TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
YC's Latest AI Startup List: Rare Earth Magnets, Offshore Computing... Deeply Aligned with U.S. National Will

YC's Latest AI Startup List: Rare Earth Magnets, Offshore Computing... Deeply Aligned with U.S. National Will

2026.11.18
Share
YC's latest two startup lists show that the AI battlefield is rapidly shifting from cloud models to the infrastructure and bottleneck layers of the physical world.

Written by: Xiaobing

In the summer of 2026, YC released a 15-item "Request for Startups" list covering lunar manufacturing, inference chips, and anti-drone defense, laying almost the entire hard-tech spectrum on the table.

Three months later, the Fall 2026 batch list was narrowed to 13 items.

Three months after that, the Winter 2027 batch came down to 11.

Each round of trimming served to distill the focus.

Looking at the four RFS lists this year (YC has switched to four batches per year), a clear evolutionary path for the investment thesis emerges: the battlefield for AI is rapidly shifting from cloud models to the infrastructure and bottleneck layers of the physical world.

Below is a complete breakdown of the two latest lists: Fall 2026 (13 items) and Winter 2027 (11 items).

A Letter from the Secretary of the Army

The most eye-opening item on the Fall 2026 list has nothing to do with AI technology: a handwritten submission from the current U.S. Secretary of the Army, Daniel P. Driscoll.

This marks the first time in YC history that a sitting high-ranking government official signed directly on an RFS list.

Titled "The Future of American Defense," it features someone from the Pentagon addressing the call for applications from a startup accelerator with a direct plea: We need you.

It signals that the U.S. military is bypassing the traditional defense contractor ecosystem (the comfort zone of companies like Lockheed Martin) and directly inviting builders into Silicon Valley garages.

Consequently, YC's role has subtly shifted: it is becoming the interface connecting entrepreneurs with national will.

Looking at the complete Fall 2026 list, Driscoll's appearance is far from an isolated incident.

Maritime compute, AI-native compliance infrastructure, operating systems for the physical world... the underlying tones of military and national security permeate multiple directions.

Fall 2026: A Three-Layer Structure Across 13 Directions

Spreading out the 13 items of the Fall 2026 list reveals a clear three-layer architecture.

The Physical Layer: AI Needs Hardware, Hardware Needs New Ground

Francois Chaubard's proposal for "Compute at Sea" is the boldest vision in this layer.

Building data centers on land is getting increasingly difficult: power approval, environmental compliance, land costs—each one slows down the expansion of AI compute capacity.

Moving modular compute capacity to sea to bypass land bottlenecks, YC alum company Atomarine is already building nuclear-powered compute barges.

Charlie Warren's "New Operating Systems for the Physical World" targets another pain point: AI can write code and chat, but still cannot directly control physical devices.

Factory robots, farm irrigation systems, and city traffic lights all require an AI-native operating system for scheduling.

"Data for the Real World" points to an even more fundamental issue: internet data for training AI is nearly exhausted, and the next wave of data must be collected from the physical world.

The data generated by sensors, cameras, and industrial equipment is the fuel for AI's next phase.

The Trust Layer: The Stronger AI Gets, The More Critical "Who You Are" Becomes

"Proving You’re Human" addresses a rapidly deteriorating problem: as AI-generated content and behavior become increasingly indistinguishable from reality, identity verification will become essential infrastructure for the entire internet.

"AI-Native Compliance Infrastructure" targets the enterprise side.

AI is starting to make decisions—who audits them? Who is liable when things go wrong? Traditional compliance frameworks were simply never designed for the AI era.

The logic of this layer can be condensed into one sentence: the greater the capability, the more urgent the need for trust infrastructure.

The Application Layer: From Tools to Services

Andrew Miklas's "The Primer" pays homage to the AI tutor in Neal Stephenson's sci-fi novel *The Diamond Age*, with the goal of providing every child with a personalized AI mentor.

Raphael Schaad's "AI-Powered Consumer Products for 1 Billion People" locks its target on consumer goods for a global audience of a billion users.

Max Kolysh's proposal for "AI for the Aging Population" touches upon a market rarely taken seriously by the tech circle: global aging.

Demographic structures in Japan, South Korea, and China are rapidly collapsing. AI caregiving, AI companionship, and AI health management represent a massive market with extremely high certainty.

Pete Koomen's "A Cloud for Small Software" and Aaron Epstein's "Multiplayer AI" return to developer tools.

The former aims to provide dedicated cloud infrastructure for small-scale apps in the AI era, while the latter focuses on AI experiences in multi-user collaboration scenarios.

Harsha Gaddipati's "Self-Maintaining APIs" proposes using AI to have APIs monitor and repair themselves, freeing up developer productivity.

The Best Time to Build in Crypto Is Now

This quote comes from YC Partner-at-Large Nemil Dalal.

He wrote a dedicated Crypto direction in the Fall 2026 RFS, titled "The Best Time to Build in Crypto."

Dalal's confidence stems primarily from stablecoins.

Global stablecoin settlement volume is closing in on Visa's. The progression of the GENIUS Act and the CLARITY Act means the U.S. regulatory framework is shifting from crackdowns to integration.

Stablecoins have found their PMF and are becoming part of global financial infrastructure.

He specifically highlighted a concept: Agentic Commerce.

As AI Agents begin replacing humans to handle purchasing and transactions, they won't apply for bank accounts or credit cards.

On-chain payments are the most natural settlement method for AI Agents.

This cross-narrative (AI × Crypto) is emerging as a shared frontier for both industries.

The wave of institutionalization is also accelerating.

ETFs, custody, compliant trading platforms... traditional financial institutions' interest in Crypto assets has moved from exploration to strategic allocation.

More interestingly, YC itself is taking action.

In February 2026, YC partnered with Coinbase to begin funding all YC startups using USDC stablecoins across Ethereum, Base, and Solana, with transaction costs under $0.01 and settlement times measured in seconds.

Note: This option is available to all YC startups, not limited to Crypto ventures.

On October 28, Dalal posted an even more detailed "Crypto Request for Startups" on X, further breaking down specific tracks into sub-directions like stablecoin infrastructure, on-chain trading tools, and institutional DeFi products.

Winter 2027: The Bottleneck Isn't in Models, It's in Transformers and Rare-Earth Magnets

If Fall 2026 was drafting a blueprint for AI's march into the physical world, Winter 2027 presses directly on the most painful spot: bottlenecks.

Six partners, 11 directions, almost each answering the same question: models are powerful enough now; what's stopping them from being deployed?

The answer is surprisingly "low-tech."

Jared Friedman's "Data Center Supply Chain" provides a staggering figure: delivery cycles for large power transformers are three years.

No matter how many GPU chips there are, without power access they're scrap metal.

The entire AI industry's expansion speed is being choked by a power equipment category from the last century.

Diana Hu's "Robots Made in America" points to another chokehold: the U.S. deployed only 38,400 robots in 2025, while China controls 94% of the global sintered rare-earth magnet production capacity.

Rare-earth magnets are core components of robot motors. If the supply chain is held by others, the U.S. robotics industry is built on sand.

Transformers, rare-earth magnets—these sound like topics from the past century, yet they are exactly what YC wants to fund most in 2027.

Brad Flora proposed two directions, forming an interesting contrast.

"The $15-an-Hour Expert" depicts an AI-driven class mobility scenario: when AI enables average workers earning $15/hour to complete tasks that previously required professional certification, vocational education and the labor market as a whole will be rewritten.

"Wall Street for the Physical Buildout" notes that heavy-asset projects like robotic factories, GPU clusters, and data centers require an entirely new suite of financial products for financing and pricing.

One liberates labor, the other liberates capital.

Another direction from Diana Hu, "Vertical Foundation Models," steps out of the general-purpose LLM arms race and looks toward smaller, specialized models for vertical sectors like manufacturing, energy, and biotech.

Jared Friedman's "LLMs That Write Well" points to a counterintuitive phenomenon: models that get stronger at coding and math through training actually see their writing capabilities degrade.

The impact of this issue on the content industry and marketing is tangible.

Gustaf Alströmer wrote two government-focused directions back-to-back.

"Plaid for Government Benefits" builds a unified API for government welfare systems, while "AI for City Hall" uses AI to reshape local government administrative efficiency.

These directions aren't sexy, but the market is huge: the U.S. government spends over $100 billion annually on IT, with the vast majority consumed by legacy systems from the last century.

Jon Xu's "Own Your Intelligence" raises a strategic question: almost every company using AI is essentially "renting intelligence," outsourcing core capabilities to a handful of model providers.

Open-source models have caught up in performance; enterprises should start considering "owning" their own AI capabilities.

Ankit Gupta's "Startups in a World of Abundant Space Access" looks at the space cost revolution brought by SpaceX's Starship.

With launch costs dropping 100-fold, tasks once exclusively reserved for NASA are now within reach for startups.

"AI for the Air-Gapped Economy" focuses on environments that cannot connect to networks, such as military bases, nuclear facilities, and hospital operating rooms. Offline inference and localized deployment remain largely uncharted territory.

Two Lists, One Main Thread

Overlaying the Fall 2026 and Winter 2027 lists reveals an evolving trajectory for the investment thesis.

YC's focus is undergoing a fundamental gravitational shift: from software to hardware, from models to infrastructure, from the digital world to the physical world.

Fall 2026's maritime compute, defense tech, and OS for the physical world sent the initial signals; Winter 2027's transformer supply chains, rare-earth magnets, and domestic U.S. robotics manufacturing turned those signals into marching orders.

These directions share a common trait: each involves supply chain security and geopolitical maneuvering, meaning YC's investment thesis is now deeply intertwined with U.S. industrial policy.

For the Crypto industry, these crossover signals are equally worth noting.

The dedicated Crypto RFS for Fall 2026 and YC's USDC funding plan have already made their stance clear.

Although Winter 2027 doesn't explicitly name Crypto, at least three directions naturally point to on-chain use cases:

Decentralized AI capabilities called for by "Own Your Intelligence" are already being pursued by DePIN and decentralized inference networks.

The new financing tools needed for "Wall Street for the Physical Buildout" could find their most efficient answer in on-chain RWAs.

"Proving You’re Human" (Fall 2026) combined with offline AI and the Air-Gapped Economy (Winter 2027) forms a complete demand map for decentralized identity and privacy computing.

YC is doing subtraction; Crypto founders should do addition at these intersections.