Deepgram Buys Pipecat’s Maker to Own the Voice Stack, but Its India Endpoint Still Sends the LLM Abroad

Deepgram said on 29 October that it is acquiring Daily, the company behind open-source Pipecat, at its first Speak event in San Francisco. Neither company disclosed the price or published terms.

Pipecat stays open source under BSD 2-Clause, and the companies plan an independent Pipecat Foundation. Daily’s team, including co-founder Kwindla Hultman Kramer, joins Deepgram, where Kramer becomes Head of Open Source. Deepgram says the framework remains vendor neutral while it sells Daily’s WebRTC, telephony and Pipecat Cloud products.

The catch is in the word open. The framework stays free, but the stack around it is commercial. The new Real-Time Context Standard is version 0.1 alpha, the foundation has no named members or legal structure, and the price is undisclosed. For Indian teams, a second catch: the India endpoint keeps speech processing in Hyderabad, but a managed third-party LLM handles the reasoning step outside the country.

Quick answer: Deepgram is buying Daily, the maker of Pipecat, and will keep Pipecat open source under BSD 2-Clause while forming an independent foundation. The price is undisclosed. RCS 0.1 is alpha and may change incompatibly before 1.0. Indian developers can keep speech data in India through api.in.deepgram.com, but managed LLM calls and billing metadata can still leave.

What Deepgram actually bought

Daily built a 75-point real-time network, WebRTC and telephony libraries, and Pipecat Cloud. Kramer says 18 engineers built Pipecat, open sourced in 2024, with more than 200 integrations and almost 400 contributors.

The GitHub numbers back the reach. On 1 November the pipecat-ai/pipecat repository had 16,974 stars, 3,020 forks and a BSD 2-Clause licence.

Deepgram brings the models and the customer list. It says more than 5,000 organisations and 650,000 developers have used its speech APIs, processing more than 200,000 years of audio. Both companies went through Y Combinator in 2016, and Deepgram has powered transcription inside Daily’s audio and video APIs since 2019.

That explains the logic. Deepgram already sold speech models and a managed agent, while Daily brings transport and orchestration. One vendor can now sell the network, runtime, models and glue.

The open-source promise has fine print

Keeping Pipecat under BSD 2-Clause is a real commitment. Anyone can use, modify and redistribute the code commercially. Deepgram cannot easily close it later without a fork.

The foundation is the part to watch. Kramer says it will be independent, managed by a steering committee representing the wider real-time AI ecosystem, including companies that compete with Deepgram. Deepgram says it will invest in Pipecat and keep it vendor neutral.

What is missing is the paperwork. As of 1 November, neither company had published the foundation’s legal form, members, funding, bylaws or start date. Kramer is joining Deepgram as Head of Open Source, so the person most associated with Pipecat’s neutrality will also be a Deepgram employee.

There is also a commercial boundary. Pipecat Cloud is part of the acquisition, and Deepgram will sell and support it. The framework is free, the production path is not. Do not confuse the two when planning costs or exit routes.

RCS 0.1 is a useful idea, not a finished standard

Deepgram also introduced the Real-Time Context Standard, or RCS, co-developed with the Pipecat team. The problem is real. A voice agent passes text from speech-to-text to a language model to text-to-speech, and the tone, pace and delivery of the original audio get lost in the handoff.

RCS carries two kinds of context in the same text stream. Perception describes how the user sounded, such as ([tone frustrated=0.86, neutral=0.10],[speech_rate value=4.1]). Expression describes how the reply should be spoken, using short tags such as <<warm>> or [[laugh]].

The idea is sound: fewer adapters and a common way for models from different vendors to preserve context. The maturity is not. Pipecat’s documentation labels RCS 0.1 an alpha and warns that until 1.0, formats, vocabularies and context types may change in ways that are not backward compatible. Who is speaking and turn intent are still on the roadmap.

The reference implementation is early. The realtime-context-standard-demo repository was created on 28 October, had 32 stars on 1 November and no licence file. Fine for a demo, not yet a standard to design around.

What the stack costs

Voice agents are priced per minute while customers count outcomes. A five-minute Voice Agent call at the Standard rate costs $0.375 before telephony or recording. Building the same call from separate STT, TTS and Pipecat Cloud pieces has a different profile, and the split is where margin leaks. Deepgram now owns every line of that bill.

Confirmed vs not confirmed

What it means for Indian developers

The deal has an immediate residency angle in India. Deepgram made its India endpoint generally available on 15 September. It runs in AWS ap-south-2 in Hyderabad, uses the same API and pricing as the global endpoint, and needs only a host change to api.in.deepgram.com. Storage and inference for speech-to-text and text-to-speech happen in India, and mip_opt_out=true keeps content out of model improvement.

The boundary is explicit in Deepgram’s documentation. A managed third-party LLM in a Voice Agent processes the reasoning step outside India, and operational metadata and billing are processed in the US. For a bank, insurer or payments company, the voice audio can stay in Hyderabad while the transcript and the model’s reasoning leave it. The endpoint is a strong building block, not a complete residency answer.

Pipecat helps because it is designed to swap providers. A team can keep Deepgram for speech in India, route the LLM to an in-country or self-hosted model, and keep orchestration portable. The same design is an exit route if pricing changes after the acquisition. BSD 2-Clause means the framework can be forked, but forking a fast-moving project with hundreds of contributors is expensive. Portability at the provider layer is the cheaper hedge.

Three checks before the next procurement cycle. Pin the Pipecat version and test upgrades rather than tracking main. Map every hop, including the LLM, logging, analytics and billing, against the residency clause you signed. Model unit cost at your real call length, because the $0.075 per minute Voice Agent rate and the $0.01 per minute Pipecat Cloud rate solve different parts of the problem and are easy to double count.

The deal makes sense for Deepgram: it owns the models, network and orchestration layer and can sell one production path. The risk is familiar. The community keeps the licence, while the roadmap, foundation and standard sit close to the acquirer.

Frequently asked questions

Is Pipecat still open source after the Deepgram acquisition?

Yes. Both companies say Pipecat remains under BSD 2-Clause, and the repository still carried that licence on 1 November. Deepgram promises an independent foundation.

Did Deepgram disclose the acquisition price?

No. The announcement gives no price, and TechCrunch reported the terms as undisclosed. Any specific valuation is unconfirmed.

Has the acquisition closed?

The deal was announced on 29 October, with no closing date or regulatory approvals published. Treat it as announced, not complete.

Does Deepgram’s India endpoint keep all voice data in India?

For Deepgram’s speech-to-text and text-to-speech, storage and inference run in AWS ap-south-2 in Hyderabad. A managed third-party LLM used by a Voice Agent processes the reasoning step outside India, and operational metadata and billing are processed in the US.

Related: Grok Voice 3.0 and the fight over per-minute voice pricing, SquadStack’s Arth V2 benchmark for Indian sales calls and Block’s Dessa speech model that runs on a MacBook.

Sources

• Deepgram, acquisition announcement, 29 October 2026. deepgram.com
• Daily, Kwindla Hultman Kramer, foundation announcement, 29 October 2026. daily.co
• Pipecat documentation, Real-time Context Standard, RCS 0.1 alpha. docs.pipecat.ai
• Deepgram, India endpoint generally available, 15 September 2026. deepgram.com
• Deepgram pricing. deepgram.com
• Daily, Pipecat Cloud and Video SDK pricing. daily.co
• GitHub, pipecat-ai/pipecat metadata. github.com
• TechCrunch, Ivan Mehta, Deepgram acquires Pipecat maker Daily, 29 October 2026. techcrunch.com