Compute Independence Will Determine Whether AI Becomes Economic Capacity

Sarah Guo thinks the defining question is not merely which labs build the strongest models. It is whether the United States can build the energy, compute, supply chains, and deployment paths needed to turn frontier capability into broad economic capacity.

That is a consequential distinction because Guo does not believe American competitiveness is inevitable. She argues that automation will be necessary if the United States wants a more resilient industrial base while labor remains expensive and people reasonably reject low-paid, dehumanizing work. But automation requires compute, and compute is an industrial system: GPUs, fabs, power, cooling, data-center construction, labor, raw materials, financing, and global supply chains.

“If we don't have it,” Guo says of compute, “we are naturally not competitive or we're at least like not independent.”

Her concern is not that the country lacks technical talent, entrepreneurial capacity, or capital. It is that physical construction and political alignment move more slowly than software. Building data centers requires communities to accept them. Making nuclear power competitive as baseload generation requires both public confidence in safety and enough construction for costs to decline. The relevant bottleneck is accumulated capability: trained labor, tacit knowledge, manufacturing capacity, materials, and the ability to execute repeatedly.

A hyperscaler infrastructure leader told Guo that nothing likely to move the needle at sufficient scale would arrive before 2030. She found that depressing precisely because the underlying technologies are not beyond reach. The problem, in her account, is that the physical supply chain cannot advance at the speed of software or executive decision-making.

Guo calls the strategic objective “compute independence,” analogous to energy independence. Working backward from a data center full of GPUs used for training and inference reveals dependence on a global supply chain with exceptionally narrow links. Some of those links sit in places that may not be reliably stable or accessible to the United States and its allies.

Compute independence does not mean every component must be produced domestically. Guo explicitly accepts comparative advantage. The goal is redundancy: having more than one source for critical inputs. She cites PAC Silica, an effort associated with Jacob Helberg, as an example of working through supply-chain layers and identifying alternative capacity paths.

Conviction has invested in the labor gap for data centers and robotics, nuclear energy, and alternative chip architectures. It has also examined data-center builders and solar-and-battery installers, though Guo says the firm has not made an investment in that category. The reason is not hostility to operational businesses. Data-center development is simply an unusual mix of operations, technology, real estate, and—most importantly—financing, while Guo remains fundamentally drawn to durable technology products and assets.

The danger, as she sees it, is political as much as industrial. Fear of job displacement, resentment toward rents captured by a small number of technology firms, and a broader anti-capitalist orientation could slow the buildout of energy, infrastructure, and industrial capacity. That would leave the country not simply compute-constrained but dependent.

Patrick O'Shaughnessy characterizes the relevant frontier community as roughly 250 researchers and entrepreneurs doing the most interesting work in AI. Guo agrees with the general description, but not as a formal census or closed network. Her focus is on staying close to the people whose research, companies, and operating decisions are advancing the field.

Guo says the landscape has become “violently competitive” and globally consequential. A belief that has become more common among researchers in the past year is that recursive self-improvement—models improving the research process that produces better models—could put AI one or two years from some kind of exponential intelligence.

She is careful to present this as a belief, not a settled forecast. Guo notes that Andrej Karpathy has self-consciously said he has thought such a moment was two years away for roughly a decade, and thinks it again now. Her point is not that the date is known. It is that the possibility is now psychologically salient among people doing frontier work.

The size and capital requirements of major labs have also changed how researchers experience their own contribution. At a 200-person lab with relatively few researchers, an individual might plausibly see the next breakthrough as partly dependent on their work. If the perceived path forward requires hundreds of billions of dollars in compute and thousands of employees, the locus of agency appears to shift.

Guo says a substantial group of researchers now feel one of two disempowering things: either their work will not matter because models will eventually do it, or compute scale is the only thing that matters. Neither view sits comfortably with people who want to make an important contribution.

Her answer is not that people have ceased to matter. Guo believes in a “great man and great woman” theory of history: high-agency people, given the right risk capital, network, talent, and environment, can change outcomes. She works with, co-invests alongside, and has friends at both large and small labs. Her concern is not competition between them. It is the prospect of an economy effectively consumed by the owners of one, two, or three frontier models.

That belief informs her interest in a competitive Western ecosystem: one able to build open models, support companies outside the largest labs, and develop the industrial base required to compete at the frontier.

Sarah Guo begins the open-model debate with what she considers an already-established reality: increasingly competitive open-source models have emerged over the past three years, particularly from China but also from the United States and Europe. Powerful Western open models already exist, she says, and open capability is in use broadly enough that “the cat is out of the bag.”

Her argument for open models is not simply ideological. It is operational. In many settings, using a frontier provider’s model is too expensive, too slow, or too sensitive from a data perspective. Those constraints should become more significant as companies find more ways to use AI. Businesses want control over their economics, their capacity, and their own technological destiny.

Open models give companies and developers room to post-train, build task-specific harnesses, and tailor systems around particular workflows. Guo sees that as necessary because an economy contains far more use cases than researchers at a small number of frontier labs can anticipate. The relevant question is not whether a lab can imagine every application. It cannot. The question is whether firms and individuals can adapt capable systems to the variety of work they actually face.

Guo expects broad access to intelligence that eventually becomes “too cheap to meter,” borrowing Sam Altman’s phrase. Open models, in her view, will be part of how that happens. But she does not treat openness as an answer to safety problems.

A model capable of defensive cybersecurity work or beneficial biology research may also enable related offensive cybersecurity or biological misuse. That safety profile has to be understood. Guo’s position is that frontier systems should be tested and that concerns about possible backdoor behavior in Chinese-developed models should receive rigorous empirical investigation rather than speculative discussion alone.

The work is not trivial, she says, but it can be done. Her objection is to treating domestic restrictions on open-model use as an effective substitute for it. Restricting those models inside the United States would principally constrain law-abiding American firms, she argues. Adversarial users would not be deterred by rules they already intend to violate, while domestic companies would face slower deployment, fewer permitted uses, and potentially less economic value.

The tension is therefore not simply open versus closed. Guo wants direct testing and an understanding of model risks, alongside a deployment environment that does not handicap domestic users without materially limiting malicious ones.

Sarah Guo sees robotics as an example of how model progress collides with the stubborn constraints of the physical world. The core challenge is not merely producing a more capable robot. It is obtaining enough relevant data for systems to generalize reliably across real environments and tasks.

She points to Tony Zhao and Cheng Chi, founders of Sunday Robotics, as researchers who approached that problem directly. The pair had worked at Toyota Research, DeepMind, and Tesla before starting the company while they were PhD students at Stanford. Guo says that after studying their work, she concluded they had collectively contributed an unusual share of the important ideas in robotics AI over the preceding four years.

A broad belief in robotics, as Guo describes it, is that general-purpose robots would be much closer if the field had an internet-scale corpus of robotics data. The practical issue is that physical-world data is costly to collect, environments vary, and a company cannot afford to capture every task and every edge case through brute force.

What impressed Guo about Zhao and Chi was their effort to treat collection itself as a technical problem. Their questions include how to collect data as cheaply as possible, how the data should be structured, whether it represents the distribution of real-world tasks and environments, how collection interacts with model training, and how far lower-cost data can transfer into useful learning.

Guo says Sunday Robotics moved from early work in a Stanford basement to a full-stack system manufactured in the United States in just under two years. The team had completed hundreds of iterations spanning hardware, model development, data collection, task design, and testing in real-world environments.

She emphasizes that “nothing is true until it is shipped.” Still, the company believed general semi-humanoid robots could enter home beta deployments by the end of the year, or the following year if not. Guo finds the speed of that progression striking. In her account, a meaningful part of robotics has shifted from asking whether broadly capable robots will work to asking when—though the remaining work is still constrained by manufacturing, data, deployment, and real-world reliability.

Sarah Guo describes Conviction’s early advantage as relatively simple: a massive technological transition was underway, early-stage venture was not yet maximally competitive, and investors willing to understand both the technology and the community from first principles could develop better access and make better decisions. The hard part was execution—sustained focus, effort, and a high bar for the people with whom the firm worked.

That approach led Conviction to look for application markets based partly on what models could newly do, rather than only by working backward from an established customer problem. Guo does not reject customer-driven thinking; she wants both. But she believes model capability can reveal valuable workflows before ordinary demand signals make them obvious.

Harvey, the legal AI company founded by Winston Weinberg and Gabe Pereyra, was an early example. In late 2022, Guo saw law as a plausible fit for language models because legal work involves reading large document sets, retrieving relevant precedent and internal knowledge, and generating structured text. The technical fit mattered, but so did the founders’ ambition: a progression from answering a narrow question about a California landlord-tenant agreement to doing a substantial share of the work in a transaction as complex as the Activision Blizzard acquisition.

That progression was uncertain in timing and scope. What gave Guo confidence was the combination of a technological premise, a view of where value would sit if the premise held, and founders who genuinely believed in the more expansive future.

Her decision process starts from an instinctive judgment but is designed to expose that judgment to failure. If she knows a founder’s prior work, understands enough of the relevant domain, and finds the idea coherent, she may quickly place the opportunity at an eight or nine on a ten-point internal scale. The work afterward is to find what could make that judgment wrong: gaps in her technical understanding, misunderstood market premises, opposing views from informed people, or mistaken assumptions about the founders themselves.

She writes memos, gets second reads from partners and trusted outsiders, and tries to state clearly what evidence would change her mind. In Conviction’s early days, she would send full memos to investors and operators such as John Lilly or Dylan Field for outside criticism.

The distinctive principle is not that conviction excludes doubt. It is that one person must own the decision after taking in other people’s information. Guo finds it difficult to understand an investment model with no individual ownership, because somebody ultimately has to decide whether they believe.

What she considers dangerous is making large technology or research bets mainly on pedigree, references, referral sources, or the identity of other investors. Those signals can be useful. But they do not substitute for an investor’s own intuition about the technical theory and the business that could emerge from it.

Exceptional founders are not separable from their ideas in Guo’s view. When she says someone has extraordinary judgment, she means that person has repeatedly been right about how an industry works and has made correct decisions even when others disagreed. If an investor cannot understand what that person is trying to build, Guo argues, the investor cannot really assess that judgment; they can only admire the reputation.

Sarah Guo says Conviction’s internal debates often begin with a question that old venture heuristics cannot answer: is this market newly capable of supporting a venture-backed business?

The firm’s podcast is called No Priors because Guo thinks some lessons inherited from earlier technology cycles no longer apply cleanly. Semiconductors are a central example. Venture-backed semiconductor companies were, in her words, a “godawful business” for a long time. But concentrated, at-scale demand for accelerators and growing buyer interest in supply-chain independence can change the equation. If major customers do not want to depend on a single line at TSMC, alternative capacity becomes more valuable.

The same debate appears in solar and batteries, nuclear, turbine manufacturing, robotics, and biology. These are not the software markets venture investors most favored a decade ago. The issue is not whether the technological direction is interesting. Conviction often begins with broad agreement about that. The issue is whether the market structure can produce venture-scale returns, or whether the distribution of outcomes is sufficiently attractive even when the median path is difficult.

Biology has moved Guo strongly toward the affirmative. She believes models can create and capture substantial value in the field, including through AI software businesses, not only through conventional drug-development companies. Conviction was the first check into Chai Discovery, which Guo says is working with several top-10 pharmaceutical companies on significant efforts to accelerate parts of their research-and-development processes.

The conventional view when Conviction invested, according to Guo, was that the only dependable way to make money in biotech or pharma was to make drugs: advance candidates through a costly and risky process, build milestone-based deals, and decide how much clinical risk to take. Software investors, by contrast, were often told they could not build enduring platform businesses for pharma.

Conviction invested before that question was resolved. Guo says a concrete sign of value can be as simple as a $10 million contract, paired with conversations with the scientists and operating leaders using the product.

The customer’s view matters because the customer can see whether a tool changes their work. Guo expects a clearer industry inflection when an AI-enabled process demonstrably changes the trajectory of, or produces, a new drug or indication. That could trigger a much larger investment wave. But she already expects major acceleration in cures, while recognizing that regulation, safety, and physical-world timelines remain constraints models cannot simply wish away.

Sarah Guo calls the name Conviction aspirational. Her ideal version of early-stage investing is simple: invest early, hold a meaningful position, do not sell it, and work with the company until it succeeds, is acquired, or fails.

That model is built for companies whose value may take time to become legible. Guo cites Figma, Notion, Rippling, and Base10 as examples from her experience of businesses that were not obvious at the beginning or took years to find the thing that worked. An investor would not choose for a company to need four or five years to find its footing. But those experiences support taking a view that the market does not yet share and remaining a committed partner until it is validated or disproved.

Guo does not frame this as contrarianism for its own sake. The task is to decide what is true without allowing dominant narratives or powerful ecosystem players to dictate the answer. If an investor finds a truth the market has mispriced and can hold it before consensus forms, that creates opportunity.

The best source of that information, in Guo’s view, is people making the future: founders producing behavior she did not expect, researchers advancing an unfamiliar direction, companies building concrete AI plans, and customers revealing whether a product is valuable.

She offers Suno, the AI music-generation company led by Mikey Shulman, as a case where her own intuition was wrong. Guo had passed because she doubted enough people wanted to make music and was uncertain about the scale of creation and consumption. Talking with Shulman and observing the business changed her view. She had underestimated the demand for expression, entertainment, and creation across AI products.

That lesson is more useful to her than a sweeping theory about which layer of the AI stack will win. Frontier labs may pursue AGI or ASI while prioritizing nearer-term products such as chat, coding, advertising, richer interfaces, and more complex tasks. But Guo does not think a grand strategic theory of the labs resolves the prospects of any specific application company. The relevant work is assessing the actual scope and competitiveness of a particular lab effort against a particular business.

She expects the next year to bring more visible examples of the productivity effects already evident in software engineering. At one portfolio company, she says, a marketing function serving many customers and handling ordinary work such as sales enablement and content was effectively run by a person and a half after its marketing lead built what he called an autonomous marketing department.

Greater productivity does not necessarily mean people work less. O’Shaughnessy says AI has made him work more, and Guo says the same. Her expectation is that people will redirect time previously spent on mundane work into more activity—provided they have access to the tools and education to use them.