AI Startups: From Data Infrastructure to Products That Learn

Why a new AI lab still needs dbt

Within weeks of Thinking Machines forming, dbt was already an important part of its stack, Sarah Catanzaro recalls. A frontier AI lab still needs to manage training datasets and understand how people use its products. Analyzing interactions with agents and large language models adds complexity—and sustains demand for data tooling even when data teams grow less quickly than expected.

That practical need underlies the opening disagreement in this year-end 2025 conversation. Asked whether the dbt–Fivetran merger marks the end of the modern data stack, Catanzaro of Amplify Partners rejects the premise. She discloses that her firm funded dbt and says both companies were growing healthily and beating revenue targets. In her account, combining them accelerated the path to liquidity as the market demanded greater revenue scale for an initial public offering (IPO).

She estimates that IPO revenue bar at above roughly $600 million and believes the combined business would be close to $600 million. These are approximate figures, with no revenue period or metric specified. Her broader point is that demand for tools need not track headcount: moderately sized data teams can still support substantial analytical work.

Changing workloads, familiar infrastructure

One technical change is workload predictability. Traditional analytics queries often came from deterministic systems such as Tableau or Looker dashboards. Dataset analysis, curation, and preparation are more ad hoc. Catanzaro expects that variability to complicate learned optimizers, whose designs benefit from predictable workloads, while leaving open whether it requires a broader rethink of data infrastructure.

Data catalogs, which help users discover available data, illustrate how a real need can fail to sustain a standalone product category. Catanzaro spent substantial time working on catalogs as a data scientist and believed strongly in dedicated products. She now calls that an investment thesis she got wrong: cataloging features inside Snowflake and dbt proved good enough for human users, weakening the case for a separate purchase.

Her tentative alternative changes both the user and the job. Metadata services for machines could serve microservices and agents. Governance might also have offered a stronger business than the human discoverability many vendors emphasized. These are possible directions for the category, not established replacements.

Closer to the hardware, the constraint is direct: a GPU waiting for data is an idle GPU. Catanzaro says labs scrutinize discovery, preparation, and loading because inefficient pipelines waste compute capacity. An investor highlights portfolio company Spiral’s Vortex file format for efficient loading onto GPUs, though without benchmarks or workload conditions.

Existing infrastructure has also held up surprisingly well. One participant recalls expecting a friend working on transactional databases at OpenAI to describe a new database; the answer was RocksDB. Agents interacting with one another could eventually drive much larger transaction volumes, but another participant counters that AI is still smaller than advertising. Future growth is a reason to watch the infrastructure, not yet a reason to assume it needs replacing.

A $100 million seed round needs a plan

Looking back at 2025, the investor describes repeatedly encountering seed rounds above $100 million for companies with a long-term vision but little near-term roadmap. Several proposed frontier labs for more secure, humane, or data-efficient AI, yet offered only tentative ideas about the next six months—perhaps a consumer app. The concern is the gap between the capital requested and the work it would fund.

The proposed fundraising method starts with execution: 1. Define milestones for the next 12–24 months. 2. Estimate the headcount, compute, and equipment needed to reach them. 3. Add roughly a 20% buffer.

The investor acknowledges backing familiar founders before they had clear plans. But an unfamiliar founder requesting a commitment to a $100 million—or sometimes $1 billion—raise within seven days leaves little time to build conviction. The deadline cuts both ways: founders also need time to assess whether an investor will be a useful working partner.

Large announcements nevertheless send a useful signal. Antithesis illustrates how funding can overshadow product understanding: one participant remembers a reported $100 million seed round led by Jane Street more readily than the product. Another, disclosing an investment, describes probabilistic simulation testing and names Palantir as a user. The first participant concedes that Antithesis may be a poor example of an unsubstantiated business; a deeper conversation with General Intuition likewise made its reported $100 million raise understandable. Sometimes the missing piece is a clear explanation of the business.

Some companies also have genuine capital requirements at this scale. Building a wet lab for high-throughput biology can be expensive and essential to the product. The criticism is of raises detached from milestones, especially when a large valuation becomes a recruiting tool and candidates are sold on the nominal dollar value of their equity.

For engineers evaluating offers, private valuation is not demonstrated exit value. A hypothetical agreement to value Latent Space at $5 billion makes the distinction: a negotiated price does not establish what an eventual buyer will pay. Another participant mentions employee equity buyback promises, without naming companies or terms. The investor warns that a company can spend its funding, sell for less than it raised, and leave employees with nothing. Potential upside and conviction in the vision are stronger job-selection criteria than unicorn status alone.

World models need a defined use case

The host introduces world models as a strong bet, despite skepticism among a researcher group they meet at NeurIPS. The response identifies a missing prerequisite: an agreed definition. Video models have visible applications such as editing; autonomous driving and possibly coding enter under broader definitions. Evaluating the opportunity requires specifying what the system does and which task it serves.

The consequential research question is transfer across settings. A model developed for video-game generation might not generalize to factories or robotics. That is an unresolved research problem, not necessarily a permanent barrier. The host points to evidence discussed in a General Intuition episode, but offers no experimental details that settle the question here.

Memory must earn retention

Memory management and continual learning enter through a concrete product problem: fast growth can coexist with weak retention. If coding-tool users switch whenever a competitor releases a new feature, personalization could give them a reason to stay. Looking ahead to 2026, the discussion places this opportunity in both consumer products and enterprise coding assistants.

The participants find rule-based memory such as Cursor rules inadequate and judge implementations including ChatGPT’s harshly. Beyond those product judgments lies a debugging problem: behavior depends on the core product plus accumulated user memory. When something goes wrong, the team must determine which caused it, while a frustrated user may have little patience for the distinction.

The counterargument is pragmatic: users already tolerate products with little or no memory, so imperfect improvements can still help. As AI’s novelty fades, familiar acquisition and retention practices become more important. Personalization is one of them.

Continual learning, as envisioned here, goes beyond remembering facts and preferences. A system would acquire skills through interactions and adapt as languages, frameworks, and repositories change. Implementing that adaptation through weight updates creates a serving challenge: weights become mutable state. Infrastructure must load, unload, and cache weights that change through learning, rather than simply serve a fixed set. Personalization therefore creates both a product opportunity and a systems problem.

RL environments: clones or real feedback?

The next debate concerns reinforcement learning (RL) environments: the applications, tasks, and evaluation rubrics used for agent training. One speaker predicts that the category is a fad, explicitly accepting the possibility of being wrong. Another counters that labs pay seven- or eight-figure sums for environments they could build internally. The skeptic recalls similarly large spending on poor data annotation: present demand does not establish lasting value.

The proposed alternative is to learn from real applications and their activity. Why buy a DoorDash clone if DoorDash’s own logs and traces are available? The counterpoint is practical: a separate environment can support faster parallel rollouts—running multiple attempts at once. The tradeoff remains unresolved.

Cursor’s feedback loop supplies a concrete example. A participant describes real platform activity being used to improve coding agents and tab completion, without quantifying the improvement. But real applications do not remove environment-design work: teams still need to choose tasks and define rubrics for judging results. The narrower objection is to treating an application clone itself as the valuable product.

The application that makes research necessary

Advances in RAG are argued to benefit applications especially strongly when retrieval enables a critical product function. Harvey is cited as an application company hiring strong researchers to improve that capability. Sierra offers a parallel case: solving difficult rule-following problems can enable better customer support. The speaker sees that focus as a contributor to Sierra’s success.

For memory and continual learning, the test is equally concrete: does the implementation make the experience dramatically better, or enable something previously impossible? The question can also run in reverse. Start with a desired user outcome, then identify the hard technical problem blocking it. Runway is offered as an example, with the speaker speculating that it would not have built models unless its product required them. The favored startup delivers a useful experience by solving the research and engineering problems necessary to make it possible.