgeist. Our fears become their fears. Our hopes become their hopes. The stories we tell ourselves about tomorrow are becoming the machines that will build it. As we forge our next chapter, believing in what’s next has never mattered more.

This piece is an exploration of that future, and of where great businesses will be built inside it.

Where in history are we?

“History doesn’t repeat itself, but it often rhymes.”

For twenty years, venture’s most reliable playbook was software as a service. Build a product once, sell it to thousands of companies, and collect rent on it every month at incredible gross margins. It was one of the best business models venture ever had, and it shaped an entire generation of investors.

Looking back, the generation found its voice in 2011, around the time Marc Andreessen published his thesis “Why Software Is Eating the World” in the Wall Street Journal. Founders Fund’s manifesto lamented, “We wanted flying cars, instead we got 140 characters” and questioned “What happened to the future?”

Fifteen years later, the irony is hard to miss: the future is here, software is eating itself, and the best minds of a generation are building flying cars.

Andreessen was right, so right that software is now eating itself: language models are reshaping the industry that built them.

And the future Founders Fund asked about is here too. For a while, the smartest people of a generation built accounting software. Now they’re building power plants, robots, therapeutics, and flying cars.

Venture has changed

The prevailing attitude in venture today is that software is dead, overthrown, like Kronos, by its own child.

It’s easy to hear that and brush it off. Sure, software is dead; everyone’s been saying that for over a year. Here’s the thing - working from the inside of the industry, it’s shocking how different venture looks today from even two years ago. People knew these things were happening, but the effects weren’t fully felt yet. When it’s suddenly easier to get funding for a magnet refinery than an Uber competitor, it does leave you scratching your head and looking around asking what just happened?

I’ll tell you what happened. Capital is rotating. Software funds are raising money to deploy into factories, defense, energy and space. Investors are scrambling to up-skill on enormously complex categories such as molecular biology. The best young engineers are choosing materials, pharmaceuticals, and robotics over another business app.

The move from bits to atoms has been building for years, and recent months have been peak mania.

While the change has been fast, we’re not entirely in uncharted territory. Venture was born in atoms. Silicon Valley is named after an element. One of the deals that started the valley, Arthur Rock funding Fairchild Semiconductor in 1957, was a bet on hardware. Bessemer, one of the oldest firms in the business, is named after a way of making steel. When Don Valentine told Jensen Huang “If you lose my money, I’ll kill you”, he was investing in a new type of chip.

Venture is coming home - and futurists should be happy about that.

But we must tread lightly: home may be harder than we remember. SaaS was an easy button, and it’s gone.

What’s actually worth investing in right now?

What software lost

To answer that, it’s worth backing up for a second. Why exactly do investors think SaaS is dead?

The short answer is that coding models “solved” software. An app that used to take a team of 12 engineers a month can effectively be completed by a single developer in a day. Moreover, the task of engineering itself is different. Writing software used to be an ivory-tower discipline for the “gifted programmer”. Now it means popping in a Zyn and, between prompts, wondering if you should’ve studied biology instead.

Expensive code

Cheaper code breaks one of software’s oldest moats. Software was always really expensive to build. Good engineers command high salaries, are hard to recruit, and (deep in the weeds of bugs and code) can only ship so much. You needed a lot of them to build anything good, so for most companies, buying software was the only sane option.

Coding agents are undoing that. Starbucks, which spends about $400 million a year on software, is already exploring replacing software it buys from Microsoft and IBM with tools built in-house. Smaller teams that used to rent generic software like Notion can now vibe-code their own, built around exactly how they work. The question “Why don’t we just build this ourselves?” is getting harder to dismiss.

Of course, building the software is the easy part of replacing it. Years of integrations, compliance requirements, and tribal knowledge won’t disappear overnight, so the incumbents will be fine for a while. New companies aren’t so lucky. They start without those advantages, in a market where every competitor can build quickly, and the window to make something durable is narrowing. The code alone buys less time than it used to.

Ambitious talent

There is a second shift, and in early-stage investing it may matter even more: what ambitious people want to work on.

Writing software isn’t sexy anymore. The smartest people in the world want to work on hard, impactful problems. Having spoken to talented young people, telling Claude what to code just isn’t all that inspiring. Among brilliant young minds, there is a renaissance of interest in biology, materials science, robotics, manufacturing, space, and the frontier of model research.

This should worry anyone investing in early-stage software. Early-stage investing is foremost a people business. The main thing you underwrite is the team. When talent moves, capital has to follow.

Schumpeterian profits

So software is under scrutiny, the best minds are heading back to the frontier, and capital is following them there. That’s good news for the future. But as history reorganizes, two lessons are worth carrying with us.

The first is about what happens when a new technology shows up. People get excited, and then they fund and build everything it makes possible. That’s a good thing. It’s how the world gets better. Most of the value from a new technology ends up with the people who use it: cheaper power, faster travel, longer lives. Economists call this consumer surplus.

It’s also why innovation, on its own, is a terrible business.

Nobel laureate William Nordhaus estimated that from 1948 to 2001, American innovators captured about 2.2% of the value their innovations created. Being first buys you a temporary reward, what Nordhaus called “Schumpeterian profits,” and then competitors copy you and compete it away.

The intelligence race may produce the largest Schumpeterian profits in history, and some of the shortest-lived. Work that used to require expensive human labor can now be done with compute, and right now that spread earns a premium: companies such as AI-powered rollups that replace labor with models are posting remarkable margins. But as the whole economy shifts from labor to compute, the price of intelligence will fall toward the price of compute, and the price of compute toward the price of energy. Without a moat, you don’t have a generational business, even if you have God in a B300.

Directionally correct isn’t good enough

The second lesson is about manias. When a story about the future is true, capital floods in to fund it, and people stop being discerning. Walking around the valley today, people love to talk about being directionally correct. We’re seeing this lack of discernment right now.

We’ve also seen this before. In the 1840s, Britain went all in on railways, the hottest market of its era. At the peak in 1847, railway construction took up nearly 7% of GDP. Between 1845 and 1850, railway shares lost more than half their value, and some of the biggest railways turned out to be paying shareholders with new investors’ money rather than profits.

The track still got built. The lines approved in those few years add up to more than half the length of Britain’s rail network today. The railways changed the world, and many of the families who funded them lost their savings anyway.

A century and a half later, it happened again. In the five years after 1996, telecom companies spent more than $500 billion on fiber and networks, close to a trillion dollars today. Everyone was right that the internet would need bandwidth. Still, by 2001, about 95% of that fiber sat unused, and Global Crossing, once valued at $47 billion, was bankrupt.

That same fiber now carries YouTube, Netflix, and much of the cloud. The investors were right about where things were going. They never asked how much, and how fast?

Once again, history rhymes. The economist Carlota Perez found the same pattern in every technological revolution since the 1770s. A frenzy of capital builds the infrastructure, a crash follows, and the real returns come in the deployment after it.

Here we find the trap of peak mania. The railway investors were directionally correct. They lost anyway. When everyone can see the writing on the wall, being directionally correct isn’t good enough.

Where should we look to invest?

We believe that great investing sits at the intersection of understanding the future technically and the shape of a good business.

Here is our view of where the future is going. Great businesses will be built inside these themes. Topology exists to back them.

Meet us at the edge

We invest in people hellbent on ushering in new ages of progress. If you’re building at the edge, we want to hear from you.

Over the past several months, a rapid succession of releases including DeepSeek V4, MiniMax M3, Tencent’s Hy3, NVIDIA’s Nemotron 3 Ultra, and GLM 5.2 have steadily narrowed the performance gap with frontier closed models, making it increasingly difficult to ignore the open-weight ecosystem. The real inflection point came in July when Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that reset expectations for what “open” could mean at scale. It was not only the largest open-weight model ever released, but also marked the first time since DeepSeek R1’s emergence that an open-weight model achieved near-performance parity with leading closed-weight models (as of its release, nearly on par with Fable 5). And perhaps most importantly, K3 was able to achieve this performance at ~30% of the cost per output token of comparable closed models. All of this activity culminated yesterday in Meta readopting a pro-open-weight position and announcing their new Muse models as open weight.

The flurry of new models is the latest iteration in what has been a two-year-long cat-and-mouse game between US frontier models and Chinese open-weight models: a leading US lab like Anthropic or OpenAI releases a new model that breaks all prior performance benchmarks, and just a few months later, a Chinese competitor like DeepSeek, Z.ai, or Moonshot releases an open-weight model that closes the gap. The broader trend, which the graph above shows, is what’s really interesting. Not only are the open-weight models becoming more and more performant, but the open-weight gap is at the closest point of parity since the DeepSeek upset, which put China on the map.

A year ago, a common belief among many Western investors and researchers was that U.S. export controls on advanced GPUs would create a widening gap between American frontier labs and Chinese competitors. Instead, Chinese labs have repeatedly demonstrated that they can produce frontier-class open-weight models despite those constraints. How exactly this has been possible is the subject of much debate: one proposed contributor is distillation, although the extent remains disputed.

While the frontier models still outperform on most tasks, it has become clear that open-weight models like Kimi are good enough for many tasks, and the pareto-optimal approach is to use the cheapest intelligence capable of completing the work reliably.

In practice, we’ve observed that engineering teams often prefer to use SOTA closed models to handle more complex, planning-oriented, or general-purpose tasks, while open-weight alternatives like Kimi handle smaller-scoped, repetitive, or predictable workstreams. The data clearly supports that the general market agrees. OpenRouter data shows that open-weight models account for ~63% of OpenRouter token usage in the US. One application-layer unicorn founder we spoke with told us that 98%+ of their token usage is now running through open-weight models. DoorDash, Pinterest, Coinbase, and Cursor have all confirmed using open-weight models internally for low-level tasks.

The Hugging Face data above measures distinct accounts publishing tagged derivatives of selected base families—not inference usage, downloads, or every distillation relationship. Publishing is concentrated around accessible families such as Qwen, while frontier-scale families remain smaller: the chart now calls out Moonshot’s Kimi activity explicitly, and shows Muse Glimmer in a separately labeled August snapshot because the month is still incomplete. We expect Kimi K3, Muse Glimmer, and DeepSeek V4 to broaden the fine-tuning surface as smaller checkpoints, adapters, and distilled descendants become available.

But perhaps more important than just the usage data, we’re seeing that the zeitgeist is starting to shift meaningfully towards an open-weight future, with proponents arguing that it is the only way to ensure safe, aligned models and American competitiveness. NVIDIA CEO Jensen Huang published an open letter at the end of July endorsing the importance of open-weight models. It was then signed by several of the leading AI companies, including OpenAI, Microsoft, and Google. Even Anthropic, who has been vocally cautious of open-weight models, finally endorsed their value. And most recently, Meta CEO Mark Zuckerberg’s manifesto paints open-weight models as vital to ensuring AI maximally helps humanity. Whether or not we believe the marketing, it has become clear to us that open-weight models are not only here to stay, but will become a vital part of the technology stack for any AI company going forward.

What an open-weight future means

So if we believe that an open-weight future is inevitable, what does this mean for startups and for the tech ecosystem at large? Below are a few of our predictions:

Open-weight models will lead the pack when it comes to intelligence per dollar. Given the cost advantages of serving these models and their likely continued improvement over time, open-weight models will continue to eat up more and more AI workloads, particularly for lower-level or specialized tasks. Additionally, we believe this trend will put downward pricing pressure on frontier models. OpenAI recently announced that they were dropping the price of Luna and Terra models by up to 80%, making them pareto optimal. While frontier labs are unlikely to open-weight their flagship models, they will likely continue competing aggressively on price for lower-tier offerings. Their ability to do so will depend not only on pricing strategy, but also on whether they can maintain meaningful advantages in inference efficiency, infrastructure, and systems engineering that are difficult for the broader ecosystem to replicate.

The foundation labs’ core business is under threat. We believe more performant open-weight models will continue to eat into the frontier labs’ enterprise API business. While these businesses will continue to have the advantage of brand and distribution, in order to stay relevant, we will see them make investments in new models and business lines. We are seeing early signs of this in Anthropic’s and OpenAI’s expansion into products and programs for science, as well as in their investments in forward-deployed services and implementation work. We wrote a previous article highlighting why R&D may be a special market for SOTA advantage.

Value accrues to the application-layer leaders. Today, our data shows us that most AI apps are hovering at around 40% GM, with a huge portion of COGS being inference costs. With cheap, fine-tuned models replacing more and more of their workloads, we can expect this to creep closer to traditional SaaS margins, something like 70%+. This advantage will particularly accrue to the leading players at the application layer, who already have moats in the form of trust, integrations, deeply embedded workflows, and proprietary data.

Model-agnostic ecosystems will thrive. In a world where companies have a plethora of model choices and are optimizing for some combination of performance, latency, and cost, model-agnostic “middleware” solutions will become increasingly important. This includes model-routing companies like OpenRouter and increasingly Cursor in addition to sandboxes for fine-tuning and evaluating models such as Prime Intellect and Braintrust.

The US-China proxy war will heat up. Chinese companies will continue to promote an open-weight ecosystem, given their bet that intelligence becomes a commodity, with the physical supply chain as the limiting factor. Export controls on China will increasingly be an important policy question.

Overall, we believe that the future is open weight. This will encourage more competition and transparency in the AI ecosystem and will ultimately ensure that AI diffuses to every sector of our economy. If you’re leveraging open-weight models or building infrastructure for this ecosystem, we’re excited to meet you.

The prevailing bear case for model labs is straightforward: intelligence becomes abundant, models converge in quality, inference prices collapse, and economic value migrates upward to the applications that own customers and distribution. You've probably heard it: “the model layer commoditizing.”

This is already happening, partially. But it does not follow that the entire model layer will commoditize.

A worthwhile upstream question is: Which kinds of intelligence commoditize?

The intelligence floor

Most enterprise tasks have a minimum level of intelligence required to perform them adequately.

Imagine an email extraction task: find the date of an appointment in an arbitrary email. Below a certain capability threshold, the model is unusable. It picks dates incorrectly, misses important context, and creates more work than it saves through inaccuracy. Once the model reliably clears that threshold, however, additional intelligence may not matter very much.

A model that is 20 percent “smarter” does not necessarily create 20 percent more value. If several models can all perform the task with sufficient accuracy, buyers begin optimizing for price, latency, reliability, privacy, and ease of integration.

The task has crossed its intelligence floor.

After that point, intelligence becomes interchangeable. It’s commoditized. And the race to the bottom begins.

Competition pushes the model’s price toward the underlying cost of inference, just as standardized compute, storage, and bandwidth gradually became utility-like inputs.

The empirical evidence for this process is strong. Stanford’s AI Index found that the inference cost of achieving roughly GPT-3.5-level performance fell more than 280-fold between late 2022 and late 2024. Smaller models continue reach performance levels that previously required much larger systems.

Open-weight models exacerbate this dynamic. Over the past year they have consistently trailed the frontier. Once an open-weight model clears a given intelligence threshold, competition to serve intelligence at that level explodes. Possible oligopolistic pricing structures that preserve model-provider margins become nearly impossible to sustain.

This suggests a general rule:

Once several models exceed the intelligence requirement for a task, that task’s model economics begin to commoditize.

Much of the existing enterprise market may fit this description. Document extraction, summarization, basic customer service, routine content generation, classification, translation, and constrained workflow automation do not require unlimited intelligence. They require sufficiently high accuracy at sufficiently low cost.

As every major lab improves, more tasks fall behind the frontier and enter this commodity zone.

But not every task has a floor

There is another class of work for which better intelligence continues to produce more economic value.

This type of work is that of a generally intellgent worker who faces open-ended tasks with undefined ceilings of excellence.

Consider software engineering. A model that can reliably resolve a one-hour ticket by writing slop is useful, kind of. On the other hand, a model that can resolve a one-hour ticket by reasoning across a large codebase, can coordinate its own tests, write maintainable code, identify forward facing bugs, and recover from mistakes is not a substitute for the first model. Even if the task is the same - additional intelligence improves the work.

The same logic applies to scientific research, complex engineering, strategic decision-making, legal reasoning, and general-purpose autonomous work. In these domains, improvements in intelligence can expand the set of economically valuable tasks a model can perform.

There is no obvious point at which the market says: “This model is intelligent enough to conduct research.”

A more capable research model might generate stronger hypotheses, design better experiments, interpret more ambiguous results, or operate over a longer research program. A more capable coding agent might move from completing tickets to building products. A more capable general agent might progress from assisting an employee to substituting for increasingly large portions of the employee’s work.

For these tasks, the capability curve may not contain a fixed intelligence floor. It may instead resemble an expanding frontier:

This is the portion of the market that may not commoditize.

A model market split in two

The model layer is therefore likely to divide into two markets.

The first is the saturated-intelligence market. These are tasks for which many models have cleared the required capability threshold. Performance differences become less important, switching costs are limited, and providers compete on cost and operational characteristics. Intelligence in this market trends toward the price of compute.

The second is the frontier-intelligence market. These are tasks where the most capable model can perform work that weaker models cannot, or can perform it with materially higher reliability and autonomy. Here, access to the frontier may command a substantial premium because the product is not tokens. The product is access to a larger domain of economically useful cognition which is not widely held.

The 'time arbitrage' of this held advantage is exceedingly important: a small capability lead can produce a large economic advantage when it unlocks a new category of work.

How large is the frontier market?

The model-commoditization thesis depends on an implicit assumption: that most economic demand for AI will eventually consist of bounded tasks with attainable intelligence floors.

This isn't unreasonable. Most business processes are repetitive, constrained, and tolerant of “good enough” performance. The enormous volume of these tasks could make commodity inference a very large market even at low margins.

But a disproportionate share of economic value may sit in the opposing category.

The world spends trillions of dollars on software engineering, scientific research, engineering design, product development, and other forms of difficult, open-ended knowledge work. As we've explored, these are domains where better intelligence can continue to improve the quality of the outcome rather than merely clear a fixed threshold.

If each generation of frontier models can automate a larger share of this work, the leading model is not competing only for existing AI spend. It is continuously expanding the addressable market; creating entirely new categories of value.

Take drug discovery. A frontier model that identifies a previously unknown drug target may make a new therapeutic market addressable; a model that helps invent a new computing interface, material, or physics breakthrough could create an entirely new market. In both cases, SOTA intelligence expands the economy rather than merely capturing existing AI spend.

This creates an unusual industry structure:

The lagging edge commoditizes while the leading edge discovers new demand.

Every capability eventually becomes cheaper and widely available. But the frontier may continue moving outward faster than the commodity market catches it.

Under this model, claims that “models are commoditizing” can be simultaneously correct and incomplete. They describe what happens to a fixed level of intelligence over time. They do not acknowledge the economic value of the moving frontier.

A note on the frontier vs commoditization gap

It's worth noting that the progress of frontier model capabilities may take two dramatically different trajectories in the future: a plateau or recursive intelligence improvement.

In the plateau scenario, progress in frontier intelligence slows despite continued investment. Successive model generations produce smaller capability gains, even as the capital required to train and operate them increases by orders of magnitude.

As frontier progress slows, cheaper and open-weight models have more time to catch up. The gap between frontier and commoditized intelligence narrows. Eventually, the leading model may remain technically superior without being sufficiently better to command a meaningful economic premium.

Note that this is not the trajectory we currently appear to be on, given METR’s measurements of model task-completion horizons.

Recursive intelligence offers a fundementally different outcome: labs at the forefront may capture an enduring and rapid capture of frontier market value.

If frontier models become capable of meaningfully contributing to the research required to build better frontier models, progress may accelerate rather than plateau. A leading lab could use its current capability advantage to improve its research process, produce a stronger successor, and then use that successor to accelerate the next generation.

The gap between frontier and commoditized intelligence could therefore substantially widen rather than narrow.

In my view, this is the bull-case justification for the extreme valuations commanded by frontier labs today. The risk profile is extraordinarily asymmetric: the frontier-intelligence model market either enters a race to the bottom, or the leading labs capture a large share of an unprecedented expansion in economically useful intelligence.

The frontier model layer does not commoditize

In conclusion, the strongest version of the commoditization thesis is probably wrong.

Models do not move uniformly from differentiated products to interchangeable utilities. Instead, each level of intelligence appears to pass through a lifecycle.

At first, a capability exists only at the frontier and commands a premium. Competitors catch up. Smaller and cheaper models reproduce it. The capability becomes standardized, and its price collapses toward inference cost.

But while this is happening, the frontier models may unlock a new capability above it.

The model market is therefore better understood as a cascading system:

The unresolved question is not whether today’s intelligence will commoditize. It almost certainly will.

The question is whether the frontier continues to move into economically valuable territory quickly enough to create a persistent premium for whoever leads it.

My base case is that it does. Routine intelligence will become abundant and cheap. But the most valuable portion of the model market may remain attached to a moving boundary of autonomy, reliability, and task complexity. That boundary could represent a large share of model-layer value even if it represents a minority of total inference volume.

Models commoditize, but frontier intelligence may not.

The strategic consequence is that the model layer may support very different businesses at the same time: low-margin utilities serving saturated tasks and extraordinarily capital-intensive frontier labs selling access to the next tranche of automatable work and discovery.