Flapping Airplanes

Flapping Airplanes is a San Francisco AI research company founded in 2025 by brothers Ben and Asher Spector and Aidan Smith. It studies ways to make AI use data and computing resources more efficiently. Its research combines model development with software systems for graphics processing units, including a megakernel framework.^[1][2]

The company launched publicly on January 28, 2026 with a $180 million seed round and announced a $1 billion Series A on October 21, 2026. Bloomberg reported that the Series A valued the company at $6 billion; the company itself did not state a valuation. As of October 2026 it had not released a commercial product.^{[3][4][8][9][10][11]}

History

Founding and name

Sequoia Capital's company profile gives 2025 as the founding year, and The Next Web described the lab as founded in late 2025.^[1][10] Before the company existed, Index Ventures partner Mark Xu had known Asher Spector since they were students at Harvard and knew Ben Spector through Prod, the student-run incubator Ben co-founded.^[3]

The name refers to the ornithopter, an aircraft design with flapping wings. Asher Spector told Bloomberg that it stands for combining the strengths of natural and artificial intelligence, according to The Next Web.^[10] In a February 2026 TechCrunch interview, Ben Spector compared current AI systems to Boeing 787s and said: "We're not trying to build birds. That's a step too far. We're trying to build some kind of a flapping airplane."^[12] Sequoia described the name as "a nod to a more biological notion of machine intelligence."^[6]

Launch and first year

The lab launched publicly on January 28, 2026, when Index, GV and Sequoia published notes on their seed investments.^{[3][6][7][11]} In February, TechCrunch's Russell Brandom interviewed all three founders for the Equity podcast and a published Q&A about the company's research agenda and hiring.^[12][22]

On October 21, 2026, the company said on X that it had "spent the past year developing new primitives for GPUs" and had raised a $1 billion Series A co-led by Index Ventures and Lux Capital.^[4]

Founders

Ben Spector is the company's chief executive. He co-founded Prod, a student-run startup incubator. Asher Spector completed a doctorate in statistics at Stanford University, while Smith worked at Neuralink for three years while studying at Georgia Tech.^[2][3]

According to Index and GV, Prod played a part in the formation of companies including Cursor, Mercor, Etched and Decart.^[3][7] The Hertz Foundation, which lists Ben Spector in its fellows directory, says the not-for-profit accelerator's cohort companies raised $25 million under his leadership.^[14] GV wrote that Ben Spector's Stanford doctoral work was done under Christopher Ré.^[7] Stanford's statistics department lists Emmanuel Candès as Asher Spector's dissertation advisor.^[15]

In his January 2026 investment note, Sequoia partner David Cahn gave the brothers' ages as 25 (Ben) and 26 (Asher) and said they were "building the 'young person's AGI lab.'"^[6]

Funding

Flapping Airplanes announced its seed financing in January 2026. Index Ventures said it co-led the investment alongside GV and Sequoia Capital.^[3] GV described its own role as co-leading with Sequoia, with participation from Index and Menlo Ventures.^[7] TechCrunch later reported that the seed round had valued the company at $1.5 billion, a figure also given by Harvard's statistics department.^[8][16]

In October, the company announced a larger Series A co-led by Index and Lux Capital. Index said partner Mark Xu would join the board.^[2][4] Earlier backers Sequoia and GV participated, along with Kleiner Perkins.^[8][10] The company did not disclose a valuation. The Next Web and SiliconANGLE attributed a $6 billion valuation to Bloomberg, and TechCrunch described the round as "reportedly at a $6 billion valuation"; none of these accounts said whether the figure was pre-money or post-money.^[8][9][10] SiliconANGLE put total investment in the company at $1.18 billion after the round.^[9]

The amounts are the company's announced financing figures.^[4][5] Its website also lists its backers, including NVIDIA, Jump Trading and XTX Ventures.^[5] As of late October 2026, that list named Index, Google Ventures, Sequoia, Lux Capital and Nova Global as its main backers, with further support from Valor, Atreides, Kleiner Perkins, Fundomo, Steve Cohen, BCV, Menlo Ventures, Hummingbird, Victor Lazarte, Hof, Argo, Absolute Ventures, A*, CRV, Definition, Seven Stars, Genesis, Jump Trading, XTX Ventures, Conviction, Duke, Snowy Leopard and Nvidia. The site also named individual supporters including Jeff Dean, Stanley Druckenmiller, Katie Haun, Andrej Karpathy, Henry Kravis, Micky Malka and Christopher Ré.^[5]

Team and hiring

Index described an October 2026 team of roughly two dozen people working together in San Francisco. The investor said the company hired for ability and potential rather than formal credentials. Its account described work spanning GPU kernels, pretraining, post-training and inference, with each layer developed in conjunction with the others.^[2]

Index gave two examples of that hiring approach: one engineer who joined while still in high school built the agent infrastructure that automates much of the team's work, and another who had never trained a model before joining went on to lead the company's pre-training run.^[2] GV said the founders had recruited "high-school prodigies, math olympians, and gritty researchers," and Cahn wrote that he had met nearly every candidate who interviewed at the company.^[6][7]

In the TechCrunch interview, Aidan Smith said creativity was "the number one thing we look for," and Ben Spector said the company had also hired people with experience on large-scale systems.^[12]

Research approach

The company's initial research agenda centered on data efficiency: learning useful capabilities from fewer examples. In its January investment announcement, Index described the founders' hypothesis that the gap between human learning and the data requirements of current AI models leaves room for different learning methods. This was a rationale for research, rather than an announced result showing a particular improvement.^[3]

Sequoia described the company as a laboratory for research over long time horizons, with independence resembling doctoral research. Its investment account framed biological learning as inspiration for more data-efficient models and said the company would explore several possible directions. That description did not identify a single model architecture as the company's settled approach.^[6]

The founders set out the same thesis in their February 2026 interview. Ben Spector said that "the current frontier models are trained on the sum totality of human knowledge, and humans can obviously make do with an awful lot less."^[12] Smith called the brain "an existence proof" that other learning algorithms exist, contrasted its methods with gradient descent, and said that "the brain, in many ways, is the floor" rather than the ceiling for AI capability. He added that the team was inspired by the brain but did not see itself as "tied down by it."^[12] Asher Spector said the company was "looking for 1000x wins in data efficiency."^[12]

Asher Spector offered three hypotheses about what more data-efficient training might produce, while saying that he did not know the answer: models forced to learn from less data might develop deeper understanding and reason better while knowing fewer facts; post-training could become efficient enough to move a model into a new domain with a few examples; and data-limited fields could open up. On robotics, he said: "My opinion is that it's a limited data problem, not a hardware problem."^[12]

Ben Spector said he would not "frame us as the antithesis of scale," calling scale "an important tool in the toolbox." He argued that radical ideas are cheaper to test than incremental ones: an incremental change has to be checked far up the scaling ladder, while a radical idea that does not work will probably fail on its first small run.^[12] On commercialization, Asher Spector said the company could not give a timeline and wanted to start with research, because signing large enterprise contracts early would distract it.^[12] In October, he told Bloomberg that the company planned eventually to sell its technology to make AI cheaper for other companies, according to The Next Web.^[10]

By October, Flapping Airplanes described a broader effort to develop models and computing systems together. The company argued that making GPUs more efficient and more flexible could enable learning algorithms that are difficult to express with conventional systems. It identified robotics and scientific discovery as intended applications where improved economics could expand AI use.^[5] Its website says the company believes the only way to do this is to "co-evolve models and systems, bridging the traditional divides between software and hardware."^[5] Asher Spector summarized the commercial thesis to Bloomberg as "demand for AI is going to outpace chip and data supply," adding that getting 10% more performance out of each chip would be "tremendously economically valuable."^[10]

GPU systems work

Flapping Airplanes says it has developed a megakernel framework. Index describes a megakernel as running a whole program on a GPU within one kernel, with the aim of reducing the gaps between separately executed steps.^[2] The company says its framework can run a complete training process, serve inference and run video games, including combinations of these activities within a single kernel.^[5]

A specific company claim is that the framework can overlap inference limited by memory access with training limited by arithmetic computation. It presents this capability as a way to explore unusual model architectures and optimizers. These are descriptions of the company's internal research system, not independently established measurements of performance or data efficiency. The cited October announcements do not provide a reproducible benchmark comparison or a public release of the described framework.^[2][4][5]

The website says that "in recent months, the majority of our work has built on top of these new primitives."^[5] Smith told Bloomberg that the company had been writing its own version of the software that runs Nvidia's graphics chips, according to The Next Web.^[10] Index described the systems work as following a pattern the founders see in machine-learning history, in which the biggest algorithmic breakthroughs are preceded by new ways of interacting with hardware.^[2]

Founders' academic research

At Stanford's Hazy Research group, Ben Spector worked on software for GPU kernels. He was the lead author of ThunderKittens, a framework for writing high-performance AI kernels that was released in May 2024 and described in an October 2024 paper with Simran Arora, Aaryan Singhal, Daniel Y. Fu and Christopher Ré. The paper reported matching cuBLAS and FlashAttention-3 on matrix multiplication and attention inference, and outperforming the strongest baselines by 10-40% on attention backward passes, 8 times on state space models and 14 times on linear attention.^[17][24]

He was also a joint first author of Hazy Research's megakernel posts. In May 2025, he and Jordan Juravsky, with other Hazy Research authors, described merging the entire forward pass of Llama 3.2 1B into a single megakernel. On an NVIDIA H100, they reported using 78% of memory bandwidth and outperforming existing systems by more than 1.5 times, compared with at most 50% bandwidth use by vLLM and SGLang on that workload.^[18] A September 2025 follow-up released a throughput-oriented megakernel for tensor-parallel Llama-70B inference that overlapped compute, memory and communication operations; integrated into the Tokasaurus inference engine, it outperformed SGLang by more than 22% on end-to-end throughput in the authors' benchmark.^[19] Public descriptions of Flapping Airplanes' framework do not say whether it is derived from this academic code.

Asher Spector's research was in statistical methodology. With Lucas Janson he wrote on constructing model-X knockoffs, a method for feature selection that controls the expected proportion of false discoveries, showing that the standard construction can be powerless in some easy settings.^[20] He later proposed masked likelihood ratio statistics for knockoffs with William Fithian, and the mosaic permutation test, a goodness-of-fit test for financial factor models, with Rina Foygel Barber, Trevor Hastie, Ronald N. Kahn and Emmanuel Candès.^[21][23]

Reception

GV's and Sequoia's investment notes presented the company as a departure from scaling-focused AI development. GV's Dave Munichiello criticized a "scale orthodoxy" in major labs, and Cahn wrote that "data is the bottleneck today to further AI scaling, not compute."^[6][7] TechCrunch's Russell Brandom called the lab's goal "a particularly interesting one" and rated it "Level Two" on his five-level scale of how hard new labs are trying to make money, a level he defined as "We have the outlines of a concept of a plan."^[11][13]

Coverage of the Series A noted both the size of the valuation increase and the absence of a product. TechCrunch noted that the valuation had quadrupled in nine months, SiliconANGLE wrote that no commercial product had yet come out of the lab, and The Next Web reported that it had no product on sale.^[8][9][10] The Next Web described Flapping Airplanes as one of a growing group of AI "neolabs" that focus on research before products, alongside AMI Labs and Mirendil.^[10] Debarghya Das's "The Ultimate Neolab List" includes the company and rates it a "Clean" fit for that category.^[25]

Continue reading

• Index VenturesIndex Ventures is an international venture capital firm founded in 1996 in Geneva, Switzerland, that backs technology startups from seed through late-stage growth and today operates from dual hubs in London…Mentioned 6 times above
• Lux CapitalLux Capital is an American venture capital firm, founded in 2000 by Josh Wolfe, Peter Hebert, and Robert Paull, that invests in emerging science and "deep technology" startups across artificial intelligence…Mentioned 5 times above
• Christopher RéChristopher Re (his surname is sometimes written with an acute accent over the "e") is an American computer scientist and a professor of computer science at Stanford UniversityMentioned 3 times above
• NNeolabA neolab, also written neo-lab or NeoLab, is an informal term for an AI research company organized around a technical research program.Mentioned twice above
• TThunderKittensThunderKittens (often abbreviated TK) is an embedded C++ domain-specific language and header-only library for writing high-performance AI kernels on modern NVIDIA GPUs.Listed in See also
• AMI LabsAMI Labs (Advanced Machine Intelligence) is a Paris-based artificial intelligence research company co-founded by Yann LeCun, the Turing Award winner and former Chief AI Scientist at MetaListed in See also

More from See also

• List of neolabs

More to read

• Graphics processing unit
• Nvidia
• Sequoia Capital
• Kleiner Perkins
• Menlo Ventures

Topics:AI CompaniesAI Research

References

1. ^1 ^2 ^3Sequoia Capital. Flapping Airplanes. Company profile, accessed October 23, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Mark Xu and Shardul Shah. Toward More Efficient AI: Leading Flapping Airplanes' $1B Series A. Index Ventures, October 21, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Shardul Shah and Mark Xu. Taking Flight: Our Investment in Flapping Airplanes. Index Ventures, January 28, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6Flapping Airplanes. Series A announcement. X, October 21, 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Flapping Airplanes. Official website. Research description and funding announcements, accessed October 23 and 27, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8David Cahn. Partnering With Flapping Airplanes. Sequoia Capital, January 28, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Dave Munichiello. Better Wings: Why We Invested in Flapping Airplanes. GV, January 28, 2026.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Julie Bort. Year-old AI lab Flapping Airplanes raises $1B at a $6B valuation. TechCrunch, October 21, 2026.
9. ^1 ^2 ^3 ^4 ^5Duncan Riley. Year-old AI research lab Flapping Airplanes raises $1B at a reported $6B valuation. SiliconANGLE, October 21, 2026.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Ana-Maria Stanciuc. AI lab Flapping Airplanes is now worth $6B, four times its January value. The Next Web, October 21, 2026.
11. ^1 ^2 ^3Russell Brandom. Flapping Airplanes and the promise of research-driven AI. TechCrunch, January 29, 2026.
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Russell Brandom. Flapping Airplanes on the future of AI: 'We want to try really radically different things'. TechCrunch, February 16, 2026.
13. ^Russell Brandom. A new test for AI labs: Are you even trying to make money?. TechCrunch, January 24, 2026.
14. ^1 ^2Hertz Foundation. Benjamin Spector. Hertz Fellows Directory, accessed October 27, 2026.
15. ^1 ^2Stanford University Department of Statistics. Asher Spector. Doctoral student profile, accessed October 27, 2026.
16. ^1 ^2Harvard University Department of Statistics. Asher Spector Featured in the Wall Street Journal. February 2026.
17. ^Benjamin F. Spector, Simran Arora, Aaryan Singhal, Daniel Y. Fu and Christopher Ré. ThunderKittens: Simple, Fast, and Adorable AI Kernels. arXiv:2410.20399, October 27, 2024.
18. ^Benjamin Spector, Jordan Juravsky, Stuart Sul, Owen Dugan, Dylan Lim, Dan Fu, Simran Arora and Chris Ré. Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B. Hazy Research, Stanford University, May 27, 2025.
19. ^Benjamin Spector, Jordan Juravsky, Stuart Sul, Dylan Lim, Owen Dugan, Simran Arora and Chris Ré. We Bought the Whole GPU, So We're Damn Well Going to Use the Whole GPU. Hazy Research, Stanford University, September 28, 2025.
20. ^Asher Spector and Lucas Janson. Powerful Knockoffs via Minimizing Reconstructability. arXiv:2011.14625, November 30, 2020.
21. ^Asher Spector and William Fithian. Asymptotically Optimal Knockoff Statistics via the Masked Likelihood Ratio. arXiv:2212.08766, December 17, 2022.
22. ^1 ^2Russell Brandom and Theresa Loconsolo. This Sequoia-backed lab thinks the brain is 'the floor, not the ceiling' for AI. TechCrunch Equity podcast, February 10, 2026.
23. ^Asher Spector, Rina Foygel Barber, Trevor Hastie, Ronald N. Kahn and Emmanuel Candès. The mosaic permutation test: an exact and nonparametric goodness-of-fit test for factor models. arXiv:2404.15017, April 23, 2024.
24. ^Benjamin Spector, Aaryan Singhal, Simran Arora and Chris Re. GPUs Go Brrr. Hazy Research, Stanford University, May 12, 2024.
25. ^Debarghya Das. The Ultimate Neolab List. Updated October 24, 2026.

Show all 25 references