Suggestions
Scott Stephenson
Language A.I. | CEO at Deepgram
Scott Stephenson is a Venture Partner at Pioneer Fund and the Co-Founder and CEO of Deepgram, a company specializing in AI-based speech recognition technology. He has been with Pioneer Fund since 2018 and founded Deepgram in 2015. Stephenson holds a Ph.D. in Particle Physics from the University of Michigan and has a background in physics research, including work on the Large Underground Xenon Detector (LUX) for dark matter detection.12
At Deepgram, he focuses on developing scalable AI solutions for audio processing, offering APIs for speech-to-text and other language understanding tools. His expertise combines scientific research and entrepreneurial leadership, making him a prominent figure in AI and voice technology.3
Highlights
What is behind the obsession with the “overspending in AI” narrative right now. Did Musk overspend in building the gigafactories? No. Did Nikola (the scam electric car company) overspend and lose tons of money? Yes.
This is a disruption, people. The real fear investors should be having is choosing business models that don’t work or are a scam, not doubting whether you need to spend to discover the models and power the revolution. Pick the right visionaries and you’ll be fine.
I encourage everyone to zoom out a bit. All of the value ever created in totality by mankind will be dwarfed by the value generated by AI in the next 3 decades. Yes, we have to invest in infrastructure for that to happen.
Here is a data point: 9 years ago Deepgram built it’s first AI data center. When we first built it, literally every GPU in the cluster was dedicated to training foundation models in audio (our first product—speech to text). We also were nobody physicists and everyone doubted that our end to end deep learning based method would work (I.e. similar doubt by VCs in the air back then).
There was no revenue, so no inference workload. When we released the product a couple years later (yeah, not in six months like a SaaS company, AI is not SaaS), utilization shifted down from 100% to 80% in research and up from 0% to 20% in production inference (I.e. revenue generating). We then bought a lot more machines and shifted the ratio back to 95% research, and 5% inference. Once the revenue grew, inference creeped back up to 30% of our datacenter FLOPS. Then we bought more and pushed the inference ratio backed down to 5-10% to give research room to breath. We did this many times. We also still use GPUs from the original build out (they don’t expire for 5-10 years).
What was aggregate usage doing? Scaling to make Deepgram’s STT API the number one in the world (what it is now). What else is happening? The same thing at Deepgram with TTS and the infra and models powering the total conversation via a voice AI agents platform (lots of R&D spend to build it effective, right, and with good COGS). Even with a scaled business, our research compute is still the biggest fraction of our data centers by far.
Moral of the story? You have to significantly spend out front to be relevant in the AI wave as a foundational AI company in the research phase of an explosively disruptive market. (Note: This does not apply to the companies not doing foundational AI, i.e. those building Apps on top of what the model + infra people build, There’s no need for them to spend massively ahead on research, they aren’t doing research).
—
Mantras folks can lean on as they learn how foundational AI businesses work:
- Foundational AI is not SaaS.
- A hot GitHub repo does not a good business make.
- Good fused science + GTM business models will win, not flashiness.
- The research phase of a revolution needs large fractional R&D spend.
- We are in the research phase of the Intelligence Revolution
Choose between a) fast and affordable TTS but sounds like a robot or b) quality TTS that is too slow and expensive? Not anymore.
Today, the Deepgram team releases Aura into general availability.
Aura is a fast, affordable, and human-like Text-to-Speech API for real-time voice conversations. What to expect: AI conversations will speedup drastically and be more lifelike.
Re: cost; I know that paying $3/hr just for conversational TTS does not make sense, and that's why Aura costs 5x less. This is absolutely necessary to enable voice AI agents at massive scale.
Aura is a crucial addition to our Voice AI Platform. Now developers can bring their own LLM and Deepgram will take care of the audio: listen with Nova-2 and speak with Aura.
The products using Deepgram's Voice AI platform (Aura and/or Nova) are so good. Check out Daily, Five9, Twilio, Vonage, Vapi (YC W21), Vocode, Retell AI, Groq and many others (it's thousands now) to see Deepgram's voice AI in action.
Quick info: • Aura is available to anyone with a Deepgram API key. • Aura currently has a dozen unique voices. • All new signups can take advantage of $200 in free credits. • That's ~13M characters worth of voice synthesis. • Both current and new customers can start building with TTS right away.
Frederic Lardinois from TechCrunch covered the Aura launch here 🎉 https://lnkd.in/g4qM8jSh.
Our customers gave insightful and candid feedback during development 🙌 and are shipping scaled traffic to production already 📈. And the Deepgram team built an amazing product that users find value in right away 💪.
Thank you, respect, and congrats!
You can try Aura working with Nova-2 and OpenAI yourself using this open source tech demo: https://lnkd.in/ghsZxNNW
The future will be fun.
