Modular (legally Modular Inc, also referred to as Modular AI) is an American AI infrastructure software company that builds tools for running AI models on chips from many vendors without rewriting code for each one.[25] Its main products are the Mojo programming language, the MAX model-serving and compiler framework, and the hosted Modular Cloud inference service.[29][35] Chris Lattner, the creator of the LLVM compiler infrastructure and the Swift programming language, and Tim Davis, a former product leader at Google Brain, founded the company in January 2022 after meeting at Google.[1][2] Modular raised $380 million across three funding rounds, the last of which valued it at $1.6 billion in September 2025.[19]

Qualcomm agreed to acquire Modular on June 21, 2026, announced the deal on June 24 and completed it on July 28, 2026, paying mainly in Qualcomm stock.[25][26][30][31] The deal was valued at about $3.9 billion when it was announced, based on Qualcomm's share price the day before; Qualcomm's own quarterly filing later put the value at about $3.1 billion, based on the share price on the closing date.[27][30] Modular now operates as a wholly owned Qualcomm subsidiary, and Qualcomm said Mojo, MAX and Modular Cloud would continue as products and brands.[29][31] Lattner became Qualcomm's Executive Vice President of Advanced AI Software and Platforms, and Davis became Senior Vice President and General Manager of Modular.[29][33]

Overview

Field	Detail
Legal name	Modular Inc[30]
Founded	January 4, 2022[2]
Founders	Chris Lattner (CEO), Tim Davis (President)[1][12]
Headquarters	Los Altos, California (earlier Palo Alto); offices in San Francisco, Boston and Edinburgh[1][8][23]
Products	Mojo, MAX, Mammoth, Modular Cloud, BentoML[19][22][35]
Funding	$380 million across three rounds (2022 to 2025)[19]
Last private valuation	$1.6 billion (September 2025, company figure)[19]
Employees	About 25 (June 2022); about 70 (August 2023); more than 130 (September 2025); almost 200 at acquisition[2][3][8][19]
Parent	Qualcomm (since July 28, 2026)[30][31]
Website	modular.com

Founders and background

Lattner designed and built most of the fundamental parts of LLVM while a graduate research assistant at the University of Illinois at Urbana-Champaign from 2000 to 2005.[2] His 2004 paper with Vikram Adve at the International Symposium on Code Generation and Optimization (CGO) described LLVM as a compiler framework built around a common low-level code representation.[45] At Apple he co-founded the Clang compiler and the Swift programming language, and from January 2013 he ran the company's developer tools department.[2] He then spent five months in 2017 as vice president of Autopilot software at Tesla.[2] From August 2017 to January 2020 he was a senior director at Google, where he led the TensorFlow infrastructure team (including the XLA compiler team), drove the bring-up of the TensorFlow Tensor Processing Unit platform for Google Cloud, and co-founded MLIR, a compiler infrastructure for heterogeneous hardware.[2] The MLIR paper, posted in 2020 with Lattner as first author, framed the project as a response to "software fragmentation" and to the cost of building compilers for new hardware.[46] From January 2020 to January 2022 he was president of engineering and product at the RISC-V chip designer SiFive.[2]

Davis grew up in Australia.[44] Modular describes him as a product leader who helped build Google's AI infrastructure at Google Brain and Core Systems, covering TensorFlow APIs, the XLA and MLIR compilers, server runtimes for CPUs, GPUs and TPUs, and TensorFlow Lite for mobile devices.[1] TechCrunch reported that from 2020 to early 2022 he was the product lead for Google's machine learning APIs, compilers and runtime infrastructure for server and edge devices.[3] In a 2023 interview, Davis said the two met at Google and started Modular because they felt AI "was being held back by overly complex and fragmented infrastructure."[44]

Lattner has said the company was founded on bets that were unusual in early 2022: that years of research and development could produce a large step forward in AI software, that inference would matter more than training, and that the industry would need many different kinds of accelerators.[2] In a 2025 essay series, "Democratizing AI Compute", he argued that the main problem with CUDA was the complexity it had driven into the AI software stack, and that the industry needed something "inspired by the design of LLVM" for AI hardware.[18]

History

Founding and seed round (2022)

Modular started on January 4, 2022.[2] It came out of stealth on June 30, 2022, announcing a $30 million seed round led by GV, with Greylock, The Factory and SV Angel participating.[3][4] At the time it had about 25 employees and described a platform that would unify front ends for popular AI frameworks and let developers target many kinds of hardware. TechCrunch wrote that the details were "a bit murky", and Lattner cautioned that it was early days.[3] Lattner said the main competition was "the mindset that dominates AI software development within Big Tech, and Big Tech itself."[3]

Mojo and the inference engine (2023)

On May 2, 2023, Modular held its first product keynote and announced two products: the Modular Inference Engine, a runtime for PyTorch and TensorFlow models offered to a limited number of early-access partners, and Mojo, a new language that it said combined "the usability of Python with the performance of C."[5] Modular claimed Mojo ran a Mandelbrot benchmark 35,000 times faster than standard Python on an AWS server, and that all of the engine's in-house kernels were written in Mojo.[5] The Register called the speedup "alleged" and noted that the language was not yet under an open-source license.[6] Mojo was first offered through a hosted notebook playground; a local toolchain followed for Linux on September 7, 2023, and for macOS on October 19, 2023.[10][11]

On August 24, 2023, Modular announced a $100 million round led by General Catalyst, with GV, SV Angel, Greylock and Factory participating, bringing its total raised to $130 million.[8] The company did not disclose a valuation; Reuters noted that The Information had reported a $600 million figure earlier that month.[9] TechCrunch described Modular as Palo Alto-based with roughly 70 employees, and reported Lattner's claim that more than 120,000 developers had joined its community in the four months after the May keynote.[8]

MAX and the move to GPUs (2023 to 2025)

At ModCon 2023, Modular's first developer conference, held on December 4, 2023, the company announced MAX (Modular Accelerated Xecution), a package made up of the MAX Engine (the renamed inference engine), MAX Serving and Mojo.[12] It also announced a partnership with Amazon Web Services to bring the paid MAX Enterprise Edition exclusively to AWS production services, and a technology partnership with Nvidia.[12] MAX became available as a developer preview on Linux on February 29, 2024, supporting Intel, AMD and Arm CPUs, with GPU support described as "coming soon."[13] On March 28, 2024, Modular released the core modules of the Mojo standard library under the Apache License 2.0 with LLVM exceptions and began accepting outside contributions to it.[14]

MAX 24.6, released on December 17, 2024, previewed what Modular called MAX GPU, a serving stack that the company said removed the dependency on vendor-specific computation libraries such as CUDA by using GPU kernels written in Mojo for Nvidia hardware.[15] In April 2025 Modular simplified its license, making MAX and Mojo free for non-production use and for production use on CPUs and Nvidia GPUs, and free on other accelerators up to eight devices.[16] On June 10, 2025, it announced a partnership with AMD and general availability of its platform on AMD's MI300 and MI325 accelerators, running the same container as on Nvidia hardware.[17] The same day it released a public preview of Mammoth, a Kubernetes-native platform for serving models across large distributed deployments.[48] Modular said it reached state-of-the-art performance on AMD's MI355 within two weeks of that chip's launch, and published a four-part series on matrix-multiplication kernels for Nvidia's Blackwell GPUs.[21]

Third funding round (2025)

On September 24, 2025, Modular announced a $250 million round, its third, led by Thomas Tull's US Innovative Technology (USIT) fund, with DFJ Growth joining as a new investor and GV, General Catalyst and Greylock also participating.[19][20] The company said the round valued it at $1.6 billion, "almost tripling its valuation from its last raise", and brought its total funding to $380 million.[19] SDxCentral described it as a Series C.[20] Modular said it had more than 130 employees and named Inworld, SF Compute, Jane Street, Oracle, AWS, Lambda Labs, TensorWave, AMD and Nvidia among its partners and customers.[19]

BentoML acquisition (2026)

On February 10, 2026, Modular announced that it had acquired BentoML, the company behind the open-source BentoML model-serving framework, after months of working together on customer deployments.[22] The announcement said BentoML was used by more than 10,000 organizations, including more than 50 Fortune 500 companies, and that the open-source project would continue under the Apache 2.0 license; it did not disclose terms.[22] Modular presented the purchase as adding production deployment infrastructure to its stack and expanding the team behind its cloud service.[22]

In April 2026 Modular opened an office at the University of Edinburgh's Bayes Centre and a second Bay Area office in San Francisco's Jackson Square neighborhood, alongside its Los Altos headquarters.[23]

Acquisition by Qualcomm (2026)

Bloomberg reported on June 23, 2026 (UTC) that Qualcomm was in advanced talks to buy Modular in a deal valuing it at about $4 billion.[24] Qualcomm announced the deal the next day, June 24, and its filing showed that the definitive agreement had been signed on June 21.[25][26] In that filing, Qualcomm said it expected to issue up to 19.2 million shares of its common stock to Modular's equity holders in a private placement.[26] Bloomberg put the value at about $3.9 billion based on Qualcomm's closing price of $204.13 on June 23; Reuters calculated $3.92 billion and reported that Qualcomm's shares fell about 4% in late-morning trading on the day of the announcement.[27][28] Qualcomm CEO Cristiano Amon said the industry was "moving toward disaggregated, multi-vendor architectures" and that "the future belongs to developer-friendly, horizontal platforms that can run across diverse compute environments."[25]

The deal closed on July 28, 2026, when a Qualcomm subsidiary, Mod Acquisition Sub, merged into Modular; Qualcomm announced completion on July 29.[29][31] Qualcomm's quarterly report for the period ending June 28, 2026, said the transaction valued Modular at approximately $3.1 billion based on the closing price of Qualcomm stock on the acquisition date, with consideration consisting primarily of about 18 million Qualcomm shares. That total included 4 million shares, with an estimated fair value of about $700 million, issued to certain executives and subject to a four-year service requirement.[30] A resale prospectus filed on July 31 registered 17,826,566 shares for former Modular stockholders and gave Qualcomm's last sale price on July 30 as $151.60.[32] A separate registration covered up to 1,438,756 Qualcomm shares for Modular stock options and restricted shares that Qualcomm assumed.[31]

Date (2026)	Event
June 21	Qualcomm and Modular sign the merger agreement[26]
June 23	Bloomberg reports advanced talks at about $4 billion[24]
June 24	Qualcomm announces the deal; 8-K discloses up to 19.2 million shares[25][26]
July 28	Merger becomes effective[31]
July 29	Qualcomm announces completion; 10-Q values the deal at about $3.1 billion[29][30]
July 31	Resale prospectus for 17,826,566 shares; S-8 for assumed equity awards[31][32]

The four largest selling stockholders named in the resale prospectus were USIT II Holdco PC1 LLC (2,487,472 shares, managed by an entity whose manager is Thomas Tull), GV 2021 (2,206,989), General Catalyst Group XI - Endurance (1,700,124), and GV 2023 (989,272); 92 other holders accounted for the remaining 10,442,709 shares.[32] Lattner later wrote that the company had "nearly $150M in the bank" when it was acquired and that it had grown to almost 200 people.[2]

Under Qualcomm (2026)

Mojo reached version 1.0 on August 11, 2026. Modular described it as a stable foundation and said changes during the 1.x series should be mainly additive.[34] At ModCon 2026 in San Francisco on August 18, Modular released the full Mojo compiler and toolchain under Apache 2.0 with LLVM exceptions, made Modular Cloud generally available, and announced support for AWS Trainium, Google TPUs, and Qualcomm's Cloud AI 100 Ultra and Dragonfly accelerators.[35][36][38] It also said Microsoft's Windows team was working with it on native Windows support for Mojo, and that Modular would keep supporting hardware "that competes directly with Qualcomm Technologies' platforms."[35] Modular began accepting outside contributions to the Mojo compiler with Mojo 1.1 on September 17, 2026.[39]

On October 22, 2026, Modular made the rest of the MAX source code, including its graph compiler, runtime and hardware backends, publicly available under a new MAX License.[41][42] The license allows free commercial use and modification but bars using MAX as input to an AI system to reimplement MAX, bars distributing modified source except to contribute changes back, and requires commercial hosted inference services built on MAX to display a "Modular Accelerated" wordmark.[41] Modular said the restrictions were meant to stop MAX from splitting into incompatible versions, and that it would launch an alliance program for hardware vendors, model developers and data-center operators.[41][42]

Products and technology

Mojo

Mojo is a compiled systems programming language with Python-like syntax. Modular built it on MLIR and introduced it in 2023 as a way to write high-performance CPU and GPU code without switching to C++ or CUDA.[5] It was originally presented as intended to become a superset of Python.[6] Reviewing the first beta of Mojo 1.0 in May 2026, InfoWorld wrote that Mojo "is not a drop-in replacement for Python" and had become a systems language with Rust-like ownership, strong types and pointer types.[43] Modular writes the kernels in MAX in Mojo.[19] The standard library was opened in 2024, the compiler in August 2026, and the compiler began taking outside contributions in September 2026.[14][36][39] Modular said that by Mojo 1.0 nearly 200 contributors had landed more than 1,100 pull requests to the standard library.[34] Mojo 1.2, released in October 2026, added pattern matching.[42]

MAX

MAX is Modular's framework for compiling and serving models. In its current form it combines a graph compiler, a runtime, hardware backends, a library of Mojo kernels and a serving layer with an OpenAI-compatible API.[19][41] The company has published comparisons claiming 20% to 50% higher performance than vLLM and SGLang on Nvidia B200 and AMD MI355 hardware; these are Modular's own benchmarks.[19] Its kernels, models, serving code and Python APIs were released under Apache 2.0 during 2025, and the remaining core components became source-available under the MAX License in October 2026.[21][41]

Mammoth and Modular Cloud

Mammoth is a Kubernetes-native control plane for distributed inference, with routing that takes the prefill stage into account and support for disaggregated compute and cache.[19] Modular Cloud, the company's managed service, offers shared endpoints with per-token pricing and dedicated deployments on Modular's compute or a customer's own cloud.[35] Before its general launch in August 2026, Modular said, the service had been handling OpenRouter traffic under the name ModelRun. It named MiniMax as a flagship customer serving MiniMax M3 on a dedicated deployment.[35]

Hardware support

Modular's stack targets Nvidia, AMD and Apple GPUs as well as CPUs, and in 2026 it added Google TPUs, AWS Trainium and Qualcomm neural processing units.[35][42] Lattner wrote that by ModCon 2026 the company could serve models on accelerators from six vendors.[2] Engineers from HTEC brought up TPU support themselves, with Modular in a supporting role.[35] Modular said the new non-GPU platforms would move into production on Modular Cloud over the following months.[35]

Modular described the Qualcomm work in detail. Mojo compiles to Qualcomm's Hexagon LLVM backend, and the Cloud AI 100 plugs into the MAX graph compiler and serving framework the same way Nvidia, AMD and Apple GPUs do. Modular called the Cloud AI 100 the first ASIC integrated into its stack.[37] The team ran GPT-2 end to end first, then served Gemma 4 31B across the card's four chips using tensor parallelism, and later added speculative decoding and the Gemma 4 mixture-of-experts variant. Modular said the work took under six months from the first Gemma 4 kernel to a live endpoint. Work on the next-generation Qualcomm Dragonfly AI200 had reached the point of running GPT-2, with the rest of the stack still being validated.[37]

Funding

Date	Round	Amount	Lead investor	Other participants	Valuation
June 2022	Seed	$30 million	GV	Greylock, The Factory, SV Angel[3][4]	Not disclosed
August 2023	Second round	$100 million	General Catalyst	GV, SV Angel, Greylock, Factory[8]	Not disclosed; The Information reported $600 million (per Reuters)[9]
September 2025	Third round (Series C)	$250 million	US Innovative Technology fund (Thomas Tull)	DFJ Growth, GV, General Catalyst, Greylock[19][20]	$1.6 billion (company figure)[19]
July 2026	Acquired by Qualcomm	About 18 million Qualcomm shares	Not applicable	Not applicable	About $3.9 billion at announcement; about $3.1 billion at closing (Qualcomm 10-Q)[27][30]

Leadership

Modular's leadership page listed the following people and their roles after the acquisition.[1]

Name	Role after acquisition	Background (per Modular)
Chris Lattner	Co-founder; EVP, Advanced AI Software and Platforms, Qualcomm[29]	LLVM, Clang, Swift, MLIR, Cloud TPUs; Apple, Google, Tesla, SiFive
Tim Davis	Co-founder; SVP and GM of Modular, Qualcomm[33]	Google Brain and Core Systems product leader (TensorFlow, XLA, MLIR, TF Lite)
Mostafa Hagog	VP of Engineering	Former Nvidia engineering director (cuDNN, CUTLASS); former SiFive VP of Software
Eric Johnson	VP of Product Management	Led TensorFlow API, compiler and runtime teams at Google Brain
Kalor Lewis	Modular finance leader	Former VP of Finance at Fivetran; earlier at Palantir
Mike Edwards	Business operations	Former Apple; LLVM Foundation board member

Reception

Early coverage of Mojo focused on its claimed speed and on Lattner's track record.[6] Jeremy Howard of fast.ai, who disclosed that he was an advisor to Modular, wrote at launch that "Mojo may be the biggest programming language advance in decades."[7] Other early coverage was more cautious. The Register described the 35,000x figure as "alleged", and TechCrunch's Kyle Wiggers wrote that he was "less convinced" Modular could drive wide adoption of a new language when Python was so entrenched among machine learning developers.[6][8]

Coverage of the Qualcomm deal centered on CUDA. Reuters wrote that buying Modular "pits Qualcomm against CUDA," and quoted Emarketer analyst Jacob Bourne saying Qualcomm was "betting that by owning software that squeezes more inference more efficiently out of hardware, it can stake a claim in the data center market."[28] A September 2026 analysis in ITdaily called the "Qualcomm attacks CUDA" framing an oversimplification. It noted that MAX was not open source at the time, that no independent benchmarks of MAX on Qualcomm silicon existed, and asked whether a team of Modular's size could keep six hardware architectures at top performance.[40] ITdaily also attributed the gap between the $3.9 billion headline and the $3.1 billion booked value to the fall in Qualcomm's share price.[40]

Debarghya Das's "Ultimate Neolab List" includes Modular but marks it "Not a neolab," describing it as "compiler and infrastructure tooling rather than a model lab."[47]

See also

References

Modular. "About Modular". Modular. Retrieved 27 October 2026.
Chris Lattner. "Chris Lattner: More than you ever wanted to know". Personal website, last updated September 2026. Retrieved 27 October 2026.
Kyle Wiggers. "Modular closes $30 million seed round to simplify the process of developing AI systems". TechCrunch, 30 June 2022.
Maria Deutscher. "Modular raises $30M to ease AI development and deployment". SiliconANGLE, 30 June 2022.
Chris Lattner, Tim Davis and Eric Johnson. "A unified, extensible platform to superpower your AI". Modular, 2 May 2023.
Thomas Claburn. "Modular finds its Mojo, a Python superset with C-level speed". The Register, 5 May 2023.
Jeremy Howard. "Mojo may be the biggest programming language advance in decades". fast.ai, 4 May 2023.
Kyle Wiggers. "Modular raises $100M for AI dev tools". TechCrunch, 24 August 2023.
Reuters. "AI startup Modular raises $100 mln in General Catalyst-led funding" (archived copy). Reuters, 24 August 2023.
Eric Johnson. "Mojo🔥 - It's finally here!". Modular, 7 September 2023.
Shashank Prasanna. "Mojo🔥 is now available on Mac". Modular, 19 October 2023.
Shashank Prasanna. "Key announcements from ModCon 2023". Modular, 4 December 2023.
Modular Team. "Announcing MAX Developer Edition Preview". Modular, 29 February 2024.
Modular Team. "The Next Big Step in Mojo🔥 Open Source". Modular, 28 March 2024.
Modular Team. "Introducing MAX 24.6: A GPU Native Generative AI Platform". Modular, 17 December 2024.
Modular Team. "A New, Simpler License for MAX and Mojo". Modular, 23 April 2025.
Modular Team. "Modular + AMD: Unleashing AI performance on AMD GPUs". Modular, 10 June 2025.
Chris Lattner. "Modular's bet to break out of the Matrix (Democratizing AI Compute, Part 10)". Modular, 8 May 2025.
Modular Team. "Modular Raises $250M to scale AI's Unified Compute Layer". Modular, 24 September 2025.
Ben Wodecki. "Modular raises $250M for AI's 'unified compute layer' at $1.6B valuation". SDxCentral, 26 September 2025.
Michael Dunn-OConnor. "Modular 2025 Year in Review". Modular, 19 December 2025.
Chris Lattner, Chaoyu Yang and Tim Davis. "BentoML Joins Modular". Modular, 10 February 2026.
Conor Bronsdon. "Modular Opens Edinburgh & San Francisco Offices". Modular, 10 April 2026.
Liana Baker and Ryan Gould. "Qualcomm Is Said to Near Deal for AI Software Firm Modular". Bloomberg News via Bloomberg Law, 23 June 2026.
Qualcomm. "Qualcomm to Acquire Modular". Qualcomm press release, 24 June 2026. Also published by Modular at modular.com/blog/qualcomm-to-acquire-modular.
Qualcomm Incorporated. "Form 8-K (Item 3.02)". U.S. Securities and Exchange Commission, filed 24 June 2026.
Ian King and Liana Baker. "Qualcomm to Buy Modular for $3.9 Billion to Help AI Push". Bloomberg News via Bloomberg Law, 24 June 2026.
Anhata Rooprai. "Qualcomm to buy startup Modular for $4 billion in AI software push". Reuters via The Star, 24 June 2026.
Qualcomm. "Qualcomm Completes Acquisition of Modular". Qualcomm press release, 29 July 2026.
Qualcomm Incorporated. "Form 10-Q for the quarterly period ended June 28, 2026". U.S. Securities and Exchange Commission, filed 29 July 2026.
Qualcomm Incorporated. "Form S-8 registration statement". U.S. Securities and Exchange Commission, filed 31 July 2026.
Qualcomm Incorporated. "Prospectus supplement (Rule 424(b)(7)): up to 17,826,566 shares of common stock". U.S. Securities and Exchange Commission, 31 July 2026.
Modular. "Qualcomm Completes Acquisition of Modular". Modular Forum, 29 July 2026.
Modular Team. "Modular 26.5: Mojo 1.0 is here!". Modular, 11 August 2026.
Modular Team. "ModCon 2026: Open source, open cloud, open silicon". Modular, 18 August 2026.
Modular Team. "Mojo🔥 is now open source!". Modular, 18 August 2026.
Ruyman Reyes et al. "Modular and Qualcomm: Same code, new silicon". Modular, 18 August 2026.
Bobby Borisov. "Mojo Programming Language Goes Fully Open Source". Linuxiac, 19 August 2026.
Modular Team. "Modular 26.6: Open compiler contributions, audio generation, and expanded model support". Modular, 17 September 2026.
Cédric Van Loon. "Why Qualcomm is paying 3.9 billion dollars for a programming language almost no one speaks". ITdaily, 24 September 2026.
Modular Team. "MAX is now source available and open to everyone". Modular, 22 October 2026.
Modular Team. "Modular 26.7: Opening up MAX". Modular, 22 October 2026.
Serdar Yegulalp. "First look: Mojo 1.0 mixes Python and Rust". InfoWorld, 20 May 2026.
Antoine Tardif. "Tim Davis, Co-Founder & President of Modular - Interview Series". Unite.AI, 14 September 2023.
Chris Lattner and Vikram Adve. "LLVM: A Compilation Framework for Lifelong Program Analysis & Transformation". Proceedings of the 2004 International Symposium on Code Generation and Optimization (CGO'04), March 2004.
Chris Lattner, Mehdi Amini, Uday Bondhugula, Albert Cohen, Andy Davis, Jacques Pienaar, River Riddle, Tatiana Shpeisman, Nicolas Vasilache and Oleksandr Zinenko. "MLIR: A Compiler Infrastructure for the End of Moore's Law". arXiv:2002.11054, 25 February 2020.
Debarghya Das. "The Ultimate Neolab List". Retrieved 27 October 2026.
Modular Team. "Introducing Mammoth: Enterprise-Scale GenAI Deployments Made Simple". Modular, 10 June 2025.

Editing: Modular (company)