Every software startup now faces a new test: whether one of the big AI labs could ship it as a feature next quarter. For a growing number of products, the answer is yes.
In mature markets like the US, that is pushing the next serious competition toward the physical world. In startup language, it is the move from bits to atoms: from code to robots, power, factories and supply chains. The things a lab can't push out as an update.
Every release moves a market
In February 2026, Anthropic released plug-ins for its Claude Cowork agent, covering work in legal, sales, marketing and data analysis. Within days, Thomson Reuters, the owner of Westlaw, fell nearly 18% in one session, which Reuters reported as its biggest daily loss on record. RELX, the parent of LexisNexis, fell 14%. Wolters Kluwer fell about 13%. LegalZoom fell 19.7%.
In April, the same thing happened in another part of the market. Anthropic launched Claude Managed Agents, which bundles code execution, credentials and hosting. Akamai fell 16.6%, Cloudflare 13.5% and DigitalOcean 13.4% in one session.
Analysts disagreed on how much of this was justified. A plug-in that reviews contracts is not a legal research database, and some argued the market moved too fast. The stock moves themselves matter less than the belief behind them: a single release from a lab can now open up an entire software category.
For founders, that leads to a hard test. If a lab can ship your product as a feature, your product is a feature.
Why it keeps happening
The labs build the platform, and more and more of what sits on top of it. Most AI products run on their models. The labs can see what people build, they already have users in their own apps, and the base model improves every few months. Work that needed a startup last year can be a prompt this year.
They also hold an unusual share of the capital. PitchBook and NVCA counted $267.2B in US venture deals in Q1 2026. The top 5 deals, for OpenAI, Anthropic, xAI, Waymo and Databricks, took 73.2% of it. OpenAI's round alone was $122B.
The first post in this series made the point that a small team can now build working software with AI doing much of the work. The other side of that point is less comfortable. Everyone else can build it too, the labs included.
Software still has room
None of this means software is finished. The defensible part has become narrower, and it tends to sit in specific places.
Y Combinator's latest Requests for Startups gives a useful map. Even its software ideas sit where the labs are less likely to go. YC argues that models got worse at writing as they improved at code, and that the labs will keep prioritising code. It asks for AI that runs inside defense contractors, banks, hospitals and nuclear plants, where data can't leave the building. It asks for specialised models for manufacturing, energy and biology, for tools that let companies own their models instead of renting them, and for software that helps cities process permits.
The common thread is reach. The data is locked away, the site has no internet, the customer is regulated and slow to buy, or the work is simply not a priority for the labs. That is where software keeps its room.
The list points to atoms
The stronger signal is on the physical side. YC publishes this list before each batch, and the Winter 2027 version starts from the idea that the bottleneck in AI is no longer the models. 5 of its 11 requests are physical, or close to it.
Robotics comes first. The US installed 38,400 industrial robots in 2025, up 12%, and YC sees the bottleneck in robotics moving from the models to the hardware. YC wants founders building robots and the parts inside them: magnets, motors, gears, sensors, battery packs.
The data center supply chain is under similar strain. Large power transformers have a 3-year lead time. Switchgear and liquid cooling equipment are sold out. The steel for transformer cores comes from a single US plant that can't keep up.
The remaining requests cover space companies built for a world of 10,000 Starship launches a year, AI that runs on-site with no internet connection, and the financing needed to pay for all of it before it earns anything. One YC partner summed up the problem: "Software is cheap to start. Robots, GPU clusters, and data centers aren't."
Read as a whole, the list describes a world where intelligence is getting cheap and everything around it is scarce: parts, power, permits, capital and trust.
The money is already moving
The funding data points the same way. Crunchbase groups robotics, autonomous vehicles, aerospace, drones, industrial automation and sensors as physical AI. Startups in that group raised $47.4B across 521 deals in the first half of 2026, almost 4 times the $12B of the previous six months. Waymo's $16B round accounts for nearly a third of that total, so the headline is inflated. Even without it, the increase is large.
Robotics alone raised $18.8B by June, already above the $15B raised in all of 2025. At the earliest stage, Carta's 2025 data shows hardware and biotech as the second and third largest pre-seed industries by cash raised, up from third and fifth a year earlier.
Hard is the point
Investors have long been cautious about hardware, and with reason. It needs money before revenue. Cycles are long. There are suppliers, certifications and regulators to deal with. No prompt makes a transformer arrive faster.
Those same constraints are also what protect a company. A lab can release a legal assistant overnight. It cannot release a magnet factory, a trusted supplier network or a permit to build next to a substation. Cloud credits pay for compute, not for a production line.
The trade has shifted. Software became cheap to start and easy to copy. Physical businesses are still expensive to start and hard to copy. As AI makes the first part easier, the hard part becomes the moat.
A few checks help separate a company from a feature. Whether a lab could ship its core in one release. Whether it sits somewhere the labs can't easily reach, such as a factory floor, a closed network or a regulated workflow. Whether it improves with every deployment. And, if it needs real money before revenue, whether that has been planned for from the start. None of these is a rule, but together they say a lot.
Two clocks
The shift is not happening at the same speed everywhere.
In the US, software is mature. Most businesses already run on it, and AI is now absorbing much of that layer. That is why the frontier there is moving to robots, power and supply chains.
In Azerbaijan and much of the Caucasus, Central Asia and MENA, the software layer is still being built. Many businesses and public services are only part of the way through going digital, and the region is still developing the skills to build software well. Here, software is still a real need, not a solved problem.
So the region runs on a different clock. For now, much of the opportunity at home is still in software, especially where local knowledge matters: language, regulation, government, payments and the way business actually works in each market. Global labs are unlikely to build for those details first.
That window is real, but probably shorter than it looks. The tools the labs release are mostly global from day one. Local software that is only a thin layer on a model will face the same test as anywhere else. What lasts is software tied to local data, local distribution and local trust.
Over the next 5 years, the direction here is likely to follow the US, only later. As the software layer matures, competition will move toward the physical side as well. The region already has some of the inputs: energy, heavy industry and engineering talent.
The gaps are real as well. Hardware needs patient capital, industrial partners, testing facilities and financing that much of the region does not have yet. Local investors are used to small software rounds. Physical companies will need larger cheques, longer timelines and closer work with industry and governments. Those take years to build, which is a reason to start before the clock catches up.
The aim is not to copy YC's list today. It is to build the software the region still needs, and to prepare for the physical side while there is time.
The takeaway
AI has made software cheap to build and easy to copy, and markets have started to price that in. In mature markets like the US, the frontier is already moving to atoms: robots, energy, data centers, space and the capital behind them. YC is saying so openly, and the funding data agrees.
Our region is on a different clock. Software is still needed here, and local software with real local advantages has room to grow. But the direction is the same, and over the next 5 years competition is likely to move toward the physical side here too.
The companies best placed for that cycle are the ones whose core can't be shipped as an update.



