AI’s Capital Flywheel—and the Businesses It Might Leave Standing

Six months old, negotiating hundreds of millions

A company barely six months old may already be negotiating hundreds of millions of dollars in compute commitments. Sarah Wang’s example captures how AI changes early company building: choosing a compute supplier can also mean negotiating equity, selecting a strategic partner and arranging distribution. Those decisions arrive while the investment is still fundamentally a bet on a technical founder.

That mismatch opens the Latent Space conversation with Martin Casado and Sarah Wang. Wang argues that young AI companies need early-stage judgment and growth-stage support simultaneously. Casado adds that a company without revenue can already have enough users and capital requirements to demand sophisticated quantitative analysis. He contrasts earlier $20–60 million Series A or B checks with rounds whose strategic investments come attached to compute contracts that take months to negotiate.

Asked about circular strategic funding, Casado focuses on whether the purchased GPUs serve real customers. He contrasts their utilization with unused fiber during the internet build-out and argues that AI lacks a comparable supply overhang. Wang adds the critical condition: spending on research must produce better capabilities, and those capabilities must produce willingness to pay. The financing logic depends on scaling continuing to improve capability. If that relationship breaks, so does a central premise of the investment.

When capital becomes a product-development loop

Model companies also blur the boundary between infrastructure and applications. They conduct core research, supply platforms for other developers and reach users directly. Wang describes an emerging financing sequence: 1. Raise capital and buy compute. 2. Convert compute into a capability breakthrough. 3. Put that breakthrough into an integrated application, such as ChatGPT or Claude Code. 4. Gain users and market share, potentially with subsidies. 5. Raise again at peak momentum and repeat.

One investor argues that a model team of roughly 10–20 people can turn funding into a better model within a year, then use the resulting demand to finance another cycle. Traditional software development has coordination bottlenecks that hiring alone cannot remove. This capital-to-capability flywheel could make money a more direct input to product improvement.

A host extends the idea into a speculative future: spend $1 million on inference—running models—today and obtain a product tomorrow. If model outputs could become useful products with so little friction, early-stage investing could become much more iterative. The imagined funding extremes, $100,000 of inference or a $100 million seed round, describe that possibility rather than demonstrated startup economics.

The uncomfortable consequence is competition between a platform and its customers. Wang points out that model providers sell access to developers while competing with them through applications. Will AI develop the stable boundaries that allowed separate companies to win different layers of the cloud stack? In a hypothetical Anthropic scenario, each successive funding round is three times larger. If the lab can outspend its entire application ecosystem, it could expand into those applications’ markets. The participants call this a bitter lesson for startups: capital turns into growth, which finances still more capital. Whether that cycle can sustain such dominance remains unresolved.

Character: research and product compete for the same GPUs

When a host suggests that Character demonstrates a frontier lab overwhelming an application through spending, two respondents reject the explanation. An investor involved in the deal dates the firm’s investment to January 2023 and Character’s Google intellectual-property licensing agreement to August 2024. Their account emphasizes artificial general intelligence (AGI) versus product operation: they characterize Noam Shazeer’s ambition as reaching broadly capable intelligence, with Character’s human-centered product and data collection serving as a route toward it.

The conflict is concrete. A lab must divide limited GPUs among serving users, near-term research and long-term research. Product revenue helps fund research, but serving the product consumes resources researchers also want. Fundraising can relax that constraint, provided progress continues and capital remains available. The investor stresses that fundraising conditions in 2023 were different from those surrounding the later frontier-lab rounds.

AGI ambition can also weaken attachment to a particular company. One participant argues that researchers may prioritize whichever organization seems most likely to advance that goal, contributing to what they perceive as unusually high founder movement. Intense recruiting and media scrutiny add to the pressure.

Participants tentatively view Meta’s 2025 recruiting surge as a possible one-off team-building episode, while arguing that it raised the broader compensation baseline. One hiring participant reports candidates with competing offers around $10 million a year. A host observes that a $5–6 million job changes the incentive to become a founder. An investor offers the counterweight: abundant strategic capital and acquisitions made primarily to recruit teams can produce strong venture outcomes. The same talent competition can destabilize a startup and reward its investors.

Overlooked software and the economics of robots

Asked what deserves more investment, one participant points to databases, monitoring, logging and developer tooling. Investor attention, they argue, has concentrated on fashionable startups and deep technology while overlooking durable software businesses. AI can create new demand for these products even when their core business sits outside the token path: they remain software businesses rather than model-output businesses.

The complaint is that growth has become a substitute for investment analysis. The participant objects to dismissing a company growing fivefold in a large market because it misses a fashionable zero-to-100 trajectory. Neither comparison specifies enough metrics, periods or valuations to calculate returns. The useful point is that headline growth says little about margins, and rapid company growth alone does not establish an attractive investment.

Robotics presents a different problem. An investor says the team is studying hardware but has not yet seen the ChatGPT-like breakthrough that funding levels appear to assume. Another draws on the drone and autonomous-vehicle waves: an agricultural robot inherits agriculture’s competition, pricing and supply chains; a mining robot inherits mining economics. A robot must compete economically in its customer’s industry, including against human workers.

Believing robotics will succeed is therefore different from knowing how to evaluate a particular robotics business. Applied Intuition, DeepMap and Scale AI are cited as opportunities in software serving multiple customers or industries. Physical deployments require more sector-specific investigation. The American Dynamism team is described as equipped for hardware, regulatory compliance and government sales, while the speakers’ own strengths lie in core technologies that serve many industries.

Custom chips: the arithmetic can work before the schedule does

For a custom chip tailored to a model, one investor puts the rough economic threshold at a $1 billion training run. The qualification is timing: the chip must be ready while it is still useful for that model.

The illustrative arithmetic assumes roughly $1 billion of inference spending. Cutting that cost by 20% saves $200 million, which the investor says could fund a chip tape-out; halving it saves $500 million. These are hypothetical savings, not chip measurements. The argument does not clearly distinguish inference revenue from inference cost, so its spending base remains an assumption. Economic justification and timely delivery are separate hurdles.

What investors can automate—and what still needs a network

Company building also depends on less tangible infrastructure. An investor argues that geographic concentration compounds effectiveness through hiring networks, customer introductions and local presence. Historically, early-stage companies needed more of that help than mature businesses. AI complicates the distinction because companies raising large rounds can still require extensive operational support. Local headquarters do not imply local markets: another participant emphasizes that portfolio companies sell globally and sometimes use international supply chains.

A growth investor describes much of their help as finding the right person to shorten a process, with no clear AI workflow yet for that relationship-driven work. Analysis feels different. Three team members gave Claude Cowork a raw file and reportedly received results within seconds that they checked and found accurate. Customer database analysis and cohort retention—tracking whether groups of customers stay over time—are the motivating tasks. This was a successful team trial, rather than a controlled accuracy evaluation.

Two futures for the model industry

A host speculates that an Anthropic consumer push could threaten OpenAI; a respondent emphasizes Anthropic’s stated enterprise focus. That disagreement leads to two possible industry structures: - Fragmentation: competitors quickly catch up after model releases, and many companies capture value as software is rebuilt. - Concentration: additional training investment produces models general enough to enter adjacent markets, leaving a few dominant providers.

Model lifetime is central to the economics. One investor argues that revenue can look attractive against the cost of training the previous model while failing to cover current spending on its successor. Without company figures or a defined accounting method, this is an argument about financing rather than a reported margin calculation. Short-lived model advantages require repeated investment; failure to fund the next run could let competitors catch up. Falling compute prices could ease that constraint.

Expectations have already swung sharply. One participant recalls GPT-4 appearing to lead for roughly nine or ten months, benchmark-leading open-source releases seeming almost daily around March 2024, and the market later feeling concentrated again. These approximate recollections underline how difficult it is to predict a durable structure while capability progress continues.

One route to more durable economics is task-level saturation: a model becomes good enough that further improvement adds little value for a particular job. Its useful life could lengthen, spreading its development cost over more time. Opus 4.5 is offered as a hypothetical example. Specialized enterprise companies might then compete through implementation and services that make the capability useful. Legal work is suggested, then qualified because models are still improving in that domain.

The counterargument is that a provider need not achieve AGI to threaten its customers. Usage of its application programming interface (API) can reveal where demand exists, while large funding rounds can finance entry into those markets. Task saturation might stabilize the technology without resolving competition from the supplier.

A coding assistant does more than code

Even a supposedly narrow task can require broad capabilities. One participant describes coding sessions that involve compliance, web searches, historical knowledge and ongoing conversation. Perhaps the generally strongest model wins because real work keeps crossing the nominal task boundary.

In that participant’s personal experience, Codex finds difficult bugs better than Opus 4.5, while Opus has a conversational manner that makes it a stronger brainstorming partner on complex projects. The preference illustrates a distinction between solving the technical problem and collaborating through it; it is not a benchmark result.

The participants disagree about whether this leaves a durable category for specialized coding models. One argues that a coding assistant must be good at much more than programming. Another expects general GPT and Codex variants to persist, perhaps differing along only a few dimensions, such as coding performance and conversational manner. Their practical suggestion is to notice how much of a coding session is not code.

World Labs: useful scenes need useful representations

Asked what they are coding, an investor describes helping World Labs, whose foundation model generates 3D scenes represented as Gaussian splats. These scene representations need rendering support in an industry historically organized around meshes. The speaker argues that radiance-field representations can reconstruct scenes without requiring mesh topology; the practical obstacle is getting the resulting scenes into usable graphics workflows.

Spark.js addresses that obstacle with an open-source JavaScript renderer for Gaussian splats in the Three.js ecosystem. The contributor builds world-building demos to exercise the library and expose scaling problems. They describe a two-person effort with Andreas Sundquist as the primary developer, an acquaintance from their PhD years. Having paid for undergraduate study through game-engine work in the late 1990s, the contributor credits AI assistance with making development accessible again: it handles framework work while they focus on algorithms and scaling.

Using a large language model to write spatial software raises a different question from whether the model itself represents space well. A participant challenges the need for separate systems by citing DeepMind’s reported mathematics-competition gold performance: they contrast the 2024 AlphaProof/AlphaGeometry approach with what they describe as a single DeepThink language model using extended thinking. The comparison comes without evaluation details. The response allows that architectures may converge while maintaining that different problems need different representations.

The spatial example is a dark room. Verbal directions can tell someone to move left or duck beneath an obstacle. Turning on the lights conveys distances, shapes and possible movements much more directly. The speaker uses this contrast to explain why ordinary language is inadequate for precise spatial reasoning, offering an interpretation of Fei-Fei’s position rather than quoting her.

The investment case starts with reducing the marginal cost of something people already buy. Images, speech, films and 3D scenes have paying markets. An investor imagines commissioned room modeling costing $4,000–$10,000 through Fiverr or roughly $30,000 from a professional, then asks what becomes possible if comparable output can be generated from an image for less than a dollar. The figures illustrate a potential cost reduction; they are not a matched comparison of quality or production costs.

Generation differs from reconstruction. Many scene-reconstruction apps collect photographs from multiple viewpoints. Marble is described as accepting a single 2D image and generating regions it cannot see, such as the back or underside of a table. Those regions are generated rather than recovered from additional observations.

Selective bets, capability breakthroughs and public noise

The foundation-model investment thesis combines two beliefs: exceptional founders can be identified through prior achievements, and specialization leaves room for multiple valuable companies. ElevenLabs is offered as a specialist that has retained a leading position despite many competing audio-model launches. The investor’s description of it as number one has no specified ranking metric; the underlying argument is that general-purpose labs have not automatically won every domain.

Compute helps explain the size of these bets: one investor estimates that it typically consumes about 80% of a round. The expected payoff is a capability breakthrough that unlocks demand quickly. They cite an unnamed company reaching tens of millions in revenue within weeks around its general release, contrasting that with software companies taking seven years to reach a similar level. The revenue measure and exact interval are unspecified, but the anecdote captures the investment thesis: capability can unlock unusually fast commercial uptake, including for specialists.

Asked whether a founder split changes the Thinking Machines thesis, an investor expresses greater confidence, anticipates a significant 2026 and expects recovery from January’s events. They refer to undisclosed work without presenting results or explaining the split.

Another participant explicitly declines to comment on Thinking Machines specifically, then describes a broader problem: rumors about events they witnessed become distorted through repetition, acquire credibility and leave founders managing invented problems alongside real ones. Their advice, associated with the Cursor team, is to keep attention on building the business.

Cursor and the margin question

Cursor closes the discussion by illustrating two routes to vertical integration—owning both models and applications: - Start with a foundation model and build applications above it. - Start with an application and its product data, then develop models underneath it. An investor presents Cursor as following the second route. They report that it developed a near-frontier coding model for one hundredth of the cost or less, and that the model was briefly the most popular coding model. No model name, cost baseline, benchmark, popularity measure or dates accompany those comparisons.

The strategic question is how to capture margin on the tokens flowing through the product. Cursor is praised for knowing what business it is in: developer tools for professional developers, focused on coding. Its Graphite acquisition illustrates that broader tools ambition. A focused application can support a substantial business while creating an opportunity to build proprietary models. The investor nevertheless expects models from other providers to remain in use.

The host advances a separate thesis for agent labs, companies building products on other providers’ models. They may price against human labor or time saved while purchasing intelligence at falling per-token prices. If that gap widens and applications capture the value, their margins could exceed those of model suppliers.

The response ends on the supplier’s power to compete. A model provider can launch its own application, subsidize it and charge third-party applications more. Cloud infrastructure and operating systems offer historical parallels, but the participants see this version unfolding unusually quickly. Pricing against labor may create room for an application business; its durability still depends on what happens when its supplier becomes its competitor.