YC's new Requests for Startups points to bottlenecks AI has yet to solve
Hardware, data, financing and specialized knowledge are gaining attention. For Brazilian founders, the question is which problem is already within reach.
Access to a good AI model does not, on its own, solve the challenge of running a company. That is the reading that connects the 11 areas in the Winter 2027 edition of Requests for Startups, Y Combinator's list of problems it would like to see founders tackle. The opportunity may lie in equipment, reliable data, financing or knowledge of an industry. For people building in Brazil, the list's value lies in finding a hypothesis they can investigate with real customers.
What YC is looking for
The official Winter 2027 list, consulted on October 29, 2026, includes these areas. The table summarizes YC's stated interests without turning its ideas into market forecasts or investment recommendations.
If needed, swipe to see the whole table.
| Area | Proposed problem |
|---|---|
| Abundant access to space | Businesses enabled by more affordable launches. |
| AI-augmented professionals | People supported by AI in specialized work. |
| Robots made in the US | Locally produced robots and components. |
| Models that write better | Quality of language, voice and style. |
| Integration of government benefits | Verification and reuse of data with consent. |
| Financing physical infrastructure | Credit and financial instruments for equipment and computing capacity. |
| Industry-specific models | Models designed for specific domains. |
| AI for city governments | Municipal processes and administrative work. |
| Data center supply chains | Equipment and supplies for infrastructure. |
| AI in isolated environments | Operation without an internet connection within the customer's infrastructure. |
| Owning your intelligence | Companies adapting and improving models. |
A list of ideas cannot replace a customer
In its article about the 2024 RFS, YC explains that the categories are a starting point and do not represent the only ideas it accepts. This context helps frame the current edition: there is interest in investigating these areas, not a promise of selection or a check for any company that includes them in its pitch.
A founder's first decision should be to choose a problem they know well. Access to users, an understanding of how purchases happen and the ability to recognize a recurring failure can be more useful than adapting a pitch to the latest category. A hypothesis suggested by YC still needs someone willing to test and pay for the solution.
First reading: the opportunity may lie in what enables AI to work
In the editorial team's reading, some areas shift attention from the end application to the conditions it needs to operate. A service can have a convincing demonstration and still be unviable if equipment takes too long to arrive, capacity costs too much or installation requires a process nobody has mastered. Investigating these obstacles calls for a different conversation from a presentation of features.
To explore a hypothesis like this, start by asking where a project gets stuck. Is a component missing? Can the customer finance the deployment? Does maintenance cost more than the expected benefit? The first product could be a narrow solution to that bottleneck, before attempting to serve the entire chain. YC's interest in a category does not eliminate the industrial and commercial risks of pursuing it.
It is also useful to distinguish contexts. An idea tied to manufacturing in the United States responds to conditions in that market. Repeating the same proposal in Brazil without checking local buyers, suppliers and requirements could produce a business with no fit. The useful question is which part of the problem exists for a customer the team can reach.
Second reading: knowing the work may be worth more than adding a chat interface
Another way to organize the list is to look at tasks that require specific knowledge. A test does not need to start with the team's own model. It can start with a bounded activity, a set of information the team is allowed to use and a clear criterion for checking the answer. The founder gets a testable question: does the solution help someone do this work better?
This requires care with the commercial promise. Supporting a professional does not mean replacing a required qualification, removing responsibility or allowing unsupervised decisions. In a pilot, define who can use the tool, which cases are excluded and when a person needs to review the result. The greater the consequences of an error, the more important it is to make that boundary explicit.
A hypothetical example would be organizing document checks within an operation. Before announcing full automation, the team could measure which inconsistencies the tool finds, which it misses and how much review work remains. This learning helps determine whether there is a product or just an interesting demonstration.
Third reading: the company's knowledge needs to become a verifiable asset
The area YC calls Own Your Intelligence proposes tools that let companies adapt and improve models with their own knowledge. The idea connects workflows, training data and evaluations. It is a proposal from the authors, not proof that every company should train or host a model.
For the editorial team, the question before choosing an architecture is what the company knows and can verify. Are there good and bad examples of the task? Can the data be used for this purpose? Who confirms that an answer is correct? Without this, changing models may change the technology without improving the process.
The same reasoning applies to language. More fluent writing can still be wrong, generic or distant from the brand. To assess a writing tool, use real texts, verifiable facts and criteria for usefulness to the reader. A preference for a particular style needs to be tested separately from factual accuracy.
How to turn one of these areas into an investigation in Brazil
Choose a concrete task and describe who struggles with it, how often and how they handle it today. Then talk to the people doing the work and those who authorize the purchase. These people may assess the same solution differently. Also record why the current alternative remains in use.
Next, propose a small pilot. Define which data will be available, what result would be useful and what would make the team stop the test. Compare quality, time, cost and the need for review against the previous baseline. Do not change several parts of the process at once if that prevents you from understanding what improved.
Only then ask what sustains the business: access to customers, accumulated knowledge, integration, operations or the ability to solve a problem that is difficult to copy. YC's list can open the conversation. The decision to build needs to rest on what the founder can learn and verify beyond it.
Frequently asked questions
What is Requests for Startups?
It is the initiative through which Y Combinator presents problems and ideas it would like to see founders investigate. The edition consulted is Winter 2027.
Does building one of these ideas guarantee investment from YC?
No. The list expresses interest in the areas without replacing selection, execution or validation with customers.
How can the list be used in the Brazilian context?
Choose a problem the team can access, check its local conditions and test a bounded task before expanding the proposal.