Hard-Tech Startups Rise as AI Lowers the Cost of Building

Diana Hu says YC’s recent accepted-company data shows a shift in both what startups are building and how quickly they can sell it. Hard-tech companies—businesses that “touch atoms and not just bits,” as Harj Taggar puts it—have risen from 8% to 20% of the batch over roughly the past year to 18 months. At the same time, the median YC company, which enters pre-product and pre-revenue, is reaching about $20,000 in monthly revenue by the end of the batch, up from roughly $8,000.

The connection between those trends is AI leverage. Jared Friedman argues that the case for hard tech is not merely that investors are moving away from software. More capable models can accelerate scientific research and reduce the software labor needed to build physical businesses. A company working across hardware, supply chains, and research still needs strong engineers, Garry Tan says, but it may no longer need to recruit engineering organizations at the scale of Google or Meta to build a sophisticated full-stack system.

At the same time, software products that take responsibility for an outcome rather than supporting an employee’s workflow can be worth more to customers. Hu says the share of accepted companies doing full-stack, end-to-end work on a task has risen from 10% to more than 25% of the batch. The agent does the work—such as clinical intake, insurance brokerage, or medical billing—rather than functioning as a point solution someone must operate.

The composition of YC’s batches reflects the research burden of physical technology. Friedman says one in six founders in the current summer batch has a PhD, substantially more than historically. Fields such as silicon photonics require serious research backgrounds, and he says the technically specialized founders YC has funded have performed disproportionately well.

Hu breaks the hard-tech increase into several categories: robotics has risen from around 1% of the batch to 6–7%; industrial manufacturing from about 4% to 10%; defense from 1.5% to 5%; compute infrastructure from roughly 1% to nearly 4%; and power infrastructure from around 1% to nearly 3%. Across these atom-focused categories, she says, batch representation has generally tripled or quintupled.

She identifies three demand sources: founders building across the space stack, a generation motivated to work on defense and dual-use products, and the physical buildout needed to supply AI compute. YC showed Exosat, which describes itself as building low-orbit satellites for sovereign direct-to-cell and broadband operators, and Beyond Reach Labs, which describes its work as solar infrastructure for orbital power. Those companies sit at the overlap of the space and compute buildouts Hu describes.

Diana Hu describes bringing data centers online as a physical process involving site construction, planning software, equipment deployment, power, batteries, and the underlying semiconductor stack.

Hu says even Nvidia A100 GPUs, now an older generation, have been appreciating in hourly cost because demand exceeds available compute. That shortage creates openings throughout the stack: alternative processors, power systems, data-center construction, and the networking equipment that connects chips.

YC displayed Lamb Labs’ description of its model processing units: processors that hardcode a model and its weights into silicon, making “the model” the chip. The company says this approach targets the memory-bandwidth constraints of GPUs. Hu also points to Baud’s attempt to build custom hardware using ternary model representations. Her premise is that large-language-model workloads do not necessarily require full floating-point precision; Nvidia architectures have moved from FP32 toward lower-precision formats including FP16 and FP8. Friedman adds that even FP2 is “somewhat usable.”

The bottleneck can be between GPUs rather than inside them. Jared Friedman describes data-center switches as routing systems that let GPUs communicate. GPU speed has continued to improve, while electronic switches can constrain workloads. Dipole Labs, displayed during the discussion, describes itself as building photonic hardware for AI and quantum infrastructure. Friedman says its product is a fully optical switch intended to carry information from one GPU to another as photons rather than through conventional electronic switching.

This buildout also creates demand for basic industrial inputs delivered at an unusual pace. Friedman describes Nox Metals as an effort to rebuild domestic metal-manufacturing capacity in Detroit. Its site advertises aluminum plate and bar cut to size with instant pricing and delivery options online; the company’s displayed announcement said it had raised an $11.5 million seed round to “Reindustrialize America.” Friedman says its customers include defense-technology startups that need material quickly and find established suppliers unable to keep up.

He compares the dynamic to the early web era, when startups preferred newer vendors such as Stripe because those vendors worked at startup speed. Nox Metals, in his account, could become that kind of supplier for the defense-technology ecosystem.

Tan makes a similar argument about defense products themselves. He cites Icarus, which he describes as building a solar-powered high-altitude aircraft for surveillance and communications, and says the company has reached seven-figure contracts with the Department of War. He also cites 9 Mothers, whose displayed site identifies its first product, Edda, as a system designed to stop high-speed FPV suicide drones. Tan characterizes it as a computer-vision-enabled shotgun turret intended to protect Special Forces from commodity drone attacks.

Tan argues that startups can use AI and newer development methods to build capabilities that established defense primes, organized around cost-plus contracting, struggle to produce. Friedman adds an important constraint: capital-intensive companies still need downstream investors and cannot generally be bootstrapped as SaaS companies can. The claim is not that hard tech has become cheap, but that smaller teams can now undertake more of the technical work and face more immediate demand for what they build.

The software implication is not that SaaS is dead. It is that systems which merely store, display, or organize work face pressure to become systems in which agents actually perform it.

Harj Taggar argues that software remains valuable when agents can use it. Agents may use software more intensively than people do, which gives systems of record a way to retain their importance. But owning data alone may not be sufficient.

Tan frames the choice starkly. A system of record can expose its data through MCPA protocol that lets AI systems connect to external tools and data sources. and risk making that data portable enough that switching becomes easy. Or it can become a “harness”: the place where people and agents not only read and write records but actually do their work. He sees Slack as potentially advantaged because it already sits inside how many AI-forward companies collaborate, giving it workflow position and collaboration data.

The discussion used an OpenAI post on ARC-AGI-3 as evidence that the surrounding environment can materially change a model’s output. The post said GPT-4o struggled on the benchmark because the official harness did not let it remember what it had learned. OpenAI said that enabling two API settings tripled its score while using six times fewer output tokens: the displayed comparison showed a score of 28 under the official ARC harness and 177 under the Responses API harness.

Taggar’s conclusion is that the relevant unit is not the model alone, but the model plus the harness. Tan adds that, with a custom harness, Astra reportedly reached north of 90% on ARC-AGI-3 after performance had been in the low double digits a few months earlier.

Hu says the faster revenue figures are concentrated among startups designed around that fuller responsibility for a workflow. Friedman acknowledges that some early enterprise spending may reflect AI hype. His stronger explanation is that a product which completes a job can command more value than a system that tracks the job without doing it. In companies he works with, he says, that value proposition can be strong enough for large enterprises to write substantial checks very early.

Taggar uses Juicebox to distinguish the two models. The company began as an LLM-powered recruiting search tool: recruiters could describe a candidate profile and receive relevant people to contact. Its newer agent product also contacts candidates and may eventually schedule interviews and handle further steps. Taggar expects that agents could double or triple Juicebox’s revenue per account as customers use more of them.

He does not treat that as the elimination of recruiters. Reaching out to hundreds of candidates is rote work that recruiters may be glad to hand off, he says, while judging culture fit and handling the human side of hiring remain more difficult and distinctive tasks. The product changes the allocation of work—and what customers will pay for it—rather than simply replacing the recruiter.

Jared Friedman says companies selling training data and reinforcement-learning environments to model labs have become a significant but comparatively quiet category. When YC funded Scale in 2016, he says, the field was a small niche. In the past two years, YC has funded more than a dozen companies that each make more than $10 million annually selling data or RL environments to labs. Tan says some are making hundreds of millions of dollars despite being only a few years old.

The category is quiet partly because companies doing well may not want to advertise it, Friedman says. But Tan’s reasoning is that data is one leg of the scaling law: more compute alone does not produce better models without material on which to train them. RL environments provide structured, specialized settings in which models can learn and be evaluated. Finance is one example Tan gives of a domain where a company can build deeply customized environments.

Hu says major labs are reportedly spending about $1 billion on the category, including long-horizon RL tasks. She sees a related market forming around robotics, where labs need egocentric data, teleoperation, and real-world task environments to make AI operate in the physical world.

The Human Archive site displayed during the discussion describes the scale that such collection can require: custom rigs, gloves, and wrist cameras; more than 100 full-time staff; data from more than 100,000 contributors and 500 industry partners; and work across more than 30 countries. Hu also names Praxis Robotics and Deep Reach among companies operating in the emerging physical-world data category.

For Hu, robotics differs from language modeling because it must represent and act in three-dimensional physical space, with more degrees of freedom and much tighter timing requirements. She explicitly presents the resulting claim as a hypothesis, not a settled result: general robotics foundation models may be useful starting points, but deployments in particular verticals may work better with models trained on custom data from their operating environment.

Diana Hu gives the example of Boost Robotics, which builds data-center robots for inspection and maintenance, including cabling work. A robot in that setting must respond in real time to changes around it; unlike an LLM, it cannot simply run in the background and return later with an answer. A disruption at the wrong moment could mean connecting the wrong cable to the wrong port.

Friedman says that, to his understanding, YC companies deploying Physical Intelligence’s Pi models fine-tune them rather than using them out of the box. Ultra, a warehouse-robotics company, starts with Pi but has thousands of hours of footage of putting items into boxes. That task-specific material, he says, makes its system good at that specific task.

Tan expects companies with distinctive proprietary data to increasingly train or fine-tune models for work that generic frontier systems do not perform well enough. He points to coding transcripts as one potential feedback loop: a system could identify strong coding behavior in its own records and use it to train a better coding model. His expectation is not that every company will train a model, but that tools for doing so will become more consequential as open-weight systems approach frontier performance.

Diana Hu says solo founders represented about 5% of YC’s accepted companies a year ago and now account for 18–19%, the largest spike YC has seen.

Tan’s explanation is that startups traditionally required a difficult combination of selling, persuasion, recruiting, and world-class technical execution. Co-founders made it easier to cover those functions. Now, he argues, knowing what to prompt and what to build is becoming more important than personally writing every line of code.

Taggar says YC has always had successful founders who began alone, citing Instacart, Brian Armstrong at Coinbase, and Parker Conrad. The bar was exceptionally high because one person needed to conceive of the company, sell it, and build it. AI lowers the building threshold enough for more people to get started.

But the speakers do not treat that as a case against co-founders. Tan says strong co-founders remain valuable and improve the odds of success. Taggar expects many companies that begin with one founder to add co-founders after gaining traction, rather than starting with a conventional 50/50 partnership. Friedman says he has also observed more people adding co-founders later in the company lifecycle.

The related trend is a resurgence of experienced founders. Garry Tan points to Peter Steinberger, a builder in his early 40s who had managed developers, worked at startups, adopted AI tools early, and tried many ideas before finding what to build. In Tan’s view, experience provides taste: an ability to recognize problems, understand where they fail, and form stronger opinions about which opportunities matter.

Friedman adds that managing coding agents may reward some of the same skills as managing engineering teams. Builders who have delegated work, supervised output, and organized people may be able to run large collections of coding agents more effectively than someone with less operating experience.

Tan’s practical advice is to begin using the tools. New models can suddenly solve bugs and unresolved problems that were beyond reach a month earlier, he says. He does not know whether that pace will last 18, 24, or 36 months. His point is narrower: founders who can identify a real problem and use the available leverage have more ability to get started than they did before.